Predicting Time Series Energy Consumption Based on Transformer and LSTM

Haitao Wang, Jiandun Li, Liu Chang

Abstract: Energy is crucial to economic and social development. Accurate energy consumption forecasting is essential to effective energy management, reasonable energy layout planning, and ensuring the sustainable and healthy development of the energy industry. Nevertheless, precisely and efficiently forecasting energy consumption remains to be a challenge. Previous studies have proposed solutions mainly from traditional machine learning and mathematical statistics, which can effectively forecast short-term energy consumption in small-scale data. However, it is still challenging to explore the characteristics of high-dimensional and large-scale energy data and predict medium- to long-term energy consumption and its fluctuation trends. In this paper, a new time series energy consumption prediction model is proposed, which combines the attention mechanism of the Transformer model with the natural language processing ability of LSTM, based on a dataset of Spain’s energy production and climate change from 2015 to 2018. Compared with the state-of-the-art models such as RNN, GRU, and LSTM, our model achieves better performance in the 7-day energy consumption prediction task (RMSE = 0.7, MAE = 0.5).