Five months quasi-operational forecasting of atmospheric constituents - comparison of results to data from official monitoring networks

Transcription

1 Five months quasi-operational forecasting of atmospheric constituents - comparison of results to data from official monitoring networks S. Tilmes, J. Rifimann, I. Jacobsen, J. Zimmermann Deutscher Wetterdienst, Germany. Abstract From May to September 1999 quasi-operational forecasts of atmospheric trace gases for Central Europe were performed at the Deutscher Wetterdienst (DWD) using a comprehensive model system developed in of a cooperation between the DWD and several universities. In parallel to the forecasting an operational real-time verification of model results was established using the measurements of the environmental agencies of Germany and its Federal States at about 360 sites. Subsequently an evaluation of the model performance for the complete summer season was carried out. The aim of this was twofold. On the one hand the focus lies on the climatological properties of the modelling system. On the other hand, monitoring the underlying observations is essential for the assessment of the overall interpretation. 1 Introduction The main objective of the German "Tropospheric Research Program" (TFS, funded by the German government) was to enhance the understanding of the formation of photochemical oxidants in the troposphere. Especially, it was expected to improve the forecasting quality of chemistry transport models (CTMs) for the troposphere. Thereupon research groups from the DWD and the Universities of Cologne, Karlsruhe, and Stuttgart built up the TFS model network to set-up and test a comprehensive modelling system for air quality forecasting. The aim of the research project at the DWD

2 52 Air Pollution VIII was to demonstrate that a daily real-time 48 hour forecast of atmospheric trace constituents for Central Europe with this modelling system is feasible until afternoon of the first day of the forecast. This was shown day by day from May to September Accompanying this activities a second project was concerned with the development of methods for the verification of the forecasts and the evaluation of the modelling system using routine air quality data. In the next section we will describe the state of the modelling system during the summer This will be followed by a presentation of the operational results and their verification. Examples for the evaluation of the model system with data from the whole season and the assessment of the observation data will be given afterwards. Finally we will summarise and give an outlook on our further plans. 2 The forecasting system Figure 1 gives an overview over the state of the modelling system as it was in operation in the testing phase in summer The modelling domain covers Central Europe, from Ireland to Poland and from Italy to Southern Scandinavia. The horizontal resolution is , about 21 km, with 109 x 109 gridpoints. In the vertical there are 36 layers used for the meteorological forecast and 16 for the CTM with a coarser resolution in the upper troposphere and stratosphere. In both models the lowest layer is about 35 m thick and the model top is at 20 hpa. A terrain following cr-coordinate is used. 2.1 The weather prediction part The nonhydrostatic Lokal-Modell (LM) of the DWD was used to provide the meteorological input to the system (Doms and Schattler [1]). The forecasts were driven by the analysis of the Europa-Modell (EM) and by boundary data provided by EM forecasts. The LM ran on a CRAY T3E. For the future it is planned to use data from the new global model GME of the DWD as initial and boundary data or to utilize the interpolated operational weather forecast by the LM. 2.2 The emission forecasting part At the University of Stuttgart the CAREAIR-ECM was developed to predict emission factors with the desired spatial and temporal resolution for all the species calculated by the CTM (Friedrich et al. [2]). In 1999 climatological data were used to describe the dependency of the emissions on the meteorology. Additionally, during this summer the biogenic emissions were calculated by the CTM itself. The ECM was operated on a SGI workstation. The next steps in the development of the system will include the consideration of the current weather forecast and the operation of a module for the biogenic emissions.

3 Air Pollution VIII 53 STATE OF THE FORECASTING SYSTEM (modifications for operational tests in summer 1999) EM data meteorological boundary data LM meteorological forecasts v modelling of anthropogenic emissions without actual meteorology PPC preprocessor for meteorological input, photolysis rates CTM EEM chemistry preprocessor for transport forecast data flow { ':» input for different module ^- forecast output DWD: Deutscher Wetterdienst EEM: EL RAD emission model EM: Europa-Modell IER: Inst. f. Energiewirtschaft; Univ. Stuttgart IGM: Inst. f. Geophys. + Meteorol.; Univ. Koln GME: Global-Modell LM: Lokal-Modell ECM: Emissions-Modell CAREAIR-ECM PPC: CTM preprocessor CTM: chemistry transport model PST: post-processor PST verification, products routine testing of the forecasting system, build up of operational visualization and verification Figure 1: Sketch of the modelling system in summer The chemistry transport forecasting part The EURAD-CTM from the University of Cologne was applied to calculate the distribution of the atmospheric trace substances (Hass [3]). As chemistry scheme the RADM2 mechanism was used. Transport was calculated by a fourth order Bott algorithm. In order to meet the requirements of the LM the CTM was extended by a nonhydrostatic option and the opportunity to utilize a rotated geographical coordinate system. Chemical initial data were provided by the results of the previous day 24 hour forecast. The operational runs with the CTM were carried out on a CRAY C90.

