Deep learning architectures for autonomous feature selection in residential energy modelling

Increasing volumes of building operational data provide substantial opportunities for data-driven energy performance assessment. While supervised learning models are prevalent, their efficacy is frequently constrained by manual feature engineering. Extracting salient features from high-dimensional, noisy datasets remains a critical bottleneck that hinders model scalability, generalizability, and interpretability. This study addresses this gap by systematically evaluating three deep learning architectures: TabNet, a Deep Feature Selection (DFS) neural network, and a conventional Artificial Neural Network (ANN) with permutation importance, for automated feature selection from residential building data. All models demonstrated robust predictive performance and the capacity to autonomously rank influential features. Specifically, the TabNet model achieved a validation accuracy of 94.36%, while the DFS and ANN architectures reached overall accuracies of 93% and 94%, respectively. Furthermore, all three methodologies converged on carbon emissions (total and per floor area) and current energy consumption as the primary determinants of energy performance. This work contributes a scalable framework for developing transparent and robust models, facilitating building energy optimization at scale.