| The stability of the power system is closely related to the development of the national economy.Only on the premise of ensuring the stable supply of electricity can we ensure the sustainable and stable development of the economy.However,due to the inability to store power on a large scale at present,it is necessary for the power supply department to make a reasonable power outage plan to ensure the normal life of residents while avoiding the waste of power resources.Therefore,it is of great significance to forecast the future electricity use.However,due to the characteristics of nonlinear and unstable load data,it is difficult to improve the validity of load data.The factors affecting the variation of electric load vary from place to place,and it is difficult to determine the parameters of the forecasting model,which usually leads to the poor forecasting effect of the trained load model.Aiming at the above problems,this thesis proposes a combination model to improve the accuracy of load prediction.The first is to build the VAR-LASSO influencing factor selection model.VAR algorithm was used to determine the lag order of the influencing factors,so as to alleviate the influence of historical data of each influencing factor on the current power load value.The LASSO algorithm is adopted to determine the dimensions of influencing factors to ensure that the dimensionality of input data is reduced while the original information is preserved to the maximum extent,which alleviates the problems of heavy computation and easy loss of important information during the dimensionality reduction of high-dimensional data by data dimensionality reduction methods such as cluster analysis and factor analysis,and effectively improves the prediction accuracy of neural network and the interpretability of the model.The second is to construct CEEMDAN-TSO-LSTM prediction model.(1)The CEEMDAN algorithm is used to decompose the original load data for noise reduction,and the sequence with similar complexity is reconstructed based on the perarrangement entropy method,which effectively overcomes the shortcomings of wavelet analysis and other signal decomposition methods that rely on subjective experience,and effectively solves the influence of EEMD algorithm’s modal aliasing phenomenon and white noise interference on the prediction effect.(2)In view of the shortcomings of the LSTM neural grid,such as the difficulty in determining the hyperparameters and the slow convergence rate of the model,the TSO algorithm was proposed to optimize the LSTM parameters to effectively overcome the problems of large randomness of parameter selection and slow convergence rate of the model.(3)Combine CEEMDAN,TSO and LSTM algorithms to form a CEEMDAN-TSO-LSTM combination model,and improve the effect of load prediction from the two aspects of load data decomposition and noise reduction and parameter optimization.The power load data of Yixing City was used for empirical analysis,and the data after determining the influence factors were input into CEEMDAN-TSO-LSTM model for fitting,and the data was compared with BP model,LSTM model,The fitting effect of CEEMDANLSTM model and CEEMDAN-PSO-LSTM model was evaluated by MAE,MPAE,RMSE and goodness of fit.The experimental results show that the CEEMDAN-TSO-LSTM model has a better fitting effect in the case of high correlation of short-term power load in Yixing City.The model can effectively improve the accuracy of load prediction and provide a basis for the power supply department’s power scheduling plan. |