4 54 Air Pollution VIII 3 Routine operation of the system The LM run set up on the routine 00 UTC run of the EM which was available at about 4:15 UTC. Afterwards the LM data were extracted from the data base and processed to the format desired by the CTM. Anthropogenic and especially biogenic emissions were calculated using this preprocessed meteorological data. At about 8:00 UTC the CTM calculation started producing a 48 hour forecast until about 14:15 UTC. Consequently the prediction of the maximum ozone concentrations for the following day were available at about 13:30 UTC. The forecasts of the LM, ECM, and CTM were archived routinely during the operation of the model chain. With the current setup about 0.9 GByte output was produced for each daily 48 hour run. The complete model chain was controlled by a script running on the CRAY C Routine products Accompanying the model chain a second script running on the workstation controlled the preparation and distribution of products. This included the generation of plots and MPEG films of the forecasted meteorology and the ground-level ozone field. These were distributed to the network partners via ftp. Additional output was generated for special purposes. For example meteorological and chemical initial and boundary data were extracted and delivered via ftp for subsequent forecasts with the KAMM/DRAIS model system at the University of Karlsruhe (Nester et al. [4]). Another example was the preparation of data and plots as a supporting tool for the routine observation programme at the meteorological observatory Hohenpeifienberg which is operated by the DWD as part of a GAW global station. 3.2 Routine verification The forecasts were verified on a routine basis. Meteorological parameters were verified against the analyses of the operational weather prediction system which has a horizontal resolution of about 7 km. The main focus was put on the verification of the chemical forecasts. For this purpose we received once a day data from the routine monitoring networks of the German environmental protection agency (UBA) and the corresponding institutions of the Federal States. These data consisted of half hour averaged groundlevel ozone values from more than 360 sites all over Germany. The data presented in this paper are of preliminary status. Figure 2 gives an example for the verification of the 14 hour forecast valid for September 10 14:00 UTC Ozone mixing ratios are shown on the left (filled contours of the model results for the lowest model layer, filled circles for the observed data) and the differences between forecasts interpolated to the station locations and observations on the right hand side. Below a statistical evaluation of the data is presented including a frequency distribution of the deviations modelled minus observed data. For that time

5 Air Pollution VIII 55 Figure 2: Operational verification for September UTC no data were available for Bavaria with the exception of observations from two UBA stations. In this verification no quality control was performed on the observations nor any preselection of sites was done. Thus, the results must be interpreted taking into account that to a great extend the sites are only of limited spatial representativeness and that the inhomogeneity in the spatial distribution of the sites has effects on the calculated scores.

6 56 Air Pollution VIII Verifcation period: September 6-16,1999 observations: thin; frcst. from today: thick; frcst. from yesterday: dashed Figure 3: Spatial verification for the period September Evaluation of the model performance Up to now the evaluation of the model performance within this five month of operation concentrated on ground level ozone data. This was mainly due to the lack of other routinely available data of appropriate quality. Since 48 hour forecasts were performed, two model time series can be constructed and compared. The time series of thefirstday (the forecast "from today", forecast length 1-24 hours) and the one of the second day (the forecast "from yesterday", forecast length hours). 4.1 Spatial statistics Figure 3 shows time series of the spatial statistical measures. The selected time range is September The interesting feature of this episode

7 Air Pollution VIII 57 is its pronounced photosmog character, which is unusually late in the year and is following a summer with rather low ozone mixing ratios. In the first row the number of observation sites entering the statistics is shown. For each hour there were data from at least 270 sites available. The next three rows present the mean values, the 98 % percentile, and the standard deviation for the two forecasted time series and the observed time series. Bias (model minus observations), root mean square error (rmse), and spatial correlation coefficient can be seen in the lower rows of the figure. The beginning and the intensity of the ozone episode was well forecasted by the model as can be seen from the mean and 98 % percentile values. Only at the end of the episode (September 15-16) the model showed some over predict ion. About ten German sites can be classified as remote. Those sites are in general situated on elevated mountainous terrain. Due to the coarse horizontal resolution the model orography is mostly smoother. Thus, for these stations it is questionable to compare the observations to data from the lowest model layer. The effect is an underestimation of the 98 % percentile values during night time. Another effect of the coarse horizontal resolution is that the model cannot reproduce the very locally influenced ozone concentrations which are observed at traffic related sites. This effect is also visible mainly during the night-time hours when at these stations most of the ozone is titrated to N02- This is the reason why the night-time spatial standard deviation in the observations is much larger than in the model forecasts. This is also valid for the scores of the bias, rmse, and correlation. On the other hand thefigureshows an overall good model performance for the times of the daily maximum concentrations in the afternoon. In Figure 3 a marked difference can be observed between the performance of the two forecast series. As mentioned above, the anthropogenic emission data are calculated using a climatological meteorology. Thus, most of these differences are due to the different integration periods of the meteorological forecasts by the LM. Further evaluation studies clearly show a pronounced negative influence on the mean model performance. This is valid for the rmse and the correlation but not for the bias what indicates a "sane" model behaviour without significant drifts. 4.2 Temporal statistics A second step in evaluating the model performance was the computation of temporal statistics, the calculation of statistical measures for the observed and modelled time series at each single station location. Since most of the sites are under urban influence the ozone concentrations show a marked diurnal variation. To avoid that this extenuates the statistical scores only the data for UTC are evaluated to compute the scores presented in Figure 4. The spatial distribution of the bias, rmse, and correlation coefficient are shown accompanied by the frequency distributions for the

