基于SVMD-ISSA-CNN-TGLSTM的供热负荷预测模型  

Heat load prediction model based on SVMD-ISSA-CNN-TGLSTM

作　　者：薛贵军 牛盼 谢文举 李水清 XUE Guijun;NIU Pan;XIE Wenju;LI Shuiqing(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China;Intelligent Factory,North China University of Science and Technology,Tangshan 063000,China)

机构地区：[1]华北理工大学电气工程学院,河北唐山063210 [2]华北理工大学智能仪器厂,河北唐山063000

出　　处：《现代电子技术》2024年第11期131-139,共9页Modern Electronics Technique

基　　金：河北省自然科学基金项目(F2021209006);河北省高等学校科学技术研究项目。

摘　　要：针对目前集中供热负荷预测的研究中极少考虑换热站内部因素以及供热负荷预测精准度较低的问题,提出一种基于SVMD-ISSA-CNN-TGLSTM的混合预测模型。首先,利用卷积神经网络和转换门控长短期记忆神经网络构建具有空间提取能力的CNN-TGLSTM模型;其次,考虑到负荷序列的非平稳特征,采用SVMD分解,并引用改进的麻雀搜索算法来优化模型的参数,避免调参陷入局部最优;最后,将不同模型之间的预测效果与经济效益进行对比。结果表明:SVMD-ISSA-CNN-TGLSTM模型经济效益最高,评价指标RMSE、MSE、MAE相比ISSA-CNN-TGLSTM模型分别降低了35.7%、59.0%、32.7%,且均优于其他不同模型,预测效果最佳。In the current research on centralized heating load prediction,the internal factors of heat exchange stations and the low accuracy of heating load prediction are rarely considered,so a hybrid prediction model based on SVMD-ISSA-CNN-TGLSTM is proposed.A CNN-TGLSTM model with spatial extraction capability is constructed by convolutional neural network(CNN)and transformation-gated long short-term memory(TGLSTM)neural network.In view of the non-stationary characteristics of the load sequence,the SVMD(successive variational mode decomposition)is adopted and the improved sparrow search algo-rithm(ISSA)is invoked to optimize the parameters of the model,so as to avoid the parameter adjusting from falling into local optimum.The prediction effects and economic benefits of the different models are contrasted.The results show that the SVMD-ISSA-CNN-TGLSTM model has the best economic benefit,and its evaluation indexes RMSE(root mean square error),MSE(mean square error)and MAE(mean absolute error)are reduced by 35.7%,59.0%and 32.7%,respectively,in comparison with those of the ISSA-CNN-TGLSTM model,and all of the results are better than the other models,so its prediction effect is the best.

关 键 词：供热负荷预测 逐次变分模态分解 改进的麻雀搜索算法 卷积神经网络 转换门控长短期记忆神经网络 空间提取能力 

分 类 号：TN919-34[电子电信—通信与信息系统] TH16[电子电信—信息与通信工程] TP391.9[机械工程—机械制造及自动化]