Flood Classification and Improved Loss Function by Combining Deep Learning Models to Improve Water Level Prediction in a Small Mountain Watershed

Abstract

Floods are major natural disasters that present considerable challenges to socioeconomic and ecological systems. Flash floods are highly nonlinear and exhibit rapid spatiotemporal variability. Existing methods struggle to capture these features, leading to suboptimal long-term and peak flood prediction accuracy. This study proposes a hierarchical flood prediction model based on clustering to enhance forecasting accuracy in the Heshengxi watershed. We employ STGCN and GWN models with the spatiotemporal attention mechanism. Enhanced loss functions further refine flood prediction accuracy. Results show that the hierarchical prediction method is an effective means of extracting flood features by addressing the variability of prediction parameters for different flood magnitudes. The integration of Graph Convolutional and Time Aware models enables the model to recognize the spatiotemporal flood characteristics, overcoming limitations of prevailing methods and ensuring long-term forecast accuracy. The optimized loss function further improves the prediction performance, resulting in a significant improvement in the accuracy of flood peak prediction, with a reduction of 0.26% in the relative error of the peak prediction by the GWN model. This framework provides an effective solution for flood warning, emergency response, and optimal scheduling. It also demonstrates the potential of deep learning models in the field of intelligent hydrological forecasting.