Multi-model ensemble (MME) prediction of rainfall using neural networks during monsoon season in India

Transcription

1 METEOROLOGICAL APPLICATIONS Meteorol. Appl. 19: (2012) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: /met.254 Multi-model ensemble (MME) prediction of rainfall using neural networks during monsoon season in India Ashok Kumar, a *A.K.Mitra, a A. K. Bohra, a G. R. Iyengar a and V. R. Durai b a National Centre for Medium Range Weather Forecasting (NCMRWF), A-50, Institute Area, Sec - 62, Noida, UP , India b India Meteorological Department, Mausam Bhawan, Lodhi Road, New Delhi , India ABSTRACT: The prediction of Asian summer monsoon rainfall on various space and time scales is still a difficult task. Compared to the mid-latitudes, proportional improvement in the skill in prediction of monsoon rainfall in medium range has been less in recent years. Global models and data assimilation techniques are being further improved for monsoons and the tropics. However, multi-model ensemble (MME) forecasting is gaining popularity, as it has the potential to provide more information for practical forecasting in terms of making a consensus forecast and reducing the model uncertainties. As major centres are exchanging the model output in near real time, MME is a viable, inexpensive, way for enhancing the forecast skill. Apart from a simple ensemble mean, the MME predictions of large-scale monsoon precipitation in the medium range was carried out during the 2009 monsoon at NCMRWF/MoES, India. The neural network weights were obtained and a neural network was trained based upon forecast data from four global models for the 2007 and 2008 monsoons in order to develop the multi-model ensemble system. The skill score for country and sub-regions, indicates that a multi-model ensemble forecast has a higher skill than individual model forecasts and also higher skill than the simple ensemble mean in general. Although the skill of the global models falls beyond day 3, a significant improvement could be seen by employing the MME technique up to day 5. Copyright 2011 Royal Meteorological Society KEY WORDS monsoon-rainfall; global-models; muti-model-ensemble(mme); neural-networks Received 25 May 2010; Revised 28 December 2010; Accepted 31 January Introduction The Asian Monsoon is one of the major components of the Earth s climate system. Realistic modelling, simulation and prediction of the monsoon are challenging scientific tasks for the world s Earth system science community. For India, the monsoon rains are of enormous importance giving shape to its agriculture, economy and rhythms of life. The science pertaining to the monsoon has progressed significantly in the last two decades due to an increased wealth of new data from satellite observations, understanding the processes and enhanced computing power. Numerical models and data assimilation algorithms have further improved at all major international centres across the globe. The accuracy of the weather forecasts has improved steadily in last three decades, and the systematic errors with forecast length in the medium range have reduced. However, in general, the forecast skill in the tropics is still lower compared to that of mid-latitudes and is particularly of concern for rainfall forecasts over the Indian monsoon region. The errors in the rainfall forecast are due to uncertainties in the assimilation process as well as in the physical parameterization. These uncertainties could be due to the errors in the prescribed initial states, which may arise from observational instruments, satellite estimates and the data assimilation methods. The forecast skill is also dependent on the synoptic situation, flow regime, region, and known (or unknown) teleconnections. The accuracy of a model also varies depending on Correspondence to: A. Kumar, National Centre for Medium Range Weather Forecasting (NCMRWF) A-50, Institute Area, Sec - 62, Noida, UP , India. ashok@ncmrwf.gov.in the formulation, horizontal-vertical resolutions and the parameterization schemes representing the small-scale processes in the model. The process of improving forecast skill of an individual Numerical Weather Prediction (NWP) system through research and development in modelling and data assimilation is a rather slow process. For example, at the European Centre for Medium Range Weather Forecasts (ECMWF), the research and development of 10 years has resulted in an improvement of skills of 1 day. Therefore, from a practical forecasting aspect one has to learn to live with these uncertainties and the current available skill of the models. The point is how to make best use of the forecasts from any (or many) centres, when each centre has data from many deterministic models. In the context of the availability of model data from different centres and each centre producing many ensemble members, the use of ensemble methods in the short and medium ranges, and even for short-term climate prediction, has become popular in recent years. Due to the uncertainties in the modelling system, the forecast errors increase with the forecast length, until it becomes no longer useful. If one has an ensemble of forecasts (many forecasts) one can say something extra about the reliability of the forecasts to the user. Clustering (or tubing) of several similar forecasts is also useful (Atger, 1999). From many ensemble forecasts one can obtain warnings for possible extreme/severe events. One single forecast may fail to catch the extreme event, but the ensemble might give some extra information on the extreme (or anomalous) episode. The current practices at major centres are: (1) to perturb the initial conditions (scientific based) and make many member runs from these different initial conditions; (2) to make many runs by Copyright 2011 Royal Meteorological Society

