Rainfall Forecasting Based on Spatio-Temporal Information Fusion Using Informer

Frequent flood disasters in China cause substantial economic losses and casualties each year, endangering lives and properties. However, existing machine learning-based flood prediction models encounter challenges in effectively extracting relevant data features, leading to suboptimal prediction accuracy and a higher risk of overfitting. To address these issues, this study proposes a deep learning approach that leverages diverse datasets to predict rainfall indicators. The research commences with an analysis of the topography of Zhejiang Province and the geographical distribution of monitoring stations, followed by data preprocessing. Subsequently, an enhanced Informer model is employed, integrating the Prob-attention module to overcome the limitations of previous models concerning long prediction lengths in data and to extract temporal and spatial features more effectively. Experimental results show that the model significantly outperforms traditional time series prediction algorithms, achieving better accuracy with average RMSE, MAPE, and R2 values of 0.39, 9.51%, and 0.91, respectively.