Sub-seasonal Forecast And Error Diagnosis Of Temperature And Precipitation In East Asia Based On Multi-model Integration

Bridging the gap between weather forecasting and climate prediction,subseasonal forecasts are of great importance in the seamless forecasts aiming at disaster prevention and mitigation,yet currently of relatively poor quality.In recent years,corresponding multimodel numerical weather prediction products have been released via the Subseasonal to Seasonal(S2S)Prediction Project jointly established by the World Weather Research Programme(WWRP)and the World Climate Research Programme(WCRP).However,a mature framework of associated statistical post-processing has not yet been established.In this study,aiming at more skillful subseasonal forecasts of the surface air temperature and precipitation over East Asia,multiple post-processing models have been constructed towards deterministic and probabilistic forecasts with lead times of 8–42 days using forecasts from the S2 S multimodel daily outputs and observations from the U.S.Climate Prediction Center.Various verifications and evaluations are carried out to examine the forecast experiments and to take diagnostic analysis on the biases.The obtained results are elaborated as follows.(1)Forecast experiments of individual models indicate that performances of subseasonal forecasts differ with each other to a certain extent among different models,demonstrating the necessity of effective integration of information from multiple models.Since the subseasonal forecast is characterized by a fairly long time range,in procedures of traditional multimodel ensemble models,the systematic bias averaged from the training period could not be considerably continued till the forecast period,which leads to the generally limited capability of the traditional models in improving skills of the subseasonal forecast.Correspondingly,a Kalman filter based superensemble(KFSUP)model is established in this study,which emerges instant updates and maximum complements of various latest and effective forecast information from multimodel outputs via iterative algorithms between forecasts and observations over the training period.On the subseasonal timescale,remarkable improvements are reported towards both of the temperature and precipitation forecasts by the KFSUP post-processing model.For temperature forecasts,the KFSUP model efficiently constrains the large biases occurring at the longer lead times and over the high-altitude areas in the raw forecasts.It yields greater advancements in forecasting the daily maximum temperature(Tmax)than the daily minimum temperature(Tmin).But the calibrated forecast skill of Tmax is still widely inferior to that of Tmin.With respect to the precipitation forecasts,the KFSUP model shows most conspicuous forecast ameliorations for lower-level precipitation,while the calibration capability gradually weakens as the precipitation threshold increases.(2)Considering the superiority of the Kalman filter algorithm on the subseasonal timescale,the pattern projection method concentrating on the single-model forecast calibration is optimized in this study.The Kalman filter based pattern projection method(KFPPM)is therefore proposed,which is demonstrated with stable capability on improving the temperature and precipitation forecasts for all lead times with magnitudes distinctly greater than the models of gridly calibration and the raw pattern projection.Moreover,a more advanced Kalman filter based pattern projection superensemble(KFPPSUP)is further constructed combining conceptions of both pattern projection and multimodel ensemble.The KFPPSUP model inherits advantages of both KFPPM and KFSUP,which not only considers the effective integration of multimodel information but also takes the synergistic impacts of surrounding factors to the forecast target into account.Such procedures reduce the multimodel biases from aspects of both time series of spatial distribution to a large extent.For temperature forecasts,the KFPPSUP model exhibits conspicuous improvements over the whole East Asia with great magnitudes.Over a large portion of the area,the longest lead times with biases below 2°C reach 25 days and 42 days for Tmax and Tmin,respectively.With regard to forecasts of precipitation,the KFPPSUP model also fruitfully ameliorates skills of the other post-processing procedures.Although the optimization in forecasting precipitation with large thresholds is still relatively limited,it is characterized by an efficient capability to calibrate the subseasonal forecasts of precipitation structures and locations.(3)Examining the probabilistic forecasts on the subseasonal timescale,the model characteristics are considered composed of three items: reliability,resolution and uncertainty.For temperature forecasts,the Ensemble Model Output Statistics(EMOS)model has remarkably positive and continuous advancements in improving forecast skills over the whole East Asia,which shows general superiority to the Bayesian Model Averaging(BMA)model.The two post-processing models both effectively enhance the forecast reliability and resolution,with the EMOS model being more intelligent.After the post-processing calibrations,an intensified reliability is obtained in the Tmax forecast,which becomes consistent with that in the Tmin forecast,while the Tmax shows higher resolution of forecast than Tmin.However,due to the significantly higher inherent uncertainty of Tmax forecasts,the forecast skill of Tmax is still inferior to Tmin for all cases in spite of the greater magnitudes of forecast calibrations to Tmax.Regarding the probabilistic forecasts of precipitation,the BMA and EMOS generally raise the accuracy and reduce the biases.The BMA model always enhances the forecast reliability but shows a limitation in resolution improvements,whereas the EMOS model shows noticeable and stable dominance than the raw ensemble system and the BMA model for either the reliability or the resolution of precipitation forecasts.In addition,forecasts of the low-level precipitation are always characterized by relatively high inherent uncertainty accompanied with the high resolution,which shows impacts mainly from the uncertainty of the variable itself and the reliability of the forecast system.By contrast,forecasts of high-level precipitation exhibit low uncertainty and high reliability,and therefore predominantly depend on the resolution capability of the forecast system on the subseasonal timescale.In view of the presented analyses,the study would emerge important technical and theoretical supports for the operational subseasonal forecasts and provide fundamentals for the promotions and applications of the multiple deterministic and probabilistic forecast approaches on the subseasonal timescale,which feature the thesis with great scientific and practical significances.