ANN-based model predictive control for optimizing space cooling management

This study integrates artificial neural networks into a simulation and optimization framework to implement model predictive control (MPC) for residential space cooling. Based on weather forecasts, the framework provides the optimal setpoint scheduling over a daily planning horizon to reduce energy consumption, costs, and occupant thermal discomfort. A multi-objective optimization approach is adopted, targeting the minimization of system operating costs and of a novel function, defined as comfort penalty, which quantifies potential occupant discomfort hours throughout the day. A genetic algorithm is employed for optimization, while feedforward neural networks are trained to replicate and predict the behavior of the building-plant system. The feedforward neural networks are trained to predict both indoor temperature and cooling loads, demonstrating promising accuracy when compared to building model outputs. Upon obtaining the Pareto front, the optimal solutions are compared with a typical summer control strategy. Results show potential savings of up to 49 % without compromising the other objective, or simultaneous improvements of both objectives, with reductions of 30 % in cooling costs and 27 % in comfort penalty (utopia criterion). These findings highlight that, when properly designed, metamodels can accurately predict building-plant dynamics and deliver reliable optimization results with minimal computational effort.