Novel hybrid machine learning-based prediction of building space heating load: a comprehensive study.

Space heating load is the main energy consumption agent for residential buildings. Its prediction is crucial for the efficient and cost-effective management of energy suppliers. In this study, nine machine learning methods, including a novel hybrid natural gradient boosting (NGBoost) and residual-based mixture density network (MDN) model, are evaluated to predict the space heating load of a building using a dataset composed of seven independent input parameters. The evaluation of the methods is carried out based on the performance criteria of mean absolute error, mean squared error, root mean squared error, mean absolute percentage error, and the coefficient of determination. The results show that the smallest mean absolute percentage error, with a value of 0.05, belongs to the novel hybrid NGBoost + MDN model, which is more effective and competitive compared to histogram-based gradient boosting, elastic net, light gradient boosting machine, and extreme learning machine approaches. The best model in prediction accuracy is NGBoost + MDN because it has the lowest MAPE. All the methods achieve values above 0.99 for the coefficient of determination, however the decision tree has the highest score of 0.9989. Moreover, extreme gradient boosting and elastic net techniques perform smaller squared errors with 0.063 and 0.065. In addition, it has been determined from the uncertainty analysis that the newly developed hybrid model demonstrates better prediction stability, as evidenced by its consistently maintaining prediction residuals within a 'Low Error Region' area of the residual plot. The Residual Prediction Deviation (RPD) index measurement indicates the model's very high degree of robustness and reliability for managing stochastic variations in heating load at an RPD score of 25.845. These results are relevant to the particular case studied and should be analyzed considering the choice of the building, as well as the settings used for validation purposes.