Factor decomposition and prediction of solar energy consumption in the United States

Given advances in methodology of deep neural networks, this study used the LMDI (logarithmic mean Divisia index) to decompose 1983-2017 United States solar energy consumption data and identified four driving factors. These factors were considered as a group to provide a single variable input, and as four individually decomposed effects used to combine with LSTM (long short-term memory) to predict changes in solar energy consumption. Compared with the autoregressive integrated moving average (ARIMA) method, the results show that the proposed approach combined with LSTM has better feasibility. First, the structural effect accounts for the largest proportion of the total contribution in consumption, reflecting the significance of the growth of solar energy. Second, multi-variable LSTM for a non-stationary time series is better than single-variable LSTM. Finally, the prediction accuracy of LSTM is better than that of classical time series ARIMA, for both training and test data. These findings provide insights into future demand for solar energy in the United States. (C) 2019 Elsevier Ltd. All rights reserved.