Building Energy Performance Certificate Labelling Classification Based on Explainable Artificial Intelligence

Abstract

In this study, the machine learning based classification approach is aptly applied to classify the building energy performance certificate (EPC) rating levels instead of applying the traditional approach of direct measurements of the annual building energy consumption and its effective floor areas. Accurately acquired EPC ratings deliver a successful building EPC programme which enables future regulatory actions for the buildings nationwide. Historical data and experiences from the countries who have implemented the building EPC programme can help to accelerate the development of such a programme in the developing countries like South Africa. With these concerns, an artificial neural network (ANN) classification model together with the explainable artificial intelligence (XAI) tools are adopted in this study to obtain the building EPC rating levels. In this study, an ANN model is trained from the historical registry of the building EPC best practices in Lombardy, Italy. The ANN classification model is calibrated and validated with optimal number of neurons. Results show that probability of detection for the ‘G’ labelled buildings is 0.9655 with a precision of 0.8547. In addition, the Local Interpretable Model-Agnostic Explanation (LIME) is adopted to explain the inherent properties of the identified ANN classification model. The LIME XAI shows that opaque surface and average U values for walls are the most influential features on the EPC rating classification while the degree day feature has the least influence during the ANN based EPC rating classification process.