Explainable artificial intelligence for building energy performance certificate labelling classification

Abstract

The building energy performance certificates (EPC) are widely adopted for sustainable development and improvement in building energy efficiency. Different from the conventional direct measurement based approach of acquiring a building’s EPC label, this study proposes a novel and alternative approach to classify a building’s EPC label using artificial neural network (ANN) models. Given the extensive best building EPC practices in developed countries, historical building EPC data and experiences can expedite the development and improvement of this procedure in developing countries. This study first develops the ANN classification model to attain the building EPC label. The classification result shows that the building EPC classification can achieve a 99% precision with sufficient input data. With the assistance of explainable artificial intelligence (XAI) tools such as the Local Interpretable Model-Agnostic Explanation (LIME) and SHapley Additive exPlanation (SHAP), some less important input features for the ANN classification models can be removed without severely influencing the ANN model’s accuracy. In the case studies, the EPC best practices historical registry data from Lombardy, Italy are used in training the ANN model. The ANN models’ accuracy for the case study 1 is 93% with 14 input features where CO2 emissions and net surface area are the two most influential features. The most influential input feature for case study 2 is the winter AC non-renewable energy performance, and the accuracy of the case study 2 ANN model is 89% with 26 input features.