A hybrid hydrodynamic–machine learning framework for optimal control of unsteady flow in large irrigation canals

Unsteady flow dynamics in large-scale irrigation canals pose significant challenges for real-time operational control, particularly in arid regions where precise water delivery is critical for agricultural productivity and resource efficiency. This study presents a hybrid framework that integrates a one-dimensional Saint-Venant hydrodynamic model with long short-term memory (LSTM) neural networks and model predictive control (MPC) to optimize gate scheduling and flow regulation in a branched irrigation network spanning over 400 km. The hydrodynamic module solves the full Saint-Venant equations using an implicit Preissmann scheme, providing physically consistent state estimates and training data for the machine learning component. The LSTM network learns the residual error dynamics between the physics-based simulation and observed field data, effectively compensating for model uncertainty arising from uncharacterized bed roughness, lateral inflows, and structural losses. A rolling-horizon MPC controller integrates the hybrid surrogate model to determine optimal gate positions that minimize water delivery deficits while respecting operational constraints. Validation on the Fergana Valley Main Canal system in Uzbekistan demonstrates that the proposed framework reduces root-mean-square error (RMSE) of flow depth predictions by 43% compared to the standalone hydrodynamic model and achieves a 28% reduction in water delivery inefficiency relative to conventional rule-based operations. The results confirm the efficacy of physics-guided machine learning for scalable, adaptive irrigation management in data-sparse environments.

Перевод пока недоступен