Urban Rainfall Runoff Prediction Using LSTM with DTW-Enhanced Clustering

Urban rainfall-runoff prediction (URRP) serves as a critical methodology for urban flood mitigation and integrated water resource management. Nevertheless, the inherent complexity of urban hydrological systems that are characterized by spatially heterogeneous precipitation patterns, highly nonlinear rainfall-runoff relationships, and strong temporal dependencies poses significant challenges in developing accurate URRP models. To address these challenges, we propose an innovative hybrid deep learning framework [dynamic time warping (DTW)-long short-term memory (LSTM)] that synergistically combines DTW with developed LSTM neural networks. The proposed methodology features two key innovations. First, we developed a DTW-enhanced clustering algorithm that effectively groups rainfall events based on their temporal similarity patterns, thereby addressing the spatial variability challenge. Second, we designed and developed LSTM architecture specifically tailored to capture both the nonlinear transformations and long-term dependencies in urban hydrological systems. Experimental validation in Yuelai New City, Chongqing, demonstrates the model’s superior performance, achieving real-time prediction capability (runtime<3 s) while maintaining computational efficiency (training time<380 s). Comprehensive benchmarking against eight baseline models [feedforward neural network (FNN), recurrent neural network (RNN), gated recurrent unit (GRU), LSTM, DTW-FNN, DTW-RNN, DTW-GRU, and LSTM with K-means clustering] also confirmed the superior performance of the DTW-LSTM model. The experimental results suggested that this model has potential applications in real-time flood early warning systems and smart city management.