Rolling recursive prediction model for water level prediction in gate-controlled water delivery canal systems

Study region This study focuses on a cascade canal section of the Middle Route of the South-to-North Water Diversion Project (SNWDMRP), China, characterized by large flows and frequent dynamic gate regulations and pronounced hydraulic lag effects over long distances. Study focus Accurate multi-step water-level prediction is vital for large water diversion systems. Existing data-driven models often use static strategies unsuited for artificial gate operations. We propose a novel rolling recursive prediction method for gate-controlled canals that explicitly integrates planned gate operation data to enable scenario-based water-level prediction. Additionally, to overcome imbalanced historical datasets lacking extreme events, we employ a hydrodynamic model for model-based data augmentation to simulate high-fluctuation scenarios. New hydrological insights for the region This study proposes a rolling recursive prediction model for multi-step water-level prediction in gate-controlled canal systems. At each prediction step, the model updates hydraulic state variables with the most recent observations and gate operation inputs, capturing the spatiotemporal coupling among water levels, gate controls, and boundary conditions over extended horizons. The LSTM-Rolling model maintained NSE above 0.8 across all gate sites over a 72 h horizon, while static models dropped below 0.5. Hydrodynamic model-based data augmentation supplemented the training data with physically plausible scenarios that are under-represented in the historical record, reducing the 24 h RMSE from 0.024 m to 0.013 m. Together, the rolling recursive model and targeted data augmentation provide a practical approach for scenario-conditioned water-level prediction in engineered water diversion canals.