A multi-objective dynamic water allocation framework for large-scale hierarchical canal networks

Canal water-allocation optimization is essential for sustainable irrigation-district development under water scarcity. However, strong spatiotemporal variability in eco-hydrological conditions makes water supply and demand difficult to predict, thereby weakening allocation effectiveness. This study proposes a MARL-based strategy for multi-objective water allocation in large-scale hierarchical canal networks, termed MWAHC, using rolling updates of dynamically update irrigation-district information. The framework represents the hierarchical water-delivery and allocation process from the main canal to branch canals and irrigation subareas, while coordinating the objectives of water conservation, crop production, groundwater-depth regulation, and root-zone salinity control. A centralized shared team reward coordinated local-agent decisions, while updated eco-hydrological feedback guided subsequent monthly adjustments. A hard constraint mapping based on the maximum conveyance capacities of canals at each level ensured the operational feasibility of the water allocation scheme. Data from 2000 to 2014, 2015–2017, and 2018–2022 were used for training, validation, and independent testing, respectively, in the Hetao Irrigation District, China. Compared with historical schemes, MWAHC reduced annual diversion by 0.63–5.18 × 10⁸ m³ , with water-saving ratios of 1.39%–9.85%, while crop production was maintained. Groundwater depth remained within the 2.5 m ecological threshold, and the growing-season root-zone soil salt accumulation increment remained below 1.2 g/kg. Furthermore, comparisons with historical mean meteorological inputs and perfect forecasts confirmed that updated forecasts improved water-allocation performance, while substantial water savings and simulated constraint satisfaction were retained despite forecast errors. Temporally, the learned strategy moderately increased spring irrigation in April and substantially reduced autumn irrigation in October–November. Spatially, water was shifted toward high-efficiency canals and high-risk irrigation subareas. Across both temporal and spatial dimensions, allocations were further refined using updated rainfall, evapotranspiration, and drainage information. These results demonstrate that MWAHC provides a decision-support framework for intelligent dynamic allocation in complex canal networks under coordinated multi-objective regulation.