基于多特征提取和多层级迁移学习的电动汽车充电站充电量预测  

Electric vehicle charging station charging forecasting based on multi-feature extraction and multi-level transfer learning

作　　者：李振华[1] 张成浩 刘奕舟 魏伟 LI Zhenhua;ZHANG Chenghao;LIU Yizhou;WEI Wei(College of Electrical Engineering&New Energy,China Three Gorges University,Yichang 443002,China;State Grid Hubei Electric Power Co.,Ltd.Marketing Service Center(Measurement Center),Wuhan 443080,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002 [2]国网湖北省电力有限公司营销服务中心(计量中心),湖北武汉443080

出　　处：《电力系统保护与控制》2025年第6期150-162,共13页Power System Protection and Control

基　　金：国家自然科学基金项目资助(52277012);武汉强磁场学科交叉基金项目资助(WHMFC202202)。

摘　　要：电动汽车充电站充电量预测对于充电站规划、建设、充电管理平台营销等有着实际意义。但新建、改造的充电站可能会面临部分时段数据缺失、历史数据不足和浅层神经网络模型难以捕捉等多变且复杂的输入特征的问题。因此,提出了一种基于多特征提取和多层级迁移学习的电动汽车充电站充电量预测方法。首先,使用K-Means算法对所有用户在不同时间段的充电次数进行聚类,得到4类充电行为特征,将其和其他影响特征融合作为模型的输入特征集。其次,设计并行连接的多尺度混合时间卷积网络(temporal convolutional network,TCN)层作为特征提取器,两层BiLSTM层将提取到的特征进行深层学习,Attention层加强个体特征选择。最后,将源充电站数据进行相关性等级划分,按照相关性由弱到强输入到模型中进行多层级迁移学习,保留损失函数最低的训练权重,得到最终的预测结果。算例分析表明,多层级迁移学习可以弥补新建、改造的充电站数据样本不足的缺陷。与直接迁移相比,多层级迁移平均绝对误差(mean absolute error,MAE)降低了10.75%,均方根误差(root mean square error,RMSE)降低了13.73%,拟合优度R2提升了0.4%。Electric vehicle charging station charging forecasting is crucial for charging station planning,construction,and marketing strategies in charging management platforms.However,newly built or upgraded charging stations may face problems of missing data for some time periods,insufficient historical data,and difficulties in capturing complex input features with shallow neural network models.Therefore,a charging forecasting method for electric vehicle charging stations based on multi-feature extraction and multi-level transfer learning is proposed.First,the K-Means algorithm is used to cluster users’charging frequency over different time periods to obtain four types of charging behavioral features.These are integrated with other influencing features to form the model’s input feature set.Next,a parallel-connected multiscale hybrid temporal convolutional network(TCN)layer is designed as the feature extractor,followed by two BiLSTM layers for deeper feature learning.An Attention layer is added to strengthen individual feature selection.Finally,charging station data from source locations are classified into correlation levels,and data are fed into the model in a weak-to-strong correlation order through multi-level transfer learning.The training weights with the lowest loss function value are retained to obtain the final prediction results.Case study results show that multi-level transfer learning can compensate for the lack of data samples in new or upgraded charging stations.Compared to direct transfer learning,the proposed method reduces the mean absolute error(MAE)by 10.75%,decreases the root mean square error(RMSE)by 13.73%,and improves the R2 by 0.4%.

关 键 词：充电量预测 迁移学习 特征提取 时间卷积网络 双向长短期记忆网络 注意力机制 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]