Daily pattern prediction based classification modeling approach for day-ahead electricity price forecasting. (February 2019)

• Record Type: Journal Article
• Authors: Wang, Fei; Li, Kangping; Zhou, Lidong; Ren, Hui; Contreras, Javier; Shafie-khah, Miadreza; Catalão, João P.S.
• Is Part Of: International journal of electrical power & energy systems. Volume 105(2019)
• Page Start: 529
• Page End: 540
• Publication Date: 2019-02
• Subjects: Electricity price -- Day-ahead forecasting -- Classification modeling -- Daily pattern prediction -- Weighted voting mechanism; Electrical engineering -- Periodicals; Electric power systems -- Periodicals; Électrotechnique -- Périodiques; Réseaux électriques (Énergie) -- Périodiques; Electric power systems; Electrical engineering; Periodicals; 621.3
• DOI: 10.1016/j.ijepes.2018.08.039
• Languages: English
• ISSNs: 0142-0615

Abstract

Highlights: Daily pattern prediction-based modeling approach for electricity price forecasting. Weighted voting mechanism (WVM) to better generate the final day-ahead DPP. The proposed approach is validated in the case study using the data from PJM. Abstract: Day-ahead electricity price forecasting (DAEPF) plays a very important role in the decision-making optimization of electricity market participants, the dispatch control of independent system operators (ISOs) and the strategy formulation of energy trading. Unified modeling that only fits a single mapping relation between the historical data and future data usually produces larger errors because the different fluctuation patterns in electricity price data show different mapping relations. A daily pattern prediction (DPP) based classification modeling approach for DAEPF is proposed to solve this problem. The basic idea is that first recognize the price pattern of the next day from the "rough" day-ahead forecasting results provided by conventional forecasting methods and then perform classification modeling to further improve the forecasting accuracy through building a specific forecasting model for each pattern. The proposed approach consists of four steps. First, K-means is utilized to group all the historical daily electricity price curves into several clusters in order to assign each daily curve a pattern label for the training of the following daily pattern recognition (DPR) model and classification modeling. Second, a DPP model is proposed to recognize the price pattern of the next day from the forecasting results provided by multiple conventional forecasting methods. A weighted voting mechanism (WVM) method is proposed in this step to combine multiple day-ahead pattern predictions to obtain a more accurate DPP result. Third, the classification forecasting model of each different daily pattern can be established according to the clustering results in step 1. Fourth, the credibility of DPP result is checked to eventually determine whether the proposed classification DAEPF modeling approach can be adopted or not. A case study using the real electricity price data from the PJM market indicates that the proposed approach presents a better performance than unified modeling for a certain daily pattern whose DPP results show high reliability and accuracy.