Optimal management of unsteady water flow in large main canals using artificial intelligence: a case study of the Amu-Zang main canal

Effective management of unsteady flow in large irrigation main canals is a fundamental challenge in arid-zone water resources engineering. The Amu-Zang Main Canal (AZMC), a 312 km gravity-fed system in southern Uzbekistan, presents unique difficulties owing to its high sediment load, frequent demand fluctuations among 23 offtake nodes, and limited telemetry coverage across its lower reaches. This study proposes an integrated artificial intelligence framework combining an ensemble of Temporal Convolutional Networks (TCN) with a Reinforcement Learning (RL) agent trained via Proximal Policy Optimization (PPO) for real-time gate scheduling and discharge regulation. Unlike residual-correction approaches, the TCN ensemble is trained end-to-end to predict multi-step flow states from raw sensor observations, eliminating the dependency on a pre-calibrated physics-based model. The RL agent interacts with the TCN environment model to discover adaptive control policies that minimize both water delivery deficit and sediment-induced scouring risk – a dual objective not addressed in prior canal control studies. The framework is evaluated on three years of operational data (2021-2023) from the AZMC telemetry network. Results show that the TCN ensemble achieves a mean absolute percentage error (MAPE) of 4.7% for 6-hour ahead flow depth forecasting, outperforming persistence (12.3%) and ARIMA (9.1%) baselines. The RL-based controller reduces the cumulative seasonal delivery deficit by 34% compared to the existing supervisory control protocol, while simultaneously reducing gate velocity-induced bed scour events by 61%. The proposed methodology offers a scalable, model-free pathway toward intelligent canal automation in regions where physics-based calibration data are scarce.

Перевод пока недоступен