Research On Short-term Electricity Load Forecasting Based On Attention Mechanism And Temporal Convolutional Network

Short-term load forecasting is an important part of the energy industry.Accurate power load forecasting can greatly reduce the operation cost of the power grid,which is an important guarantee for the stable and efficient operation of the power system.With the popularity of smart meters and smart sensors,power load data is more easily accessible,which provides large data support for the application of deep learning in power load forecasting.Electricity load data is a kind of time series data,which not only has time series characteristics but also has its own unique characteristics.This paper defines the size of the lag window according to autocorrelation.It forms the input of prediction model through selecting historical load data for 24 hours,seven days,eight weeks,and three months in advance,and combines holidays,temperature and other characteristics.In this paper,a prediction method based on convolutional neural networks and long short-term memory networks(CNN-LSTM)is firstly proposed to predict shortterm electric load more accurately.CNN has the advantage of extracting the main features from the input model,while LSTM is able to capture long-term dependencies in the time series.The model combines the advantages of the two previous models.A better set of hyperparameters in the CNN-LSTM model is derived using the control variables method.To demonstrate the effectiveness of the model,the model is compared with five models,i.e.,a single CNN,LSTM,RNN,GRU,and BP,performed by designing comparison experiments based on a publicly available load dataset.The experimental results show that the CNN-LSTM model reduces the MAPE values by2.2%,3.3%,5.1%,37.5%,and 35.6% compared to the single LSTM,GRU,RNN,CNN,and BP models,respectively.Based on the above research,to further improve the prediction accuracy,the CNNLSTM model is improved,and a prediction method based on attention mechanism and temporal convolution(TCNA-LSTM)is proposed.TCN can receive long-distance information,and reduce redundant features,and has a longer storage space.Residual connections can transfer shallow information to deeper layers.Attention can capture key information in the training data.To demonstrate the effectiveness of the model,a comparison experiment was designed based on the same dataset and the results showed that the TCNA-LSTM model reduced the MAPE values by 73.7%,74.3%,74.6%,75.1%,83.1%,and 83.6% compared to the CNN-LSTM,LSTM,GRU,RNN,CNN,and BP models,respectively.