Analysis of hourly cooling load prediction accuracy with data-mining approaches on different training time scales

Abstract

Data-mining approaches for improving building cooling load predictions are presented and analyzed on different training time scales (T-1 to T-6) in this paper. Multiple linear regression (MLR), auto regression (AR), autoregressive exogenous (ARX) model, multiple nonlinear regression (MNR), and a back propagation neural network (BP). In this study, a real case in Guangzhou was examined to verify the performance of the prediction models on different time scales using measured cooling load and weather data. Based on the sensitivity analysis, the variable with largest sensitivity is selected as the input to the proposed MNR model, which can ensure prediction accuracy while increasing prediction speed. The case study shows that the prediction from the MNR model is comparable to or better than that from the BP model on different training time scales. The results also show that the MNR model could provide satisfactory prediction accuracy with less data (e.g., 3 d of weather and historical load data), while the BP model requires at least one month of training data to provide the same prediction accuracy. An analysis of the prediction performance results shows that the prediction accuracy from the regression models will decrease and the prediction accuracy from the BP model gradually increases as the training time scale increases. Surprisingly, the prediction accuracy with the MNR model is superior to that from the BP model with T-1 to T-4 training time scales. Finally, it is shown that the MNR model with the limited variable as inputs provides higher prediction accuracy than four other models with training time scales T-1 to T-4. The MNR model requires less training data, simpler hardware, and low computational complexity, making it preferable for quick on-site load prediction.