Benchmarking Machine Learning Algorithms for Predicting Building Energy Performance

Current data-driven load-forecasting models for residential buildings often overlook the impact of evolving occupancy patterns, relying on default hyperparameters. This oversight may result in reduced predictive accuracy with behavioural changes. This paper introduces a benchmarking approach evaluating common data-driven models-random forests, multilayer perceptron neural networks, extreme gradient boosting, and support vector machines-under concept drift. The day-ahead heating load prediction models are trained based on a series of input variables, including, inter alia, weather conditions, calendar information, and historical data. Optimal training periods and hyperparameter tuning are determined using a development set, while predictive accuracy is assessed through cross-validation on an evaluation set. Simulation results demonstrate that machine learning models trained on datasets with varying occupancy profiles perform comparably to those trained and evaluated with consistent profiles. Although optimal hyperparameters may vary, the models maintain consistent performance across datasets with diverse occupancy patterns, indicating resilience to changes in occupant behaviour.