Attention Mechanism-Based Transfer Learning Model for Day-Ahead Energy Demand Forecasting of Shopping Mall Buildings

Abstract

The forecasting performance of a data-driven model decreases rapidly with limited training data set. Herein, we solve this problem by developing an attention mechanism-based transfer learning model and compare its predicting ability in day-ahead energy consumption to those of three direct learning models: (1) Artificial neural networks with Auto-regressive (AR-ANN), (2) Random Forest with Auto-regressive (AR-RF), and (3) long short-term memory neural network (LSTM). Our target building is a large-scale shopping mall in Harbin with two-year monitored data. The 2-months to 1-year data selected from the first year and all data from the second year are used as the training and testing set, respectively. Those models are applied to the target building's peak electricity demand (PED) and total energy consumption (TEC). Results show that the proposed transfer learning model outperforms these 3 direct learning models when insufficient data is available in the training set. The lowest prediction errors of AR-ANN, AR-RF and LSTM are 34.34% and 25.28% for the next day’s PED and TEC, respectively, with 2-month training data available. In comparison, the corresponding prediction errors of the proposed model are only 11.48% and 9.43%, respectively. The case study demonstrates the great performance of the proposed model with limited data.

Keywords: transfer learning, Attention Mechanisms, shopping mall building, Energy Prediction