Xinjiang Oil & Gas ›› 2025, Vol. 21 ›› Issue (3): 31-40.DOI: 10.12388/j.issn.1673-2677.2025.03.004

• NEW ENERGY • Previous Articles     Next Articles

Method of AI-Assisted Photovoltaic Power Forecasting for Peak Regulation

YANG Xiaoya1CHEN Xiangyu1,2HAN Leitao1,3WANG Keqin1,4QIU Xiaolong1,5ZHU Peiwang1XIAO Gang1   

  1. 1.Zhejiang University,Hangzhou 310027,Zhejiang,China; 2.PetroChina Xinjiang Oilfield Company,Karamay 834000,Xinjiang,China; 3.Haining Product Quality Inspection Institute (Zhejiang Solar Product Quality Inspection Center),Haining 314400,Zhejiang, China; 4.China Huaneng Group Clean Energy Technology Research Institute Co.,Ltd.,Changping 102209,Beijing,China;5.Qingshan  Lake Energy Research Base,Zhejiang University,Hangzhou 311305,Zhejiang,China
  • Received:2025-05-28 Revised:2025-08-04 Accepted:2025-08-10 Online:2025-09-05 Published:2025-09-05
  • Contact: XIAO Gang

面向调峰需求的AI辅助光伏发电预测方法

杨小丫1陈香玉1,2韩雷涛1,3王柯钦1,4仇晓龙1,5祝培旺1肖刚1
  

  1. 1.浙江大学,浙江杭州 310027; 2.中国石油新疆油田分公司,新疆克拉玛依 834000; 3.海宁市产品质量检验检测所(浙江省太阳能产品质量检验中心),浙江海宁 314400; 4.中国华能集团清洁能源技术研究院有限公司,北京昌平 102209;5.浙江大学青山湖能源研究基地,浙江杭州 311305
  • 通讯作者: 肖刚
  • 作者简介:杨小丫(1999-),浙江大学能源动力专业在读博士,目前从事多能互补能源系统中的数据驱动建模与人工智能方法 研究。(E-mail)12427157@zju.edu.cn
  • 基金资助:

    浙江省市场监督管理局科技计划项目“光伏电热耦合热水系统综合性能测试研究”(ZD2025020);嘉兴市科技计划项目“电热联产联储一体化光储系统技术研究”(2025AC034)。

Abstract:

With the increasing penetration of photovoltaic (PV) generation into power systems,the randomness and uncertainty of its output have raised higher requirements for the flexible peak regulation capability of grids. To offer more accurate predicted scenarios and facilitate flexibility oriented dispatch,an intelligent method integrating fuzzy clustering,similar day extraction,and probabilistic prediction was developed. Highly correlated meteorological variables including temperature,humidity,global horizontal irradiance,and tilted irradiance were first identified using the Pearson correlation coefficient. Fuzzy C-means (FCM) clustering was then applied to classify weather types. Feature weights were determined using the CRITIC method,and similar days within each weather category were extracted based on weighted Euclidean distance to construct a high quality training dataset. A quantile regression long short-term memory (QRLSTM) network was subsequently employed to perform short-term probabilistic forecasting of PV output. Simulation results demonstrated that the proposed approach achieved high prediction accuracy across various weather conditions,with confidence interval coverage rates exceeding 90% and significantly reduced confidence interval ranges compared to those of benchmark models. It was concluded that the proposed method effectively enhances the reliability and robustness of PV power prediction and provides high quality scenario support for uncertainty aware dispatch in multi-energy complementary systems.

摘要:

随着光伏发电在电力系统中渗透率的持续提升,其出力的随机性与不确定性对电网的灵活调峰能力提出了更高要求。为灵活调峰提供更加精准可靠的预测场景与支撑,提出一种融合模糊聚类、相似日提取与概率预测的智能预测方法。首先,基于皮尔逊相关系数筛选与光伏出力密切相关的气象特征变量,包括温度、湿度、总水平辐照度和倾斜辐照度;随后,采用模糊C均值(FCM)算法对天气类型进行聚类并基于CRITIC方法确定气象变量权重,通过加权欧氏距离在同类天气中提取相似日,构建高质量训练样本;最终,利用结合分位数回归的长短期记忆神经网络(QRLSTM)实现光伏功率的短期概率预测。仿真实验结果表明,该方法在典型晴天、阴天、间歇性多云等多种天气条件下均实现了高精度的预测性能,预测区间置信覆盖率超过90%,平均区间宽度显著低于传统方法,展现出良好的稳定性和泛化能力。研究结论表明,该方法可有效提升光伏功率预测的可信度与鲁棒性,可为多能互补系统中的调度优化提供高质量的不确定性场景支撑。

CLC Number: