Enhancing Precipitation Nowcasting using ConvLSTM model with multiple meteorological factors

The short-term warning of meteorological disasters has become increasingly important with advancements in meteorological observation technology. However, accurate precipitation forecasting remains a significant challenge that meteorological researchers need to address. Currently, traditional precipitation forecasting methods suffer from low computational efficiency and cumulative error issues. Moreover, most deep learning approaches primarily focus on extracting future characteristics of precipitation using precipitation data alone, neglecting the interrelationships among various meteorological variables within the atmospheric system. To address the various issues present in existing methods, this study proposes a multifactor deep learning architecture that integrates ConvLSTM networks with multiple meteorological elements. Additionally, a hybrid loss function is employed to enhance the effectiveness of precipitation forecasting. All experiments conducted in this study utilize the publicly available SEVIRE dataset. Overall comparative results demonstrate that our methodological strategy outperforms other approaches in terms of the CSI metric and the HSS metric, achieving improvements of up to 16.98% and 17.49%, respectively. Ablation studies further confirm that incorporating additional meteorological variables into the learning process can enhance the predictive performance of the model, underscoring the significance of these extra meteorological variables in precipitation forecasting.