FINAL PROJECT REPORTS

Transcription

1 FINAL PROJECT REPORTS DECEMBER 2007

2 Authors: Bertrand Timbal (Bureau of Meteorology) Final report for Project Development of the analogue downscaling technique for rainfall, temperature, dew point and pan evaporation Principal Investigator: Dr. Bertrand Timbal, Centre for Australian Weather and Climate Research (CAWCR), Bureau of Meteorology Tel: , Fax: , GPO Box 1289, Melbourne 3001 Co-Authors: Dr. Bradley Murphy (National Climate Centre, BoM) Elodie Fernandez and Zhihong Li (Centre for Australian Weather and Climate Research) Completed: 29 February 2008 Confidential Page 1 29/02/08

3 Authors: Bertrand Timbal (Bureau of Meteorology) Project Abstract - Executive summary Initial Project objectives: Set up a statistical downscaling technique to relate large-scale changes to local variations in south-eastern Australia Proposed methodology: Existing downscaling methodology will be expanded to include humidity variables (dew point temperature and pan evaporation) in addition to proven dataset (rainfall and daily temperature extremes). Large-scale predictors will be tested and the spatial variation of skill across south-eastern Australia will be assessed for all calendar seasons and the suitable stations data, following on recommendation from milestone Methodology will be optimised for south-eastern Australia using identified coherent climatic regions Summary of the findings: This project has seen the development and validation of a single downscaling method based on the idea of meteorological analogues to the entire SEACI regions and for all existing high quality climate surface networks: rainfall, temperature, dew-point temperature and pan-evaporation. The SEACI domain was divided into three climate entities: the Southern part of the Murray-Darling basin (SMD); east of the SMD, on the coastal side of the Great Dividing Range, the South-East Coast (SEC) and west of the SMD, the South-West coast of Eastern Australia (SWEA). Individual Statistical Downscaling Models (SDMs) were optimised for each region, each calendar season and each predictands; a total of 72 SDMs (3 regions * 4 seasons * 6 predictands). The optimisation comprised two steps: the selection of the best combination of predictors (step 1) and then setting up other critical parameters of the SDM. The skill of the SDMs are fairly consistent across the three regions, the four seasons and the six predictands, thus confirming that the analogue approach is a suitable downscaling method for mid-latitude temperate climate. The reproduction of the observed Probability Distribution Functions (PDFs) was assessed by checking the two main moments of the reconstructed series: mean and variance. The mean of the observed series is very well reproduced with the exception of rainfall but all the reconstructed series do under-estimate the observed variance, this underestimation varies from one predictand to another and is largest for rainfall. In the case of rainfall, because the daily PDFs is not near normally distributed the reduced variance leads to a dry bias, this dry bias can be reduced with a very simple and robust inflation factor. The skill of the SDMs was assessed by looking at the ability of the technique to reproduce day-to-day variability using correlation and Root-Mean Square Errors (RMSEs): the best correlations tend to be achieved for most variables during the transition seasons autumn and spring, correlations in winter are often low but with low RMSEs (i.e. not less skill), in contrast, for all variables but daily temperature extremes, the model tends to have less skill (low correlation and high RMSEs) in summer. Confidential Page 2 29/02/08

4 Authors: Bertrand Timbal (Bureau of Meteorology) Technical details Expand existing downscaling model to humidity dataset The Australian Bureau of Meteorology has developed a SDM using the idea of meteorological analogues (Timbal and McAvaney, 2001). This is one example of a more general type of SDM based on weather classification methods in which predictands are chosen by matching previous (i.e., analogous situations) to the current weather-state. The method was originally designed for weather forecasting applications but was abandoned due to its limited success and lack of suitable analogues for systems with large degrees of freedom. The popularity of the method has recently increased with the availability of longer time-series datasets following the completion of reanalysis project and the recognition that the size of the search space must be suitably restricted when identifying analogues. Even so, the analogue method still performs poorly when the pool of training observations is limited and/or the number of classifying predictors is large. The Bureau SDM was first developed for daily temperature extremes (T min and T max ) across the Murray-Darling Basin (MDB) (Timbal and McAvaney, 2001). It was then extended to rainfall occurrences (Timbal et al., 2003) and amount (Timbal, 2004). As part of this SEACI project, the Bureau of Meteorology existing downscaling technique has being tested for new surface variables to complement previous work done on rainfall and temperature. These new surface variables are the most recent addition to the Bureau High Quality (HQ) climatological networks. Dew point stations were homogenised (Lucas, 2006) and are available in the HQ dataset from 1957 to stations across the SEACI domain were considered. At each location, daily maximum, daily minimum, and 9am dew point temperatures are available but the optimisation of the SDM was applied only to daily extreme dew point temperature. Pan evaporation HQ stations have been also been assembled across Australia (Jovanovic et al., 2008) from 1975 to stations are scattered across the SEACI domain. The Bureau pan-evaporation HQ dataset is a monthly dataset, the quality control was extended to daily values across the SEACI region, using monthly corrections for non-homogeneities at stations which required such correction (as part of project 1.1.1). The application of a SDM to these moisture variables is a very novel research as there is currently very few examples in the literature of fitting a statistical downscaling model to surface moisture variables (Huth, 2005 for a case study fro dew point across the Czech republic) and none as extensive as our study. Overall, applying a single technique across a large region such as the SEACI domain and across a large range of predictands is a very large undertaking. This extensive work (a total of 72 individual SDMs were optimised: 3 regions * 4 seasons * 6 predictands) was possible due to the simplicity of the chosen downscaling method. The analogue approach used here is one of the simplest existing downscaling methods. Despite its simplicity which was paramount to be able to perform this work, this method has been shown to compare well with more advanced techniques (Zorita and von Storch, 1999). The simplicity, flexibility and robustness of the technique were important to ensure that a single technique could be used across a range of variables and several climatic regions. Choice of coherent climatic regions In order to apply the Bureau of Meteorology SDM to the SEACI domain, surface observations were gathered into three distinct climate entities (Fig. 1), roughly following the Confidential Page 3 29/02/08

5 Authors: Bertrand Timbal (Bureau of Meteorology) rotated Empirical Orthogonal Functions (EOFs) for rainfall suggested by Drosdowsky (1993): (1) the South-West of Eastern Australia (SWEA): southwest of a line roughly from Melbourne to the south of the Flinders ranges and following the end of the Great Dividing Range (GDR) over Western Victoria, (2) the southern half of the Murray-Darling Basin (SMD) south of 30ºS in the north, limited in the west by SWEA and in the east by the GDR, and (3) the South-East Coast (SEC), a coastal band east of the GDR from Wilson Promontory in Victoria in the south all the way along the New South Wales coast up to the Hunter valley in the north. Figure 1: Locations of the station data chosen for the SEACI program (upper right map) and for the three climatic entities used to optimise the statistical downscaling model: the Southern Murray-Darling Basin (SMD), the South-West of Eastern Australia (SWEA) and the South-East Coast (SEC). Different symbols are used for different surface predictands: D for dewpoint temperature, E for pan-evaporation, T for temperature and the small points are rainfall station). The number of surface predictands available in each climatic region is summarised in Table 1. Although a small part of the Australian continent is covered by these three climatic regions, together they cover a large proportion of the number of HQ observation sites across Australia (between 30% for dew point temperature and 54% for rainfall. It underlines the fact that these regions are amongst the most populated in Australia: hence the relatively denser network of observations in particular for rainfall. A logical consequence is that they are amongst the most important for human related activities (e.g. agriculture). For rainfall, a large number of additional Confidential Page 4 29/02/08

6 Authors: Bertrand Timbal (Bureau of Meteorology) stations have been used as well to apply the SDM to: the number of additional rainfall stations per climate regions is also shown (Table 1). Predictands SWEA SMD SEC Temperature (T max & T min ) Rainfall HQ network additional stations Pan-Evaporation Dew point (dt max & dt min ) Table 1: Number of stations considered in each climatic region for the four types of predictand For temperature some stations as well were added but only a handful (they are included in the number of temperature stations provided. Overall not all the SEACI relevant stations identified in project (and shown in the top right of Fig. 1) are included in one of the three regions as the original stations lists cover a wider geographical area than the three climatic entities (inserts in Fig. 1). Optimization of the predictors The choice of the optimal combination of predictors constitutes the first step in the optimisation of the individual SDMs. The predictors considered were chosen on the basis of previous experience while developing the BoM SDM (Timbal and McAvaney, 2001; Timbal et al., 2003; Timbal, 2004), and evidences in the literature from other studies in similar areas. The optimum combination of predictors varies across regions, seasons and predictands (Table 2). The optimal number of predictors is often three, apart from pan evaporation where most frequently only two predictors are used and for rainfall where four predictors are often required. The need for a large number of predictors for rainfall shows that it is a difficult predictand to capture from largescale analogues. Some general patterns are emerging from the optimum combinations of predictors: Mean sea level pressure (MSLP) is the most frequently chosen predictor. It is used for all individual SDMs in the case of rainfall, T max and dt min but is picked up far less often for pan evaporation. This feature suggests that MSLP is a critical predictor for a synoptically driven technique such as the analogue approach. Thermal predictors are very important, especially for T max and T min. In general, T 850 is the most important thermal predictor, although T min is more important for dt min and dt max ; thermal predictors rarely matter for rainfall. Moisture variables are also important predictors across all predictands with the notable exception of T max. Specific humidity is almost always picked up apart from pan evaporation for which relative humidity is more skilful. Rainfall is often part of the optimised predictor s combination to downscale rainfall. Some measure of the air flow (either the zonal or meridional component of the wind) is often added to the optimised combination. It is an additional predictor to the de-facto combination of Confidential Page 5 29/02/08

7 Authors: Bertrand Timbal (Bureau of Meteorology) synoptic-thermal-moisture. It is most useful for rainfall and then T max, and least useful for dewpoint temperature and pan evaporation. The zonal component is the most frequently used. Variable Season SWEA SEC SMDB Summer MSLP & T 850 MSLP & T max MSLP & T max Maximum Autumn MSLP & T max MSLP & T max MSLP & T max Temperature Winter MSLP & T 850 & T max & U 850 MSLP & T max MSLP & T 850 & T max & U 850 T max Spring MSLP & T 850 MSLP & T 850 & T max & U 850 MSLP & T 850 & U 850 Summer MSLP & T 850 MSLP & T 850 & Q 850 T 850 & Q 850 Minimum Autumn MSLP & T 850 & Q 850 MSLP & T 850 & Q 850 T 850 & Q 850 Temperature Winter MSLP & T 850 & Q 850 MSLP & T 850 & T min & U 850 MSLP & T 850 & Q 850 T min Spring MSLP & T 850 & Q 850 MSLP & T 850 & Q 850 MSLP & T 850 & Q 850 Summer MSLP & PRCP & T 850 MSLP & T max & Q 850 & U 850 MSLP & PRCP & V 850 Rainfall Autumn MSLP & T max & Q 850 & U 850 MSLP & PRCP & Q 850 & U 850 MSLP & PRCP & V 850 PRCP Winter MSLP & PRCP & V 850 MSLP & PRCP & U 850 MSLP & PRCP & V 850 Spring MSLP & PRCP MSLP & PRCP & Q 850 & U 850 MSLP & PRCP & V 850 Maximum Summer MSLP & Q 925 MSLP & Q 925 & T min MSLP & Q 850 & T min & V 850 dew-point Autumn MSLP & Q 925 & T min & V 850 MSLP & Q 925 & T min MSLP & Q 850 & T min & V 850 Temperature Winter MSLP & Q 925 & T min MSLP & Q 925 & T min T min dt max Spring MSLP & Q 925 & T min MSLP & Q 925 & T min MSLP & Q 850 Minimum Summer MSLP & Q 925 & T min MSLP & Q 925 & T min MSLP & Q 850 & T 850 dew-point Autumn MSLP & Q 925 & T min MSLP & Q 925 & T min MSLP & Q 850 & T min & U 850 Temperature Winter MSLP & Q 925 & T min MSLP & Q 925 & T min MSLP & Q 850 & T min & V 850 dt min Spring MSLP & Q 925 & T min MSLP & Q 925 & T min MSLP & Q 850 & T 850 Summer T max & R 925 T max & R 925 T max & R 850 Autumn T max & R 925 T max & R 925 T max & R 850 Winter T max & R 925 MSLP & T max & R 925 T max & R 925 Pan- Evaporation P-Evap Spring T max & R 925 T max & R 925 & U 850 T max & R 850 Table 2: Optimum combination of predictors for each calendar seasons and the six predictands in three regions: SWEA, SEC and SMD. The predictors are defined as follow: MSLP is the Mean Sea Level Pressure; T max and T min are the surface min and max temperature; PRCP is the total rainfall; Q is the specific humidity; R is the relative humidity; T is the temperature; U and V are the zonal and meridional wind components; and subscript numbers indicates the atmospheric level for the variable in hpa. The second step of the optimisation of individual SDMs was to set up some critical parameters of the analogue model. The SDM includes a large number of tuneable parameters; however previous studies have shown that only three parameters are critical and therefore only these three where systematically explored: 1. The size of the geographical domain used for the predictors (latitude and longitude); in general two domain sizes were tested; these domain sizes are region dependent 2. The calendar window from which analogues are found. Three periods were tested; 15, 30 and 60 days prior to or after the date for which an analogue is searched for. 3. The way the daily anomalies are calculated using either three monthly means or a single seasonal average. Confidential Page 6 29/02/08

