Ensemble Forecast Bias Correction Based On Machine Learning Methods

Under the escalating impact of global warming,the world is experiencing a surge in extreme weather events and climatic disturbances.The accuracy of predicting such meteorological disasters has become paramount for safeguarding human lives,protecting valuable assets,and ensuring societal stability on a national scale.Presently,ensemble forecasting stands as the prevailing technique in weather and climate prediction.Nevertheless,due to the inherent imperfections in model physical processes and boundary conditions,ensemble forecasts are prone to errors and uncertainties.Hence,it becomes imperative to employ bias revisions in order to enhance the forecast accuracy of these ensembles.In recent years,the integration of machine learning algorithms into various domains has revolutionized forecasting and data analysis.These algorithms possess the ability to extract invaluable insights from vast datasets,enabling the creation of highly accurate forecasting models.This breakthrough holds immense scientific and practical significance.Implementing machine learning algorithms for bias revision in ensemble forecasts represents a pioneering approach aimed at elevating the accuracy and reliability of such predictions.This study sets out to explore the feasibility and effectiveness of utilizing machine learning algorithms to enhance the Nonlinear Local Lyapunov Vector(NLLV)ensemble forecasts,initially based on the renowned Lorenz96 theoretical atmospheric model.Subsequently,the study moves beyond theoretical realms and delves into practical operational forecasting,employing complex numerical models to explore and validate this innovative methodology from multiple perspectives.The primary objective and conclusions of this paper are summarized as follows.(1)This paper rigorously examines and confirms the efficacy of machine learning algorithms in enhancing NLLV ensemble forecasts based on the Lorenz96 model.The results indicate that the machine learning model(Ens-Rid)utilizing the Ridge regression(Ridge)algorithm,in conjunction with the outcomes of NLLV ensemble forecasts,effectively enhances the overall forecast accuracy,surpassing the ensemble average forecasts(Ens Ave),control forecasts(Ctl),and machine learning models based on Ctl results(Ctl-Rid).(2)The extent of improvement in Ens-Rid forecast accuracy is contingent upon the number of ensemble members,with an increased number of members contributing to enhanced overall forecast accuracy of the Ens-Rid model.Further analysis of the attractors of Ens-Rid,Ctl-Rid,and Ens Ave forecasts reveals that their probability distributions exhibit a reduced value domain,increased kurtosis,and a convergence towards the mean,particularly pronounced in Ens-Rid.(3)Building upon the theoretical underpinnings,machine learning algorithms are further employed to enhance practical operational ensemble forecasts.Four machine learning algorithms,namely Ridge,extreme gradient boosting(XGBoost),artificial neural network(ANN),and random forest(RF),are utilized to refine typhoon intensity forecasts within the Typhoon Ensemble Assimilation Forecasting System(TEDAPS)of the Shanghai Typhoon Institute,China Meteorological Administration.The findings underscore the superiority of the XGBoost algorithm over other algorithms,with revisions conducted using XGBoost resulting in a reduction of forecast errors in maximum wind speed(MWS)by more than 35% and in minimum sea level pressure(MSLP)by 15%,in comparison to the Ens Ave approach.(4)Overall,the XGBoost algorithm outperforms the Ens Ave method.However,a few samples exhibit forecast errors larger than those generated by Ens Ave.Bias analysis reveals that TEDAPS typically underestimates typhoon intensity,under-predicts MWS,and over-predicts MSLP.Nonetheless,the application of the XGBoost algorithm mitigates these biases and improves the accuracy of typhoon intensity forecasts.(5)The XGBoost algorithm shows significant variability in the revisions of tropical storms(TS,MWS ≤ 32.6 m/s)and strong typhoons(TY,MWS > 32.6 m/s).Although the Ens Ave method forecasts TS better than TY,the revision of TY is better using the XGBoost algorithm.This indicates that the XGBoost algorithm is more advantageous for improving the forecasts of strong typhoons.In addition,the XGBoost algorithm also performs well in revising the pooled forecast samples of typhoon intensity that are extremely difficult to forecast.