An adaptive hybrid ensemble with pattern similarity analysis and error correction for short-term load forecasting

Abstract

Forecasting future electricity consumption is one of the critical processes for addressing energy management and supply–demand balance in modern electrical systems. In this paper, an adaptive hybrid ensemble, CMKP-EG-SVR, is presented to solve the short-term load forecasting problem. For this purpose, the merits of the median filtering are exploited at an initial phase of data preprocessing to reduce the high-frequency parts and to detect and correct outliers and missing data. Besides, Pearson’s correlation, k-nearest neighbors algorithm and a particular calendar grouping based on day-type encoding are used as efficient tools for selecting relevant features and extracting similar patterns. To achieve diversification gains and improve the overall performance, several prediction engines based on support vector regression are proposed, in an adaptive weighting procedure, by our ensemble forecasting scheme. Finally, an error correction strategy based on a modified Gaussian function is also included for a further enhancement in prediction accuracy. The main idea behind this integration is to meet the multiple requirements and characteristics of accurate prediction using a unique and adaptive forecasting scheme. The electricity demand data from the Australian state of New South Wales are employed to describe an extensive evaluation under both normal and anomalous load conditions. The obtained results of the proposed CMKP-EG-SVR model demonstrate: (1) an average computation time of 4.47 s; (2) mean absolute percentage error values of the range of 1.03%–4.26%; (3) a superior performance over 24 state-of-the-art forecasting techniques.