Enhancing Multidimensional Rainfall and Reservoir Flow Forecasting: Innovations and Improvements in the Informer Model

In China, the annual occurrence of flooding events inflicts significant economic damages and human casualties, posing a serious threat to both life and property. Despite the adoption of machine learning techniques in flood forecasting, these models often struggle with the effective extraction of critical data features, resulting in less than optimal predictive accuracy and an increased propensity for overfitting. This study introduces a novel deep learning methodology that utilizes a variety of data sets for the prediction of rainfall metrics, aiming to surmount these hurdles. The research begins by examining the topographical features of Zhejiang Province and the spatial arrangement of its monitoring stations, followed by meticulous data preprocessing. An advanced version of the Informer model, incorporating a Prob-attention mechanism, is then deployed. This enhancement is designed to address the previous models' shortcomings related to long-term prediction intervals and to more proficiently capture both temporal and spatial data characteristics. The experimental outcomes demonstrate the superiority of our model over traditional time series forecasting methods, with it achieving notably higher accuracy, evidenced by average RMSE, MAPE, and R^2 values of 0.35, 8.75%, and 0.925, respectively.