8 Authors: Bertrand Timbal (Bureau of Meteorology) Once these two steps were completed, the optimised SDMs were validated their skills were systematically evaluated. Skill of the SDM The evaluation of the skill of the SDMs was done using a fully cross-validated approach to ensure that no spurious skill was taken into account. The model was first optimised on one half of the existing dataset and then applied to the other half (the length of these halves varies for each predictands according to the length of the available record). When applied to the validation part of the dataset, analogue are searched for in the development half of the dataset to ensure a fully crossvalidation. Hence if the climate has recorded a shift during the two periods, the method has to be able to reproduce that shift, thus adding confidence in the ability of the technique to reproduce non-stationarity in the climate system now (useful for detection and attribution study) and in the future (useful for generating regional projections). However in this project, only the ability of the technique to reproduce the simultaneous observations was analysed; the ability of the method to reproduce observed changes is currently underway as part of the project Here, the evaluation of the SDMs focused on the ability of the technique to reproduce the main characteristics of the observed series and that it is doing it for the right reasons. A range of metrics was used. First the ability of the technique to reproduce the observed probability distribution functions (PDFs) was evaluated by looking at the first two moments of the PDFs: the mean and the variance. Further than the ability of the technique to reproduce the observed shape of the PDFs as defined by the first two moments of the series, it is important to ensure that the technique is skilful in reproducing day-to-day variability that is driven by largescale synoptic changes. As indeed a random choice of analogue would reproduce perfectly the observed mean and variance but is not a skilful model. To do so, the Pearson correlation between daily observed and reconstructed series was calculated separately per region, per season and for each predictand. Each number is an average across all observations available in each region. Alongside the correlation coefficient, Root-Mean-Square errors (RMSEs) were also used to complement the evaluation of the skill of the SDMs to reproduce day-to-day variability. The motivation to use both correlation and RMSE is to be able to differentiate between cases where correlation is low but RMSEs are also low: indicative of an observed series with little variability and hence difficult to reproduce very well, and cases where correlation is low and RMS is large indicative of a less skilful SDM. Results are detailed in the appendix attached to this report. Overall it was found that the analogue approach was successful across the SEACI domain, as results are fairly consistent across the three regions, the four seasons and the six predictands. It was found that the mean of the observed series is very well reproduced for all variables with the exception of rainfall, but the reconstructed series does under-estimate the observed variance in all cases. This underestimation of the variance varies from one predictand to another. In the case of rainfall, because the underestimation is very large and the daily PDFs do not have a near-normal distribution, the reduced variance leads to a dry bias. The variance reduction issues and subsequent dry bias can be addressed using a very simple (i.e. a single parametric coefficient applicable to all stations and all seasons) and robust (i.e. it only depends on the pool of observed rainfall occurrences and hence is applicable to the downscaling of climate models) inflation factor. In terms of ability to reproduce day-to-day variability, it was found that the analogue method was overall quite successful. The lowest skill was observed for rainfall, the best for daily temperature extremes. The best correlations tend to be achieved for most variables during the transition Confidential Page 7 29/02/08

9 Authors: Bertrand Timbal (Bureau of Meteorology) seasons autumn and spring, correlations in winter are often low but with low RMSEs (i.e. not less skill), in contrast, for all variables but daily temperature extremes, the model tends to have less skill (low correlation and high RMSEs) in summer. Additional information Acknowledgement This work was funded by the South Eastern Australia Climate Initiative. References Drosdowsky W., 1993: An Analysis of Australian seasonal Rainfall anomalies: I: Spatial Patterns. Int. J. Climatology, 13, 1-30 Huth R., 2005: Downscaling of humidity variables: a search for suitable predictors and predictands, Int. J. of Climatology, 25, Jovanovic, B., Jones, D. A., and Collins, D., 2008: A High Quality Monthly Pan-Evaporation Dataset for Australia. Climatic Change (published on-line) Lucas, C., 2006: A high quality humidity database for Australia. Abstracts for 17 th Australia New Zealand Climate Forum, Canberra, September 2006, p 35 Timbal B. and B.J. McAvaney, 2001: "An Analogue based method to downscale surface air temperature: Application for Australia", Clim. Dyn., 17, Timbal B. A. Dufour and B.J. McAvaney, 2003: "An estimate of climate change for Western France using a statistical downscaling technique", Clim. Dyn., 17, Timbal B., 2004: "South West Australia past and future rainfall trends", Clim. Res., 26(3), Zorita, E., and H. von Storch (1999), The analog method as a simple statistical downscaling technique: comparison with more complicated methods, J Climate, 12, Confidential Page 8 29/02/08

10 Authors: Bertrand Timbal (Bureau of Meteorology) Outputs from this project Publications: Timbal, B., E. Fernandez and Z. Li, 2008: Optimisation of a statistical downscaling model and development of a graphical user friendly interface to provide downscaled climate change projections on a continental scale. Env. Mod. & Software, (submitted). Conference papers: B. Timbal and Z. Li, 2007: A user friendly interface to provide point specific climate change projections, MOSDIM07, Christchurch, December Project Milestone Reporting Table Milestone description 1 Performance indicators 2 Completion Budget 4 date 3 for Mileston e ($) Progress 5 Recomme nded changes to workplan 6 1. Test statistical downscaling on humidity dataset Quantify skill of this technique for moisture field 01/01/07 30 k$ Large-scale predictors have been tested. RH is the only surface predictand s not used. 2. Identify coherent climatic regions SE Australia subdomains defined 01/05/07 25 k$ 3 sub-domains have been identified across the S.E.A. None 3. Optimize choice of predictors Define the optimal model in all cases 01/09/07 25 k$ All individual models have been optimised. None. Work completed 4. Evaluate skill of the technique across the area of interest A 6-page report compiling results from a range of metrics 01/01/08 25 k$ This work is now completed. The 6-page report is near completion and will be attached to the project final report None. Confidential Page 9 29/02/08

11 Authors: Bertrand Timbal (Bureau of Meteorology) Appendix: Technical report on the evaluation of the technique (Milestone 4). Methodology: Following on the work done to select suitable stations for all surface predictands: maximal and minimal daily temperature (T max, T min ), rainfall, pan-evaporation and maximal and minimal daily dew-point temperature (dt max, dt min ) as part of project 1.1.1, statistical Downscaling Models (SDMs) were formed by first identifying coherent climatic regions (milestone 2 of this project) and optimizing the choice of predictors (milestone 3). This report summarizes the evaluation of the skill of individual SDMs across the South-East of Australia (SEA) for the six surface predictands (milestone 4). The evaluation of the SDMs is done in a fully cross-validated manner. The model is first optimised on one half of the existing dataset and then applied to the other half. When applied to the validation part of the dataset, analogue are searched for in the development half of the dataset to ensure full cross-validation. Hence if the climate has recorded a shift during the two periods, the method has to be able to reproduce that shift, thus adding confidence in the ability of the technique to reproduce non-stationarity in the climate system now (useful for detection and attribution study) and in the future (useful for generating regional projections). The evaluation was carried out using a range of metrics. First the ability of the technique to reproduce the observed probability distribution functions (PDFs) was evaluated by looking at the first two moments of the PDFs: the mean and the variance. Besides the ability of the technique to reproduce the observed shape of the PDFs as defined by the first two moments of the series, it is important to ensure that the technique is skilful in reproducing day-to-day variability that is driven by large-scale synoptic changes. Indeed, a random choice of analogue would reproduce perfectly the observed mean and variance but is not a skilful model. The Pearson correlation between daily observed and reconstructed series was calculated separately per region, per season and for each predictand. Each number is an average across all observations available in each region. Alongside the correlation coefficient, Root-Mean-Square Errors (RMSEs) were also used to complement the evaluation of the skill of the SDMs to reproduce day-to-day variability. The motivation to use both correlation and RMSE is to be able to differentiate between cases where correlation is low but RMSEs are also low: indicative of an observed series with little variability and hence difficult to reproduce very well, and cases where correlation is low and RMSE is large indicative of a less skilful SDM. Temperature: In the case of daily extreme temperature, the development period is 1958 to 1982 and the validation period is 1983 to 2006 (i.e. analogues to reproduce 1983 to 2006 are picked up from 1958 to 1982, a notably cooler period in many instances). The reproduction of the mean values for both predictands (Fig. 1) is very accurate. In each graph, points correspond to a single location for a single season with the observed mean value on the x-axis and the reconstructed mean along the y-axis. The number of points in each graph is equal to the total number of stations in one of the three climate regions times four seasons (224 in the case of temperature). Results for the mean are not far from a perfect match (especially for T max ), with points aligned with the diagonal. Confidential Page 10 29/02/08

12 Authors: Bertrand Timbal (Bureau of Meteorology) Reconstructed Mean Tmax Reconstructed Mean Tmin Observed Observed Reconstructed Tmax variance Reconstructed Tmin variance Observed Observed Correlation Summer Autumn Winter Spring Correlation Summer Autumn Winter Spring Tmax SWEA SMD SEC 0.1 Tmin SWEA SMD SEC Root Mean Sqaure errors Tmax Summer Autumn Winter Spring SWEA SMD SEC Root Mean Square Errors Tmin Summer Autumn Winter Spring SWEA SMD SEC Fig 1: Scatter plot of the reconstructed versus observed mean (top row) and variance (second row) and correlations (third row) and RMSEs (fourth row) between the two series for T max (left) and T min (right). On scatter plots, there is one point per station and per season, the colour-code refers to season: winter (blue), spring (green), summer (red) and autumn (orange). The diagonal is the line of perfect fit. Correlations and RMSEs are averaged across all stations per region (name on X-axis) and specified by season (coloured bars). Units for mean, variance and RMSE are ºC. Confidential Page 11 29/02/08

13 Authors: Bertrand Timbal (Bureau of Meteorology) Furthermore there is no evidence that the SDMs have more difficulty at reproducing mean observed values at either hand of the spectrum (large or small values). Similarly results are shown for the reproduction of the standard deviation. The technique appears to have a tendency to underestimate the observed variance: points are aligned below the diagonal for most cases. This is particularly true in summer for T max but obvious across all seasons for T min. Average across all stations the reduction of variance ranges from 11.8% in winter to 0% in spring for T max and 11.8% in winter and 5.3% in summer for T min. This variance underestimation is relatively small with the analogue approach (which does not require any linear assumption) compared to many other techniques (in particular linear techniques), but it remains an issue across all statistical downscaling technique (von Storch, 1999). For temperature, as daily values are not far from being normally distributed, the underestimation of the variance does not have a flow-on effect on the reproduction of the mean (unbiased as noted earlier). Finally the ability of the SDM to skilfully reproduce day-to-day variability (i.e. the ability to reproduce the right PDFs for the right reasons) appears very successful based on correlations for both T max, T min. There is a marked seasonal cycle in correlation coefficients: lower values are observed in winter and highest values during autumn and spring. However in most instances, in particular for T max, winter corresponds also to the lowest RMSEs. Therefore the lower correlation does not imply less skill but a season where day-to-day variability is less marked and hence harder to capture. On the contrary in autumn and spring the high correlation values are help by large dayto-day variability during the transient seasons. In the case of T min, it is not obvious as RMSEs are fairly similar across all seasons and hence the lower correlations in winter suggest that the model is less skilful. Overall no particular region stands out as a climatic entity where the SDM skill in reproducing day-to-day variability is consistently lower or higher across all seasons. This result vindicates the fact that the model is applicable to the entire South-Eastern part of Australia where the climate by and large is driven by synoptic disturbances. Rainfall: Similar graphs were generated for rainfall (Fig. 2), as for temperature, the development period is 1958 to 1982 and the validation period is 1983 to The reproduction of the first two moments of the series (mean and variance) is less successful than for temperature (left plots in the first two rows of Fig. 2). The underestimation of the variance is much larger: ranging from 27% in autumn to 45% in summer. The consequence of the reduction of variance, in the case of rainfall, can be seen on the reproduction of the mean; points are located below the diagonal, indicating a bias toward drier values for reconstructed series. In the case of rainfall, the reproduction of the mean is dependent on the ability of the technique to reproduce the observed variance as rainfall is not normally distributed. For this reason, a correction factor to adjust the reconstructed rainfall series and enhance the variance and improve the reproduction of the mean was introduced in earlier applications of the analogue approach to rainfall series in Western Australia (Timbal et al., 2006). The rationale for the applied correction is that the analogue reconstructed rainfall is affected by the size of the pool of analogues which becomes smaller in the case of rare large rainfall events. Therefore, the error in finding the best matching analogue increases and the chances are that the Confidential Page 12 29/02/08

