Prediction-based control of energy storage systems using dynamic accuracy weighting

Abstract

Integrating domain knowledge into artificial intelligence models is increasingly recognized as essential for improving energy storage system control based on load predictions. Commonly used accuracy metrics for load prediction models, such as mean absolute percentage error, coefficient of variation of mean absolute error, and coefficient of variation of root mean squared error, are not monotonically correlated with final control performance; in other words, the model with the highest prediction accuracy does not necessarily yield optimal control outcomes. This study introduces a dynamically weighted error metric, which incorporates the attributes of energy storage systems and the temporal dynamics of prediction-based control by leveraging domain knowledge from heating, ventilation, and air conditioning systems. The proposed dynamically weighted error metric enhanced the selection of load prediction models, and these models reduced the operating cost of six energy storage systems by up to 6.5 % compared to those using traditional prediction accuracy metrics. The scalability of dynamically weighted error metric was further validated across 10 energy storage capacities and 18 Time-of-Use tariffs in the six building cases, achieving 93.9 %–97.2 % of the ideal cost reductions and outperforming traditional metrics (86.4 %–95.4 %). The applicability of dynamically weighted error metric to common energy storage systems is discussed and confirmed. Additionally, a web-based tool was developed to facilitate dynamically weighted error calculation in practical applications. This study demonstrates that incorporating domain knowledge through dynamic accuracy weighting evidently enhances the whole-process performance of artificial intelligence in energy storage system control.