Distribution-Adaptive Graph Attention Networks for Flood Forecasting

Flood forecasting is an important task for disaster prevention and mitigation. Many recent researchers intend to utilize data-driven deep-learning models to improve their prediction accuracy. Deep-learning technology commonly assumes that the time series data is independently and identically distributed. However, as time goes on, environmental changes can cause the distribution of temporal data to change. Neglecting considerations of distribution changes can lead to a decrease in prediction accuracy. In addition to distribution changes, the accuracy of flood forecasting is also influenced by the spatiotemporal relationships among the flood factors. This paper proposes a flood forecasting model based on Distribution-Adaptive Graph Attention Networks (DAGAT). DAGAT can extract spatiotemporal information from flood data and capture the spatial relative importance among flood factors. In the meantime, it also uses the distribution adaptation mechanism of the Boosting algorithm to train weight parameters, enabling the reduction of distribution differences among different segmented periods and effectively improving the accuracy of flood forecasting. Through comparative experiments, this method’s effectiveness and superiority of this method are validated, demonstrating the potential application value in flood forecasting.