14 Authors: Bertrand Timbal (Bureau of Meteorology) best analogue found would describe more frequent but less intense rainfall events thus underestimating the rainfall in the reconstructed series. It is assumed that the size of the pool depends on the ratio of rain days over dry days and that is valid across the range of climates encounter in SEA as it was the case in W.A. Reconstructed Mean Rainfall Observed Reconstructed Mean rainfall Observed 150 Rainfall variance 150 Rainfall variance Reconstructed Reconstructed Observed Observed Correlation Summer Autumn Winter Spring Root Mean Square Errors Summer Winter Autumn Spring 0.1 Rain SWEA SMD SEC 0.0 Rain SWEA SMD SEC Fig 2: As per Fig. 1 but for rainfall. The additional two scatter plot of the reconstructed versus observed mean and variance (in the right column) are for rainfall with an inflation factor applied to the reconstructed series (see main text for details). Units are mm. It was decided that the same very simple factor should be applied without further adjustment to limit some of the danger linked to artificially inflating the variance when using downscaling techniques (von Storch, 1999). The following single factor was used Confidential Page 13 29/02/08

15 Authors: Bertrand Timbal (Bureau of Meteorology) N dry 1 And C 1. 5 factor N C = factor wet Where N dry and N wet are the numbers of dry and wet (> 0.3mm) days observed for the season at an individual location. These numbers are station and season dependent. They are calculated on the available observations from which analogues are drawn and are therefore independent of the series being reconstructed. These ratios are equally applicable when developing the downscaling model (and hence evaluating their impact) or when downscaling climate simulations. The impact of the inflation factor is clear (right plots in the first two rows of Fig. 2). It has dramatically reduced the variance bias and lead to an un-bias reproduction of the mean (as was the case for temperature) and un-biased reproduction of the variance (therefore better than for temperature). However, the spread of the reconstructed versus observed mean of the series is unchanged with the uncorrected series and is larger than with temperature. For rainfall, correlations are by far lower than for temperature, although due to the very large sample considered (about 2000 days); all these correlations are significant at least at the 95% level, indicating some level of skill. Correlations peak in winter (up to 0.3 to 0.4) and are particularly low for the dry season: summer and autumn in the case of SWEA. These low correlations are confirmed by the high values for RMSEs. The largest errors are seen in summer and the smallest in winter thus confirming that the SDMs are more skilful for rainfall in winter. Dew-point temperature: Finally the skill of the model on newly formed high-quality dataset is evaluated, starting with dew-point temperature (Lucas, 2006): daily maximal (dt max ) and daily minimum (dt min ). In the case of daily extreme dew-point temperature, the development period is 1958 to 1982 and the validation period is 1983 to 2003 (as the high quality dataset has not been updated past that point). Results are very similar than for temperature, the SDMs are able to reproduce the mean of the observed series very accurately (with slightly larger errors for dt min ). The underestimation of the variance is again visible: ranging between 11.6% in summer and 4.4% in autumn for dt max, and from 16.7% in winter and 7.7% in spring for dt min. As for temperature, daily values of dew-point temperature are not far from being normally distributed, and hence the underestimation of the variance does not have a flow-on effect on the reproduction of the mean. Correlations between reconstructed and daily series for dt max have a lot in common with results for temperature, albeit with correlation being overall lower. Lowest correlation are in winter but corresponds to low RMSEs as well, hence not suggesting less skilful SDMs. Highest correlation are seen during the transition season autumn and spring and largest RMSEs tend to be in summer. In the case of dt min results are very homogeneous across seasons and regions. The exception being SMD during the spring and summer when RMSEs are much larger and correlation are rather low thus suggesting that the model is less skilful during the warmer seasons in this region. Confidential Page 14 29/02/08

16 Authors: Bertrand Timbal (Bureau of Meteorology) Reconstructed Mean dtmx Reconstructed Mean dtmn Observed Observed Reconstructed dtmx variance Reconstructed dtmn variance Observed Observed Summer Autumn Winter Spring Summer Autumn Winter Spring Correlation Correlation dtmax SWEA SMD SEC 0.1 dtmin SWEA SMD SEC Root Mean Square Errors dtmax Summer Autumn Winter Spring SWEA SMD SEC Root Mean Square Errors dtmin Summer Autumn Winter Spring SWEA SMD SEC Fig 3: As per Fig. 1 but for daily maximum dew point temperature (left) and daily minimum dew point temperature (right). Units for mean, variance and RMSE are ºC. Confidential Page 15 29/02/08

17 Authors: Bertrand Timbal (Bureau of Meteorology) Pan-evaporation: Finally in the case of pan-evaporation, for which the high quality dataset span a shorter period (Jovanovic et al., 2008), the development period is 1975 to 1988 and the validation period is 1989 to 2003 (as the high quality dataset has not been updated past that point). As for the other variables, the mean of the reconstructed series is very accurate although some errors up to 1 mm.day -1 are noticeable during the warmer seasons (spring and summer). Errors on the variance can be quite large, up to 3 mm.day -1 in some instances, but there is a slightly lesser bias toward a reduction of the variance. The mean variance bias is between 13.7% in winter and only 1.6% in spring. The skill of the SDMs in reproducing day-to-day variability varies a lot form one season to another according to the correlation: it is high for the transitions seasons and low in winter (especially in SMD and SWEA) and in summer (especially in SEC). However the RMSEs suggest that the low correlations in winter are partly due to very small day to day variability thus giving very small RMSEs. On the contrary in summer, the low correlation and high RMSEs suggest that the SDMs have less skill. Reconstructed Mean p-evap Reconstructed p-evap variance Observed Observed Correlation Summer Autumn Winter Spring Root Mean Square Errors Summer Autumn Winter Spring 0.1 p-evap SWEA SMD SEC 0.0 p-evap SWEA SMD SEC Fig 4: As per Fig. 1 but for pan-evaporation. Units for mean, variance and RMSE are mm.day -1. Confidential Page 16 29/02/08

18 Authors: Bertrand Timbal (Bureau of Meteorology) Conclusions: Overall the evaluation of the skill of the SDMs has shown that: The results are fairly consistent across the three regions thus confirming that the analogue approach is a suitable downscaling method for mid-latitude temperate climate; The reproduction of the mean of the observed series (in a fully cross-validated sense) is very accurate with the exception of rainfall; For all variables, the reconstructed series does under-estimate the observed variance, this underestimation varies from one predictand to another and is largest for rainfall; In the case of rainfall, because the daily PDFs is not near normally distributed the reduced variance leads to a dry bias, this dry bias can be reduced with a very simple and robust inflation factor; Best skills tend to be achieved for most variables during the transition seasons autumn and spring; Correlations in winter are often low but this is often because the day-to-day variability in winter is low rather than because the model is less skilful; and In contrast, for all variables but daily temperature extremes, the model tends to have less skill (low correlation and high RMSEs) in summer. References: Jovanovic, B., Jones, D. A., and Collins, D., 2008: A High Quality Monthly Pan-Evaporation Dataset for Australia. Climatic Change, (published on-line) Lucas, C., 2006: A high quality humidity database for Australia. Abstracts for 17 th Australia New Zealand Climate Forum, Canberra, September 2006, p 35 Timbal, B., J. Arblaster, and S. Power (2006), Attribution of the late 20th century rainfall decline in Southwest Australia, J Climate, 19(10), Von Storch, H. (1999), On the Use of Inflation in Statistical Downscaling, J Climate, 12, Confidential Page 17 29/02/08

19 Final report for Project Further Development of Statistical Downscaling Methodology Principal Investigator: Steve Charles CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Tel: Fax: Co-Authors: Guobin Fu CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Completed: January 2008

20 Abstract The stochastic downscaling framework implemented in this, and related, projects simulates rainfall as a two stage process, for rainfall occurrences and amounts respectively. The Nonhomogeneous Hidden Markov Model (NHMM) simulates multi-site daily rainfall occurrences. Then a regression approach is used to model a station s empirical rainfall distribution conditional on neighbouring station rainfall occurrences. We assess an extended amounts model that includes atmospheric predictor covariates in the regression model, in addition to neighbouring station rainfall occurrences. We also add resampling from a gamma distribution, rather than the empirical distribution. Results indicate that the extended amounts model versions often better reproduce observed validation period rainfall statistics. Another limitation of the current NHMM is a calibration restriction to 30 stations or less. A version of the NHMM developed separately by collaborators at the International Research Institute for Climate and Society (IRI) does not have this restriction, allowing calibrations to networks with a much larger numbers of stations. Here we show the IRI-NHMM can reproduce validation period rainfall statistics for a network of 132 stations. Significant research highlights, breakthroughs and snapshots Extended versions of the rainfall amounts component of the stochastic downscaling model, using atmospheric predictors in addition to neighbouring station rainfall occurrence, and a gamma distribution instead of the empirical distribution, are often able to better reproduce the daily rainfall distribution of the validation period. These extended rainfall amounts model versions will be used in the generation of projected rainfall series and the differences assessed in Project The IRI-NHMM provides the ability to downscale to a much larger network of stations, although the algorithms it uses for reproducing between-site spatial correlations are not as advanced as those of the CSIRO-NHMM. Testing on a network of 132 stations produced adequate results. Statement of results, their interpretation, and practical significance against each objective Objective 1: Determine which, if any, deficiencies in NHMM performance require further model development. Significant changes can occur in individual station daily rainfall probability density functions (PDFs) over time. Thus a key measure of the performance of any statistical downscaling approach is its ability to reproduce the rainfall statistics of a period not used in calibration. Here we use the period to assess the NHMMs that were calibrated for in Project Calibration of Statistical Downscaling Models. Six variants of methods for conditioning multi-site daily rainfall amounts on neighbouring station rainfall occurrences and atmospheric predictors, resampling from gamma or empirical distributions, were assessed to determine how well such changes are captured. The original CSIRO-NHMM used only neighbouring station rainfall occurrences and resamples station amounts from observed empirical distributions (Charles et al. 1999).

21 In addition, here we include two modifications: (i) conditioning on atmospheric predictors and (ii) sampling from a gamma distribution fitted to observed station daily rainfall (by weather state). Thus the six variants involve conditioning on: (i) neighbouring station rainfall occurrences only with empirical distribution (ii) neighbouring station rainfall occurrences only with gamma distribution (iii) neighbouring station rainfall occurrences and atmospheric predictors with empirical distribution (iv) neighbouring station rainfall occurrences and atmospheric predictors with gamma distribution (v) atmospheric predictors only with empirical distribution (vi) atmospheric predictors only with gamma distribution Note the predictor sets used for conditioning these amounts models are independent of those used for NHMM calibration. The candidate set of predictors for amounts conditioning was restricted to specific humidity and dew point temperature depression at 850, 700 and 500 hpa. A stepwise multiple linear regression procedure was applied to select the conditioning predictor from these candidates, for each station and for each weather state. Figures 1 to 5 present, for five example stations, quantile-quantile plots of observed versus downscaled validation period daily rainfall. The versions using the fitted gamma distributions, rather than the empirical distributions, often better reproduce the upper tails of the distributions (i.e. the very high daily rainfalls that occur rarely). The gamma distribution can also sometimes improve the overall distribution, but more often the addition of the atmospheric predictors to the covariates used, in combination with the gamma distribution, improves performance the most. Objective 2: Assess the CSIRO NHMM against a NHMM version produced by collaborators at the International Research Institute for Climate and Society (IRI), New York, USA. We investigate the IRI-NHMM performance for a 132 station network in terms of reproduction of probability of rainfall occurrence (i.e. wet-day frequencies) and logodds ratio. The log-odds ratio is similar to a measure of correlation but measures association in binary (i.e. wet or dry) data. We compare the winter weather states of the CSIRO-NHMM (Project report Figure 4) against that of the IRI-NHMM in Figure 6. The rainfall patterns of the 132 station model correspond to the states seen in the original 30 station model, as do the resultant composite predictor fields. Figure 7 shows that the 132 station IRI-NHMM can reproduce the winter rainfall probabilities of the fitting period with a slight bias for the validation period. A similar bias was seen in the original 30 station CSIRO-NHMM (Project report Figure 8). Similarly, log-odds ratios for the 132 stations are adequately reproduced although the non-spatial IRI-NHMM parameterisation leads to underestimation for station pairs with high correlation (Figure 8, compare to Project report Figure 9). Overall the IRI-NHMM can perform reasonably well for a 132 station network. More detailed investigation of inter-site rainfall statistics would be needed if particular applications where high spatial consistency is crucial, such as sub-catchment hydrological modelling, were envisaged.

