Optimizing Power Forecasting Models with Customized Features for Academic and Industrial Buildings

Power consumption prediction is a crucial component in enhancing the efficiency and sustainability of building operations. This study investigates the impact of data collection frequency and model selection on the predictive accuracy of power consumption in two distinct building types: an Academic one with 15-min interval data and an Industrial one with hourly data. Various machine learning models, including Support Vector Machine (SVM) with Radial and Sigmoid kernels, Random Forest (RF), and Deep Neural Networks (DNNs), across different data splits and feature sets, were considered. Our analysis reveals that higher data collection frequency generally improves model performance, as indicated by lower RMSE, MAPE, and CV values, alongside higher R² scores. The inclusion of more historical power consumption features was also found to have a more significant impact on the accuracy of predictions than including climate condition features. Moreover, the SVM-Radial model consistently outperformed others, particularly in capturing complex, non-linear patterns in the data. However, the DNN model, while competent in some metrics, showed elevated MAPE values, suggesting potential overfitting issues. These findings suggest that careful consideration of data frequency, features, and model selection is essential for optimizing power prediction, contributing to more efficient power management strategies in building operations.