1、第 47 卷 第 4 期 电 网 技 术 Vol.47 No.4 2023 年 4 月 Power System Technology Apr.2023 文章编号:1000-3673(2023)04-1653-10 中图分类号:TM 614 文献标志码:A 学科代码:47040 考虑风电时序特性的深度小波 时序卷积网络超短期风功率预测陈海鹏1,李赫1,阚天洋2,赵畅3,张忠4,于海薇1(1 现代电力系统仿真控制与绿色电能新技术教育部重点实验室(东北电力大学),吉林省 吉林市 132012;2国网冀北电力有限公司承德供电公司,河北省 承德市 067000;3国能吉林江南热电有限公司,吉林省 吉林
2、市 132002;4鲁能新能源集团有限公司甘肃分公司,甘肃省 兰州市 730000)DWT-DTCNA Ultra-short-term Wind Power Prediction Considering Wind Power Timing Characteristics CHEN Haipeng1,LI He1,KAN Tianyang2,ZHAO Chang3,ZHANG Zhong4,YU Haiwei1 (1.Key Laboratory of Modern Power System Simulation and Control&Renewable Energy Technology
3、(Northeast Electric Power University),Ministry of Education,Jilin 132012,Jilin Province,China;2.Chengde Power Supply Company,State Grid Jibei Electric Power Company Limited,Chengde 067000,Hebei Province,China;3.CHN Energy Jilin Jiangnan Thermal Power Co.,Ltd.,Jilin 132002,Jilin Province,China;4.Gans
4、u Branch of Luneng New Energy(Group)Co.,Ltd.,Lanzhou 730000,Gansu Province,China)ABSTRACT:Ultra-short-term wind power forecasting is of great significance for the formulation of power system production scheduling plans,but wind power output is largely affected by weather factors with the characteris
5、tics of strong randomness,volatility,and uncontrollability.At the same time,the impact of wind power uncertainty on the wind power time series relationship poses a challenge to the accuracy of wind power prediction.Fully extracting the wind power time series characteristics has become an important w
6、ay to solve this problem.Aiming at the this problem,an ultra-short-term wind power prediction based on the DWT-DDQN-TCN-Attention(DWT-DTCNA)network is proposed,consisting of the Discrete Wavelet Transformation(DWT),the Double Depth Q Network(DDQN),and the Temporal Convolutional Network(TCN)and the A
7、ttention Mechanism.First,the DWT is used to decompose the wind power data series into the wind power data sets of different frequencies.The autocorrelation function analysis is performed on the wind power data sets of different frequencies.The wind power training subset is extracted with high autoco
8、rrelation as the input of the prediction model.Secondly,according to the different frequency wind power datasets obtained after the DWT decomposition,the corresponding TCN-Attention wind power ultra-short-term 基金项目:国家自然科学基金项目(51777027)。Project Supported by National Natural Science Foundation of Chin
9、a(51777027).prediction model is trained,and the wind power timing relationship is deeply excavated.In order to reduce the influence of the parameters of the deep learning model on the prediction accuracy,the parameters of the prediction model are optimized by the DDQN algorithm.Finally,the ultra-sho
10、rt-term wind power prediction results of different frequencies are reconstructed by the DWT to obtain the wind power sequence of the forecast day.Taking the measured data of a wind farm in northwest China as an example,the simulation and analysis results show that the proposed method fully explores
11、the timing relationship of wind power sequence,optimizes the internal parameters of the model,and effectively improves the ultra-short-term wind power prediction accuracy.KEY WORDS:discrete wavelet transform;time series convolution network;deep reinforcement learning;ultra-short-term prediction;atte
12、ntion 摘要:超短期风电功率预测对于电力系统生产调度计划的制定具有重要意义,风电出力具有较强的随机性、波动性、不可控性。风电不确定性对风电时序关系的影响,给风电功率预测精度提出了挑战。针对上述问题,提出了基于离散小波变换(discrete wavelet transformation,DWT)、双深度 Q 网络(double depth Q network,DDQN)、时序卷积网络(temporal convolutional network,TCN)和注意力机制(Attention)的DWT-DDQN-TCN-Attention(DWT-DTCNA)超短期风功率预测方法。首先,利用 DW
13、T 将风电数据序列分解为不同频率的风电数据集,对不同频率的风电数据集做自相关函数分1654 陈海鹏等:考虑风电时序特性的深度小波时序卷积网络超短期风功率预测 Vol.47 No.4 析,提取高自相关性的风功率训练子集作为预测模型的输入。其次,根据 DWT 分解后得到的不同频率风功率数据集分别训练相应的 TCNA 的风电超短期预测模型,深度挖掘风电功率时序关系,获得精度更高、更稳定的预测结果。为减少深度学习模型的参数对预测精度的影响,采用 DDQN算法优化预测模型的参数。最后,利用 DWT 将不同频率超短期风功率预测结果进行重构,获得了预测日的风电功率序列。以西北部某风电场实测数据为例进行仿真分
14、析,结果表明所提方法能够充分提取风电功率序列的时序特征,优化模型内部参数,有效提高了超短期风电功率预测精度。关键词:离散小波变换;时序卷积网络;深度强化学习;超短期预测;注意力机制 DOI:10.13335/j.1000-3673.pst.2022.1019 0 引言 我国“3060”目标的提出明确了风电发展的方向,风电大规模利用将是实现“双碳”目标的主力军1。然而,在电力系统中,大规模风电联网增加了电力系统波动性,给电力系统生产调度计划制定带来了困难。准确的风电功率预测可降低风电联网带来的影响2,对于促进风电功率消纳、优化电力系统调度具有重大意义3。现今,常用的风电功率预测方法包括物理预测法
15、和统计预测法两大类4。其中,物理预测主要是利用数值天气预报(numerical weather prediction,NWP)的预测结果得到气象数据,利用风机的功率曲线计算得出风机的实际输出功率。然而,由于NWP 更新频率较低(13h),难以满足超短期风电功率预测的要求。统计法最具代表性的为机器学习中的浅层学习模型和深度学习模型。浅层学习模型主要有支持向量机5、随机森林等6,可以直接从气象和风功率历史数据中挖掘出深层变化规律进行预测,但由于此类模型网络较浅,面对复杂曲线拟合能力有限,预测精度较低;深度学习7-9模型主要有卷积神经网络、长短期记忆网络(long short-term memory
16、,LSTM),文献10在利用卷积神经网络深挖数据特征的基础上实现了对风电功率的爬坡预测。文献11指出 LSTM 能够较好地提取风电序列的时间依赖性特征,在一定程度上提高风功率预测精度。上述风电功率预测模型均为简单模型,可实现多场景下风电功率预测的部分需求。但新型电力系统中新能源渗透率逐步增加,风电功率序列非线性特征增强,也更加难以实现特征提取及精准预测。常见的单一的风电预测模型不能充分提取风电功率序列的时序特征,无法实现精准地预测风电功率12;深度学习预测模型内部参数对预测精度的影响较大,导致模型预测结果不稳定。文献13采用 LSTM 算法搭建区间预测模型建立了具有一定置信度的最佳预测区间,但 LSTM 参数变化使得预测模型容易陷入局部最优,增加了预测区间的宽度;文献14应用循环神经网络证明了深度学习模型能够提取风电功率序列中的时序关系,同时说明了深度学习模型的参数变化对于预测精度的稳定性具有重要意义。针对上述问题,文献15采用信号处理中应用较广的经验模态分析与支持向量机(support vector machines,SVM)相结合提出风电场短期风电功率组合预测方法,相比较单一 SV