22 Summary of methods and modifications (with reasons) Six versions of the daily multi-site amounts model were assessed and the version using additional atmospheric covariates as well as resampling from a gamma distribution was often found to perform the best. The NHMM version developed by the IRI has the advantage that it is not limited to 30 rainfall stations, as is the CSIRO NHMM code. However its ability to model the between-site spatial characteristics is parameterised in a less rigorous way and hence such characteristics may not be adequately reproduced for all applications. Here the IRI-NHMM calibrated to the same 30-station network, and also to a 64 and 132 station network (necessarily accepting stations of lesser data quality) performed well. Only results from the 132 station version are presented here. Summary of links to other projects The statistical downscaling model improvements produced in are essential for providing the best NHMM that will be applied to: o understanding current regional hydroclimate (Projects 1.4.2, 1.4.3, 1.5.2, 1.5.3) o climate change projection at the regional scale (Projects 2.1.3, 2.1.4). The improved NHMM will be compared to the models developed in Projects and Publications arising from this project Kirshner, S Learning with Tree-Averaged Densities and Distributions (conference paper, uses SEACI dataset and acknowledges S.Charles) Acknowledgement The IRI version of the NHMM used in the project was developed by Dr Sergey Kirshner, University of Alberta, Canada in collaboration with Dr Andrew Robertson, IRI for Climate and Society, New York, USA. Their on-going assistance is gratefully acknowledged. This project is funded by the South Eastern Australian Climate Initiative. Recommendations for changes to work plan from your original table None. References Charles SP, Bates BC, Hughes JP A spatio-temporal model for downscaling precipitation occurrence and amounts. Journal of Geophysical Research Atmospheres 104(D24):

23 Figure 1. Station daily amounts quantile-quantile plots for amounts model versions using: (a) neighbouring station rainfall occurrences only with empirical distribution; (b) atmospheric predictors only with empirical distribution; (c) neighbouring station rainfall occurrences and atmospheric predictors with empirical distribution; (d) neighbouring station rainfall occurrences only with gamma distribution; (e) atmospheric predictors only with gamma distribution; and (f) neighbouring station rainfall occurrences and atmospheric predictors with gamma distribution. Figure 2. As Figure 1, for Station

24 Figure 3. As Figure 1, for Station Figure 4. As Figure 1, for Station

25 Figure 5. As Figure 1, for Station

26 Figure 6. Winter weather states of 132 station IRI-NHMM.

27 Figure 7. Probability 132 stations (a) validation period (b) fitting period. Figure 8. Log-odds 132 stations (a) validation period (b) fitting period.

28 Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) 1. Assess deficiencies in CSIRO NHMM performance 2. Modify CSIRO NHMM to address deficiencies 3. Calibrate IRI NHMMs Performance indicators 2 (1-3 dot points) Properties of rainfall occurrence and amounts requiring improvement identified. Modified NHMM calibrated and assessed. Calibration of IRI NHMMs performed. Completion date 3 xx/xx/xxxx Budget 4 for Milestone ($) (SEACI contribution) Completed at each Milestone date Progress 5 (1-3 dot points) 1/8/ Completed. 31/12/ Completed. 1/11/ Completed. Recommended changes to workplan 6 (1-3 dot points) 4. Compare CSIRO NHMM performance to IRI NHMM performance Relative performance of NHMMs assessed. Written report (4-6 pages). 31/12/ Completed (this report is the written report)

29 Final report for Project Comparison of Observed and Reanalyses Downscaled Synoptics and Precipitation Principal Investigator: Steve Charles CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Tel: Fax: Co-Authors: Guobin Fu CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Completed: January

30 Abstract A stochastic downscaling model that relates multi-site daily rainfall to large scale atmospheric circulation, by conditioning a small number of discrete weather states (rainfall occurrence spatial patterns) on a set of atmospheric predictors, is used here to investigate variability and trends in rainfall over the lower MDB. The weather state time-series are analysed in relation to the corresponding atmospheric predictor and station rainfall time-series. Consistent and meteorologically realistic relationships between the series are evident. This provides confidence in the downscaling model s adequacy across the range of natural climate variability experienced for the period of available data (1958 to 2006). There is strong evidence of trends in the selected atmospheric predictors related to changes in positioning and frequency of synoptic systems and, in particular, a decrease in saturation of the middle atmosphere over the region. These are shown to link explicitly to the observed station rainfall trends through changes in weather state frequencies. Significant research highlights, breakthroughs and snapshots There is a high level of consistency between the NNR and ERA40 downscaled weather state and rainfall time-series, giving confidence that the downscaled atmospheric rainfall relationships are robust across re-analysis products. A small number (three to four) of regionally and seasonally specific atmospheric indices ( predictors ) are sufficient to adequately downscale to multi-site daily rainfall for a 30 site network across the lower MDB, reproducing station rainfall interannual to interdecadal variability and long term trends. Observed changes in predictors indicate drying has occurred in the middle atmosphere over the lower MDB in recent decades. Statement of results, their interpretation, and practical significance against each objective Objective 1: Compare observed south-eastern Australian synoptic and precipitation variability and trends to statistically downscaled weather-state and precipitation time-series for 1958 to In Project Calibration of Statistical Downscaling Models, Nonhomogeneous Hidden Markov Models (NHMMs) were calibrated to data for 30 sites for summer (Nov-Mar) and winter (Apr-Oct) seasons. The adequacies of these calibrated NHMMs were assessed for spatial and individual station precipitation statistics. The NCEP/NCAR re-analysis (NNR) for and the European ERA-40 re-analysis (ERA40) for (as ERA40 ends mid 2002) are used here to drive these NHMMs, producing multiple stochastic realisations of 30-site daily precipitation conditional on the single daily atmospheric predictor series extracted from these two re-analysis products. This allows assessment of how rainfall variability and trends (for which we have only one realisation, the observed record) relate to synoptic drivers, as captured through the atmospheric predictors. One key benefit of the NHMM stochastic downscaling approach is the classification of regional daily rainfall patterns into a discrete set of weather states, allowing analysis of their time-series properties and relationship with synoptic drivers. Recall, from Project 1.3.4, that the selected summer (Nov-Mar) NHMM has 6 states and 3 2

31 atmospheric predictors [mean sea level pressure (MSLP), 700 hpa dew-point temperature depression (DT d ), and East West 500 hpa geopotential height (GPH) gradient]. The winter (Apr-Oct) NHMM has 5 states and 4 predictors [North South MSLP gradient, 700 hpa and 850 hpa DT d, and North South 700 hpa GPH gradient]. For both summer and winter, State 1 is dry everywhere. States 3 and 2 are wet everywhere for summer and winter, respectively. The other states have varying spatial patterns of rainfall occurrence, as presented in the Project report. Figure 1 presents the seasonal cycles of weather state mean probabilities. For summer (Figure 1a for NNR and 1b for ERA40) the dominance of the dry State 1 is evident, particularly for January March. Correspondingly for winter (Figures 1c and 1d for NNR and ERA40, respectively), State 1 also dominates with a maximum in April followed by a steady decline until July August and then a more gradual increase to October. The wetter states, which occur at much lower frequencies than the dry states, have opposite cycles to the dry states. For example, the wetter summer States 3 ( wet everywhere ) and 5 ( moderately wet everywhere ) peak in November and the wetter winter States 2 ( wet everywhere ) and 5 ( wet everywhere, moderate in northwest ) peak around July August with minimums in April. Interestingly the moderately wet winter State 3 peaks in October, as States 2 and 5 are declining. The weather states are related to the multiple atmospheric predictors in a nonlinear way, so direct correspondence between individual predictors and weather state properties is not expected. However in most cases the seasonal cycles of the atmospheric predictors also show strong seasonality. For example, both the summer MSLP (Figure 2), winter North South MSLP and winter 700 hpa GPH gradient (Figure 3) show strong seasonality related to the annual latitudinal movement of pressure belts. The summer 700 hpa DTD is at a maximum (i.e. driest middle atmosphere) in February March. Winter 700 and 850 hpa DTD minima (i.e. moistest middle and lower atmosphere) occurs around July, with maxima (driest) in April and October respectively. Summer East West 500 hpa GPH gradient shows large within season variability but no seasonality. Figure 4 (summer) and Figure 5 (winter) differentiate the probability density functions (PDFs) of the predictor series by weather state. Although differences are subtle, they correspond to the resultant weather state s rainfall patterns (as shown in the Project report). For example the winter predictors for State 1 (dry everywhere) are noticeably different to those of State 2 (wet everywhere). The timeseries properties of these predictors are discussed later. Figure 6 shows the time-series of weather state probabilities from 1958 to 2006 for NNR and to 2001 for ERA40. The two reanalysis products produce very similar series, with slight differences in some annual peaks or troughs but the same interdecadal variability and long-term trends. The summer dry State 1 is more frequent since the late 1970s (Figures 6a and 6b). The wet south summer State 2 is less frequent and less variable since the mid 1980s, as is summer State 5 moderately wet everywhere since the mid 1990s. In contrast summer State 3 wet everywhere is slightly more variable with a couple of higher peaks since the early 1990s. Summer State 4 wet in the east and south is stable since the late 1970s after an increasing trend. State 6 wet SW to dry NE gradient is also stable since the late 1960s. For winter, the dry State 1 has an overall increasing trend particularly evident since the 3

32 early 1980s. Correspondingly the other winter states, representing various spatial patterns of wetness (see their patterns in Project report), have decreasing trends particularly since the 1980s (Figures 6c and 6d). Table 1 highlights these trends and interdecadal variability, by summarising the weather state s mean probabilities for key periods corresponding to the trends seen in Figure 6. For summer, the trend of increased dry everywhere State 1 is evident although the percentage differences are small, e.g. comparing the 1985/6 to 2004/5 NHMM calibration period to the earlier 1958/9 to 1984/5 period the mean probability increases from 53% to 56%. Figure 7 shows the differences in the predictor distributions for these two periods. There are slight changes in the tails of the MSLP and East West 500 hpa GPH gradient, while the 700 hpa DTD is higher for the later period. This suggests that for the same synoptic pressure patterns, the middle atmosphere is drier in recent decades. The recent 1996/7 to 2005/6 decade has a summer State 1 mean probability of 58%, compared to the previous 1986/7 to 1995/6 decade of 54% (Table 1). Figure 8 indicates that MSLP has increased in the middle of its distribution, so mean MSLP has increased even though the higher values are less frequent. DTD at 700 hpa has seen a distributional shift to become slightly drier across most of the distribution. Summer State 2, 3, 4 and 6 mean probabilities have not changed appreciably across these periods. In the last decade State 5 has decreased from 14% to 10% (Table 1). Recall State 5 has a moderately wet rainfall occurrence pattern across the entire region. The winter decadal changes are similar to summer (Table 1). The four wetter states have decreased by 1 or 2% between 1958 to 1985 and 1986 to 2005 and the dry State 1 has increased by 4% (42% to 46%). Figure 9 shows no appreciable change to N-S MSLP gradient, slight changes to N-S 700 hpa GPH gradient, and increases in DTD, particularly at 700 hpa. There has been a larger relative increase in the dry State 1 for the recent decade, with the mean probability increasing from 44% for 1987 to 1996 to 51% for 1997 to 2006 (Table 1). The corresponding predictor changes indicate a slight increase in DTD at both 850 and 700 hpa levels and decrease in N-S MSLP gradient (Figure 10). Classifying years as wet or dry, based on whether the majority of stations had winter rainfalls above or below the long-term mean, produces a large difference for State 1 mean probabilities of 52% in dry years compared to 35% in wet years (Table 1). The 1997 to 2006 State 1 mean probability of 51% is very close to this dry year mean of 52%, emphasising the dryness of the last decade. The mean probability of State 2 wet everywhere increases from 8% to 16% between dry and wet years, and State 5 which is also wet everywhere (although only moderately wet in the NW of the study region) increases from 11% to 16%. Figure 11 shows shifts in the distributions of all predictors between the dry and wet year composites consistent with these state probability differences. Looking at differences between individual years allows better appreciation of the variability in rainfall patterns and associated synoptics. Comparing 1982 (a dry year) to 1983 (a wet year) the State 1 probability changes from 60% to 41% and State 2 ( wet everywhere ) increases from 3% to 17% (Table 1). Figure 12 shows that some 4

33 predictors do not shift, rather the shape (standard deviation and skewness) of the distributions change. For example, the wet 1983 has a more variable 700 hpa DTD distribution with more low and high values, whereas in contrast the 850 hpa DTD is shifted to lower values for the wet year. Comparisons between other years can show larger relative changes in other weather states, for example comparing 2000 to 2001 State 5 decreases from 16% to 12% (Table 1). The rainfall time-series of individual stations show characteristics consistent with the weather state and predictor time-series. Figures 13 to 21 show seasonal rainfall timeseries and corresponding NHMM simulations for a sub-set of 9 of the 30 stations. The interannual variability is captured well by the NHMM simulations. Overall winter (the lower panel of each plot) is better reproduced than summer (the upper panel of each plot). Reproduction of the dry period experienced since the mid to late 1990s shows that temporal changes in weather state frequencies (i.e. regional rainfall patterns) and station rainfall are adequately captured by the NHMM selected large-scale atmospheric predictors. To investigate these trends further, the weather state, atmospheric predictor, and station rainfall time-series are analysed on a monthly basis. Figure 22 presents the weather state time-series by month. With a large amount of interannual variability, State 1 (dry everywhere) does not show an obvious trend for January or February however it has been trending up in March since a low point in the mid-1970s. There is a similarly timed upward trend for April, and a larger upward trend and decrease in interannual variability for May. In June there is no obvious trend, however since the lowest frequency in the early 1990s there have been no further years of low frequency. The upward trend returns for July and August, is not present for September, and for October increased in the late 1970s, and then has been steady until a record high value for the last year. It has also been trending up, slightly, for November and December since the late 1970s. These trends have significant implications for water resources, as the increase in dry days in early winter (Mar-Apr- May) reduces catchment saturation, which combined with more frequent dry days in mid-winter (Jul-Aug) would be expected to lead to reduced runoff across the entire winter period. Correspondingly, winter State 2 (wet everywhere) is trending down in May and, since the early 1980s, trending up in June. State 4 (wet in the south) is trending down in June. The monthly time-series of the atmospheric predictors (Figure 23) highlight the consistent trend of increasing 700 hpa DTD. MSLP is reasonably stationary for January and February whereas there is a noticeable step-change to high MSLP in March since the early 1980s. November MSLP also shows a recent gradual increasing trend since the late 1980s. The North South MSLP gradient predictor used in winter months has been decreasing for April, May, July, August, September (less so) and October, whilst June stands out as the only month when it is increasing. The corresponding plots of North South 700 hpa GPH gradient show recent increasing trend for April, June, September and October and decreasing trends for May, July, and August. The 850 hpa DTD predictor increases for all winter months, though often not as strongly as the corresponding 700 hpa DTD increases. The contrast between May and June is interesting. Both show increases in 700 and 850 hpa DTD whereas North-South MSLP and 700 hpa GPH gradient both decrease 5

