Research On Seasonal Prediction Of Summer Precipitation And Temperature In The Middle And Lower Reaches Of The Yangtze River Based On Deep Learning Method

At present,traditional statistical and dynamical prediction methods have low prediction skills for summer precipitation in the middle and lower reaches of the Yangtze River(MLYR).Deep learning methods have led to progress in many fields,but their applications in seasonal precipitation prediction so far have been rare.This paper proposes the TU-Net method short for the Two-step U-Net model,which contains a Western North Pacific Subtropical High(WNPSH)prediction model and a precipitation prediction model fed by the WNPSH predictions and 4-month lead oceanic heat content and surface temperature.By evaluating the prediction skills of the TU-Net model compared with several dynamical models,the paper explores whether it could improve the prediction ability of summer precipitation in the MLYR region.And then,this method is applied to predict 2-m air temperature.The main conclusions are given as follows:(1)The TU-Net produces comparable skills at a 4-month lead in forecasting geopotential height and wind at 500 h Pa and 850 h Pa levels to dynamical climate models.The TU-Net model has higher prediction skills for 500-h Pa and 850-h Pa geopotential height and 500-h Pa zonal wind in summer months but lower skills for meridional wind.Overall,the prediction skills of the TU-Net in June and July are higher than in August and boreal summer(June-July-August,JJA).(2)the TU-Net can provide moderate skills in predicting summer precipitation in the MLYR region with positive ACC prediction skills in most areas and less-than-one RMSE_nvalues.Compared with NUIST-CFS1.0 and NMME,the TU-Net has the best RMSE_w and Cor scores in each summer month,and PCC scores are slightly lower than the other models only in June and JJA.The TU-Net has the best prediction skills of August precipitation,with an RMSE_wscore of 2.04,a PCC score of 0.18,and a Cor score of 0.44,improved by 5.6%,157%,and 159%over the best of the dynamical models,respectively.(3)To continue to explore the application of the TU-Net method to other climate variables,this paper constructs a 2-m temperature prediction model by replacing the output rather than changing the architecture of the precipitation prediction model.The prediction skills of the TU-Net are superior to some state-of-the-art models,or even better than all the models.