Research On Enterprise Electricity Migration Prediction Based On Data Distribution Similarity

Electricity forecasting is an important part of power generation planning in the power system.Accurate power supply forecast can not only improve the operation efficiency and level of power supply enterprises,but also has important significance for project investment and management of production enterprises.Due to differences in factors such as industries and scales of enterprises,there are still problems such as insufficient power sample data and low generalization of prediction models in power forecasting.Based on the similarity of enterprise power data distribution,this paper uses machine learning algorithm to achieve accurate prediction of power consumption of enterprises with different power consumption scales through transfer learning.The main research contents of this paper are as follows:First,migration prediction based on similarity of data distribution.According to the scale of electricity consumption,the types of enterprises are divided,and four machine learning models of long and short term memory network,error back propagation,error back propagation and extreme learning machine are built to predict and analyze the daily electricity consumption of enterprises of three scales.The empirical results show that the LSTM model has better performance for the power prediction of enterprises of the same scale.At the same time,in order to solve the problem of insufficient enterprise samples in the target domain,through transfer learning,three similarity measures of KL divergence,JS divergence and maximum mean discrepancy are used to improve the loss function of the LSTM model,and the application of the transfer LSTM model based on MMD distance is proved.to the feasibility of enterprise electricity forecasting.There is no need to re-model the target enterprise again,and a lot of time and labor costs can be saved on the basis of maintaining the prediction accuracy of the model.Finally,by adding a penalty factor λ and improving the model loss function,a transfer LSTM model based on an improved MMD is proposed,which further improves the prediction accuracy of the model and determines the impact of the penalty factor on the model prediction performance.Second,study the influence of different migration factors.The effects of the number of samples,penalty factor λ and seasonal MMD on the transfer LSTM model were studied separately:（1）In the case of a small proportion of training set samples,transfer learning does help to improve the LSTM in the case of insufficient sample proportion.Predictive performance of electricity in large-scale enterprises.（2）The change of the penalty factor λ has a great influence on the prediction performance of the mig ration LSTM model based on the annual MMD.When the penalty factor λ∈[30,35],the prediction performance of the migration LSTM model based on the annual MMD is the best.（3）For the migration strategies of different MMD distances,the prediction effect of the migration LSTM model based on the annual MMD is significantly better than the prediction effect of the migration LSTM model based on the seasonal MMD,and it is more stable for enterprises of different scales.Third,a preliminary study on the migration prediction of enterprises with different electricity consumption scales.The above researches are all conducted under the same scale of electricity consumption.To further explore the effect of the model on the power forecasting of enterprises of different scales,standardize and improve the power data of enterprises in the source domain to reduce the difference in the distribution of power data between enterprises of different scales.The migration LSTM model of MMD realizes the effective prediction of electricity among enterprises of different scales.Under the condition of ensuring high prediction accuracy,it can not only solve the problem of insufficient electricity data collection in urban power grid planning,but also save the reconstruction model training time.In general,the research in this paper aims at the application of transfer learning based on data distribution similarity in enterprise electricity forecasting,and realizes accurate electricity forecasting from enterprises with the same electricity consumption scale and different electricity consumption scales.Management and other tasks provide a reference basis,improve the economic benefits of power companies,and provide a strong guarantee for maintaining my country’s energy security.