Key points are not available for this paper at this time.
The increasing share of intermittent renewable energy calls for intelligent building energy management systems to maintain grid stability. A widely used method for operating on-site storage is model predictive control (MPC), whose effectiveness heavily depends on forecast accuracy. This paper systematically evaluates the impact of prediction models on MPC performance in smart energy storage systems (SESS). Using a three-year, multi-building dataset with 15 min resolution, we compare five forecasting methods, linear model, XGBoost, RNN, TimeMixer, and TimesNet, for load, PV generation, and electricity price prediction. While XGBoost achieves the lowest mean squared error (MSE) and yields the highest revenue gain of 104% over a no-storage baseline during a four-month winter–spring test period, other models reveal a mismatch between forecast accuracy and control performance. Notably, the linear model, ranking mostly lowest in MSE, delivers the third-highest revenue (73%), nearly on par with the second best (79%). This illustrates that prediction accuracy alone is not a reliable proxy for control quality. Even the best realistic setup remains far from the ideal benchmark using perfect forecasts (235% gain). Daily retraining improves some models substantially (linear model to 105%) but has limited effect on others (XGBoost to 107%). These findings emphasize three key insights: (1) standard metrics like MSE may misrepresent the utility of forecasts for control, (2) errors across multiple inputs compound degradation in MPC, and (3) frequent retraining can mitigate losses. Overall, the results underscore the importance of robust forecasting and carefully chosen loss functions in the smart energy systems concept. • Evaluation of forecasting models on multi-year, multi-building data. • XGBoost achieves lowest test MSE (load: 0.372, PV: 0.287, price: 0.261). • Imperfect forecasts drop MPC revenue gain from 235% to 104%. • Daily retraining improves MPC revenue gain slightly to 107%.
Ludolfinger et al. (Thu,) studied this question.