基于ResNet-LSTM网络和注意力机制的综合能源系统多元负荷预测  

Multi-Energy Load Forecasting in Integrated Energy System Based on ResNet-LSTM Network and Attention Mechanism

作　　者：王琛[1] 王颖[1] 郑涛 戴则梅[2,3] 张凯锋 Wang Chen;Wang Ying;Zheng Tao;Dai Zemei;Zhang Kaifeng(Key Laboratory of Measurement and Control of Complex Systems of Engineering,Ministry of Education,Southeast University,Nanjing,210096,China;NARI Group(State Grid Electric Power Research Institute)Co.Ltd,Nanjing,211106,China;NARI Technology Development Co.Ltd,Nanjing,211106,China)

机构地区：[1]复杂工程系统测量与控制教育部重点实验室(东南大学),南京210096 [2]南瑞集团(国网电力科学研究院)有限公司,南京211106 [3]国电南瑞科技股份有限公司,南京211106

出　　处：《电工技术学报》2022年第7期1789-1799,共11页Transactions of China Electrotechnical Society

基　　金：国家自然科学基金(51907025,51977033);国家电网公司总部科技项目(基于大数据的电网趋势预测及操作智能预演技术研究);东南大学“至善青年学者”支持计划(2242021R41176)资助。

摘　　要：综合能源系统中多种负荷之间可能存在复杂的、较强的相互耦合关系。相对于对各类负荷进行单一独立的预测,直接开展多元负荷预测能够进一步挖掘负荷之间的内在联系,提高预测准确度。该文提出一种基于ResNet-LSTM网络和注意力机制的多任务学习模型,用于拟合多能负荷之间的空间耦合关系和时间耦合关系。首先,采用多层ResNet作为多能负荷数据的特征提取单元,挖掘多能之间的空间耦合交互特征;然后,通过双向长短时记忆网络残差结构进一步挖掘多能负荷数据的时序特征;接着,使用注意力机制实现多任务对于共享特征不同程度的关注,体现不同子任务对共享特征的差异化选择,实现多元负荷的联合预测;最后,结合亚利桑那州立大学CampusMetabolism系统的多能负荷数据,与其他预测模型进行对比分析,结果表明所提出的多元负荷预测方法具有更高的预测精度。In an integrated energy system,different types of loads,i.e.electrical loads,heat loads,cooling loads,might have complex and strong coupling relationships among them.Compared with forecasting each type of load separately,to forecast multi-energy loads together in a combined multitask model can further explore the internal connections and therefore improve the accuracy of forecasting.A multi-task learning model based on ResNet-LSTM network and Attention mechanism is proposed to fit the spatial coupling relationship and time coupling relationship between multi-energy loads.Firstly,the multi-layer ResNet is used as the feature extraction unit of the multi-energy load data to mine the spatial coupling interaction characteristics between the multi-energy;secondly,the LSTM residual structure is used to further mine the time series characteristics of the multi-energy loads data;then,the Attention mechanism is used to realize that multiple subtasks have different degrees of attention to shared features,which reflects the differentiated selection of shared features by different subtasks,and realizes joint forecasting of multiple loads.Finally,we applied the proposed method with the data at the Campus Metabolism system of Arizona State University.Compared with other forecasting models,the results show that the proposed method has higher accuracy.

关 键 词：注意力机制 残差网络 长短时记忆网络 多元负荷预测 多任务学习 

分 类 号：TM715[电气工程—电力系统及自动化]