基于AGCN-LSTM模型的海上风电场功率概率预测  

Power Probability Prediction for Offshore Wind Farm Based on AGCN-LSTM Model

作　　者：苏向敬 程子凡 聂良钊 符杨 SU Xiangjing;CHENG Zifan;NIE Liangzhao;FU Yang(Offshore Wind Power Research Institute,Shanghai University of Electric Power,Shanghai 200090,China;School of Electrical Engineering,Shanghai University of Electric Power,Shanghai 200090,China;China Resources Power(Hubei)Sales Co.,Ltd.,Wuhan 430000,China)

机构地区：[1]上海电力大学海上风电研究院,上海市200090 [2]上海电力大学电气工程学院,上海市200090 [3]华润电力(湖北)销售有限公司,湖北省武汉市430000

出　　处：《电力系统自动化》2024年第22期140-149,共10页Automation of Electric Power Systems

基　　金：国家重点研发计划资助项目(海上风电并网系统远程监测与故障诊断技术,2023YFB2406900)。

摘　　要：海上风电场各风电机组间存在较强的时空关联,现有风电功率预测研究虽对时空关联进行了探索,但未能更好地表达动态的空间关联;同时,空间关联的“黑盒”也降低了预测的可信度;另外,概率预测相比确定性预测可提供更全面的信息。为此,提出一种基于自适应图卷积网络(AGCN)-长短期记忆(LSTM)网络的海上风电场功率超短期概率预测模型。首先,通过AGCN中的自适应矩阵获取季节性变化海上风电机组间的动态空间关联,并实现空间维度的可解释,同时利用LSTM网络挖掘数据时间维度的相关性;其次,基于所获得的时空关联,通过改进的分位数回归模型构建风电场功率概率预测区间;最后,基于中国东海大桥海上风电场真实数据进行仿真验证。结果表明,相较于已有概率预测模型,动态图拓扑可以更好地揭示海上风电场变化的空间关系;改进后的分位数回归模型有效避免了分位数预测存在的曲线交叉问题;二者共同提升了模型整体预测精度和稳定性。There are strong spatio-temporal correlations among the wind turbines in offshore wind farms.Although previous research on wind power prediction has explored the spatio-temporal correlation,its dynamic spatial correlation has not been better expressed.Meanwhile,the“black box”of spatial correlation significantly reduces the prediction credibility.In addition,the probability prediction provides more comprehensive information than deterministic prediction.Therefore,based on the adaptive graph convolutional network(AGCN)and long short-term memory(LSTM)network,an ultra-short-term probability prediction model of offshore wind farm power is proposed.Firstly,dynamic spatial correlations between seasonal offshore wind turbines are obtained through the adaptive matrix in AGCN,and interpretability in spatial dimensions is achieved.Meanwhile,the LSTM network is used to mine the temporal correlation of data.Secondly,based on the obtained spatio-temporal correlations,an improved quantile regression model is used to construct a probability prediction interval for wind farm power.Finally,based on the real data from the offshore wind farm of Donghai Bridge,China,simulation and verification are conducted.The results indicate that compared with existing probability prediction models,dynamic graph topology can better reveal the spatial relationship of changes in offshore wind farms.The improved quantile regression model effectively avoids the problem of curve crossing in quantile prediction.They both improve the overall prediction accuracy and stability of the model.

关 键 词：海上风电 时空相关性 自适应图卷积 长短期记忆网络 空间可解释 概率预测 

分 类 号：TM614[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]