34 in May and increase in June. These differences result in different state trends, as previously noted, for these months and therefore different trends in station rainfall. Station rainfall, by month, for a selection of stations across the study region (Figures 24 to 32) has in most cases decreased for May, particularly since the early 1990s, and increased in June consistent with the May increase in State 1 and decrease in all other states since the 1990s (or earlier) and the June increase in State 2 (Figure 22). Summary of methods and modifications (with reasons) Atmospheric predictors required for the NHMMs developed in Project were extracted from the NCEP/NCAR reanalysis for and ERA40 Re-analysis for The NHMMs selected in Project were driven by these Reanalysis predictor series, generating multiple realisations of multi-site daily precipitation series. Weather state time-series are an implicit output of the NHMM, with each day classified into a discrete state associated with a particular rainfall occurrence and synoptic pattern. The NHMM generated weather state and rainfall time-series were compared to the historical atmospheric predictor and rainfall series. This determined the adequacy of the NHMMs calibrated on 1986 to 2005 data in reproducing the observed variability of the 1958 to 2006 period, thus determining the key regional atmospheric drivers of observed precipitation variability on daily to interdecadal time scales. Summary of links to other projects Directly follows from Project Reanalyses Archive Extraction. The NHMMs developed in Projects and were used in this Project. Results of analysis of downscaled and observed synoptic (weather-state) and multi-site rainfall time-series will be compared with results of Projects 1.3.6, and 1.5.1, 1.5.2, & Publications arising from this project None. Acknowledgement The NCEP/NCAR Reanalysis data is provided by the Earth System Research Laboratory of the National Oceanic & Atmospheric Administration, U.S. Department of Commerce. The ERA-40 Reanalysis data is provided by the European Centre for Medium-Range Weather Forecasts. Thanks to Mark Collier, CSIRO Marine and Atmospheric research, for studiously maintaining local mirrors of these re-analysis products. This project is funded by the South Eastern Australian Climate Initiative. Recommendations for changes to work plan from your original table No changes. 6

35 Table 1. Selected period s weather state probabilities. State Summer 1958/9-1984/5 1985/6-2004/5 1986/7-1995/6 1996/7-2005/ Winter Dry years * Wet years # *Dry years = 1959, 1961, 1965, 1966, 1967, 1972, 1976, 1977, 1982, 1987, 1994, 1997, 2002, #Wet years = 1958, 1960, 1963, 1964, 1974, 1978, 1983, 1986, 1989, 1990,

36 (a) NNR Summer (b) ERA40 Summer 8

37 (c) NNR Winter (d) ERA40 Winter Figure 1. Weather state seasonal cycle for (a) NNR summer, (b) ERA40 summer, (c) NNR winter, and (d) ERA40 winter. 9

38 Figure 2. Summer atmospheric predictor seasonal cycle. Figure 3. Winter atmospheric predictor seasonal cycle. 10

39 Figure 4. Summer atmospheric predictor density plots by weather state. Figure 5. Winter atmospheric predictor density plots by weather state. 11

40 (a) NNR Summer (b) ERA40 Summer 12

41 (c) NNR Winter (d) ERA40 Winter Figure 6. Seasonal weather state time-series for (a) NNR summer, (b) ERA40 summer, (c) NNR winter, and (d) ERA40 winter. NNR is to 2006, ERA40 is to

42 Figure 7. Summer atmospheric predictors 1985/6 2004/5 versus 1958/9 1984/5. X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. Figure 8. Summer atmospheric predictors 1996/7 2005/6 versus 1986/7 1995/6. X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. 14

43 Figure 9. Winter atmospheric predictors versus X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. 15

44 Figure 10. Winter atmospheric predictors versus X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. 16

45 Figure 11. Winter atmospheric predictors dry years versus wet years. X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. 17

46 Figure 12. Winter atmospheric predictors 1983 (wet year) versus 1982 (dry year). X-axis of qq-plots is solid line of density plots, y-axis of qq-plots is dashed line of density plots. 18

47 Figure 13. Station observed (solid line, years with missing days not plotted) vs downscaled (box-plots) summer (upper plot) and winter (lower plot) precipitation amounts. The box plots depict the range of 100 simulation trials (the edges of the box represent the 25 percentile and the 75 percentile of the simulations). The horizontal dashed line is the long term observed mean. The vertical dashed line delineates the fitting period (1986 to 2005) from the validation period. Figure 14. As above for Station

48 Figure 15. As above for Station Figure 16. As above for Station

49 Figure 17. As above for Station Figure 18. As above for Station

50 Figure 19. As above for Station Figure 20. As above for Station

51 Figure 21. As above for Station

52 (a) January (b) February 24

53 (c) March (d) April 25

54 (e) May (f) June 26

55 (g) July (h) August 27

56 (i) September (j) October 28

57 (k) November (l) December Figure 22. Weather state time-series by month. 29

58 (a) January (b) February (c) March (d) April 30

59 (e) May (f) June (g) July (h) August (i) September 31

60 (j) October (k) November (l) December Figure 23. Predictor time-series by month. 32

61 Figure 24. Station observed monthly precipitation amounts. Smoothed line is trend. Figure 25. As above for Station

62 Figure 26. As above for Station Figure 27. As above for Station

63 Figure 28. As above for Station Figure 29. As above for Station

64 Figure 30. As above for Station Figure 31. As above for Station

65 Figure 32. As above for Station

66 Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) 1. Drive NHMMs with Reanalysis predictors 2. Analysis of weather-state and precipitation time-series Performance indicators 2 (1-3 dot points) NHMM downscaled precipitation series created Downscaled series compared to observed. Report (4-6 pages) written. Completion date 3 xx/xx/xxxx Budget 4 for Milestone ($) (SEACI contribution) Completed at each Milestone date Progress 5 (1-3 dot points) 1/9/ Completed None 31/12/ Completed (this is the report, 4-6 pages of text plus figures) Recommended changes to workplan 6 (1-3 dot points) None 38

67 FINAL REPORT FOR PROJECT DYNAMICAL DOWNSCALING OF COUPLED MODEL HISTORICAL RUNS PRINCIPAL INVESTIGATOR: DR. JOHN MCGREGOR, CSIRO Marine and Atmospheric Research, Tel: , Fax: , Address: PB1 Aspendale, Vic CO-AUTHOR: Dr. Kim Nguyen, CMAR. Date: January 2008

68 Abstract A 40-year simulation has been performed with a resolution of about 20 km over the GRDC region, using CCAM for downscaling from a Mk 3.0 coupled GCM simulation. The CCAM present-day climatology shows realistic detail compared to the coarser Mk 3.0 simulation. Over the GRDC region, seasonal rainfall biases are generally small. The simulated average daily maximum and minimum temperatures are found to be acceptable; the maximum temperatures are up to 3 degrees too cool in autumn and winter; the minimum temperatures are a little warm, except in Victoria in winter when they are up to 2 degrees too cool. The interannual variability of the mean annual rainfall of the simulation over the GRDC region is qualitatively similar to the observations, in regard to peak values and number of wet and dry episodes. The simulation produces an increasing trend of rainfall, whereas the observations show little trend over this period. The simulation also calculates an estimate of pan evaporation. No trend is found for pan evaporation over the GRDC region for the simulation. Date: January 2008 Project Objectives: Using CCAM, dynamically downscale the CSIRO Mk 3 GCM for a period corresponding to Compare the simulated climatology, in particular rainfall and temperature, against observations taken within the study region Produce derived fields (including surface fluxes and measures of potential evaporation) and compare them against available observations and observed trends 1. Model description CCAM is formulated on the quasi-uniform conformal-cubic grid. The CCAM regional climate simulations for SEACI use a C72 global grid (6 x 72 x 72 grid points). By using the Schmidt (1977) transformation with a stretching factor of 0.15, CCAM then achieves a fine resolution of 20 km for the central panel which is located over eastern Australia. The grid is illustrated in Figure 1. The previous SEACI Project described experiments using CCAM to downscale NCEP reanalyses from This project follows on from that project, using similar techniques to downscale from the Mk 3.0 coupled climate model, for model years corresponding to This report is laid out in a similar manner to that of Project The dynamical formulation of CCAM includes a number of distinctive features. The model is hydrostatic, with two-time-level semi-implicit time differencing. It employs semi-lagrangian horizontal advection with bi-cubic horizontal interpolation (McGregor, 1993; McGregor, 1996), in conjunction with total-variation-diminishing vertical advection. The grid is unstaggered, but the winds are transformed reversibly to/from C-staggered locations before/after the gravity wave calculations, providing improved dispersion characteristics (McGregor, 2005a). Three-dimensional Cartesian representation is used during the calculation of departure points, and also for the

69 advection or diffusion of vector quantities. Further details of the model dynamical formulation are provided by McGregor (2005b). CCAM includes a fairly comprehensive set of physical parameterizations. The GFDL parameterization for longwave and shortwave radiation (Schwarzkopf and Fels, 1991) is employed, with interactive cloud distributions determined by the liquid and icewater scheme of Rotstayn (1997). The model employs a stability-dependent boundary layer scheme based on Monin-Obukhov similarity theory (McGregor et al., 1993), together with non-local vertical mixing (Holtslag and Boville, 1993) and also enhanced mixing of cloudy boundary layer air (Smith, 1990). A canopy scheme is included, as described by Kowalczyk et al. (1994), having six layers for soil temperatures, six layers for soil moisture (solving Richard's equation), and three layers for snow. CCAM also includes a simple parameterization to enhance sea surface temperatures under conditions of low wind speed and large downward solar radiation, affecting the calculation of surface fluxes. The cumulus convection scheme uses the mass-flux closure described by McGregor (2003), and includes both downdrafts and detrainment. Figure 1. The C72 conformal-cubic grid used for the CCAM simulations. 2. Simulation design The current project uses CCAM to downscale from the Mk 3.0/M20 simulation, run for years corresponding to After 2000, the Mk 3.0 simulation branches into two simulations corresponding to the A2 and A1B scenarios. This report describes the present-day part of the simulation, corresponding to It should be noted that the Mk 3.0 model does not employ flux correction; hence there

70 are some biases of sea-surface temperatures (SSTs), up to 2 degrees near Australia, compared to observations. These average monthly biases are shown for a 30-year period ( ) for January and July, in Figure 2. The CCAM simulation uses daily SSTs from Mk 3.0, but with this monthly two-dimensional bias first subtracted. Sea-ice distributions are interpolated directly from the daily values of Mk 3.0; this simulation was submitted by CMAR to the Fourth Assessment Report of IPCC. Figure 2. SST biases in the Mk 3.0 coupled climate simulation, corresponding to for January (top) and July (bottom). For the prior simulation of Project downscaling from NCEP reanalyses, global nudging of winds above 500 hpa from the large-scale fields was employed, whilst outside the central high-resolution panel, gradually-increasing far-field nudging was also employed for MSL pressures and winds between 900 hpa and 500 hpa. This technique was adopted to help ensure that the north-south shifts of the jet-stream were captured. Likewise, we wish to have jet-stream locations in this CCAM simulation that are similar to those of the Mk 3.0 simulation. A new digital-filter technique is used in the present project, whereby large-scale features of MSL pressure and the winds above 500 hpa are similar to those of Mk 3.0; large-scale here is specified as

71 a length scale approximately the width of NSW. The digital filtering is efficiently achieved by a sequence of three one-dimensional passes, two of which are shown schematically (for a 60 km grid) in Figure 3, performed twice-daily. The model output was saved twice per day at 00 GMT and 12 GMT. These data have been post-processed onto a 0.2 degree grid covering 130E - 160E and 50S - 5S. Many prognostic and diagnostic fields have been saved. Figure 3. Schematic showing two one-dimensional passes of the Gaussian digitalfilter. 3. Seasonal averages of rainfall and temperature The 40-year ( ) monthly-mean CCAM rainfall was averaged to produce seasonal averages for December, January and February (DJF), March, April and May (MAM), June, July and August (JJA) and September, October and November (SON). These averages are compared against the observed seasonal rainfall, and the maximum and minimum temperatures provided by the Bureau of Meteorology. In general, CCAM simulates well the mean seasonal rainfall patterns (Figure 4). The large rainfall along the eastern coast is well captured for all seasons. The large rainfall along the Great Dividing Range and the eastern coast is also well captured, except in winter when it is deficient. Inland of the Great Dividing Range, summer rainfall is well captured, autumn and winter are somewhat lower than observed, whereas spring rainfall is somewhat larger than observed over inland New South Wales and Queensland. Seasonal rainfall biases are shown in Figure 5 for both CCAM and the host Mk3.0 simulations. For autumn, winter and spring the biases of Mk3.0 and CCAM are both generally quite small, with CCAM biases being a little smaller near the coastline. For summer, Mk3.0 has a rather large wet bias of 1-2 mm/day over much of New South Wales and Queensland, whereas the CCAM biases are again quite small.

