An Analysis of Prediction Techniques for Model Predictive Control in Smart Energy Storage Systems

Increasing shares of intermittent renewable generation in the power grid require efficient building energy management solutions to preserve grid stability. A prominent approach for controlling on-site storage systems in such settings is Model Predictive Control (MPC), whose effectiveness depends significantly on forecast accuracy. In this paper, we systematically analyze which prediction methods are most suitable for building energy applications and how their quality affects MPC performance. We focus on three representative techniques: a linear model, the tree-based XGBoost algorithm, and TimeMixer, a state-of-the-art neural network architecture. Our findings show that specialized and carefully tuned neural networks can outperform more general-purpose methods such as XGBoost. Furthermore, we demonstrate that the mean squared error (MSE) alone is insufficient as a metric for training models intended for MPC use. Finally, we observe that prediction errors add up across multiple forecast horizons, leading to performance losses of more than one third, compared with perfect forecasts, when using an extensive multi-year, multi-building dataset. These insights underscore the need for robust forecasting and carefully selected loss functions in MPC-based smart energy storage systems.