Research On Short-term Electricity Load Forecasting Based On Feature Selection And Deep Feature Extraction

With the rapid development of social economy and the implementation of the strategy of "carbon peaking and carbon neutrality",it is of great significance to accelerate the low-carbon transformation of the energy structure of the power industry and to build a new power system mainly based on renewable energy.Under this strategy,a large amount of renewable energy is integrated into the national power grid,the scale of the power system is increasing,and the economic and stable operation of the power system is facing major challenges.Therefore,the construction of smart grid must be vigorously promoted in order to achieve scheduling optimization,stable operation and cost regulation of the power system.As an important part of smart grid,power load forecasting can provide strong strategic support for the construction of smart grid.Electricity load data is a typical time series,which is influenced by external characteristic factors such as temperature and date type,and is highly nonlinearity,stochasticity and volatility.In order to achieve effective extraction of spatial and temporal features in data samples,this paper proposes a hybrid model that takes advantage of the respective models of one-dimensional convolutional neural network,temporal convolutional networks and long short term memory(CT-LSTM).Firstly,1DCNNs with three different convolutional kernels are able to extract the interconnections between different feature factors in the data samples;Secondly,TCN has a flexible receptive field and residual structure,which can capture the temporal features of long sequences;Finally,the LSTM integrates the features extracted by both to further strengthen the long-term temporal dependence.The New England Public Dataset(ISONE),for example,compared with the five models of 1D-CNN、TCN、LSTM、TCNLSTM、1D-CNN-LSTM 和 1D-CNN-TCN,the average MAPE values of this model are reduced by 23.67%、15.25%、15.97%、15.25%、8.26% and 7.41%.High-quality input samples with higher data smoothing and stronger correlation with historical load series facilitate the training and convergence of the model.In order to further explore the local hidden features in power load data and construct high quality input data samples,this paper further proposes a feature selection method(EFS) based on empirical modal decomposition and pearson correlation coefficient method,and introduces a self-attention mechanism(SAM)to construct an innovative hybrid short-term electric load forecasting model by integrating deep learning techniques(EFS-CTA-LSTM).Firstly,the EFS method is capable of mining the data itself for local hidden features at different frequencies and eliminating redundant features without increasing the burden on the model;Secondly,SAM is capable of reinforcing the key information in the feature mapping to provide quality input to the prediction process.The New England Public Dataset(ISO-NE),for example,the EFSCTA-LSTM model further reduces the MAPE value by 25% compared to the CTLSTM model.In addition,experimental results on two public datasets show that the model in this paper has higher stability and generalization,and still shows higher superiority in the application of medium-term and long-term power load forecasting.