Short-Term Power Load Forecasting Based On VMD Integrated Deep Learning

With the popularity of renewable energy sources connected to the grid,the power system gradually transforming to a more flexible and intelligent system,and the higher penetration of renewable energy generation,short-term power load forecasting(SPLF)plays an increasingly important role in the future grid planning and operation.In order to obtain higher forecasting accuracy,many studies have attempted to construct SPLF models with high accuracy and generalizability.However,the accuracy and stability of SPLF has been a pressing issue due to the nonlinearity and volatility of load data.To address this problem,this paper is to propose a SPLF model based on data decomposition integrated with deep networks.The data decomposition technique is used to decompose the original electric load into linear and smooth subseries,and the deep network is used to deeply mine the hidden information between the external correlation factors and each subseries.A multivariate feature matrix is constructed with the decomposed electric load subseries,temperature,season,double holidays and holidays,and used as the input of the prediction model,so as to achieve the prediction requirement of high accuracy and stability.In this paper,a deep network prediction model based on CNN,TCN,LSTM and SAM is proposed(CTLA).Firstly,a three-layer one-dimensional convolutional neural network(1D-CNN)is constructed as a deep network for extracting spatial features in the feature matrix,TCN for extracting temporal information in the feature matrix,and LSTM for enhancing the long-term temporal dependence of feature information;secondly,SAM is used to dynamically adjust the relevance weights of different features and further enhance the important features;finally,the New England Public Dataset(ISO-NE),for example,compared with the five models of TCN-SAM(TA),TCNLSTM-SAM(TLA),CNN-LSTM-SAM(CLA),CNN-TCN-LSTM(CTL),and CNNTCN-GRU-SAM(CTGA),the average MAPE values of this model are reduced by39.7%,39.1%,24.5%,21.2%,and 9.6%,respectively.The experimental results show that the CTLA model with the ability of deep extraction of hidden features has higher prediction accuracy.In order to reduce the impact of characteristics such as nonlinearity,volatility and randomness of raw power load data on SPLF accuracy,a hybrid SPLF model based on VMD decomposition and deep TCN network is proposed in this paper based on the previous work(VMD-CNN-TCN-LSTM-SAM,VCTLA).Firstly,VMD is used to decompose the raw power load data into eight Intrinsic Mode Functions(IMFs)with high smoothness and stability;secondly,all IMFs with external influences are reconstructed into a multivariate feature matrix and used as the input of the prediction model;finally,the ISO-NE dataset is used as an example with TA,TLA,CLA,CTL,CTGA,and CTLA models,the average MAPE values of this model were reduced by77.0%,76.7%,71.2%,69.9%,65.4%,and 61.8%,respectively.To further validate the robustness and generalization of the VCTLA model,this paper uses two public datasets as examples,including the ISO-NE dataset and the Global Energy Load Forecasting Competition 2012 dataset(GEFCom2012),compares the proposed model with a related hybrid model based on VMD decomposition,and performs training and testing based on both datasets separately.The comparison with five models,VMD-CNN-LSTM-SAM(VCLA),VMD-TCN-LSTM-SAM(VTLA),VMD-CNN-TCN-SAM(VCTA),VMD-CNN-TCN-LSTM(VCTL),and VMD-CNNTCN-GRU-SAM(VCTGA),shows that The average MAPE values of this model on the ISO-NE dataset decreased by 70.0%,66.4%,34.7%,39.0% and 17.5%,respectively;and the average MAPE values on the GEFCom2012 dataset decreased by 40.7%,24.6%,12.7%,15.2% and 7.3%,respectively.In summary,it can be seen that the proposed model in this paper can greatly reduce the nonlinearity and volatility of the raw power load data,and has the ability to deeply extract the hidden information among the features,and thus has excellent performance in robustness and generalization.