Empirical decision model learning for multi-sector greenhouse irrigation under water supply restrictions

Existing tools for irrigation management largely overlook two critical operational challenges: the combinatorial nature of water volume allocation in multisector systems, where heterogeneous sectors must be jointly managed under shared hydraulic constraints, and the water supply restrictions imposed by collective irrigation systems, which bound the total volume available to each irrigation system over a predefined time window. We propose an instantiation of the Empirical Decision Model Learning (EDML) paradigm, introduced by Lombardi et al. (2017), to address the joint Water Volume Allocation (WVA) problem in a multisector greenhouse irrigation system: for each sector, a machine learning model is trained as a function approximator of the soil moisture response during the irrigation period, and embedded as a set of linear constraints within a Mixed-Integer Linear Programming formulation that optimizes WVA decisions across all sectors under a lexicographic multi-objective structure that pursues soil moisture adequacy across sectors while minimizing water use. Evaluated on a real multisector greenhouse dataset against three baselines of increasing sophistication, the proposed EDML framework achieves a 79.1% reduction in water consumption compared to fixed-interval irrigation while driving 74.4% of sector instances into the optimal agronomic moisture range. A scalability analysis confirms solver runtimes below 0.3 s for systems with up to 50 sectors.