72 Figure 4. Seasonally-averaged rainfall (mm/day), with observations (top) and CCAM (downscaled from Mk3.0, bottom). Figure 5. Bias of seasonally-averaged rainfall (mm/day) for Mk3.0 (top) and CCAM (bottom).

73 Figure 6. Seasonally-averaged maximum temperatures (degrees C), with observations (top) and CCAM (bottom), compared to observations. The seasonally-averaged daily-maximum surface air temperatures are shown for observations and CCAM in Figure 6. The maximum temperature patterns are generally well simulated for all seasons; terrain effects are well captured. Maximum temperature biases are shown in Figure 7 for both Mk3.0 and CCAM. Over New South Wales and Queensland the CCAM biases are generally less than 1 degree, whereas Mk3.0 maximum temperature biases are up to 5 degrees too cool, especially in autumn and winter. In summer and spring, maximum temperatures over Victoria from CCAM are up to 2 degrees too warm.

74 Figure 7. Bias of seasonally-averaged maximum temperatures (degrees C) for Mk3.0 (top) and CCAM (bottom), compared to observations. The seasonally-averaged daily-minimum surface air temperatures are shown for observations and CCAM in Figure 8. The minimum temperature patterns are also generally well simulated for all seasons, and again terrain effects are well captured. Minimum temperature biases are shown in Figure 9 for both Mk3.0 and CCAM. Mk3.0 minimum temperature biases (typically up to 1 degree too cool) are now generally smaller than for CCAM (up to 2 or 3 degrees too warm) for most seasons, and especially summer and spring. During summer, both Mk3.0 and CCAM are too warm over the southern half of the MDB

75 Figure 8. Seasonally-averaged minimum temperatures (degrees C), with observations (top) and CCAM (bottom). Figure 9. Bias of seasonally-averaged minimum temperatures (degrees C) for Mk3.0 (top) and CCAM (bottom), compared to observations.

76 4. Interannual variability of rainfall and pan evaporation The interannual variability of the simulation is illustrated by means of time series of the annual averages for rainfall and pan evaporation. The averages are taken over land points for two regions: Region 1: the GRDC southern region, 134E - 154E and 39S - 32S Region 2: a larger MDB region, 141E - 154E and 39S - 25S. Figure 10 shows the annual rainfall for region 1. While the individual high and low rainfall years do not coincide, this is not to be expected as the model is using GCM years rather than actual years. However, the peak values are similar. Both observations and CCAM have about 6 episodes with rather low (El Nino-like) rainfall, and about 6 episodes with higher rainfall. The simulation shows an increasing trend (2% per decade) for rainfall during this period, whereas the observations show no trend. The situation is similar for Region 2 (Figure 11); this region has generally higher average rainfall than Region 1, as it includes more of the tropical part of Australia Reg1 Rain CCAM Obs Figure 10. Time series of observed and simulated annual rainfall for region 1 (the southern GRDC region) for CCAM is shown for GCM model years, whilst OBS refers to actual years. Trend lines are also shown for CCAM (increasing trend) and observed (little trend).

77 3 2.5 Reg2 Rain CCAM Obs Figure 11. As for Figure 10, but showing the time series for the larger region 2. Plotted for observed and simulated. CCAM is shown for GCM model years, whilst OBS refers to actual years. Trend lines are also shown for CCAM (increasing trend) and observed (little trend). In the CCAM simulation, we also explicitly calculate pan evaporation, by predicting at each time step the water temperature for a pan located at each grid point. Sensible and latent heat fluxes are calculated as part of this process. Figure 12 shows the time series for the annual pan evaporation averaged over both regions. The Figure also shows the pan evaporation for from our prior simulation downscaling from NCEP reanalyses (Project 1.3.6). During this period, there is a negligible trend in pan evaporation simulated for either simulation or either region. Region 2 is on average warmer than region 1 and produces larger pan evaporation, as is to be expected. The current simulation produces somewhat (by up to 10%) larger estimates of pan evaporation than our prior simulation that downscaled from NCEP, for reasons we are yet to clarify (presumably related to one or more of low-level temperature, humidity or wind speed being different between the two simulations).

78 NCEP: Pan Evaporation Reg 1 Reg A2: Pan Evaporation Reg 1 Reg Figure 12. Time series of simulated pan evaporation (mm/year) averaged over regions 1 and 2 for The upper plot is for the prior CCAM simulation downscaling from NCEP reanalyses. The lower plot is for the current CCAM simulation, for Mk3.0 GCM years. Trend lines are also shown. 5. Rainfall extremes Rainfall extremes have been calculated from the daily output files for the maximum daily rainfall on a 0.5 degree latitude/longitude grid. These are compared in Figure 13 against Bureau of Meteorology values plotted on a similar 0.25 degree grid. The simulated maxima agree quite well along the coastal zone (up to about 250 mm/day). The interior extreme values are somewhat low. This will be partly due to interpolation

79 from the model 20 km grid to sparser points at 0.5 degrees (about 50 km). We plan to investigate this behaviour further. Figure 13. Maximum daily rainfall (mm/day) at each point (on a 0.5 degree grid) for the duration of the 40-year run. CCAM is shown on the left, and observations on the right. 6. Summary This report presents some results from a 40-year present-day simulation downscaling a present-day simulation of the Mk 3.0 coupled atmosphere-ocean GCM by using the CCAM variable-resolution model at 20 km resolution over eastern Australia.. Many detailed 6-hourly output fields were produced and these can be made available to other researchers. The CCAM present-day climatology shows realistic detail compared to the coarser Mk 3.0 simulation. CCAM reproduces well the mean seasonal rainfall patterns; the only deficiencies are that the winter rainfall is somewhat low along the east coast, inland autumn and winter rainfall is somewhat low, and spring rainfall is somewhat large over inland New South Wales and Queensland. The mean daily temperature patterns are also well captured, and generally the maximum temperatures have smaller errors than Mk 3.0; terrain effects are also better captured. The largest maximum temperature errors (up to 2 degrees) are in summer and spring over Victoria. The CCAM minimum temperatures are up to 2 or 3 degrees too warm, especially for summer and spring, whereas Mk 3.0 is typically about 1 degree too cool. During summer, both CCAM and Mk3.0 are too warm over the southern half of the MDB. The interannual variability of the mean annual rainfall of the simulation over the GRDC region is qualitatively similar to the observations, in regard to peak values and number of wet and dry episodes. The simulation produces an increasing trend of rainfall, whereas the observations show little trend over this period. The simulation also calculates an estimate of pan evaporation. Over the GRDC region, the values are about 10% larger than for the prior simulation downscaling from NCEP reanalyses. No trend is found for pan evaporation over the GRDC region for either of the simulations.

80 Maximum daily rainfall throughout the 40-year simulation was also calculated. Near the coastline it agrees quite well with observed values. Inland, the maximum values are somewhat lower than observed. The ongoing project 2.1.5, continues the simulation on to 2100 for both the A2 and A1B scenarios. List of publication titles: McGregor, J. L., Nguyen, K. C., and M. Thatcher, 2008: Regional climate simulation at 20 km using CCAM with a scale-selective digital filter. Research Activities in Atmospheric and Oceanic Modelling Report No. 38 (ed. J. Cote), WMO/TD. Thatcher, M., and J. L. McGregor, 2008: Using a scale-selective filter for dynamical downscaling with the conformal cubic atmospheric model. Submitted to Mon. Wea. Rev. Acknowledgements Dr. Marcus Thatcher co-developed the digital filter used in this simulation. Dr. Ian Watterson kindly produced Figure 13 displaying the extreme rainfall values. This project References Holtslag, A. A. M., and B. A. Boville, 1993: Local versus non-local boundary layer diffusion in a global climate model. J. Climate, 6, Kowalczyk, E. A., J. R. Garratt, and P. B. Krummel, 1994: Implementation of a soilcanopy scheme into the CSIRO GCM -regional aspects of the model response. CSIRO Div. Atmospheric Research Tech. Paper No. 32, 59 pp. McGregor, J. L., 1993: Economical determination of departure points for semi- Lagrangian models. Mon. Wea. Rev., 121, McGregor, J. L., 1996: Semi-Lagrangian advection on conformal-cubic grids. Mon. Wea. Rev., 124, McGregor, J. L., 2003: A new convection scheme using a simple closure. In "Current issues in the parameterization of convection", BMRC Research Report 93, McGregor, J. L., 2005a: Geostrophic adjustment for reversibly staggered grids. Mon. Wea. Rev., 133, McGregor, J. L., 2005b: C-CAM: Geometric aspects and dynamical formulation [electronic publication]. CSIRO Atmospheric Research Tech. Paper No. 70, 43 pp. McGregor, J. L., H. B. Gordon, I. G. Watterson, M. R. Dix, and L. D. Rotstayn, 1993: The CSIRO 9-level atmospheric general circulation model. CSIRO Div. Atmospheric Research Tech. Paper No. 26, 89 pp. Rotstayn, L. D., 1997: A physically based scheme for the treatment of stratiform clouds and precipitation in large-scale models. I: Description and evaluation of the microphysical processes. Quart. J. Roy. Meteor. Soc., 123, Schmidt, F., 1977: Variable fine mesh in spectral global model. Beitr. Phys. Atmos., 50, Schwarzkopf, M. D., and S. B. Fels, 1991: The simplified exchange method revisited: an accurate, rapid method for computation of infrared cooling rates and fluxes. J. Geophys. Res., 96,

81 Smith, R. N. B., 1990: A scheme for predicting layer clouds and their water content in a general circulation model. Quart. J. Roy. Meteor. Soc., 116, Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) 1. Compare alternative downscaling techniques (wind forcing from upperair versus farfield) 2. Complete the simulation from at 20 km resolution Performance indicators 2 (1-3 dot points) Complete 10- year trial simulations, and select the most appropriate downscaling methodology from the Mk 3.5 GCM. Completion of full simulation Satisfactory climatology of rainfall and maximum and minimum temperatures Completio n date 3 Xx/xx/xxx x Budget 4 for Mileston e ($) Completed at each Milestone date Progress 5 (1-3 dot points) 31/3/07 33k Trial simulations were performed, comparing methods for forcing by upper winds Downscaling was compared from both Mk 3.0 and Mk /6/07 34k A present-day climate simulation has been carried out downscaling from Mk 3.0 for years corresponding to Recommended changes to workplan 6 (1-3 dot points)

82 3. Comparison of model output with available observations Complete an analysis of rainfall, extremes and trends against available observations Produce an analysis of maximum and minimum temperatures Analyse measures of potential evaporation produced by the simulation Examine the model output with regard to behaviour during El Nino years. Produce 8 page report. 31/12/07 33k The various modelled fields have been compared against observations and a report produced.

