Predicting sectoral electricity consumption based on complex network analysis

High-frequency and unit-level consumption data collected by smart meters makes accurate and short-term predictions of sectoral electricity demand possible. To facilitate electricity market pricing, load management and demand response, models handling such high-dimensional data-sets are expected to realize effective variable selection, accurate prediction and, in the meantime, retain the economic mechanisms as much as possible. This paper attempts to propose a complex network based on a variable selection model that retains the causality relationships among the most relevant sectors and can achieve prediction accuracy that is comparable to other data-driven models. A dataset containing 266,000 industrial and commercial firms in Shanghai is employed to develop a complex network relying on Granger causality and correlation coefficients. Dominant nodes are selected based on a Planar Maximally Filtered Graph algorithm and then serve as explanatory variables in the linear regression model. Further comparison with LASSO, PCA and Ridge regression shows that this model can successfully realize dimension reduction but maintain significant economic mechanisms, and achieving unbiased estimation and acceptable accuracy.