Robust building energy consumption forecasting using an online learning approach with R ranger

Recently, the online learning-based stacking ensemble approach has yielded satisfactory short-term load forecasting (STLF) because it can effectively reflect recent building energy consumption patterns in model training. However, when buildings have large variability in building energy consumption patterns between the training data and the unseen data, this approach faces difficulties because it learns the patterns using only the predicted values of multiple base models constructed using the training data. To solve this problem, we propose a robust two-stage building-level STLF model called RABOLA (short for RAnager-Based Online Learning Approach) to enable practical and fast pattern learning for the unseen data. We first collect publicly available electricity consumption data for two office buildings and construct training and test sets by performing data preprocessing for input variable configuration. In the first stage, we construct three STLF models based on tree-based ensemble learning approaches using the training set. In the second stage, we construct a ranger-based forecasting model with a sliding window size of seven days using the predicted values of the three models and external factors such as timestamp and temperature as input variables on the test set. We demonstrated through extensive comparative experiments that the RABOLA model outperforms the prediction performance of state-of-the-art stacking ensemble and deep learning approaches in terms of mean absolute percentage error and coefficient of variation of the root-mean-square-error. In addition, we present the relationships between input and output variables in building-level STLF using two interpretability methods.