Comparison of data-driven statistical techniques for cooling demand modelling of electric chiller plants in commercial districts

Abstract

This paper models and forecasts the cooling demand of centralized chiller plants in commercial districts using different statistical techniques – including OLS, Lasso, ARX, SARIMA, SARIMAX and Cochrane–Orcutt. Direct estimation of the test error using a validation data set is used to compare different techniques and to quantify goodness of fit. As a validation, these statistical techniques are compared for forecasting the cooling demand of the largest chilled water plant on the Colorado School of Mines’ campus. Overall, Cochrane–Orcutt provides the highest accuracy among these techniques, with a lowest MSE, RMSE, CV-RMSE and MBE and highest r 2 value. As a showcase of the capabilities of the developed cooling demand, the predicted demand is coupled with Hydeman et al.’s electric chiller model to predict chiller’s electric demand. The RMSE, CV-RMSE and r 2 of the electric chiller model are 22.7 kWe, 17% and 0.84, respectively.