Hybrid ITSP-LSTM Approach for Stochastic Citrus Water Allocation Addressing Trade-Offs Between Hydrological-Economic Factors and Spatial Heterogeneity

Abstract

This study addresses the critical challenge of optimizing water resource allocation in fragmented citrus cultivation zones, particularly in Anfusi Town, a key citrus production area in China’s middle-lower Yangtze River region. To overcome the limitations of traditional deterministic models and spatially heterogeneous water supply–demand dynamics, an innovative framework integrating interval two-stage stochastic programming (ITSP) with long short-term memory (LSTM) neural networks is proposed. The LSTM component forecasts irrigation demand and supply under climate variability, while ITSP optimizes dynamic allocation strategies by quantifying uncertainties through interval analysis and balancing economic returns with hydrological risks. Key results demonstrate an 8.67% increase in system-wide benefits compared to baseline practices in the current year scenario. For the planning year (2025), the model identifies optimal water distribution thresholds: an upper limit of 3.85 × 106 m3 for high-availability zone A and lower limits of 1.62 × 106 m3 for moderate-to-low-availability zones B and C. These allocations minimize water scarcity penalties while maximizing net benefits, prioritizing local over external water sources to reduce costs. The study innovates by integrating stochastic-economic analysis with spatial prioritization of high-marginal-benefit zones and uncertainty robustness via interval analysis and two-stage decision making. By bridging a research gap in citrus irrigation optimization, this approach advances sustainable water management in complex agricultural systems, offering a scalable solution for regions facing fragmented landscapes and climate-driven water scarcity.