Resource-Constrained Intelligent Scheduling of Agricultural Irrigation Systems using Artificial Intelligence for Sustainable Water Management

This study proposes a resource-constrained intelligent scheduling framework for sustainable water resource management in order to tackle the problems of multi-resource constraints and multi-objective collaborative optimization in agricultural irrigation systems. This framework integrates mixed-integer nonlinear programming (MINLP), reinforcement learning (RL), and a super-efficiency relaxation measure (SBM) evaluation mechanism for comprehensive modeling. The study constructs a closed-loop scheduling system of "modeling-decision-evaluation-feedback." At the modeling layer, a high-dimensional continuous state-space model based on soil moisture dynamics is established. This operation can explicitly characterize multiple constraints such as water resources, energy consumption, equipment capacity, and crop growth requirements. At the optimization layer, joint control variables are designed, incorporating both discrete start-stop decisions and continuous water allocation. At the intelligent decision-making layer, the irrigation scheduling process is modeled as a Markov decision process. This process introduces a comprehensive reward function that includes yield gains, water waste, regional equity, and energy costs. At the evaluation layer, this study constructs a super-efficiency SBM model to evaluate the operational performance of each irrigation unit from multiple dimensions and uses the efficiency vector as a feedback signal to dynamically adjust the scheduling strategy. The experiment was conducted on a simulation-experiment fusion platform comprising 30 irrigation units, mixed crop types, a 120-day scheduling cycle, and an 86-dimensional system state vector. Experimental results show that, compared to rule-driven methods, RCRL-SBM reduces total irrigation water consumption by 19.3%, increases total crop yield by 11.4%, improves water productivity by 14.3%, and lowers the regional equity index (Gini coefficient) from 0.287 to 0.176 while maintaining stable yield growth. Compared to genetic algorithms and standard deep reinforcement learning methods, this result reduces the scheduling stability index to 0.143 and energy consumption to 4,310 kWh. Ablation experiments demonstrate that introducing an efficiency feedback closed-loop mechanism can further improve water productivity by 5.2% and bring the overall system efficiency into the super-efficiency range (θ_SE > 1).