Multi-Time Window Runoff Prediction Based on Multi-Head Attention Model for Domestic Watershed

Runoff prediction is essential in water resources management, flood control and disaster reduction, and ecological environment protection. However, traditional methods face challenges in capturing complex nonlinear relationships and multi-scale temporal dynamics.Therefore, we propose multi-time window runoff prediction model (MHA-LSTM) based on multi-head attention mechanism. The model combines long short-term memory network (LSTM) to capture short and long-term dynamic features in runoff time series, and introduces Multi-Head Attention mechanism to enhance the ability of the model to extract key information, thus improving the prediction accuracy. An empirical study in Qingshuijiang River Basin in China shows that MHA-LSTM performs significantly better in runoff prediction than traditional LSTM and ConvLSTM models. In single-step prediction, the Nash-Sutcliffe efficiency coefficient (NSE) of MHA-LSTM reached 0.8945, surpassing LSTM by 0.1265 and convl-LSTM by 0.1220. In the 8-step prediction, NSE is 0.8763, which is increased by 0.1745 and 0.2104 compared with the two models respectively. In the 24-step prediction, the NSE is 0.8034, which is 0.1398 and 0.1284 higher than the two models, respectively.