Hybrid modeling in the predictive analytics of energy systems and prices. (15th June 2020)

• Record Type: Journal Article
• Authors: Gulay, Emrah; Duru, Okan
• Is Part Of: Applied energy. Volume 268(2020)
• Publication Date: 2020-06-15
• Subjects: Energy markets -- Energy prices -- Price discovery -- Price forecasting -- Combination forecasting -- Residual modeling; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2020.114985
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Residuals of a forecasting exercise still consist of predictive component. For capturing patterns in residuals, a two-tier approach is required. By using EMD, residuals may be reused and reflected to predictions. Recycling residuals significantly improve predictive accuracy. Abstract: The aim of this paper is to illustrate the nature of the residuals of a forecasting process and to propose a hybrid approach with linear and nonlinear components predicted by corresponding methodologies. It is a common practice that residuals are assumed to be unpredictable or are reiterated into a model as lagged variables to capture any information remaining in the residual data. The central argument of this paper is that residuals from energy price forecasting can still carry predictive information in its complex and nonlinear form. Although the linear modeling is initially very accurate, reiterating residuals in linear structures is a mismatch of data type and methodology. In this regard, the proposed algorithm hybridizes or combines linear components captured by the Autoregressive Distributed Lag Model (ARDL) and nonlinear components processed by the Empirical Mode Decomposition (EMD) and an Artificial Neural Network (ANN) to improve post-sample accuracy. The conventional reiterative process can improve in-sample accuracy, which literally has no value for business forecasting practices. Through a fair benchmark comparison, including methodologies of other combinations, the proposed algorithm is cross-validated by predictive accuracy gain in the out-of-sample (holdout) dataset.