Transfer learning with seasonal and trend adjustment for cross-building energy forecasting

Large scale smart meter deployments have resulted in popularization of sensor-based electricity forecasting which relies on historical sensor data to infer future energy consumption. Although those approaches have been very successful, they require significant quantities of historical data, often over extended periods of time, to train machine learning models and achieve accurate predictions. New buildings and buildings with newly installed meters have small historical datasets that are insufficient to create accurate predictions. Transfer learning methods have been proposed as a way to use cross-domain datasets to improve predictions. However, these methods do not consider the effects of seasonality within domains. Consequently, this paper proposes Hephaestus, a novel transfer learning method for cross-building energy forecasting based on time series multi-feature regression with seasonal and trend adjustment. This method enables energy prediction with merged data from similar buildings with different distributions and different seasonal profiles. Thus, it improves energy prediction accuracy for a new building with limited data by using datasets from other similar buildings. Hephaestus works in the pre- and post- processing phases and therefore can be used with any standard machine learning algorithm. The case study presented here demonstrates that the proposed approach can improve energy prediction for a school by 11.2% by using additional data from other schools.