Predicting energy consumption: A multiple decomposition-ensemble approach.

Reliable energy consumption prediction plays a vital role in formulating government's energy policies. In this study, a novel multiple decomposition-ensemble method based on error compensation is proposed to stably predict energy consumption. Firstly, trend decomposition is used to decompose the energy consumption into trend-subseries and error-subseries. In view of error compensation is a feasible approach for improving the prediction accuracy, the error-subseries that mentioned above are further divided into one low-frequency approximation error-subseries and several high-frequency detailed error-subseries by wavelet transform. Depending on their different dynamic changing characteristics, this study uses a Linear Regression Model to predict the trend-subseries, employs a Triple Exponential Smoothing Model to estimate the low-frequency approximation error-subseries, and uses an Auto Regression model to find the high-frequency detailed error-subseries. Finally, the overall energy consumption is the summation of these subseries predictions. Using the energy consumption data from China in 2007–2016, empirical study is carried out. The proposed multiple decomposition-ensemble method based on error compensation achieves the highest performance, which is compared with other six models (three single models, two traditional decomposition-ensemble models, and combination model). The proposed model is also validated by the U.S. data. Forecasts indicate that the energy demand of China will increase to 4.957343 billion Tons of standard Coal Equivalent in 2021, implying that China should speed up its transition to an energy-efficiency economy. • A novel prediction model based on multiple decomposition-ensemble method is built. • Error compensation is a feasible approach for improving the prediction accuracy. • Multiple decomposition-ensemble method improves the accuracy of error compensation. • Energy consumptions of China and U.S. are forecasted to demonstrate these methods. • The proposed model achieves the highest performance compared with other six models. [ABSTRACT FROM AUTHOR]