Robust Irrigation Control Under Soil-Parameter Uncertainty: A Comparative Study of Model Predictive Control and Proximal Policy Optimization

Efficient irrigation control requires maintaining adequate soil moisture while limiting unnecessary water application and deep drainage. This task becomes more challenging when soil hydraulic parameters are uncertain, since changes in field capacity, wilting point, and drainage characteristics can substantially alter the soil-water response. This study investigates closed-loop irrigation control under explicit soil-parameter uncertainty using a three-layer soil-water model. A one-at-a-time sensitivity analysis is first performed to identify influential parameters, followed by six uncertainty scenarios involving field capacity, wilting point, and the deep-drainage coefficient. Two adaptive control strategies, nonlinear Model Predictive Control (MPC) and Proximal Policy Optimization (PPO), are then evaluated using the same soil-water dynamics, environmental conditions, and irrigation constraints. A fixed irrigation schedule is used as a reference case. Under nominal conditions, the fixed schedule produced 100.55 mm of deep drainage and a rootzone moisture RMSE of 0.0556. MPC reduced the RMSE to 0.0233 with 308.55 mm of irrigation and zero deep drainage, while PPO achieved an RMSE of 0.0248 with 324.26 mm of irrigation and zero deep drainage. Under the six uncertainty scenarios, MPC required between 243.32 and 370.08 mm of irrigation and maintained an RMSE range of 0.0214–0.0251. PPO used between 322.15 and 326.45 mm of irrigation, while its RMSE ranged from 0.0116 to 0.0388. All simulated cases satisfied the implemented soil-moisture constraints and waterbalance validation checks. The results demonstrate that the two adaptive approaches exhibit different responses to soilparameter uncertainty. MPC showed stronger scenario-dependent adaptation with relatively stable moisture-tracking performance, whereas PPO produced more consistent irrigation quantities but greater variation in tracking error across the uncertainty scenarios. These findings highlight the importance of evaluating irrigation controllers beyond nominal conditions and provide a controlled comparison between model-based predictive control and learning-based policy control under uncertain soil dynamics.