Mitigating Peak Edge Effects in Multi-Zone Irrigation: A Safety-Constrained Reinforcement Learning Approach with Short-Term Evapotranspiration Forecasting

To address peak edge operation and excessive valve switching in hydraulically coupled multi-zone campus irrigation, this study proposes a collaborative scheduling framework that combines short-term evapotranspiration (ET) forecasting with safety-constrained reinforcement learning. Temperature, relative humidity, and light intensity are used to construct vapor pressure deficit and radiation proxy features, and a lightweight predictor provides two-hour-ahead ET statistics as forward-looking disturbance information. A safety layer composed of Top-2 gating and total flow projection is then used to map policy outputs into a feasible action space under parallel irrigation and total flow constraints. Using seven consecutive days of field data from October 2025, the proposed method reduced total water consumption to 131.04 m3, corresponding to reductions of 9.13% and 6.12% relative to fixed-schedule and hysteresis threshold rotational irrigation, respectively. It also reduced the maximum total flow from 2.00 to 1.60 L/s, lowered valve switching cycles to 12, and reduced the border ratios at 0.90 and 0.95 to 0. Additional ablation, sensing noise/packet loss, and Top-K extension experiments further showed that ET forecasting improves anticipatory scheduling, whereas safety projection is essential for zero-violation operation. These results demonstrate that the proposed framework provides a practical and deployable solution for safe and water-efficient multi-zone irrigation scheduling under shared pump constraints.