基于特征迁移学习的综合能源系统小样本日前电力负荷预测
Small-sample day-ahead power load forecasting of integrated energy system based on feature transfer learning
摘要点击 190  全文点击 79  投稿时间:2020-05-18  修订日期:2020-08-19
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2020.00280
  2021,38(1):63-72
中文关键词  综合能源系统  日前电力负荷预测  特征提取  迁移学习  门控循环单元
英文关键词  integrated energy system  day-ahead power load forecasting  feature extraction  transfer learning  gated recurrent unit (GRU)
基金项目  中央高校基本科研业务费专项资金项目(2020ZDPY0216).
作者单位E-mail
孙晓燕 中国矿业大学信息与控制工程学院 xysun78@126.com 
李家钊 中国矿业大学信息与控制工程学院  
曾博 新能源电力系统国家重点实验室(华北电力大学)  
巩敦卫 中国矿业大学信息与控制工程学院  
廉智勇 太原煤气化龙泉能源发展有限公司  
中文摘要
      基于历史数据和深度学习的负荷预测已广泛应用于以电能为中心的综合能源系统中以提高预测精度, 然 而, 当区域中出现新用户时, 其历史负荷数据往往极少, 此时, 深度学习难以适用. 针对此, 本文提出基于负荷特征 提取和迁移学习的预测机制. 首先, 依据源域用户历史负荷数据, 融合聚类算法和门控循环单元网络构建源域数据 的特征提取和分类模型; 然后, 利用该模型提取当前待预测目标域小样本的特征及其类别信息, 进而给出基于特征 相似度和时间遗忘因子的特征融合策略; 最后, 依据融合特征, 给出基于迁移学习和特征输入的负荷预测. 将所提 算法应用于卡迪夫某区域的高中和住宅用电预测中, 实验结果表明了该算法在综合能源系统小样本电力负荷预测 中的有效性.
英文摘要
      Deep learning-based load forecasting with large size of historical data has been successfully used in integrated energy systems to effectively improve the prediction accuracy. However, when new users join in the system, the historical load data is often very rare, and deep learning is no longer applicable. Motivated by this, we propose a forecasting mechanism based on load feature extraction and transfer learning. Firstly, the feature extraction and classification models of source domain data are constructed through optimal clustering and gated recurrent unit (GRU) training. Then, the trained GRU classification model is used to extract the features and category information of small samples in the target domain to be predicted. A feature fusion strategy based on feature similarity and time forgetting factor is proposed. Finally, according to the fusion characteristics, the load prediction based on transfer learning and feature input is given. The proposed algorithm is applied to the electricity load forecasting of high schools and buildings in Cardiff. The experimental results show the effectiveness of the algorithm in power load forecasting with small size of samples.