摘要
新能源功率预测是实现主动配电网运行态势精确感知的关键。针对新能源功率的不确定性和波动性,提出了一种融合迁移学习(TL)和长短期记忆神经网络(LSTM)的新能源功率组合预测方法。引入k-shape聚类算法对不同区域新能源提供的时序数据进行聚类,同时利用各个聚类生成若干预训练模型;结合形态距离指标,选择与目标序列最接近的聚类作为辅助数据簇以准备迁移学习;借助辅助数据簇所对应的预训练模型来完成TL-LSTM模型的训练,且在所有模型的训练过程中利用差值化处理方法避免预测结果出现“滞后”现象。以我国某实际风电场和光伏电站为典型算例,验证所提预测方法的有效性。结果表明,该方法提升了新能源功率短期预测的精度,并能够在小样本环境下进行新能源功率的预测,有较高地泛用性。
New energy power prediction is the key to realize accurate situation awareness of active distribution network.To deal with the uncertainty and volatility brought by new energy power generation,a transfer learning and long short-term memory(TL-LSTM)-based power prediction method is proposed.First,the k-shape clustering algorithm is used to cluster the time-series data provided by new energy sources in different regions,while each cluster can generate several training models.Then,the clusters closest to the target sequence are selected as auxiliary data clusters using shape-based distance(SBD)metrics for migration learning.The training of TL-LSTM models is completed with the pre-trained models corresponding to the auxiliary data clusters,and the difference treatment is used in all the model training processes to avoid the prediction lags.Finally,the effectiveness of the proposed prediction method is verified by a wind farm and a photovoltaic power plant in China as typical examples.The results show that the method improves the accuracy of short-term prediction on new energy units’outputs,and is widely applicable to new energy power prediction under small sample size.
作者
郑真
朱峰
马小丽
田书欣
姜皓喆
ZHENG Zhen;ZHU Feng;MA Xiaoli;TIAN Shuxin;JIANG Haozhe(State Grid Shanghai Qingpu Electric Power Supply Company,Shanghai 201700,China;School of Electrical Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《综合智慧能源》
CAS
2023年第1期41-48,共8页
Integrated Intelligent Energy
基金
国网上海市电力公司科技项目(52093421N002,B30934210006)。
