Hybrid PSO-SVM and symbolic regression model for agricultural water demand prediction

Agricultural water use is a key link in ensuring food security and the sustainable utilization of water resources. Understanding its evolutionary mechanisms holds significant theoretical and practical importance. This study focuses on Bayannur City, employing the Particle Swarm Optimization-Support Vector Machine (PSO-SVM) model to identify the primary controlling factors of agricultural water demand and using symbolic regression to construct a predictive equation. The results reveal a bidirectional guiding mechanism in agricultural water use: it is driven by factors such as Effective Irrigated Area and Grain Sown Area, and inhibited by High-Efficiency Irrigation Rate, Water Stress Index, and Agricultural Water Price, which demonstrates a linear restraining effect. The prediction model indicates that from 2023 to 2035, Bayannur's agricultural water demand will remain above 5 billion cubic meters annually, peaking at 5.156 billion cubic meters in 2028 before gradually stabilizing. The symbolic regression-based equation effectively captures the nonlinear coupling between driving and restraining factors, showing that restraining forces-particularly those related to policy-are gradually becoming more dominant than production-driven indicators. This study not only provides a reliable forecast of future agricultural water demand but also contributes scientific value by integrating machine learning with interpretable modeling, offering practical support for regional water resource management under the current and future policy frameworks.