TY - JOUR ID - 29126 TI - Probabilistic precipitation forecast using post processing of output of ensemble forecasting system JO - Journal of the Earth and Space Physics JA - JESPHYS LA - en SN - 2538-371X AU - Azadi, Majid AU - Vashani, Saeed AU - Hajjam, Sohrab AD - Y1 - 2012 PY - 2012 VL - 38 IS - 3 SP - 203 EP - 216 KW - Artificial Neural Network KW - Ensemble forecasts KW - Precipitation forecasts KW - Probabilistic quantitative KW - Rank-histogram calibration technique DO - 10.22059/jesphys.2012.29126 N2 - Accurate quantitative precipitation forecasts (QPFs) have been always a demanding and challenging job in numerical weather prediction (NWP). The outputs of ensemble prediction systems (EPSs) in the form of probability forecasts provide a valuable tool for probabilistic quantitative precipitation forecasts (PQPFs). In this research, different configurations of WRF and MM5 meso-scale models form our eight member ensemble prediction system. The initial and boundary conditions come from the operational 1200 UTC runs of global forecasting system (GFS) of NCEP (National Center for Environmental Prediction). The integration period goes from first November 2008 to 30 April 2009 (182 days). Both WRF and MM5 are used with non-hydrostatic option and were run with two nested domains, with the larger domain covering the south-west Middle East from 10° to 51°north and from 20° to 80° east. The smaller domain covers Iran from 23° to 41° north and from 42° to 65° east. The spatial resolutions are 45-and 15-km for the coarser and finer domains respectively. Forecasts out to +72 hour ahead from the inner domains have been used to form the raw ensemble forecasts. The ensemble forecasts assessed in this study are as follows: raw ensemble forecasts, ensemble forecasts formed by post processing of each member in the raw ensemble forecasts using artificial neural network (ANN), ensemble forecasts formed using rank-histogram calibration technique of Hamill and Cloucci (1998) on the raw ensemble forecasts and ensemble forecasts formed by using both ANN and rank-histogram calibration methods on the raw ensemble forecasts. This research shows that ANN could decrease the error of raw ensemble so that the MAE for the first day of forecast achieves less than 1.5 mm and for second and third forecast days is about 2.5 mm. And all members errors are similar for all forecast days, but it seems that the members related to the MM5 model (members 6, 7, 8) produce slightly better forecasts, while after using the post processing methods the result of MAE are nearly similar. BS was calculated for raw ensemble forecasts, ensemble forecasts formed by post processing of each member in the raw ensemble forecasts using ANN, ensemble forecasts formed using rank-histogram calibration technique of Hamill and Cloucci (1998) on the raw ensemble forecasts and ensemble forecasts formed by using both ANN and rank-histogram calibration methods for mentioned thresholds from forecast days1-3. Having performed ANN method, the forecast quality increased significantly. for example, the amount of BS for raw ensemble 0.42 decreased to 0.32 for post processed ensemble using ANN method for the first forecast day in precipitation less than 0.1 mm. also the BS calculated before and after using rank histogram method proposed by Hamill and Cloucci (1998) shows the increasing of probabilistic forecast quality such that the amount of BS for raw ensemble 0.42 decreased to 0.29 for post processed ensemble using both ANN and method proposed by Hamill and Cloucci (1998) for the second forecast day in precipitation less than 0.1 mm. The Brier Skill Score (BSS) was 0.3 for post-processed ensemble using ANN method and 0.1 for post-processed ensemble using method proposed by Hamill and Cloucci (1998) for all three thresholds in the first day of forecast. Briefly the selection of different configurations does not have much effect on decreasing the error and difference between observation and DMO increases from the first to the third forecast days in all members. The ANN and method proposed by Hamill and Cloucci (1998) as two post processing methods can significantly decrease the systematic error of DMO, but the ANN method can remove systematic error better than the method proposed by Hamill and Cloucci (1998). We can produce more accurate probabilistic forecast using ANN for raw ensemble output and the calibrating the post processed output using method proposed by Hamill and Cloucci (1998). UR - https://jesphys.ut.ac.ir/article_29126.html L1 - https://jesphys.ut.ac.ir/article_29126_651812c41483061261866da002d34e90.pdf ER -