2 162 A. Kumar et al. altering the model formulations and physical parameterizations, and, (3) use a multi-model ensemble from many different models. A multi-model ensemble (MME) is a viable option, as many centres exchange and/or provide model data in near real time. One assumption for a MME is that the deterministic models do provide some signal and the noise (and error) is less compared to signal. Some weather/climate forecast centres perform calibration of model outputs to improve the skills of the model. Different statistical post-processing techniques are applied to model output parameters for the scale, region and phenomenon of interest. These post-processing methods enable the forecasters to obtain enhanced skill and value from models. An MME is another post-processing technique that enhances the skill of rainfall prediction. This paper describes the performance of the experimental MME forecast of rainfall during the 2009 monsoon, focusing on the large-scale aspects of monsoon rain. The difficulties in producing rainfall forecasts for smaller regions by the state of the art global models are well known. Significant errors are obviously expected if one decides to come down to smaller mesoscales, below the large scale organized convective rainfall associated with the monsoon (Chakraborty, 2010) Therefore, as a first attempt, 1 1 latitude/longitude grid rainfall data from four deterministic models and the associated multi-model products in the medium range are considered. It is well known that the simple mean of many models (simple ensemble mean, giving equal weight to each deterministic model) generally produces a higher skill score. The interest here is to show that the skill of the forecast obtained by using the MME methodology is better than the skill of simple ensemble mean and deterministic models forecasts. Neural network techniques (Masters, 1993; Muller and Reinhardt, 1991) are used for applying the MME methodology. Early work by Krishnamurti et al. (1999) showed that it is possible to obtain skill improvements both in weather and climate scales by the use of the multi-model technique known as Super Ensemble forecasting. Later it was extended for tropical precipitation by incorporating the multi-analysis concept along with the use of multi-models (Krishnamurti et al., 2000). Later, Mishra and Krishnamurti (2007) applied the superensemble algorithm for the Indian monsoon and showed the skill enhancement using data from seven global models. A multi-model multi-analysis ensemble system was reported to evaluate the deterministic forecasts from the United Kingdom Meteorological Office (UKMO) and ECMWF ensemble data by Evans et al. (2000), who showed the superiority of the multi-model system over the individual model data. Richardson (2001) used multi-model and multi-analysis data to produce both deterministic and probabilistic ensemble forecasts using four global models from the UKMO, German Meteorological Service, Meteo France and the National Centre for Environmental Prediction (NCEP) and showed that simple ensemble mean and simple bias correction produce useful products. The probability of precipitation and rainfall distribution by using multi-model data from seven global models for the Australian region was studied by Ebert (2001). Multimodel multi-analysis data were also tried by using ECMWF and UKMO ensemble outputs for quasi-operational medium range forecasting in both a probabilistic and deterministic sense (Mylne et al., 2002). They noted that the MME is more beneficial than a single model ensemble prediction system (EPS) and that the skill during winter was higher compared to summer. An operational consensus forecast by including several models is seen to outperform direct model output (DMO) and model output statistics (MOS) forecasts (Woodcock and Engel, 2005). Multi-model products prepared by using the National Centre for Environmental Prediction/Global Forecast System (NCEP/GFS) and ECMWF data also have shown improvements in week 2 forecasts (Whitaker et al., 2006), improving over the MOS forecast of individual models. Most of these studies (algorithms) used a simple ensemble mean or a mean of calibrated data from deterministic models. However, by analyzing the past performance of a model for a region it might be interesting to examine if model-dependent weights help to improve the final multi-model forecast further. Johnson and Swinbank (2009) used ECMWF, UKMO and NCEP/GFS global model data to prepare multi-model ensemble forecasts in the medium range. These forecasts were studied for bias correction, model dependent weights and variance adjustments. It was found that the multi-model ensemble gives an improvement in comparison to a calibrated single model