发电技术 ›› 2025, Vol. 46 ›› Issue (4): 778-787.DOI: 10.12096/j.2096-4528.pgt.24028

• 大规模新能源并网运行调控关键技术 • 上一篇    下一篇

基于白鲨算法与改进长短期记忆网络的光伏出力预测

闫朝阳1, 李蓝青1, 徐浩嘉2, 庄锁2, 张振华1, 戎子睿1   

  1. 1.国网江苏省电力有限公司,江苏省 南京市 210008
    2.国电南瑞科技股份有限公司,江苏省 南京市 211106
  • 收稿日期:2024-02-14 修回日期:2024-05-06 出版日期:2025-08-31 发布日期:2025-08-21
  • 作者简介:闫朝阳(1988),男,硕士,工程师,主要研究方向为电力系统调度运行,yancyjs@163.com
    李蓝青(1992),男,硕士,中级工程师,主要研究方向为电力系统智能调度,llq2011@sina.com
  • 基金资助:
    国家自然科学基金项目(61933005)

Photovoltaic Output Prediction Based on Improved Long Short-Term Memory Network Using White Shark Optimization Algorithm

Chaoyang YAN1, Lanqing LI1, Haojia XU2, Suo ZHUANG2, Zhenhua ZHANG1, Zirui RONG1   

  1. 1.State Grid Jiangsu Electric Power Co. , Ltd. , Nanjing 210008, Jiangsu Province, China
    2.NARI Technology Co. , Ltd. , Nanjing 211106, Jiangsu Province, China
  • Received:2024-02-14 Revised:2024-05-06 Published:2025-08-31 Online:2025-08-21
  • Supported by:
    National Natural Science Foundation of China(61933005)

摘要:

目的 为了保障光伏接入后电力系统的安全稳定与经济运行,提出了一种基于白鲨算法与改进长短期记忆(long short-term memory,LSTM)网络的光伏功率预测模型,并采用白鲨优化(white shark optimization,WSO)算法优化预测模型参数。 方法 为充分考虑影响光伏输出功率的环境变量,首先采用变分模态分解(variational mode decomposition,VMD)方法对环境因子序列进行分解,以降低序列的非平稳性。接着,利用核主成分分析(kernel principal component analysis,KPCA)方法提取主要影响因素的特征序列,获得最佳的气象特征序列。最后,使用LSTM网络预测多变量特征序列,并使用WSO算法优化LSTM网络的参数,以实现对光伏出力的精确预测。 结果 与传统的光伏功率预测方法相比,该模型显著提高了光伏发电功率的预测精度。 结论 采用WSO算法优化LSTM中的隐藏单元数目、最大训练周期、初始学习率,可以有效地提高优化的效率和精度,为高比例光伏接入后系统的稳定运行提供了参考依据。

关键词: 太阳能, 光伏发电, 光伏出力预测, 核主成分分析, 变分模态分解, 长短期记忆网络, 白鲨优化算法

Abstract:

Objectives In order to ensure the safety, stability, and economic operation of the power system after photovoltaic integration, a photovoltaic power prediction model based on improved long short-term memory (LSTM) network using white shark optimization algorithm is proposed, and uses the white shark optimization algorithm (WSO) to optimize the prediction model parameters. Methods To fully consider the environmental variables that affect photovoltaic output power, this study first uses the variational mode decomposition (VMD) method to decompose the environmental factor sequence, in order to reduce the non stationarity of the sequence. Next, the kernel principal component analysis (KPCA) method is used to extract the feature sequences of the main influencing factors and obtainthe optimal meteorological feature sequence. Finally, a LSTM network is used to predict multivariate feature sequences, and WSO is used to optimize the parameters of the LSTM to achieve accurate prediction of photovoltaic output. Results Compared with traditional photovoltaic power prediction methods, this model significantly improves the prediction accuracy of photovoltaic power generation. Conclusions Using WSO to optimize the number of hidden units, maximum training period, and initial learning rate in LSTM can effectively improve the efficiency and accuracy of optimization, providing a reference for the stable operation of high proportion photovoltaic systems after integration.

Key words: solar energy, photovoltaic power generation, photovoltaic output prediction, kernel principal component analysis, variational mode decomposition, long short-term memory network, white shark optimization algorithm

中图分类号: