Building thermal dynamics modeling with deep learning exploiting large residential smart thermostat dataset

In this paper, we present a deep learning approach to model building thermal dynamics with large-scale smart thermostat data collected from residential buildings. We developed a Long Short-Term Memory (LSTM) model as a baseline and compared it to a CNN-LSTM model to predict indoor air temperature in a multi-step time horizon in 164 buildings. The study showed that the proposed CNN-LSTM achieved an average of 0.26 °C Mean Absolute Error (MAE) for one-hour-ahead (12 future steps) predictions, which is over 6% of improvement comparing with the baseline. Furthermore, the results indicated that the CNN-LSTM models achieved more robust performance across different building characteristics, system configurations and locations, with a standard deviation reduction of 22%, proving the effectiveness and generalizability of the proposed approach.