ensemble. They also noted that only small improvements were achieved by using the model-dependent weights and variance adjustments. Recently, more studies for monsoon rainfall using MME methodology were reported based upon a simple linear regression approach (Roy Bhowmik and Durai, 2008, 2010; Mitra et al., 2011). Section 2 describes the model data and the observations used in the study. In Section 3 the technique followed for the MME methodology is described. Section 4 describes the results in detail and a summary and future work is described in Section Model data and observations In this study the daily medium range (days 1 5) rainfall prediction data from four state of the art operational global models (NCEP/GFS, UKMO, Japan Meteorological Agency (JMA) and National Centre for Medium Range Weather Forecasting (NCMRWF)) are used for the 2007, 2008 and 2009 monsoon seasons (June, July, August, September). These model data were available up to day 7, but the MME models were developed and tested only up to day 5. These models were being run at their respective centres (countries) at a higher horizontal and vertical resolution. NCEP/GFS model data were from runs made at 35 km horizontal grid and 64 vertical layers (Yang et al., 2006), UKMO model data was from runs made at 40 km horizontal grid and 50 vertical layers (Rawlins et al., 2007), the JMA model data were from runs made at 20 km horizontal grid and 60 vertical layers, (JMA, 2007) and the NCMRWF/GFS model data were from runs made at 50 km horizontal grid and 64 vertical layers (NCMRWF, 2010). The NCMRWF/GFS model is an adopted version of the NCEP/GFS system and was implemented in The operational global models were run at their respective centres at a higher resolution. However, rainfall forecasts provided by the different centres at a coarser resolution of 1 1 were used in this study to represent the large-scale aspect of the monsoon rainfall. The purpose of this study is to note the skill enhancement coming from the multi-model ensemble algorithm. The observed gridded rainfall analysis data used in model calibration (training) has to be of good quality, otherwise it might degrade the MME results. Hence, the corresponding daily observed gridded rainfall data at 1 1 resolution was prepared by merging rain

3 Multi-model ensemble (MME) prediction of rainfall 163 (d) (e) (f) (g) Figure 1. Total rainfall (mm day 1 ) during the 2009 monsoon (June, July, August and September) from observations, simple ensemble mean, multi-model ensemble using neural networks, and member models (d) NCMRWF, (e) UKMO, (f) NCEP GFS, (g) JMA, for day 3 forecasts. gauge values with satellite estimates (Mitra et al., 2003). As during , all the participating global models did not go through any major change in their model formulation. During the mentioned period the small changes in data usage for data assimilation and changes in co-efficients in physical parameterizations will have very small impact on the large-scale monsoon rainfall pattern (Zapotocny et al., 2007). Therefore, the intention of testing a MME algorithm for the large-scale rainfall pattern associated Indian monsoon may not be affected by minor model changes during the period. 3. MME methodology The basic technique used in this study is neural networks. Neural networks have a massively parallel, layered structure with each layer consisting of several nodes called neurons. They provide a mapping from input vector x i,i = 1, 2,...,n, to the output vector y j,j = 1, 2,...,m. Besides the input and output layers the network contains one or more hidden layers. Each neuron produces an output O = f(z), where Z = zi w i + b, z i (i = 1, 2,...,n) are the inputs to the given neuron, f(z) is the activation function and is usually taken to be the sigmoidal function 1/[1 + exp( Z)], w i are the weights associated with the network and b is the bias of the neuron. The weights w i and the bias b represent the parameters of the network which are to be determined by using the training data set of the pattern to be learned (Pankaj et al., 2002). More details on neural networks and their applications can be found in text books on neural networks (for example, Muller and Reinhardt, 1991; Masters, 1993). The training algorithm used in the neural networks for minimization of error is the conjugate gradients procedure complemented by simulated annealing to evade local minima. As the conjugate gradients method is expected to be more efficient than the more commonly used back propagation algorithm and hence the network is expected to learn faster (Masters, 1993). The simulated annealing method is necessary to escape from local minima which are usually present abundantly in the error function. The error measure was taken as the usual mean squared sum of errors. The cubic root of precipitation is computed before applying the neural network technique to the observed and the forecast rainfall data obtained from the four models. This has reduced the highly variable rainfall data to data having a near normal distribution of a lesser variability. The maximum and minimum is then obtained for each series and the data are transformed to have values between 0.1 to 0.9 as per the following formulations: x mod = (x x min )/(x max x min ) (1)

