A Comparative Study of Reliable Predictive Models for Energy-Efficient MPC in Buildings

Unlike traditional controllers, Model Predictive Control (MPC) leverages predictive models to predict future system behavior, enabling proactive and efficient decision-making over a defined time horizon. Its effectiveness depends on the accuracy of the models used. Inaccurate predictions can compromise system reliability, reduce stability, and lead to suboptimal control decisions. This work considers the problem of optimizing energy efficiency while ensuring thermal comfort inside buildings, aiming at proposing a suitable MPC strategy in this context. This paper presents preliminary results towards this goal, comparing a gray-box Resistor-Capacitor (RC) model and a black-box Artificial Neural Network (ANN) trained on Pseudo-Random Binary Sequence (PRBS) generated data. Results show that while the RC model is interpretable, the ANN achieves superior long-term accuracy. Observing the Mean Absolute Error (MAE) curves and their standard deviation, we conclude that the reliable prediction horizon, which ensures a prediction error below 1°C, is only 9.5 hours with the RC model, and increases to more than 24 hours with the ANN model, making it more suitable for MPC applications.