基于多尺度模型融合和VMD-TCN-RF混合网络的短期电力负荷预测方法

Short-term Power Load Forecasting Method Based on Multiscale Model Fusion and VMD-TCN-RF Hybrid Network

为了充分挖掘不同尺度影响因子对短期电力负荷的影响,以及解决预测精度受数据非平稳特性影响的问题,提出了一种基于多尺度模型融合和VMD-TCN-RF混合网络的短期电力负荷预测方法。该方法先采用变分模态分解(Variational Mode Decomposition, VMD)将历史负荷分解为若干平稳性好的本征模态函数(Intrinsic Mode Function, IMF)分量,再把VMD分解得到的各个历史负荷的IMF分量和气象数据分别送入时间卷积网络(Temporal Convolutional Network, TCN)进行特征提取;将所有TCN网络提取的特征融合为一个新的特征向量;最后将融合得到的特征向量与经过One-Hot编码的日期因素特征向量拼接,把拼接得到的向量送入随机森林网络进行预测。通过公开的电力负荷数据集对本方法进行验证,结果表明所提方法与现有模型相比具有更高的预测精度。

In order to fully explore the impact of different scale factors on short-term power load and to solve the problem that prediction accuracy is affected by the non-stationary characteristics of data,a short-term power load forecasting method based on multi-scale model fusion and VMD-TCN-RF hybrid network is proposed.Firstly,the historical load is decomposed into several stable intrinsic mode function(IMF)components by variational mode decomposition(VMD),and then the IMF components and meteorological data of each historical load decomposed by VMD are respectively sent to the temporal convolutional network(TCN)for feature extraction.All the corresponding features extracted by TCN network are fused into a new feature vector.Finally,the fused feature vector is spliced with the date factor fea-ture vector encoded by One-Hot encoding,and the spliced vector is sent to the random forest network for prediction.The proposed method is validated on an open power load dataset,and the results show that it has higher prediction accuracy than existing models.