基于EMA-UNet模型的次季节温度预报校正OA

To address the prediction bias in subseasonal temperature forecasting resulting from the limited capacity of traditional models and existing deep learning methods to fully capture multiva-riate spatiotemporal dependencies,an ensemble neural network model based on an improved U-Net architecture EMA-UNet)is proposed.By incorporating a multi-level residual structure,a temporal embedding module,and a cross-channel spatiotemporal attention mechanism,the model overcomes the limitations of traditional architectures and achieves an accurate characterization of dynamic tem-perature evolution.Experimental results show that the EMA-UNet consistently outperforms the tradi-tional physical model EC)and mainstream deep learning models,significantly reducing prediction errors at the subseasonal scale and demonstrating enhanced robustness during high-temperature peri-ods in summer.Ablation studies further validate the effectiveness of the core modules in optimizing multivariate modeling.