Accurate solar PV power prediction interval method based on frequency-domain decomposition and LSTM model. (1st January 2023)

• Record Type: Journal Article
• Authors: Wang, Lining; Mao, Mingxuan; Xie, Jili; Liao, Zheng; Zhang, Hao; Li, Huanxin
• Is Part Of: Energy. Volume 262:Part B(2023)
• Publication Date: 2023-01-01
• Subjects: Solar PV power Forecasting -- Frequency-domain decomposition -- Improved long-short-term-memory network -- Support vector regression -- Deep learning; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2022.125592
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: The stability operation and real-time control of the integrated energy system with distributed energy resources determines the higher and higher requirements for the accuracy of solar photovoltaic (PV) output power prediction. This paper proposes an accurate PV power prediction interval approach based on frequency-domain decomposition and hybrid deep learning (DL) model. In the proposed approach, ensemble empirical mode decomposition (EEMD) is firstly used to decompose and reconstruct the original PV energy time-series data into high and low-frequency sub-series followed by the statistical feature extraction process. Furthermore, an improved long-short-term-memory network (LSTM) model with the designed hyperparameters based on Bayesian optimization (BO) is proposed to predict the sub-series with the different minute-hour-day intervals. Moreover, support vector regression (SVR) is utilized to analyze the initial time node and reduce the fluctuation error of the prediction value near zero. Finally, a comparative study with SVR, KNN, BPNN, GRU, Stacked-LSTM, LSTM, LSTM-SVR, and LSTM-SVR-BO models is constructed by using an actual dataset collected from Arizona, US. The simulation results on the datasets show the proposed prediction model outperforms the other 7 models for PV power forecasting in 1 day, 7 days, and 14 days ahead prediction with the different minute-hour-day intervals. Especially, in the seven days ahead prediction case, the proposed model's average RMSE and AbsDEV values are as low as 4.157 and 0.116, where the prediction accuracy and prediction stability are improved by about 15% on average compared to the other prediction models. Highlights: An PV power prediction approach based on hybrid DL model is presented. EEMD is used to decompose and reconstruct the original PV energy time-series data. The hyperparameters of LSTM model are designed based on BO algorithm. SVR is utilized to reduce the fluctuation error of the prediction value near zero. A comparative test with 7 DL models is conducted in multiple time dimensions.