基于多任务双层注意力优化的TCN-BiGRU综合能源负荷短期预测  

Short-term Load Forecasting for Integrated Energy System Based on TCN-BiGRU Optimized by Multi-task Dual-layer Attention

作　　者：倪建辉 张菁[1] NI Jianhui;ZHANG Jing(School of Electrical and Electronic Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《控制工程》2024年第11期1924-1936,共13页Control Engineering of China

基　　金：国家自然科学基金资助项目(52077137)。

摘　　要：基于多元负荷预测是综合能源系统(IES)生产计划和能源调度的前提,提出一种基于多任务双层注意力优化的时序卷积网络与双向门控循环单元相结合(TCN-BiGRU)的综合能源负荷短期预测方法。首先,将特征集通过最大互信息系数法进行相关性分析,构建不同负荷的输入特征集;然后,输入多任务学习平台进行离线训练,其中的共享层采用高效通道注意力网络(ECANet)优化的TCN,特定任务层则采用自注意力机制优化的BiGRU;最后,选取亚利桑那州立大学坦佩校区冬季和夏季典型日的实际数据进行在线测试。测试结果表明,对比多种深度神经网络模型,所提方法在冬季和夏季的多元负荷加权平均绝对百分比误差分别最大降低了69.35%和73.26%,加权均方根误差分别最大降低70.11%和79.46%。Based on the prediction of multi-energy load is a prerequisite for the production planning and energy dispatch of integrated energy system(IES),a short-term load forecasting method for integrated energy system based on a combination of temporal convolutional network and bidirectional gated recurrent unit(TCN-BiGRU)optimized with multi-task dual-layer attention is proposed.Firstly,the feature set is subjected to correlation analysis using the maximum information coefficient method to construct input feature sets for different loads.Then,these sets are input into a multi-task learning platform for offline training.In this platform,the shared layer employs a temporal convolutional network(TCN) optimized with the efficient channel attention network(ECANet),while the task-specific layers use bidirectional gated recurrent unit(BiGRU) optimized with the self-attention.Finally,the online test is conducted with the actual data of Arizona State University Tempe campus,selecting typical days in winter and summer,the test results show that:compared with various deep neural network models,the weighted mean absolute percentage error of the multi-energy load in winter and summer is respectively reduced by a maximum of 69.35% and 73.26%,and the weighted root mean square error is respectively reduced by a maximum of 70.11% and 79.46%.

关 键 词：多元负荷短期预测 最大互信息系数 多任务学习 时序卷积网络 双向循环门控单元 高效通道注意力网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TM715[自动化与计算机技术—控制科学与工程]