Background and objectives: It has become an imperative to Improve energy efficiency in the building sector, especially in the early stages of design, in order to reduce greenhouse gas emissions and fossil fuel consumption. This requires accurate energy forecasts to optimise decision makings. In recent years, artificial intelligence and machine learning techniques have been proposed to predict energy consumption and building performance. The simplicity of building modelling and the use of simulated building energy consumption data has improved machine learning models.

Materials and methods: The present research uses a parametric design of the ASHRAE standard sample (Case 600) with the predicting features of building orientation, window-to-wall ratio (WWR), building dimensions ––including length, width, height and building use – and simulation using EnergyPlus and jEPlus software. Then, using several widely-used machine learning approaches in predicting and improving building energy performance – including regression approach with the help of Linear and Gaussian regression, classification approach with the help of Logistic regression, K-Nearest Neighbour, Decision Tree, Random Forests and Support Vector Machine, the Multilayer Neural Network approach, the clustering approach with the help of K-means, Hierarchical and Density-Based algorithms – the cooling and heating energy of the building is predicted, with the resulting data being classified.
The results are then combined with several boosting machine learning models – including AdaBoost, Gradient Boosting Machine and its derivatives – and a comparison is made about the accuracy of prediction models with different evaluation criteria, followed by an optimisation of the hyperparameter of prediction models, a 10-fold cross-validation, and finally an interpretation of the results with the help of interpretable machine learning in the best possible way (SHapley value).

Results and conclusion: The results show that the use of the Boosting Machine model improves the prediction accuracy, and that compared to single learning models, and among the single machine learning models, the Artificial Neural Network and the Decision Tree have the highest prediction accuracy for the regression and classification problems of building cooling and heating energy prediction after cross-validation. The effect of adjusting the Artificial Neural Network in increasing the accuracy of the model prediction is more than other models, and the Label Encoding method is recommended for predicting the cooling energy and the One Hot Encoding method for predicting the heating energy for data preprocessing.