Study On Load Forecasting For Distribution Area Based On Machine Learning

As a perceptual means to speculate and judge the future load changes,Load forecasting is an important part of the smart grid data platform.The low-voltage distribution station area is the terminal link to realize intelligent and fine management of the power grid,and it can accurately predict the load of the regional distributed station area.It can provide an important basis for distribution capacity configuration,low-voltage distribution network planning,line loss reduction,risk early warning and other work.Different from large scale load prediction at city level and county level,the load prediction in distributed small area is greatly affected by various factors,and the load sequence is non-linear and non-stationary,so the traditional load prediction method is not as effective as the large area load prediction with relatively stable load characteristic curve.In this paper,first of all,this paper analyzes the factors which affects the area load,and then the area characteristic data are kernel principal component analysis(KPCA)dimension reduction after using the fuzzy c-means algorithm(FCM)clustering was divided into different area the training sample,according to the area load characteristics,this paper proposes a machine learning area combined load forecasting model for all kinds of area for load forecasting,The deep residual network model(RESNET)which can prevent gradient dispersion is used to extract the deep characteristic information of station load,and then the combined prediction model is constructed by combining with the gated cyclic network(GRU)which can learn the time correlation of load series.Finally,the model is optimized by using the fine-tuning method of transfer learning network.The model can be applied to small sample area.In this paper,the effectiveness of the combined model is proved by experiments.In the sample clustering,the appropriate cluster number is selected according to the clustering judgment.The experiment proves that the model after reasonable sample division of the platform has higher precision.In the comparison of the prediction effect,the control experiment proves that RESNET-GRU has improved the training speed and precision compared with the single model.In order to avoid the occurrence of negative migration,the MMD difference between the source domain and the target domain data distribution is considered,and the corresponding negative migration threshold is found through the error curve.Finally,the effectiveness of the proposed model is proved by comparing the actual prediction error calculation examples of multiple types of stations.