基于EEMD-CBAM-BiLSTM的牵引负荷超短期预测  

Traction Load Ultra-Short-Term Forecasting Framework Based on EEMD-CBAM-BiLSTM

作　　者：钟吴君 李培强[1] 涂春鸣[1] Zhong Wujun;Li Peiqiang;Tu Chunming(College of Electrical and Information Engineering Hunan University,Changsha 410082 China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082

出　　处：《电工技术学报》2024年第21期6850-6864,共15页Transactions of China Electrotechnical Society

基　　金：国家重点研发计划项目(2021YFB2601504);国家自然科学基金项目(52377097)资助。

摘　　要：针对电气化铁路牵引负荷难以预测的问题,构建了一种由集合经验模态分解(EEMD)、改进型卷积块注意力模块(CBAM)和双向长短期神经网络(BiLSTM)组合成的EEMD-CBAM-BILSTM预测方法,有效地降低了牵引负荷超短期预测误差与计算成本。首先,通过EEMD将牵引负荷数据分解为多个稳定、有规律的时序模态函数,突出负荷数据的时序特征;其次,将分解后的各分量整体通入由卷积神经网络(CNN)和改进型CBAM组成的特征提取模块提取全局时序特征;最后,利用贝叶斯优化(BO)搜寻BiLSTM最优参数,并将全局特征通入优化后的神经网络进行超短期时序预测。仿真算例表明,该文所提预测框架在各预测步长下均能很好地把握牵引负荷变化趋势,显著提升了牵引负荷预测的精度。China has the world's largest rail transit network,with a total mileage of more than 150,000 kilometers,the electrification rate of more than 70 percent,and railway energy consumption is the largest single load category.With the rapid development of China's electrified railway network,the railway transportation represented by high-speed rail has become an important part of China's transportation system.With the development of"net-source-storage-vehicle"collaborative energy supply technology,the internal energy management and collaborative control of the system need the technical support of accurate second-level ultra-short-term prediction on both sides of the source and load.At the same time,accurate ultra-short-term prediction of traction load can also provide data source support for research on power quality analysis of rail transit,optimal scheduling of traction substation,location and capacity determination of traction substation,etc.This paper constructs an ultra-short-term forecast framework for traction load based on EEMD-CBAM-BiLSTM,aiming at the problem of difficulty in predicting rail transit traction load due to strong mutability and volatility.Firstly,based on the analysis of the time series characteristics of rail transit traction load,the data of rail transit traction load is decomposed into several stable and regular time series mode functions by ensemble empirical mode decomposition(EEMD)to highlight the time series characteristics of load data.Secondly,the decomposed components are integrated into the feature extraction module composed of convolutional neural network(CNN)and improved convolutional block attention module(CBAM)to extract the global timing features.After that,Bayesian optimization(BO)is used to search the optimal parameters of BiLSTM neural network to make the network structure reach the best state.In this paper,the traction load data in a single day(with a resolution of 1 s)is selected to build a prediction model,and the model is compared with the mainstream single prediction model

关 键 词：牵引负荷预测 集合经验模态分解 双向长短期神经网络 贝叶斯优化 卷积块注 意力模块 卷积神经网络 

分 类 号：TM922.3[电气工程—电力电子与电力传动]