基于迁移学习和GRU网络的新建小区负荷预测

New community load prediction based on transfer learning and GRU network

针对新建小区没有任何历史负荷数据的难点,提出了基于相似小区特征进行数据迁移的门控循环单元负荷预测算法,实现对新建小区负荷的预测。首先,利用迁移学习的思想迁移出与新建小区特征高度相似的数据信息,及预测模型参数;其次,利用特征数据集作为训练集完成极端梯度增强算法回归模型的训练;然后,采用门控循环单元神经网络对训练样本集进行建模,当模型达到预测精度时,从而完成蕴含时序关系的新建设小区的中长期负荷预测。最后,以连云港某小区为例,得到了该小区在2020年1月-2022年11月时间段内的负荷预测结果,以验证所建立的蕴含时序关系的中长期负荷预测模型的有效性。

In view of the difficulty of new community without any historical load data,a gated recurrent unit(GRU)load prediction algorithm based on data transfer of similar community characteristics is proposed to realize the load prediction of new community.Firstly,the idea of transfer learning is used to transfer the data information which is highly similar to the features of the new community and pre⁃dict the model parameters;Secondly,the feature data set is used as the training set to complete the training of extreme gradient boosting(XGBoost)regression model;Then,GRU neural network is used to model the training sample set.When the prediction accuracy of the model is reached,the medium and long term load prediction of the new community with time series relationship is completed.Finally,taking a community in Lianyungang as an example,the load predic⁃tion results of the residential area form January 2020 to November 2022 are obtained,so as to verify the validity of the established medi⁃um and longtermload prediction model with time series relationship.