Machine learning for predicting energy efficiency of buildings: a small data approach

This paper provides a method for predicting the energy efficiency of buildings using artificial intelligence tools. The scopes is twofold: prediction of the levels of the heating load and cooling load of buildings. A feature of this research is the performance of intellectual analysis in conditions of a limited amount of data when solving the stated tasks. An improved method of augmentation and prediction (input-doubling method) is proposed by processing data within each cluster of the studied dataset. The selection of the latter occurs due to the use of the fast and easy-to-implement k-means method. Next, a prediction is made using the input-doubling method within each separate cluster. The simulation of the method was performed on a real-world dataset of 768 observations. The proposed approach was found to have a high prediction accuracy in the absence of overfitting and high generalization properties of the improved method. Comparison with existing methods showed an increase in accuracy by 40-46% (MSE) compared to SVR with rbf kernel, which is the basis for the improved method, and by 5-12% (MSE) compared to the closest existing hierarchical predictor.