Machine Learning–Enhanced Runoff Prediction in Plain River Networks: Insights into Recurrent River Segments

Urbanization and hydrological modifications have significantly altered natural drainage systems, reducing river networks’ capacity to manage runoff efficiently. In low-lying regions with complex river networks, heavy rainfall and rising water levels often exacerbate flooding and, in some cases, trigger backflow. Under these conditions, recurrent flow characteristics become crucial indicators of runoff variations in local river segments. Due to the nonlinear and periodic nature of these flows, accurately predicting runoff remains a critical challenge in hydrological research. To address this, we propose a novel SARIMA-WOA-LSTM model. First, the seasonal autoregressive integrated moving average (SARIMA) model captures linear trends and seasonal components in the time series, yielding predicted values and residuals. These residuals, representing the nonlinear component, then serve as input to a long short-term memory (LSTM) network. Meanwhile, the whale optimization algorithm (WOA) is employed to globally optimize LSTM hyperparameters, enhancing the model’s predictive performance. Taking the flow variation in Changxing Port, China, as a case study, the proposed SARIMA-WOA-LSTM model is compared with various single, optimized, and hybrid models. Experimental results show that SARIMA-WOA-LSTM achieves superior accuracy, precision, and stability (MAE = 1.29, RMSE = 2.14, R2=0.986). WOA-driven hyperparameter optimization significantly reduces model errors, further improving prediction outcomes. These findings demonstrate that SARIMA-WOA-LSTM has substantial application value and offers a strong reference for runoff prediction in plain river networks.