83 Final report for Project Drive Statistical Downscaling Models with GCM Predictor Sets Principal Investigator: Steve Charles CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Tel: Fax: Co-Authors: Guobin Fu CSIRO Land and Water, Private Bag 5, Wembley WA 6913 Completed: January 2008

84 Abstract In the preceding Project Extraction of statistical downscaling predictors from GCM output and assessment of GCM performance at the regional scale, the CSIRO Mk3.5 and CCAM (boundary forced by CSIRO Mk3.0) and two international GCMs (GFDL-CM2.0, from the USA, and MRI-CGM2.3.2a, from Japan) where selected based on their current-day reproduction of the predictor fields required by the Nonhomogeneous Hidden Markov Model (NHMM) stochastic downscaling model (SDM). These predictor sets were determined in Project Calibration of Statistical Downscaling Models. In this Project (2.1.3), predictor sets have been extracted from the selected climate models for three emission scenarios (A2, A1B and B1) for future periods according to availability of daily model output. These projected predictor sets were then used to drive the previously calibrated NHMMs. Initial examination of the properties of the resultant regional rainfall projections, in terms of consistency in weather state probabilities, is investigated. Weather states are the distinct regional rainfall occurrence patterns determined by the NHMM. The followon Project Evaluation of Stochastically Downscaled Projections will examine these multi-site, daily projections in detail. Significant research highlights, breakthroughs and snapshots Predictor sets for three IPCC SRES emissions scenarios (A2, A1B and B1) for a continuous period (CSIRO climate models) or for multiple time slices (international GCMs) where used to drive NHMMs to produce multiple realisations (stochastic rainfall model conditional on input atmospheric predictor series) of multi-site daily rainfall series. Projected changes in weather state probabilities (i.e. the frequency of distinct regional rainfall occurrence patterns) are reasonably consistent between GCMs and respond according to the strength of the emissions scenarios. Overall there are projected increases in dry state and decreases in wet state probabilities. This initial, and limited, assessment of projected changes is a precursor to Project Evaluation of Stochastically Downscaled Projections, where detailed analysis of changes in hydrologically relevant properties such as return periods of low and high rainfall accumulations for multi-day to multiyear totals (e.g,. 5-day maximum rainfall; 1-in-10 year annual rainfall), as well as changes in the spatial distribution of rainfall, will be undertaken. Statement of results, their interpretation, and practical significance against each objective Objective 1: For selected GCMs and emissions scenarios, produce statistically downscaled projections of multi-site daily rainfall using the NHMMs selected in Projects and Table 1 summarises the climate model output extracted to produce the predictor sets required to drive the NHMM SDMs. The CSIRO Mk3.5 A1B run is continuous until 2200, the A2 run is continuous until 2100, and the B1 run is continuous until The CCAM runs are continuous until The international climate models have daily output for the IPCC 4AR designated twenty-year time-slices of 2046 to 2065 and 2081 to 2100 for all scenarios, and in addition post-2100 stabilisation for 2181 to 2200 and 2281 to 2300 for A1B and B1 scenarios. In addition, the climate of the 20 th century run (20C3M) for 1961 to 2000 was also downscaled. This provides a baseline

85 to compare the projections against. This multi-model, multi-scenario ensemble of projected atmospheric predictor sets was used to drive the NHMMs, producing stochastic (multiple) realisations of multi-site daily rainfall. Tables 2 and 3, for summer (Nov-Mar) and winter (Apr-Oct) respectively, summarise the mean weather state probabilities obtained from 100 stochastic realisations. Biases are evident when comparing the mean probabilities obtained using observed atmospheric predictors (from NCEP/NCAR Reanalysis, NNR, and European ERA40 re-analysis) to those obtained using climate model 20C3M (i.e. current climate) predictors. Noticeable biases for the summer include underestimation of State 3 and State 5 and for winter underestimation of State 2 and overestimation of State 4. Reassessing the climate model predictors to determine the cause(s) of these biases is ongoing and will be reported in the subsequent report for Project The projected changes in weather state probabilities are mostly consistent between GCMs and within scenarios. For summer (Table 2), most GCMs for A1B and A2 scenarios project an increase in State 1 ( dry everywhere, refer to Project report Figure 3) probabilities, the exception being CCAM where probabilities do not change. The B1 scenario State 1 probability increases for Mk3.5 and GFDL but not for CCAM and MRI. Changes in other states are often quite small, and these more subtle changes and their effect on rainfall will be investigated in the follow-on Project For winter (Table 3) State 1 ( dry everywhere, Project report Figure 4) increases for all GCMs and scenarios, although changes are not as large for the B1 scenario. State 2 ( wet everywhere ) decreases for all GCMs and scenarios, as does State 3 ( moderately wet everywhere ) and State 5 ( wet everywhere, moderate in northwest ). State 4 ( wet south ) remains almost unchanged. Again, the rainfall projections will be assessed in detail in Project Summary of methods and modifications (with reasons) Four climate models were selected on the basis of their performance over south-eastern Australia, as determined in Project For these selected GCMs, the atmospheric predictor fields required for the calibrated NHMMs were extracted for three emissions scenarios (A2, A1B, B1) for all periods available. It was initially envisioned that long stabilisation periods would be downscaled to provide stationary series to assess in the subsequent Project Evaluation of Stochastically Downscaled Projections. However long stabilisation periods were only available for the CSIRO Mk3.5 A1B and B1 scenarios. For the international climate models, only twenty year time-slices where available post-stabilisation for A1B and B1 scenarios for 2181 to 2200 and 2281 to No stabilisation runs were available for CCAM. The predictor sets extracted for all runs (Table 1) were statistically downscaled using the previously calibrated NHMMs, to produce multiple realisations of projections of multi-site daily rainfall. Weather state mean probabilities summarise the projected rainfall climatologies obtained from downscaling these projections (Tables 2 and 3, for summer and winter respectively).

86 Summary of links to other projects GCMs used in this project were selected on the basis of their performance as assessed in Project Extraction of statistical downscaling predictors from GCM output and assessment of GCM performance at the regional scale. The NHMM downscaled projections will be assessed in detail in Project Evaluation of Stochastically Downscaled Projections. The NHMM derived projections should be compared to the results and findings of Project [add title here]. Publications arising from this project None. Acknowledgement Thanks to Tony Hirst, Martin Dix, John McGregor, and Kim Nguyen of CSIRO Marine and Atmospheric Research (CMAR) for making CSIRO Mk3.5 and CCAM model output available. The Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the Working Group on Coupled Modelling (WGCM) made the IPCC 4AR multi-model dataset available with the support of the Office of Science, U.S. Department of Energy. Thanks to Siobhan O Farrell and Mark Collier (CMAR) and Aurel Moise and Lawson Hanson (BoM) for making this IPCC 4AR archive available locally. This project is funded by the South Eastern Australian Climate Initiative. Recommendations for changes to work plan from your original table No changes. Table 1. Summary of GCMs, SRES scenarios & periods downscaled Model - scenario Continuous to year Time-slice Time-slice Time-slice CSIRO Mk3.5 A1B A2 B1 CCAM A1B B1 GFDL2.0 A1B A2 B1 MRI A1B A2 B Time-slice

87 Table 2 Summer weather state mean probabilities GCM SRES Period State 1 State 2 State 3 State 4 State 5 State 6 NNR Observed ERA40 Observed Mk3.5 A1B Mk3.5 A Mk3.5 B CCAM A1B CCAM A GFDL 20C3M GFDL A1B GFDL A GFDL B MRI 20C3M MRI A1B MRI A MRI B

88 Table 3 Winter weather state mean probabilities GCM SRES Period State 1 State 2 State 3 State 4 State 5 NNR Observed ERA40 Observed Mk3.5 A1B Mk3.5 A Mk3.5 B CCAM A1B CCAM A GFDL 20C3M GFDL A1B GFDL A GFDL B MRI 20C3M MRI A1B MRI A MRI B

89 Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) 1. Extract required daily atmospheric predictor fields for selected GCMs and scenarios 2. Produce NHMM input files Performance indicators 2 (1-3 dot points) GCM output for selected GCMs and scenarios extracted NHMM input files created Completion date 3 xx/xx/xxxx Budget 4 for Milestone ($) (SEACI contribution) Completed at each Milestone date Progress 5 (1-3 dot points) 1/9/ Completed None 1/10/ Completed None Recommended changes to workplan 6 (1-3 dot points) 3. Generate multiple simulations of NHMM downscaled rainfall conditional on GCM predictor datasets Downscaled projections produced. 31/12/ Completed. This report is a summary. None

90 Final report for Project #2.2.2 Projections of Water Availability Principal Investigator: Francis Chiew Contact details: (02) , Co-Authors: Freddie Mpelasoka, David Post Contact details: Completed: 20 March

91 Abstract The first part of this project assessed simple methods for transforming climate outputs from global climate models (GCMs) to drive hydrological models. Climate change projections generally come from GCMs, but the information provided by GCMs is at a resolution that is too coarse for direct use in hydrological modelling. Statistical and dynamic downscaling methods can be used to downscale GCM simulation to catchment scale rainfall, but they involve significant development, data needs and modelling run times. The simpler empirical methods, based on scaling or transforming a historical climate series to obtain a future climate series, offer a more immediate solution. The results from this study indicate that the constant scaling method can be used in most applications to transform climate outputs from GCMs to drive hydrological models. Because the constant scaling method is very simple, it can be applied to outputs from a number of runs of different GCMs, therefore taking into account the large uncertainties associated with global warming projections and GCM modelling of local climate. However, where more detailed analyses of runoff distribution is required, the daily scaling and daily translation methods are potentially better, particularly if used to transform outputs from a GCM that has skill in modelling extreme rainfall and daily rainfall series. The second part of this project developed future projections of catchment water yield for 0.05 o (~ 5 km x 5 km) grid cells across the SEACI region. The hydrological modelling is carried out using two lumped conceptual daily rainfall-runoff models (SIMHYD and Sacramento), calibrated against streamflow data from about 219 catchments. The daily climate data from SILO Data Drill are used to model the historical runoff. The future climate series for ~2030 is obtained by scaling the historical daily rainfall and PET data using the daily scaling method, informed by the IPCC SRES A1B global warming projection (0.9 o C warmer in ~2030 relative to ~1990) and 15 GCMs. There are considerable differences in the rainfall projections between GCMs, although the majority of GCMs shows a decrease in future mean annual rainfall, particularly in southern SEACI region. The modelling results indicate that the mean annual runoff in ~2030 relative to ~1990 for the SRES A1B global warming scenario in northern SEACI region (central and northern MDB) will change by -25 percent to +20 percent (median of -5 percent), and the mean annual runoff in southern SEACI region (southern MDB and Victoria) will change by -30 percent to 0 (median of -15 percent) Significant research highlights, breakthroughs and snapshots The project assessed runoff results from a rainfall-runoff model driven with future climate inputs obtained from different simple scaling/translation methods for transforming climate outputs from GCMs to drive hydrological models. The results provide guidance on the choice of appropriate methods to obtain future climate series for climate change impact on water resources studies. Because these methods are simple to use, they can easily consider different future scenarios, therefore taking into account the large uncertainties associated with global warming projections and GCM modelling of local climate. The results from this study have been written as an international conference paper and a journal paper (see Publications). A summary of the results is given in Objective 1 below.

92 The project calibrated and parameterised two rainfall-runoff models (SIMHYD and Sacramento) to estimate daily runoff from daily rainfall and PET data for 0.05 o x 0.05 o (~ 5 km x 5 km) grid cells across the SEACI region. The parameterised models can be used for various rainfall-runoff modelling applications, including future climate change impact assessment with new and improved projections. The project estimated the likely range of changes to catchment water yield and runoff characteristics by ~2030 relative to ~1990, for 0.05 o grid cells across the SEACI region. The methods and results will be published as a SEACI report and a journal paper. A summary of the method and key results is given in Objective 2 below. Statement of results, their interpretation, and practical significance against each objective Objective 1: To assess simple methods for transforming climate outputs from GCMs to drive hydrological models Most climate change impact on water resources studies involve the use of a hydrological model where: (a) the model is calibrated against historical streamflow data; (b) the model is driven with rainfall and PET series for a future climate using the same optimised parameter values; and (c) the modelled streamflows for the future and historical climates are compared to provide an estimate of the climate change impact on streamflow. The climate change projections generally come from global climate models (GCMs). However, GCMs provide information at a resolution that is too coarse for direct use in hydrological modelling. Statistical and dynamic downscaling methods can be used to downscale GCM simulations to catchment scale rainfall, but they involve significant development, data needs and modelling run times. The simpler empirical methods, based on scaling or transforming a historical climate series to obtain a future climate series, offer a more immediate solution. Because they are simple to use, they can be applied to outputs from a number of runs of different GCMs, therefore taking into account the large uncertainties associated with global warming projections and GCM modelling of local climate. This study assessed runoff results from a rainfall-runoff model driven with future climate inputs obtained from three simple scaling/translation methods for transforming climate outputs from several GCMs transient experiments: constant scaling ; daily scaling ; and daily translation. In the constant scaling and daily scaling methods, the historical rainfall and PET series are scaled by the changes indicated by the GCM to provide the future climate inputs. In the constant scaling method, the entire climate series (in each of the four seasons) is scaled by the same factor. In the daily scaling method, the different rainfall amounts are scaled differently, as indicated by the different changes to the different rainfall ranks/percentiles simulated by the GCM. In the daily translation method, a relationship between GCM simulation of the present climate and the observed 3

93 catchment-scale climate is established (at the different rainfall ranks/percentiles) and this relationship is used to convert the future GCM time series simulation to catchment-scale rainfall time series. The constant scaling method is used in most hydrological impact modelling studies due to its simplicity. The daily scaling method is likely to be a better method because it considers different changes to different rainfall amounts. This is potentially important because many GCMs indicate that the extreme rainfall in a future climate is likely to increase, even in some regions where the mean annual rainfall is expected to decrease. As high rainfall events generate significant runoff, the runoff estimated from the daily scaling method will be higher than that estimated by the constant scaling method. However, the results from this study indicate that the difference between the constant scaling and daily scaling methods is small compared to the mean annual runoff estimated using rainfall projections from different GCMs. In addition, GCM simulation of extreme climate and daily series is relatively poor compared to its ability to simulate the mean climate. The daily translation method considers changes in the sequencing of future rainfall time series, and runoff results from the daily translation method can be significantly different to the constant scaling and daily scaling methods. In summary, the constant scaling method can be used in most applications to transform climate outputs from GCMs to drive hydrological models. However, where more detailed analyses of runoff distribution is required, the daily scaling and daily translation methods are potentially better, particularly if the GCM used have skill in modelling extreme rainfall and daily rainfall series. Objective 2: To develop future projections of catchment water yield and runoff characteristics across the SEACI region This study developed future projections of catchment water yield and runoff characteristics for 0.05 o x 0.05 o (~ 5 km x 5 km) grid cells across the SEACI region. The hydrological modelling is carried out using two lumped conceptual daily rainfallrunoff models (SIMHYD and Sacramento), calibrated against streamflow data from about 219 catchments. The runoff for each grid cell is modelled using the optimised parameter values from the closest calibration catchment. The daily climate data from the SILO Data Drill ( are used to model the historical runoff. The future climate series for ~2030 is obtained by scaling the historical daily rainfall and areal PET data using the daily scaling method, informed by the IPCC SRES A1B global warming projection and 15 GCMs. The future climate series is then used to drive the SIMHYD model (using the same parameter values for modelling the historical climate) to estimate the future runoff (~2030 relative to ~1990). The key results from the modelling are summarised below. The global warming by ~2030 relative to ~1990 in the IPCC SRES A1B greenhouse gas emission scenario is 0.9 o C. The mean annual APET in ~2030 relative to ~1990 will increase by 2 to 4 percent. 4

