Attention mechanism-based transfer learning model for day-ahead energy demand forecasting of shopping mall buildings

The forecasting performance of data-driven models decreases rapidly with a limited training dataset. Herein, we sought to solve this problem by developing an attention mechanism-based transfer learning model and comparing its predictive ability in day-ahead energy consumption with those of three direct learning models: artificial neural networks with auto-regression (AR-ANN), random forest with auto-regression (AR-RF), and long short-term memory neural network (LSTM). Our target building was a large-scale shopping mall in Harbin, with 2 years of monitored data. The 2-months to 1-year data selected from the first year and all data from the second year were used as the training and testing sets, respectively. These models predicted the target building's peak electricity demand (PED) and total energy consumption (TEC). The results showed that the proposed transfer learning model outperformed the three direct learning models when data were insufficient in the training set. Specifically, the direct prediction models' lowest PED and TEC prediction errors were 34.34% and 26.32%, respectively, with 2-month training data available. In comparison, the corresponding prediction errors of the proposed model were only 12.48% and 10.78%, respectively. This study demonstrated the excellent performance of the proposed model with limited data.