4 164 A. Kumar et al. (d) (e) (f) (g) Figure 2. Difference in observations from forecasts of total rainfall (mean error) during the 2009 monsoon (June, July, August and September) from observations, simple ensemble mean, multi-model ensemble using neural networks, and member models (d) NCMRWF, (e) UKMO, (f) NCEP GFS, (g) JMA, for day 3 forecast. The observed rainfall patterns are then considered as the output layer and the forecast rainfall patterns from the four models are taken as the input layer. A neural network with one hidden layer and three neurons is used to train the net and develop the relation between the observed rainfall and the four forecasts obtained from four different models. A simple sigmoidal function, 1/[1 + exp( Z)], is used as the activation function. 4. Results In this section the results for the Indian peninsula, Indian regions and sub-regions are presented. All deterministic global model forecast and the multi-model forecast data for the forecast length of days 1 5 are used to compute the skill of rainfall forecast during the 2009 monsoon season for the Indian region. In order to limit the number of figures the results for day 3 forecasts are shown, as the results shown are similar with directional variations for days 1 5 forecasts Indian Peninsula Figure 1 displays the total rainfall during the 2009 season for day 3 forecasts from the deterministic models, the simple ensemble mean and the multi-model ensemble. The observed rainfall is shown in Figure 1. This observed rain is produced on a daily basis by merging rain gauge and satellite estimates from the METEOSAT IR data. The rainfall of the four deterministic models is in Figure 1(d g). Two multi-model products, the simple ensemble mean and the multi-model ensemble forecast using neural networks, are shown in Figure 1 and. At first glance, all the four deterministic models seem to reproduce the observed monsoon rain. However, when the details are compared, it is found that each deterministic model differs from the observations in different ways and in different regions. The NCMRWF and NCEP models produce too much rain in the Arakan coast region to the east of Bay of Bengal. In the NCEP model on the west coast of India the north south rainfall band extends too much to the south and in the northern plains (the monsoon trough region) the model produces more rain than observations. The UKMO model is seen to predict more rain on the west coast, Himalayan foothills and northern Bangladesh. In contrast to all the models, the JMA model produces the least rainfall. The multi-model products in Figure 1 and, when compared with observations, look more realistic as compared to the deterministic models. The MME products are superior to the simple ensemble mean as the simple ensemble mean has shown more rain on the western side of west coast and in the Arakan Coast Region to the east of Bay of Bengal as compared to the MME products, when compared with the observations.

5 Multi-model ensemble (MME) prediction of rainfall 165 (d) (e) (f) (g) Figure 3. Root mean square error (RMSE) for rainfall during the 2009 monsoon (June, July, August and September) from observations, simple ensemble mean, multi-model ensemble using neural networks, and member models, (d) NCMRWF, (e) UKMO, (f) NCEP GFS, (g) JMA, for day 3 forecast. Mean rainfall biases for day 3 of the four deterministic forecasts, the ensemble mean and the MME are shown in Figure 2. The total seasonal rainfall is given in Figure 2 for reference. The difference plot is interpreted as mean error during the 2009 monsoon season. In these difference (mean error) plots the areas of positive biases (model is wetter) and negative biases (model is drier) are clearly shown. For all the days, the multi-model products such as the ensemble mean (EMN, Figure 2) and MME (upper row, Figure 2) show the least biases compared to the deterministic models (Figure 2(d g)). Among the multi-model products, the least biases are seen in the MME compared to EMN for days 1, 3 and 5 forecasts, as the simple ensemble mean has shown more bias to the western side of west coast and in the Arakan Coast Region to the east of Bay of Bengal. The root mean square error (RMSE) of the deterministic models, simple ensemble mean and the multi-model ensemble in comparison to the observations for day 3 is shown in Figure 3. The observed rainfall is shown in Figure 3. The mean error described in the above paragraph shows the systematic error of the model. However, while summing over seasons, errors of opposite sign might cancel out, rendering the bias not very informative. Hence, the mean error (bias) plots also have to be examined in conjunction with the root mean square error (RMSE) values for the regions. For reference, the observed plot is shown in Figure 3. Among the deterministic models, those of NCMRWF and NCEP have higher RMSEs compared to the UKMO model. Another feature of the RMSE is that the model errors are greater in models where the rainfall amounts are also more. For example, the west coast of India, the monsoon trough region and the Arakan coast region have higher RMSE values for the deterministic models. The forecasts from the multi-model products have smaller RMSE compared to deterministic models (Figure 3(d g)). The MME products (Figure 3) also