Reinforcement learning-enhanced PID control of irrigation canals: performance assessment under normal and water-deficit conditions

Optimizing water distribution in water conveyance and delivery systems is of paramount importance in arid and semi-arid regions such as Iran. In this research, a hybrid control framework was developed by integrating the Twin-Delayed Deep Deterministic Policy Gradient reinforcement learning algorithm (RL-TD3) with a classical proportional-integral-derivative (PID) controller and a linear Integrator-Delay (ID) canal simulator for the concrete-lined E1-R1 canal within the Dez irrigation network. The key novelty of this study lies in the development of a Lyapunov-constrained RL-enhanced PID–ID control architecture that simultaneously guarantees closed-loop stability and improves irrigation performance under both normal and severe water-deficit conditions in a real large-scale irrigation canal. The proposed framework was utilized to investigate and evaluate the hydraulic performance of the canal under multiple operational scenarios. The performance assessment employed four standard irrigation metrics—adequacy, efficiency, equity, and dependability—alongside statistical indices to evaluate eight operational scenarios encompassing both normal and water-deficit conditions. The results demonstrate that the developed ID model is computationally efficient and sufficiently accurate, achieving an average relative error of 0.066 for flow variations of up to 30%. Performance was excellent under normal operation, with adequacy and efficiency indices approaching unity and equity and dependability indices approaching zero. Under water-deficit conditions, the linear model accuracy decreased noticeably, particularly for equity (increasing up to 0.113) and dependability (increasing up to 0.032). The RL-TD3 algorithm autonomously identified optimal controller gains of \(-1.6988\) , \(-0.055\) , and \(-0.086\) for \(k_p\) , \(k_i\) , and \(k_d\) , respectively, resulting in significantly improved transient response and reduced control cost compared to manually tuned gains of \(-0.209\) , \(-0.243\) , and \(-0.086\) .