A Comparative Analysis of Multi-Target Feature Selection Methods in Data-Driven Models for Building Energy and Thermal Performance Prediction

Abstract: Building energy management increasingly utilises Machine Learning (ML) to use data from sensor-rich environments. A significant challenge in this context is managing high-dimensional data, which can affect model performance. This study addresses this by applying multi-target feature selection, an underexplored method that reduces dimensionality by analysing inter-feature relationships. From 182 features, two were key for developing three ML models predicting the energy and thermal performance of the HiLo living lab. These models achieved a robust fit with an average Root Mean Squared Error (RMSE) of 0.18 kW and 1.03 °C, demonstrating multi-target feature selection’s effectiveness in enhancing building performance predictions.