8 fl J/.J' Verification period: May - September 1999 Considered time of day: UTC jrmse (ppb) h 7^ y j correlation coefficient *' o \ J / o o ' O ' o ^ov' <p ', o o o6 c,^ ". oo 0^ 9 G* GSO o O OQcDO O > o,

9 Air Pollution VIII 59 complete suite of stations. The picture is a different one for each statistical measure. The model performance is different for different regions in Germany. For example the model shows large values for the rmse within the highly industrialised Rhein- Ruhr area in North-Rhine-Westphalia (NRW) whereas the performance is better for other also industrialised areas like the southern part of Hesse. For the southern part of Saxony the model shows less skill in terms of the correlation coefficient. There is a multitude of reasons for the different effects. The latter one might be due to the inhomogeneous terrain of the Erz Gebirge, but also due to the very limited representativeness of some sites which are to a large extend influenced by local emissions. This might also be a reason for the large errors in NRW. On the other hand the model might perform better e. g. in NRW with a better horizontal resolution especially in view of a better differentiation of the anthropogenic emissions. Nevertheless, the problem of the representativeness of the German monitoring sites is very serious. Most of the sites were set up to monitor winter smog and, therefore, are often placed close to emission sources. Thus, it is necessary to distinguish between different station classes, but this is beyond the scope of this paper. 5 Conclusions and outlook Our first intention was to show that it is possible to operate a state-of-theart chemistry transport model for the routine forecasting of air pollution. This was proved over a five month period. Forecasts of atmospheric trace constituents for the following day were available in time in the afternoon of the present day. Several products were developed and distributed to different users in a routine mode. The second task was to evaluate the model performance in terms of a verification of the ground level ozone forecasts. Especially the distribution of the daily maximum concentrations was well simulated. The model showed no significant drifts or biases during thefivemonth period. Some deficiencies could be pointed out concerning for example different regions in Germany where the model performed worse. Reasons for this are beyond others the coarse resolution and the missing coupling of meteorology and emissions. On the other hand it turned out that the monitoring of the observations is a prerequisite to the assessment of the model performance. Plans for the future development of the system include new meteorological initial and boundary data from the operational global model of the DWD and chemical data assimilation. As a basis for this more chemical observation data must be made available in real time on the European scale also for the diurnal verification. This will include data from European ozone sondes and commercial airliners. It is also planned to increase the horizontal resolution within the next few years following the evolution in the numerical weather prediction.

10 60 Air Pollution VIII The developed modelling system is a powerful tool for assessing the air quality in Europe. From a scientific point of view it is well suited to investigate the climatological behaviour of the system and the long-term testing of the state-of-the-art chemistry transport models. The broad interest in the outcome of the system documents the need for such a system and the efforts in evaluation guarantees a thorough assessment of the results. The system is well suited for supporting the air quality monitoring over Europe and also the assessment of observation networks. Acknowledgements We want to thank the members of the model network group: the EUR AD group, University of Cologne, the IER, University of Stuttgart, and the IMK, University of Karlsruhe. We are grateful to the German environmental protection agency (UBA) for providing the ozone data of its stations and the networks of the Federal States via the UMEG, Karlsruhe. This work was supported by the Bundesministerium fur Bildung, Wissenschaft, Forschung und Technologic of Germany under grants 07TFS10/LT1-A2 and 07TFS10/LT1-C4. References [1] Doms, G. & Schattler, U. The nonhydrostatic limited-area model LM (Lokal-Modell) of DWD. Part I: Scientific Documentation, Deutscher Wetterdienst (DWD), Offenbach, [2] Friedrich, R., Heidegger, A. & Kudermann, F. Development of an emission calculation module as a part of a model network for regional atmospheric modelling. Proceedings of EUROTRAC Symposium '98, eds. P. M. Borrel & P. Borrel, WITpress, Southampton, , [3] Hass, H. Description of the EURAD Chemistry-Transport-Model Version 2 (CTM2)j Mitteilungen aus dem Institut fur Geophysik und Meteorologie der Universitat zu Koln, eds. A. Ebel, F. M. Neubauer & P. Speth, No. 83, [4] Nester, K., Fielder, F. & Panitz, H.-J. Simulation of mesoscale air pollution with the model system KAMM/DRAIS. Papers of the llth World Clean Air and Environment Congress, September, 1998, Paper 10D-2, Volume 4, Durban, South Africa, 1998.