94 There are considerable differences in the rainfall projections between GCMs, although the majority of GCMs shows a decrease in future mean annual rainfall. In the southern SEACI region, more than three-quarters of the GCMs show a decrease in future mean annual rainfall. Almost all the GCMs indicate that future winter rainfall will decrease. The projections from the 15 GCMs for future mean annual rainfall for ~2030 relative to ~1990 for the SRES A1B scenario range from a 10 percent decrease to a 5 percent increase (median of 2 percent decrease) in northern SEACI region (central and northern MDB) and a percent decrease to no change (median of 5 percent decrease) in southern SEACI region (Victoria and southern MDB). The rainfall-runoff modelling using the rainfall projections indicate that future mean annual runoff for ~2030 relative to ~1990 for the SRES A1B scenario will change by -25 percent to +20 percent (median of -5 percent) in northern SEACI region and by -30 percent to 0 in southern SEACI region (median of -15 percent). The projected decrease in mean annual runoff in southern SEACI region is higher than in northern SEACI region because: the projected decrease in rainfall is slightly higher in the south; and most of the projected rainfall decrease is in winter when most of the runoff in the south occurs. The method used here is very similar to the method used for the Murray-Darling Basin Sustainable Yields Project, but the results presented here are specifically for the SRES A1B scenario. The climate change projections are also similar to the CSIRO and Bureau of Meteorology projections ( which use 23 GCMs compared to the 15 GCMs used here. It is likely that the use of fewer GCMs chosen based on their abilities to simulate present-day Australian climatology will give more consistent climate change projections, and this is being addressed in SEACI Project The method and the calibrated rainfall-runoff models can be used to estimate climate change impact on runoff for different future periods and different global warming scenarios, and to update results as climate change projections improve. Summary of links to other projects This project is very closely related to Project Climate change projections and climate inputs from other projects in SEACI Theme 2 can also be used to drive the hydrological models that have been developed for the SEACI region in this project. Publications arising from this project Chiew FHS (2006) An overview of methods for estimating climate change impact on runoff. Proceedings of the 30 th Hydrology and Water Resources Symposium, Launceston, December 2006, Engineers Australia, CDROM (ISBN ). 5

95 Chiew FHS, Post DA, Teng J, Vaze J and Perraud J-M (2008) Modelling climate change impact on runoff in southeast Australia. Journal Paper, In Preparation. Mpelasoka F and Chiew FHS (2008) Runoff projection sensitivity to rainfall scenario methodology. Proceedings of International Congress on Environmental Modelling and Software (iemss 2008), Barcelona, July 2008, In Press. Mpelasoka F and Chiew FHS (2008) Influence of rainfall scenario methodology on runoff projections. Submitted to Journal of Hydrometeorology. Post DA, Chiew FHS, Vaze J, Teng J, Perraud J-M and Viney NR (2008) Future Runoff Projections (~2030) for Southeast Australia. SEACI Report, Completed and being assessed for publication. Acknowledgement This project is funded by the South Eastern Australian Climate Initiative. 6

96 Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) 1. Rainfallrunoff modelling across study area for climate change impact assessment Performance indicators 2 (1-3 dot points) Complete set-up of rainfall-runoff model for climate change impact assessment across study area Completion date 3 xx/xx/xxxx Budget 4 for Milestone ($) Completed at each Milestone date Progress 5 (1-3 dot points) 30/06/ k The SIMHYD rainfallrunoff modelling for climate change impact assessment has been set up to simulate daily runoff for ~ 5 km x 5 km grid cells across Australia. Recommende d changes to workplan 6 (1-3 dot points) 7

97 2. Projections of catchment water yield and runoff characteristics across study area for ~2030 and ~2070 Report/Paper on methods and data files of runoff projections for MDBC and DSE 31/12/ k Assessed runoff results from a rainfall-runoff model driven with future climate inputs obtained from different simple scaling and translation methods for transforming climate outputs from GCMs. Results have been written up as two research papers. Estimated the potential range of changes to catchment water yield and runoff characteristics by ~2030 relative to ~1990, for ~ 5 km grid cells across SEACI region. A SEACI report has been completed and a journal paper on the research is being written. The method and the calibrated rainfallrunoff models can be used to estimate climate change impact on runoff for different future periods and different global warming scenarios, and to update results as climate change projections improve. The 2030 runoff projections are shown as an indication of the potential impact of climate change on future runoff. In the method used here, the potential impact in 2070 is likely to be 2.5 times bigger than the impact in 2030 (A1B warming of 2.1 o C by 2070 compared to 0.9 o C by 2030). The actual numbers are only meaningful in the context of the application, and we have provided guidance to DSE and MDBC on this. 8

98 Final report for Project #3.1.1 Statistical Seasonal Streamflow Forecasting Model Principal Investigator: Francis Chiew Contact details: (02) , Co-Authors: QJ Wang, David Robertson Contact details: Completed: March 2008

99 Abstract This project developed a joint probability approach for seasonal forecasting of streamflows at multiple sites. The approach forecasts future streamflows from antecedent streamflows and seasonal climate indices, including SOI and ENSO. The joint probability approach quantifies uncertainty and maximises the information extracted from available data, allowing the use of records with non-concurrent and missing data. Cross validation results show that the approach produces forecasts of spring streamflows for the Murrumbidgee catchment with reasonable skill using only antecedent flows and SOI as predictors. Research into the best seasonal climate indices forecasting streamflows at different forecast lead times and at locations throughout the Murray Darling Basin will continue in Project Significant research highlights, breakthroughs and snapshots A statistical multi-site seasonal streamflow forecasting model was developed. The model is a significant improvement over existing statistical models it is a multisite model, it works with non-concurrent data and it facilitates the transfer of learning acquired from regional applications. The model was successfully tested using multi-site streamflow data from the Murrumbidgee catchment. A conference paper describing the model and application to the Murrumbidgee catchment will be presented at the Hydrology and Water Resources Symposium in Adelaide in April A draft journal describing the model, its application to the Murrumbidgee catchment and analysis of its forecasting ability has been prepared. Statement of results, their interpretation, and practical significance against each objective Objective 1: To assess and further develop statistical seasonal streamflow forecasting model for forecasting reservoir inflows across the MDB. The relationship between El Niño/Southern Oscillation (ENSO) and climate is the scientific basis of seasonal climate forecasts provided by research institutions and meteorological agencies throughout the world (e.g., The correlation between Australia s hydroclimate and ENSO is amongst the strongest in the world (Chiew and McMahon, 2002). The relationship between streamflow and ENSO, and the serial correlation in streamflow, can be exploited to forecast streamflow several months/seasons ahead. The forecast can be used to help manage water resources systems, and allow decisions on irrigation water allocations and environmental flows to be more realistically based (Chiew et al., 2003). To make risk-based decisions on activities that rely on water resources systems, a probabilistic streamflow forecast is required. Several probabilistic seasonal streamflow forecast models have been developed in Australia (e.g., Sharma, 2000; Piechota et al., 2001; and Chiew and Siriwardena, 2005). These models are generally

100 based on non-parametric approaches and developed for single-site, although Sharma et al. (2006) extended the model for forecasting streamflow at multiple sites. This project developed and tested a parametric joint probability approach for seasonal forecasting of streamflows at multiple sites. A Box-Cox transformed multivariate normal distribution is used to model the joint distribution of future streamflows, ENSO signals and antecedent streamflows. A Bayesian inference of model parameters and uncertainties is implemented using Markov Chain Monte Carlo (MCMC) simulations. The approach provides probabilistic forecasts of streamflows at multiple sites, which can be used to drive river system models to provide probabilistic indications of water availability and water allocation several months ahead. The approach was implemented and successfully tested on the Murrumbidgee catchment. The cross-validation results show that the approach can forecast streamflows several months ahead with reasonable skill (relatively high LEPS skill scores and Nash- Sutcliffe model efficiencies) and maintains the cross-site correlations. The advantages of this approach are: it models inter-site correlations for forecasting streamflows at multiple sites (important for river system applications); it quantifies the uncertainty of a forecast in terms of a well defined probability distribution; it maximises the extraction of information from available data without the need to discard non-concurrent data or infill missing data; and it provides a parametric structure for quantifying relationships between variables and therefore facilitates the transfer of learning acquired from regional applications. This project therefore directly addresses the first research question in Theme 3, by developing an improved statistical multi-site seasonal forecasting model and demonstrating that multi-site streamflows can be forecast with reasonable skill several months/seasons ahead. Summary of links to other projects The improved statistical multi-site seasonal forecasting model developed in this project can be used in other projects in Theme 3 and elsewhere. This project is very closely related to Project 3.2.1/ Project 3.2.1/3.2.5 will further test the approach developed here using a larger data set and incorporate advantageous features of this statistical approach and outputs from climate models (dynamic modelling approach) to improve seasonal streamflow forecasting. Publications arising from this project Wang QJ, Chiew FHS and Roberson DE (2008) A joint probability approach for seasonal forecasting of streamflows at multiple sites. Proceedings of the 31st Hydrology and Water Resources Symposium (Water Down Under 2008), Adelaide, April 2008, Engineers Australia. Wang QJ, Robertson DE and Chiew FHS (2008) Development and testing of a multisite seasonal streamflow forecasting model. In Preparation. Chiew FHS, Wang QJ and Robertson DE (2008) SEACI Project 3.1.1: Statistical seasonal streamflow forecasting model. SEACI Final Report.

101 Acknowledgement This project is funded by the South Eastern Australian Climate Initiative. References Chiew FHS and McMahon TA (2002) Global ENSO-streamflow teleconnection, streamflow forecasting and interannual variability. Hydrological Sciences Journal, 47, Chiew FHS, Zhou SL and McMahon TA (2003) Use of seasonal streamflow forecasts in water resources management. Journal of Hydrology, 270, Chiew FHS and Siriwardena LW (2005) Probabilistic seasonal streamflow forecasting methods. Proceedings of the 29th Hydrology and Water Resources Symposium, Canberra, Australia, February 2005, Engineers Australia, CDROM (ISBN ). Piechota TC, Chiew FHS, Dracup JA and McMahon TA (2001) Development of an exceedance probability streamflow forecast. ASCE Journal of Hydrologic Engineering, 6, Sharma A (2000) Seasonal to interannual rainfall probabilistic forecasts for improved water supply management: Part 3 - A nonparametric probabilistic forecast model. Journal of Hydrology, 239, Sharma A, Ling F, Murray R, Stolp S and Allison J (2006) Multiple reservoir probabilistic inflow forecasting system: Application to the Hydro-Tasmania reservoir network. Proceedings of the 30 th Hydrology and Water Resources Symposium, Launceston, Australia, December 2006, Engineers Australia, CSDOM (ISBN ). Project Milestone Reporting Table To be completed prior to commencing the project Milestone description 1 (brief) (up to 33% of project activity) Performance indicators 2 (1-3 dot points) Completion date 3 xx/xx/xxxx Budget 4 for Milestone ($)(SEACI contributio n) Completed at each Milestone date Progress 5 (1-3 dot points) Recommended changes to workplan 6 (1-3 dot points)

102 1. Application of NSFM to forecast reservoir inflows across the MDB 2. Assessment and further development of statistical seasonal streamflow forecasting model for forecasting reservoir inflows across the MDB NSFM model set-up for forecasting reservoir inflows at key locations across the MDB Statistical model for forecasting reservoir inflows across the MDB and report on model development and testing 31/8/ k (1) Streamflow data collated for over 200 catchments in SEACI region and a range of ENSO indices prepared. (2) NSFM applied to the above data. (3) Predictors previously not used but suggested by other researchers are being investigated. (4) A parametric joint probability approach was developed for both single and multiple site applications. 31/12/ k (1) A parametric joint probability model for seasonal forecasting of streamflows at multi-sites was developed and successfully tested. (2) A conference paper describing the method was written. (3) A draft journal paper on the application of the method and SEACI final report have been prepared.

103 Final report for Project #3.2.7 Implementing Physical-Statistical Climate Models Principal Investigator: Edward P. Campbell CSIRO Mathematical & Information Sciences Private Bag 5, PO Wembley, WA tel: (08) Co-Authors: Mark J. Palmer CSIRO Mathematical & Information Sciences Private Bag 5, PO Wembley, WA tel: (08) Completed: 17 January 2008