A multivariable hybrid prediction model of offshore wind power based on multi-stage optimization and reconstruction prediction. (1st January 2023)

• Record Type: Journal Article
• Authors: Wang, Hao; Ye, Jingzhen; Huang, Linxuan; Wang, Qiang; Zhang, Haohua
• Is Part Of: Energy. Volume 262:Part A(2023)
• Publication Date: 2023-01-01
• Subjects: Offshore wind power -- Decomposition and reconstruction model -- Optimization algorithm -- Interval prediction; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2022.125428
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Offshore wind power prediction is the basis for safe operation and grid dispatch. However, it is difficult due to the high volatility. Aiming at the three shortcomings of current methods: lack of analysis of the impact of multiple variables; the reconstruction method of decomposition components often adopts the summation method; the traditional machine learning prediction methods are not accurate enough, while the deep learning methods are prone to overfitting. This paper proposes a multi-variable hybrid prediction model based on multi-stage optimization and reconstruction prediction. Firstly, the isolated forest is used for data preprocessing. Secondly, the power sequence is decomposed by the variational modal decomposition optimized by the gray wolf algorithm to reduce the non-stationarity. Thirdly, the kernel extreme learning machine optimized by sparrow algorithm is used to predict. Finally, the reconstruction prediction is carried out through the long short-term memory network. Compared with the traditional machine learning method and the deep learning method, the model is effectively improved on two European offshore datasets. Then the interval prediction based on this model further verifies the accuracy and reliability. Highlights: Analyze power factors and use isolated forests to improve dataset quality. Propose a new multi-stage decomposition and reconstruction method. The model absorbs the advantages of traditional machine learning and deep learning. The interval prediction based on point prediction are more valuable for reference.