Application of an ANN model integrating station network distribution and water level increment principle in flood season water level prediction of Taihu Lake basin

Study region The Taihu Lake basin is a typical plain river network area in eastern China, characterized by extremely gentle terrain, dense river-lake systems, intensive human activities, and complex hydrological interactions. The region is highly vulnerable to flooding during the flood season due to its flat landscape, concentrated rainfall. Study focus This study constructs controlled BP-ANN models with the spatial distribution of hydrological stations and real-time water level increment information. Five ablation BP-ANN models are established to quantify the contributions of water level increment, PLS fusion and multi-station data. SHAP method was adopted to interpret outputs for level prediction. The core focus is to find reliable method to improve water level prediction under complex hydrological conditions, reduce peak errors, and enhance the reliability of regional flood forecasting systems. New hydrological insights for the region Among all developed models, the single-station method provides the optimal compromise between prediction performance and model complexity. The multi-station method yields accuracy gains only for a small subset of stations with strong hydrological connectivity and introduces redundant features for most plain-river-network sites. The upstream and mid-plain areas show different confluence patterns affected by runoff and backwater. The findings support zoned intelligent forecasting for plain river networks, and can serve as a reliable way for flood control and regional hydrological services in the plain river network regions.