基于GWO-VMD-LSTM的热力系统单参数时间序列预测方法  

Time Sequential Prediction Method of Single Parameter for Thermal System based on GWO-VMD-LSTM

作　　者：卓越 倪何[1] 肖鹏飞 ZHUO Yue;NI He;XIAO Pengfei(College of Power Engineering,Naval University of Engineering,Wuhan,China,430033;No.703 Research Institute of CSSC,Harbin,China,150078)

机构地区：[1]海军工程大学动力工程学院,湖北武汉430033 [2]中国船舶集团有限公司第七○三研究所,黑龙江哈尔滨150078

出　　处：《热能动力工程》2024年第6期80-88,共9页Journal of Engineering for Thermal Energy and Power

基　　金：国家自然科学基金面上项目(51909254);海军工程大学自主研发基金资助项目(425317T014)。

摘　　要：针对热力系统参数运行数据预测困难、准确率低的问题,基于灰狼算法(Grey Wolf Optimizer,GWO)、变分模态分解(Variational Mode Decomposition,VMD)、长短期记忆模型(Long Short Term Memory,LSTM)提出一种单参数时序预测方法。首先,使用改进适应度函数的GWO对VMD的分解层数和惩罚系数进行寻优;其次,以最优参数对运行数据进行VMD,并将筛选出的本征模态函数(intrinsic mode function,IMF)分量作为原始数据趋势项;最后,以此运行参数趋势项作为LSTM的训练集输入特征向量,构建LSTM,LSTM超参数由北方苍鹰算法(Northern Goshawk Optimization,NGO)得到。经实际案例验证,该方法可以通过降低原始数据的噪声和扰动对LSTM的影响,增加LSTM对热力参数运行趋势的可预测时间长度和预测精度,相较于传统的LSTM,所提方法的有效预测时间长度增加约176%、预测精度提高约158%。It is difficult to predict the operation data of thermal system parameters,and the accuracy of prediction data is low.Based on grey wolf optimizer(GWO)algorithm,variational mode decomposition(VMD)and long short term memory(LSTM)model,a time sequential prediction method of single parameter is proposed.Firstly,the number of decomposition layers and penalty coefficients of VMD are optimized by GWO with improved fitness function;secondly,VMD of operation data is conducted with optimal parameters and the intrinsic mode function(IMF)components are screened out as the original data trend term;finally,the LSTM is constructed by using the running parameter trend term as the input feature vector of the training set which is used for constructing LSTM,and the LSTM hyperparameters are obtained by the northern goshawk optimization(NGO)algorithm.Verified by case studies,this method effectively mitigates the impact of noise and disturbances in the original data on LSTM,thereby enhancing its predictive time length and accuracy for thermal parameter trends.In comparison to traditional LSTM models,the proposed method demonstrates a remarkable 176%increase in effective prediction time length and a substantial 158%improvement in prediction accuracy.

关 键 词：时序预测 灰狼算法 模态分解 长短期记忆网络 

分 类 号：N37[自然科学总论] TK221[动力工程及工程热物理—动力机械及工程]