have smaller RMSEs compared to the EMN (Figure 3), especially over the western side of west coast and in the Arakan Coast Region to the east of Bay of Bengal. The anomalies of the observations and forecasts are computed from their respective seasonal means during the 2009 monsoon. The anomaly correlation co-efficients (ACCs) for day 3 are shown in Figure 4. During the monsoon, in medium range forecasts it is important and challenging to predict the day to day rainfall associated with the passing transient weather systems or the fluctuating strength of the monsoon. Therefore, the rainfall anomaly in forecasts and observations has to be examined in terms of their similarity. The observed rainfall is shown in Figure 4. The anomaly correlation co-efficients are plotted on a scale of 0 to 1. The multi-model ensemble has much higher values of ACC in almost all the regions compared to individual models. It is worth mentioning that the ACC value for the multimodel ensemble (Figure 4) are much higher compared to those of the deterministic models (Figures 4(d g)). The ACCs

6 166 A. Kumar et al. (d) (e) (f) (g) Figure 4. Anomaly correlation co-efficient (ACC) for rainfall during the 2009 monsoon (June, July, August and September) from observations, simple ensemble mean, multi-model ensemble using neural networks, and member models, (d) NCMRWF, (e) UKMO, (f) NCEP GFS, (g) JMA, for day 3 forecast. for the multi-model ensemble are approximately comparable to those of the simple ensemble mean (Figure 4), but in some regions the ACCs are marginally higher, such as near Gujarat in the Arabian Sea, along the West Coast and in the Bay of Bengal Indian region and sub regions All the above results give some general idea of the quality of rainfall forecasts in terms of error statistics for the monsoon for the deterministic models and the products from the multimodel algorithm, but there are limitations of the numerical models in simulating the rainfall quantity over the right regions. Therefore, it is relevant to examine and document the skill of rainfall forecasts in terms of rainfall amounts in different categories (different threshold amounts of rainfall) in terms of threshold statistics for India as a whole and for different regions of the country. Standard statistical parameters such equitable threat score (ETS), threat score (TS) and hit rate (HR) have been computed in different categories of rainfall amounts. A brief description of these categorical statistics is given in Wilks (1995) and Ashok et al. (2002). Four uniform rainfall categories are taken for this comparative study with threshold values as no rain (0.0), 20 and 40 mm. For a hit rate one more category is used, All which considers all the thresholds together. Thresholds greater than 40 mm of rainfall are not considered in the computation of ETS, TS and HR as the sample size becomes very small and results may not be robust (Casati et al., 2008). However, for those thresholds and above, the extreme dependency score (Stephenson et al., 2008) is used. For all India and the five different regions of India (north India, central India, west coast, peninsular India and northeast India) these categorical skill scores have been computed. The skills have been computed for days 1, 3 and 5 forecasts for all six domains during the 2009 monsoon season, although in order to limit the number of figures only the plots for day 3 are given (Figures 5 11). These scores were obtained from the data covering the daily values for the entire 2009 monsoon season of 122 days (June, July, August, September). The reason for selecting these domains is related to the known important synoptic circulation features of the large-scale Indian monsoon system and the associated rainfall over the land regions. All India covers all the land grid points between 67 and 100 E, and 7 and 37 N. The north India domain covers the monsoon trough region, extending from 70 to 85 E and 25 to 35 N. The central India domain covers the monsoon trough region, extending from 73 to 90 E and 22 to 28 N. The west coast domain of India extends from 70 to 78 E and 10 to 20 N. The peninsular India domain extends from 74 to 85 E and 7 to 21 N. The north east India domain extends from 85 to 100 E and 22 to 30 N.

7 Multi-model ensemble (MME) prediction of rainfall 167 Figure 5. Equitable threat score, threat score, hit rate for rainfall for Multi-model ensemble using Neural Networks, simple ensemble mean and member models for Day 3 forecast for JJAS 2009 for All India ( E, 7 37 N). Figure 7. Equitable threat score, threat score, hit rate for rainfall for the multi-model ensemble using neural networks, simple ensemble mean and member models for Day 3 forecasts for June, July, August and September 2009 for Central India (73 90 E, N). Figure 6. Equitable threat score, threat score, hit rate for rainfall for the multi-model ensemble using Neural Networks, simple ensemble mean and member models for Day 3 forecast for June, July, August and September 2009 for North India (70 85 E, N). Figure 8. Equitable threat score, threat score, hit rate for rainfall for the multi-model ensemble using neural networks, simple ensemble mean and member models for Day 3 forecasts for June, July, August and September 2009 for West Coast (70 78 E, N).

