Comparative analysis of machine learning models for predicting heating demand and electricity consumption using weather data: Two case studies in Norway

Reducing energy consumption in buildings is a widely pursued goal, achievable through careful planning during the early design phase. This study evaluates the accuracy of eight different machine learning models in predicting heating demand and electricity consumption in two case buildings located in eastern Norway. The primary objective was to determine which model provides the most accurate predictions. Data on hourly heating demand and electricity consumption were collected for the two buildings, followed by data mining to prepare the datasets for model training. The machine learning models tested include four artificial neural network (ANN) models—Levenberg-Marquardt backpropagation, Bayesian regularization backpropagation, Scaled conjugate gradient backpropagation, and Quasi-Newton backpropagation—along with Extreme Gradient Boosting (XGBoost), Adaptive Boosting (Adaboost), Linear regression, and Support vector regression (SVR). The ANN models were simulated using machine learning tools. Model accuracy was assessed using four quality measures: Root mean squared error (RMSE), Mean absolute error (MAE), Mean absolute percentage error (MAPE), and Quality measure parameter (R²). The results indicate that the models struggled with heating demand predictions but were more accurate in predicting electricity consumption. A significant challenge was the tendency of some models to predict values. Hyper-parameter tuning revealed that prediction accuracy varied significantly. The study concluded that the Quasi-Newton backpropagation (QNBP) model yielded the most accurate predictions for electricity consumption, while the SVR model performed best for heating demand. Compared to previous studies, the models in this study showed suboptimal performance. Weather data, weekday, and working hour indices did not sufficiently explain heating demand and electricity consumption. Furthermore, there was no clear indication that the chosen hyper-parameters consistently improved prediction accuracy across the datasets. However, hyper-parameter adjustments did impact prediction performance.