长江流域月滚动降水趋势预测的深度学习订正方法OA

Monthly rolling precipitation trend forecasting constitutes a major difficulty in sub-seasonal to seasonal(S2S)pre-diction,and its core challenge lies in detecting and correcting the systematic biases embedded in dynamical model outputs.To tackle this challenge,we develop a deep learning bias-correction framework based on a hybrid Convolutional Neural Network(CNN)-Transformer architecture(H2Former).The framework leverages CNNs to capture local fine-scale features,while utili-zing Transformers to model long-range dependencies associated with large-scale circulation patterns.To comprehensively evalu-ate its performance,taking the Changjiang River Basin as a typical study region,we apply the framework to NCEP-CFSv2 fore-casts and conduct systematic bias-correction experiments,input-factor ablation studies,and architectural comparisons.Results indicated that:① the forecast accuracy over the Changjiang River Basin improved from 62.1% to 64.8% on an independent test set covering 2022~2024,with accuracy gains of up to 8.2 percentage points in the topographically complex lower Jinsha River re-gion.② Input-factor ablation experiments revealed that the scheme using only dynamical model forecasts as inputs achieved an accuracy of 64.8%,which was significantly superior to the scheme relying solely on preceding atmospheric-circulation factors(a maximum of 56.7%).③ Architectural comparison experiments further verified the superiority of H2Former,which achieved an accuracy of 64.8%,outperforming classic models including U-Net++(64.0%)and the Transformer-based SwinUNETR(64.2%).Overall,as an intelligent post-processor for dynamical climate models,deep learning exhibits considerable potential for improving forecast skill of S2S precipitation trends.The proposed framework can provide technical support for refined climate prediction services and offer a reference for water resource management,disaster prevention and mitigation decision-making in the Changjiang River Basin.