Decision-Focused Learning for Water Treatment Plant Operations: Deterministic versus Stochastic Formulations

This paper investigates the use of decision-focused learning for receding horizon control of a wastewater treatment plant equipped with solar panels and battery energy storage, with the goal of minimizing either its operating costs or carbon emissions. We demonstrate that the structure of the underlying optimization problem depends on the chosen objective, and benchmark state-of-the-art decision-focused learning methods accordingly. Our results show that decision-focused learning can reduce operating costs by up to 4.20% and 4.60% with deterministic and stochastic decision-focused learning algorithms, respectively, compared to the traditional predict-then-optimize approach, underscoring the value of training forecasting models using task-specific loss functions.