A novel network training approach for solving sample imbalance problem in wind power prediction. (1st May 2023)

• Record Type: Journal Article
• Authors: Meng, Anbo; Xian, Zikang; Yin, Hao; Luo, Jianqiang; Wang, Xiaolin; Zhang, Haitao; Rong, Jiayu; Li, Chen; Wu, Zhenbo; Xie, Zhifeng; Zhang, Zhan; Wang, Chenen; Chen, Yingjun
• Is Part Of: Energy conversion and management. Volume 283(2023)
• Publication Date: 2023-05-01
• Subjects: Segment imbalance regression -- Sample imbalance -- Decomposition technique -- Bi-direction gate recurrent unit -- Self-attention mechanism; Direct energy conversion -- Periodicals; Energy storage -- Periodicals; Energy transfer -- Periodicals; Énergie -- Conversion directe -- Périodiques; Direct energy conversion; Periodicals; 621.3105
• DOI: 10.1016/j.enconman.2023.116935
• Languages: English
• ISSNs: 0196-8904

Abstract

Highlights: Segment imbalance regression (SIR) is proposed to address sample imbalance. SIR is improved by optimizing hyperparameters with CSO. Wind series is decomposed by EEMD to improve stability. TSA-BiGRU model is proposed to extract implicit temporal relation information. The proposed hybrid model outperforms other state-of-the-art methods. Abstract: Randomness and intermittency are common challenges in wind power prediction. Most studies focus on randomness but usually ignore the intermittency of wind power that leads to sample imbalance and impairs prediction accuracy. To address the sample imbalance problem, a segment imbalance regression (SIR) method with crisscross optimization (CSO) is proposed to proactively dig and utilize the imbalance nature of samples. By investigating the interactions among adjacent samples, SIR is employed to adaptively assign different learning weights to each sample. SIR focuses on the training samples within the segment while retaining useful information outside, and thus can facilitate the gradient descend process and achieve better training performance. On this basis, aiming to extract more feature information, a novel combination network is constructed by integrating two networks, i.e., Temporal self-attention network (TSA) with temporal self-correction and bidirectional gate recurrent unit (BiGRU) with bi-direction temporal memory respectively. The sequence features of samples are first obtained by ensemble empirical mode decomposition (EEMD) as the input of the combination network, and then the SIR-CSO method is used to train the TSA-BiGRU network and enhance the adaptive learning ability. Massive experiments are conducted, and the results demonstrate the excellent performance of SIR and the superiority of the combination network. Especially in three-step prediction, the root mean square errors of the proposed model reduce 38.97%∼47.04% compared with other state-of-the-art methods.