Calibrated BEMs and LSTM Neural Networks for Indoor Temperature Prediction: A Comparative Analysis in Pre- and Post-Retrofit Scenarios

The need to mitigate the risks of overheating in buildings due to climate change has highlighted the importance of accurate models for predicting indoor temperatures and thermal comfort, particularly after retrofitting. To this end, white-box models, such as Building Energy Models (BEMs), and black-box models, such as Long Short-Term Memory (LSTM) neural networks, have been extensively used in recent decades. While BEMs provide detailed insights through physically-based simulations, requiring calibration for enhanced accuracy, LSTMs provide a datadriven approach that captures complex thermal dynamics with greater simplicity, albeit with less interpretability. Few studies have undertaken a comparative analysis of these models in terms of prediction accuracy, especially across pre- and post-retrofit conditions and different lengths of training periods. Thus, in this study, a comparison between the predicting capabilities of calibrated BEMs and LSTM in summer was carried out using two real monitored mock-ups in Northern Italy representing both pre- and post-retrofit conditions. The results show that, for the considered limited training periods (8 and 3 days), the dataset size does not significantly influence BEM accuracy, while LSTM accuracy is more affected. Moreover, BEMs show higher prediction accuracy in scenarios with higher indoor air temperature (IAT) variability, i.e. where unseen data could be less predictable, such as in pre-retrofit conditions. LSTMs, however, excel in low-variability scenarios, such as the post-retrofit conditions in this case. This study highlights the critical need for careful model selection and calibration based on the data availability and building typology to ensure prediction reliability.