Statistical Post-Processing for Precipitation Forecast Through Deep Learning Coupling Large-Scale and Local-Scale Spatiotemporal Information

Accurate forecast precipitation is crucial for hydropower generation, drought and flood warning, and hydrological forecasting. However, raw forecast precipitation often suffers from systematic errors due to inaccurate initial conditions in numerical weather prediction (NWP) models. In this study, we develop a deep-learning-based post-processing method to correct forecast precipitation. Our method leverages convolutional neural networks (CNN) to analyze spatial features and long short-term memory networks (LSTM) to capture temporal dynamics, effectively modeling the local spatiotemporal characteristics (e.g., mean sea level pressure and elevation) of precipitation. Crucially, we also consider the impact of large-scale weather patterns (e.g., high-latitude blockings, the Meiyu trough) on precipitation by extracting relevant features through a CNN model and integrating this information with the local spatiotemporal data to improve forecast accuracy. Results indicate that the proposed CNN-CNN-LSTM method outperforms the three baselines (i.e., CNN-LSTM, CNN, LSTM) for all seasons and lead times (15 days) in the Huaihe River basin of China. Specifically, for the summer precipitation with a one-day lead time, the CNN-CNN-LSTM model achieves a 4.7% reduction in root mean square error and a 30.5% reduction in relative bias compared to CNN-LSTM alone. Furthermore, the relative importance of large-scale predictors is constantly increasing with the extension of lead times. By effectively integrating large-scale weather information and local-scale spatiotemporal information, the proposed CNN-CNN-LSTM method offers a novel approach to enhance the correction effect, providing significant valuable for hydrometeorological applications.