黄睿, 朱玲俐, 高峰, 王渝红, 杨亚兰, 熊小峰. 基于变分模态分解的卷积长短时记忆网络短期电力负荷预测方法[J]. 现代电力. DOI: 10.19725/j.cnki.1007-2322.2022.0210
引用本文: 黄睿, 朱玲俐, 高峰, 王渝红, 杨亚兰, 熊小峰. 基于变分模态分解的卷积长短时记忆网络短期电力负荷预测方法[J]. 现代电力. DOI: 10.19725/j.cnki.1007-2322.2022.0210
HUANG Rui, ZHU Lingli, GAO Feng, WANG Yuhong, YANG Yalan, XIONG Xiaofeng. Short-term Power Load Forecasting Method Based on Variational Modal Decomposition for Convolutional Long-short-term Memory Network[J]. Modern Electric Power. DOI: 10.19725/j.cnki.1007-2322.2022.0210
Citation: HUANG Rui, ZHU Lingli, GAO Feng, WANG Yuhong, YANG Yalan, XIONG Xiaofeng. Short-term Power Load Forecasting Method Based on Variational Modal Decomposition for Convolutional Long-short-term Memory Network[J]. Modern Electric Power. DOI: 10.19725/j.cnki.1007-2322.2022.0210

基于变分模态分解的卷积长短时记忆网络短期电力负荷预测方法

Short-term Power Load Forecasting Method Based on Variational Modal Decomposition for Convolutional Long-short-term Memory Network

  • 摘要: 1电力负荷序列易受多重外部因素影响而呈现复杂性,不利于精准预测。为此,提出一种基于变分模态分解(variational mode decomposition,VMD)的卷积神经网络和长短期记忆网络(convolutional neural network and long short-term memory network,CNN-LSTM)相结合的短期电力负荷并行预测方法。先采用VMD将负荷数据分解为规律性强的各本征模态函数(intrinsic mode function,IMF)及残差;再将各分量分别输入到各自对应的CNN-LSTM混合预测网络获得各初始预测值,并将该值与由气候、日期类型等组合得到的相关因素特征集相结合进一步得出修正预测值;最终叠加各分量修正预测值即得到完整预测结果。在实际负荷数据上做验证分析,结果表明,考虑相关外部因素特征集后日负荷预测平均相对误差均值可降低2.18%。与几种常规负荷预测方法进行效果对比,验证了该方法的有效性和可行性。

     

    Abstract: The power load sequence is complicated and easily affected by multiple external factors, making it difficult to anticipate with accuracy. A parallel forecasting method of short-term power load combining variational modal decomposition (VMD) and convolutional neural network and long short-term memory network (CNN-LSTM) is proposed to address the problem. Firstly, VMD is adopted to decompose the load data into various intrinsic mode functions (IMF) with strong regularity and residual error; Secondly, the obtained components are input into the corresponding CNN-LSTM hybrid prediction network to obtain each initial prediction value, and combine this value with the correlation factor feature set obtained by combining climate, date type, etc. to further obtain the revised prediction value; Finally, the revised prediction values of each component are superimposed to obtain a complete prediction result. According to the simulation on the actual load data, the average relative error of daily load forecasting can be reduced by 2.18% after taking into about the relevant external factor features set. In addition, compared with several conventional load forecasting methods, the effectiveness and feasibility of the proposed method can be verified.

     

/

返回文章
返回