Dynamic weighted networks based on transformer for flood prediction in regions with sparse data

Catastrophic floods triggered by extreme weather events have caused enormous casualties and economic losses, and the situation is particularly severe in cities with a lack of flood data. This paper addresses the issue that in flood-data-scarce regions, it is difficult to effectively analyze and predict floods by combining machine-learning methods. By leveraging the concept of transfer learning, we transfer hydrological feature data from data-rich regions to data-sparse ones for learning and training. Additionally, a dynamic weighted self-attention network is introduced to improve the original transformer encoder-decoder architecture. Based on these, the dynamic weighted net-transformer (DWN–transformer) model is constructed, which can accurately predict and fill hydrological data related to floods and is applied to flood risk assessment. Compared with the standard transformer attention mechanism, the dynamic weighted self-attention (DW– SA) network can better recognize data features and risk factors in the target domain and dynamically adjust parameters adaptively. Experimental results show that on the selected Longmen Station target domain test set, the Nash-Sutcliffe Efficiency (NSE) of the model reaches 0.8023, showing a 14.75% improvement compared to the original transformer. This model provides a highly adaptable method for fitting flood-related data and predicting flood risks in areas with missing or scarce data.