基于时域卷积网络精细化光伏发电功率预测  

A Refined Photovoltaic Power Prediction Based on Time Domain Convolutional Network

作　　者：刘文杰 陈耀 宋晓宁[3] 张万强 LIU Wenjie;CHEN Yao;SONG Xiaoning;ZHANG Wanqiang(State Power Investment Corporation Dong Fang New Energy Corporation Hengshui Branch,Hengshui 053000,China;Xinjiang Goldwind Sci&Tech Co.,Ltd.Wind Power Industry Group,Wuxi 214000,China;School of IoT Engineering,Jiangnan University,Wuxi 214000,China)

机构地区：[1]国家电投集团东方新能源股份有限公司衡水分公司,河北衡水053000 [2]新疆金风科技集团股份有限公司风电产业集团,江苏无锡214000 [3]江南大学物联网工程学院,江苏无锡214000

出　　处：《供用电》2020年第10期76-82,共7页Distribution & Utilization

基　　金：国家自然科学基金项目(61876072)。

摘　　要：针对传统光伏功率预测算法模型没有与时间序列结合而引发精度不高的问题,提出了一种基于时域卷积网络精细化的光伏发电功率预测方法。该预测模型利用卷积神经网络并融合了因果卷积和膨胀卷积;在卷积特征上,使用跨层连接;在损失函数上,增加了自适应影响因子;在特征工程上,将数值气象预报发布的多个预测数据与时间特征结合进行多特征预测。根据新疆、华东两个光伏电场数据集的预测结果显示,与支持向量机算法和未加入时域特性的神经网络算法对比,该算法在光伏功率预测精度上均有不同程度的提升,充分证明了有效性。Aiming at the problem of low prediction accuracy caused by the traditional photovoltaic power prediction algorithm not combining deep learning with time series and the prediction model better,a refined photovoltaic power prediction method based on Temporal Convolutional Network is proposed.As for predictive model,the powerful feature expression ability of convolution neural network is used in the proposed method in which causal convolution and expansion convolution are also combined.In terms of convolution feature,the cross-layer connection is used,what’s more,the self-adaptive iterative parameters are added to the loss function to make it better applicable to photovoltaic power generation data sets.In the aspect of power characteristics,the multiple prediction data released by Numerical Weather Prediction are combined with time characteristics to perform multi-feature predictions.According to the prediction results of two photovoltaic power data sets in Xinjiang and Eastern China,compared with Support Vector Machine and Neural Networks,the algorithm in this paper increases the accuracy of photovoltaic power prediction in different degrees,which fully proves the effectiveness of this algorithm.

关 键 词：光伏 卷积神经网络 时间序列 功率预测 

分 类 号：TM74[电气工程—电力系统及自动化]