Improved Domain Adversarial Neural Networks Based Generation Method of Wind-photovoltaic Power Time Series Curves for Renewable Energy Base
任佳星 1孙英云 1秦继朔 2刘栋 2郭国栋 2张柯欣2
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 华北电力大学电气与电子工程学院,北京市 昌平区 102206
2. 国网经济技术研究院有限公司,北京市 昌平区 102209
折叠
摘要
准确刻画风光时序功率曲线对于加快推动新能源大规模并网、指导联合发电系统规划运行具有重要意义.针对我国沙漠、戈壁、荒漠等地区新建大型风电光伏发电基地无历史功率数据可利用的现状,该文提出基于改进域对抗网络(improved domain adversarial neural networks,IDANN)的新能源基地风光时序功率曲线生成方法.以历史气象和功率数据充足的新能源场站作为源域,仅有气象数据的新建基地作为目标域.将源域上学习的输入气象信息到输出风光功率的非线性映射知识迁移到目标域,并添加最大均值差异(maximum mean discrepancy,MMD)作为度量域间特征分布相似性的损失函数以降低目标域泛化误差.最后采用实际风光场站算例验证所提方法的有效性,并进一步表明该方法的实用价值和意义.
Abstract
Accurately characterizing the time series power curves of wind-photovoltaic is crucial for accelerating renewable energy development and guiding the operation and security of integrated energy systems.In response to the current situation of new large-scale wind and photovoltaic bases in the desert,Gobi,and barren areas of China without available historical power data,this paper proposes a method for generating wind-photovoltaic power curves using an improved domain adversarial neural network.Firstly,Existing stations with sufficient historical weather and power data are considered as the source domain,and newly established bases with only weather data are considered as the target domain.Then,we transfer the knowledge of the non-linear mapping from input weather information to output wind-photovoltaic power learned in the source domain to the target domain and add maximum mean discrepancy as a loss function to measure the similarity of feature distributions between domains for reducing the generalization error in the target domain.Finally,the effectiveness of the proposed method is verified by actual wind-photovoltaic power station examples,and the practical value and significance of the method is further demonstrated.
关键词
风光时序功率/改进域对抗网络/沙戈荒/迁移学习
Key words
wind-photovoltaic power time series/improved domain adversarial neural networks/Gobi Desert and other arid regions/transfer learning
维普
万方数据