Deep Learning Based Cross-Location Building Load Prediction Using Domain Adaptation

Buildings consume a significant portion of a nation’s total energy, accounting for about 70 percent of electricity use in the US. Predicting building energy demand accurately is crucial for utilities to ensure efficient operation. Deep learning-based approaches have demonstrated great potential for this task. However, building loads are heavily influenced by local weather conditions. Conventional deep learning models are typically location-specific and demand extensive local data for accurate predictions. A location independent model can offer great advantage, which can be attained through application of domain adaptation. This method involves training a deep learning model initially on a source domain dataset and then use this pre-trained model to perform prediction on target domain dataset, where the probability distributions may differ but are comparable. This study applies domain adaptation techniques to develop a deep learning model that predicts building energy demand independent of location. Our proposed model uses a pre-trained deep learning architecture trained on data from one location to predict building load in another location, requiring only minimal data from the target domain. The findings demonstrate that the domain-adapted model achieves improvements in both accuracy and computational speed compared to non-adapted models.