Investigating the deviation between prediction accuracy metrics and control performance metrics in the context of an ice-based thermal energy storage system

Extensive research on cooling load prediction has been conducted to improve the prediction accuracy. The existing studies mainly use statistical metrics to evaluate prediction accuracy. However, under these metrics, accurate cooling load prediction models do not always lead to better application performance of downstream control optimization. This study demonstrates the phenomenon in a real case of an ice-based thermal energy storage system. There are 180 cooling load prediction models from six different data-driven algorithms developed to test the prediction accuracy and control performance. Four prediction accuracy metrics are investigated, including mean absolute percentage error (MAPE), coefficient of variation of mean absolute error (CV-MAE), coefficient of variation of root mean squared error (CV-RMSE), and coefficient of variation of root weighted squared error (CV-RWSE). To quantitatively evaluate the deviation between prediction accuracy metrics and control performance metrics, the Spearman correlation coefficient is applied. The Spearman correlation coefficients of MAPE, CV-MAE, and CV-RMSE are +0.467, −0.490, and −0.743, respectively, which demonstrates that the prediction model with the highest accuracy does not always lead to the best control performance. It is crucial to customize prediction accuracy metrics for specific control problems. Spearman correlation coefficient can effectively evaluate the monotonicity and concentration between prediction accuracy and control performance, which can support better design of application-oriented prediction accuracy metrics and facilitate the interaction between cooling load prediction and control tasks.