8 168 A. Kumar et al. Figure 9. Equitable threat score, threat score, hit rate for rainfall for the multi-model ensemble using neural networks, simple ensemble mean and member models for Day 3 forecasts for June, July, August and September 2009 for Peninsular India (74 85 E, 7 21 N). Figure 10. Equitable threat score, threat score, hit rate for rainfall for the multi-model ensemble using neural networks, simple ensemble mean and member models for Day 3 forecasts for June, July, August and September 2009 for North East India ( E, N). Part of the figures shows the equitable threat score, ETS, for different thresholds and different models, simple ensemble mean and multi model ensemble, followed by part for threat score, TS, and then part shows the hit rate, HR. As is well known, for any given day (forecast length) these scores generally gradually decrease with increasing threshold. Again, with increasing length of forecast period (days 1 5) for each threshold rainfall category, the skill scores fall gradually. The multi-model ensemble is giving the higher value for all the scores, over all the regions and for days 1 5 forecasts, as compared to the deterministic models and the simple ensemble mean. Although the individual models seem to have almost similar skill, with a slight edge for the GFS and UKMO models over those of NCMRWF and JMA. The extreme dependency score (Stephenson et al., 2008) is also computed for the threshold value greater than 60 mm for all India (Figure 11). The multi-model ensemble shows the higher value of extreme dependency score compared to the deterministic models and the simple ensemble mean. Hence, this comparison has clearly shown the superiority of the multi-model ensemble over the individual deterministic models and the simple ensemble mean. 5. Summary and future work During the 2009 monsoon a multi-model ensemble (MME) forecasting for large-scale monsoon precipitation in medium range was carried out at NCMRWF/MoES, India. A multimodel forecast based on weights given to deterministic models from neural networks algorithm along with simple ensemble mean was tested. The neural networks algorithm used two previous season s daily data for training and deriving the weights. From the analysis of the results of the 2009 monsoon, it is seen Figure 11. Extreme dependency score for multi-model ensemble using neural networks, simple ensemble mean and member models for Day 3 forecasts for June, July, August and September 2009 for All India ( E, 7 37 N).

9 Multi-model ensemble (MME) prediction of rainfall 169 that the multi-model ensemble forecast has higher skill than individual deterministic model products and the simple ensemble mean. The MME forecast for heavy rainfall events greater than 60 mm has also shown a much better extreme dependency score compared to individual deterministic models and the simple ensemble mean. The skill score for the whole country as a region, and other sub-regions, indicates that the multi-model ensemble product is superior to those of individual models and the simple ensemble mean. A combination of ensembles is a promising approach for further development which will give rise to improvement in the predictive skill for the tropics and monsoon region. While looking for enhanced skill in the MME, as a by-product the deterministic models were also verified and one is able to keep track of the state of the art model s performance for the monsoon. This feedback is useful for continuous model development and other modelling-related research. Recently, because of bi-lateral and multi-lateral agreements between operational weather centres and collaborative research work undertaken by international bodies, real time exchange of global and regional model outputs at coarser resolution has become a reality. Use of such MME algorithms in real-time will be realized soon for operational purposes. These MME algorithms are less expensive as they could be run on less expensive computers by relatively smaller weather centres. The World Meteorological Organization s (WMO s) The Observing System Research and Predictability Experiment (THORPEX) is planning to provide data from eight global models under the THORPEX Interactive Grand Global Ensemble (TIGGE) and each model has around 15 ensemble members, making it in total around 120 members. It is planned to make use of TIGGE type data to further experiment with MME algorithms with an operational use for improved monsoon rainfall forecast as the final goal. Acknowledgement We acknowledge the availability of operational global model medium-range rainfall forecast data from NCEP/USA, UKMO/ UK, JMA/Japan. The NCMRWF global T254L64 model is a version of NCEP GFS system adapted in 2007 at NCMRWF. Thanks are due to the MoES MME project steering committee members for their useful suggestions related to this study. References Ashok Kumar, Maini Parvinder, Rathore LS, Singh SV Skill of statistical interpretation forecasting system during monsoon season in India. Atmospheric Science Letters 3(1): Atger F Tubing: an alternative to clustering for the classification of ensemble forecasts. Weather and Forecasting 14: Casati B, Wilson LJ, Stephenson DB, Numi P, Ghelli A, Pocemich M, Damrath U, Elbert EE, Brown BG, Mason S Forecast verification: current status and future directions. Meteorological Applications 15: Chakraborty A The skill of medium-range forecasts during the year of tropical convection Monthly Weather Review 138: Ebert EE Ability of a poor man s ensemble to predict the probability and distribution of precipitation. Monthly Weather Review 129: Evans RE, Harrison MSJ, Graham RJ, Mylne KR Joint mediumrange ensembles from the Met. Office and ECMWF system. Monthly Weather Review 128: JMA Outline of the Operational Numerical Weather Prediction at the Japan Meteorological Agency. Appendix to WMO Numerical Weather Prediction Progress Report. Japan Meteorological Agency: Tokyo, Japan; 194 pp. [accessed on 10 May 2007]. Johnson C, Swinbank R Medium-range multi-model ensemble combination and calibration. Quarterly Journal of the Royal Meteorological Society 135: Krishnamurti TN, Kishtawal CM, LaRow TE, Bachiochi DR, Zhang Z, Williford CE, Gadgil S, Surendran S Improved weather and seasonal climate forecasts from multi-model super- ensemble. Science 285: Krishnamurti TN, Kishtawal CM, Shin DW, Williford CE Improving tropical precipitation forecasts from a multi-analysis super-ensemble. Journal of Climate 13: Masters T Practical Neural Network Recipes in C++. Academic Press: San Diego, CA; 493. Mishra AK, Krishnamurti TN Current status of multi-model super-ensemble operational NWP forecast of the Indian summer monsoon. Journal of Earth System Science 116(5): Mitra AK, Dasgupta M, Singh SV, Krishnamurti TN Daily rainfall for Indian monsoon region from merged satellite and raingauge values: large-scale analysis from real-time data. Journal of Hydrometeorology (AMS) 4(5): Mitra AK, Iyengar GR, Durai VR, Sanjay J, Krishnamurti TN, Mishra A, Sikka DR Experimental Real-Time Multi-Model Ensemble (MME) prediction of rainfall during monsoon 2008: large-scale medium-range aspects. Journal of Earth System Science 120(1): Muller B, Reinhardt J Neural Networks: An Introduction, The Physics of Neural Networks Series. Springer-Verlag: Berlin; 266. Mylne KR, Evans RE, Clark RT Multi-model multi-analysis ensembles in quasi- operational medium-range forecasting. Quarterly Journal of the Royal Meteorological Society 128: NCMRWF Monsoon-2009: Performance of the T254L64 Global Assimilation-Forecast System. Report No. NMRF/MR/01/2010, March 2010, Mitra AK, Das Gupta M, Iyengar GR (eds). NCMRWF: Noida, India; 131. [pdf file available at NCMRWF web site Pankaj J, Kumar A, Maini P, Singh SV Short range SW monsoon rainfall forecasting over India using neural networks. Mausam 53: Rawlins F, Ballard SP, Bovis KJ, Clayton AM, Li D, Inverarity GW, Lorenc AC, Payne TJ The Met Office global four-dimensional variational data assimilation scheme. Quarterly Journal of the Royal Meteorological Society 133(623): Richardson DS Ensembles using multiple models and analyses. Quarterly Journal of the Royal Meteorological Society 127: Roy Bhowmik SK, Durai VR Multi-model ensemble forecasting of rainfall over Indian monsoon region. Atmosfera 21(3): Roy Bhowmik SK, Durai VR Application of multi-model ensemble techniques for real time district level rainfall forecasts in short range time scale over Indian region. Meteorology and Atmospheric Physics 106(1 2): Stephenson DB, Casati B, Ferro CAT, Wilson CA The extreme dependency score: a non-vanishing measure for forecasts of rare events. Meteorological Applications 15: Whitaker JS, Wei X, Vicars F Improving week-2 forecasts with multimodel reforecast ensembles. Monthly Weather Review 134: Wilks DS Statistical Methods in Atmospheric Sciences. Academic Press: New York, NY. Woodcock F, Engel C Operational consensus forecasts. Weather and Forecasting 20: Yang F, Pan H-L, Krueger SK, Moorthi S, Lord SJ Evaluation of the NCEP global forecast system at the ARM SGP site. Monthly Weather Review 134(12): Zapotocny TH, Jung JA, Le Marshall JF, Treadon RE A twoseason impact study of in satellite and in-situ data in NCEP GADS. Weather and Forecasting 22: