Multi‐agent MPC for Irrigation Scheduling: A Learning‐based Approach

This chapter develops a unified learning-based framework that integrates mixed-integer model predictive control (MPC) and the three key machine learning paradigms for daily irrigation scheduling in spatially heterogeneous fields. It begins by presenting a three-stage management-zone delineation procedure that integrates crop information, topography, estimated soil hydraulic parameters, and k-means clustering to obtain zones that reflect within-field variability. The resulting management zones are then mapped to the resolution of the field's irrigation system to enable efficient water application. The chapter then introduces management zone–level surrogate models for soil-moisture dynamics using long short-term memory networks. In addition, a decentralized multi-agent reinforcement learning strategy is proposed in which local agents are trained to determine both irrigation timing and irrigation amounts at the management-zone level. A coordination mechanism, based on the concept of a limiting management zone, is introduced to unify the timing decisions obtained by the local agents. This unified timing sequence is then enforced within the mixed-integer MPC formulation to determine irrigation amounts across zones. The framework is evaluated on a real-field case study and is compared against a baseline scheduling approach, where it demonstrates improved irrigation water-use efficiency.