| In recent years,with the rapid development of the Internet of things,the efficiency of data acquisition,transmission and storage has been gradually improved,and the power system has gradually changed to an intelligent interactive system.Short-term load forecasting(STLF)is an important part of the intelligent power system.Improving the forecasting accuracy is conducive to the economic operation of the power grid and achieving the balance between supply and demand of the power system.However,the grid connection of large-scale and fluctuating renewable energy brings great challenges to the stable operation of power system,which brings great challenges for the higher accuracy of STLF.This paper studies STLF based on deep learning theory combined with multiple influencing factor analysis,and explores the internal relations and laws of these input variables to improve the prediction accuracy.The results of this paper will provide a useful reference for power dispatching and power grid planning.The main contents of this paper are shown as follows:(1)Process the load data and external influencing factors.Firstly,the missing values and outliers of the power load data in the dataset are corrected and supplemented.Secondly,the periodicity of load data is analyzed and the influences of external factors such as temperatures and holidays on the power load are studied.Thirdly,the correlation degrees of external factors with load data are calculated by using Spearman correlation coefficient method and then the external factors with high correlations are selected.Finally,the load data are reconstructed a feature matrix along with the external factors with high correlations.(2)A hybrid model based on convolution neural network(CNN)and temporal convolution network(TCN)for STLF is proposed.The hybrid mode can make full use of the advantages of each model.The CNN is used to extract the correlation features of load data and external factors,and the TCN is utilized to extract the long-term temporal dependencies of the load data and external factors.Experiments are performed on two public datasets and compared the proposed model with CNN,LSTM,GRU,TCN and CNN-LSTM.The experimental results show that CNN-TCN hybrid model can effectively improve the accuracy of STLF.(3)In order to further mine the hidden information of power load and external factors,an improved Dense Net-TCNA hybrid model is proposed.Firstly,the residual of TCN is improved by using a parallel structure including maximum pooling and average pooling and its activation function and loss function are optimized.Secondly,the in-depth features of the input matrix are extracted by the Dense Net,and the improved TCN is used to extract the temporal features of the load data.Finally,the self-attention mechanism can be dynamically adjust the weight of features and obtain key features.Experiments are carried out on three data sets and compared the proposed model with CNN,Dense Net,TCN,improved TCN and CNN-TCN.The three public datasets results show that the improved Dense Net-TCNA hybrid model can effectively improve the accuracy and generalization of STLF. |
