Maximum Temperature Forecasts in the Yangtze River Valley Using a Predictor‐Corrector Lightweight Deep Learning

Abstract Accurate maximum temperature forecasting is vital for heatwave management and agricultural planning in the Yangtze River Valley. Numerical weather prediction methods often face challenges from atmospheric chaos, limited physical understanding, and external forces, resulting in uncertainties and biases. Deep learning offers a promising alternative with faster processing and higher accuracy. Because individual predictors, whether convolutional or transformer‐based networks, can fall short, this study introduces a lightweight forecasting framework that combines a ResNet‐based predictor with a multilayer‐perceptron‐based corrector, achieving up to 7.8% accuracy improvement. Compared to European center for medium‐range weather (EC) forecasts, this framework achieves similar RMSE but has better spatial correlation and tolerance‐based accuracy. In addition, it outperforms EC forecasts by predicting heatwave onset 9 days earlier with a one‐day lead. These results demonstrate that coupling a corrector with a predictor significantly enhances the forecasting capabilities of lightweight models, offering a reliable and cost‐effective tool for extreme weather forecasting.