Precise Time-space Load Characteristic Analysis Based On Smart Meter Data Analysis

Power industry is one of the basic industries in the whole country,so that the importance of power industry is self-evident.Electrical energy cannot be stored in large quantities which requires that the production and demand of electric power are approximately at the same time.Therefore,accurate forecasting of the load is necessary.With the emergence and development of active distribution network,the smart meters installed in households is becoming more and more popular,and the data collected by the smart meters is more and more fine,which results in the increase of load data scale.Moreover,the emergence and development of distributed photovoltaic power generation and distributed wind power,and their access and consumption,have all put forward higher requirements for the accuracy of load forecasting.A new feature selection method based on random forest for short-term load forecasting is proposed for forecasting demand of the city load.Firstly,243 related features are extracted from the historical load data and the time information of the prediction point to form the original feature set.Secondly,the original feature set is used to train a random forest as the original model.After the training process,the prediction error of the original model on test set is recorded,and the permutation importance value of each feature can be obtained.Then,the improved sequential backward search method is used to select the optimal forecasting feature subset according to the PI value of each feature.Finally,the optimal forecasting feature subset is used to train a new random forest model as the final prediction model.The experiments show that the prediction accuracy of random forest trained by the optimal forecasting feature subset is higher than the original model and comparative models based on support vector regression and artificial neural network.As it is benefit from the research conclusion that random forest can accurately analyze the importance of all the features in the original feature space,when it needs to forecast the load of the distribution network which contains a large number of residents,the importance of features can be used as the basis for power consumer clustering.Firstly,data pre-processing is done for the original smart meter load data.Secondly,338 related features are extracted from the load data which have been processed and the time information of the prediction point to form the original feature set.Then,use the original feature set as input vector to train a random forest for each power user.After the training process,the permutation importance value of each feature can be obtained,and then the k-means algorithm is used to cluster all the users.Finally,feature selection is conducted for each cluster to select out the optimal feature subset to train a new random forest for prediction,and the prediction results of all clusters are summed up as the final load forecasting results.The experiments show that the prediction accuracy of the model with cluster analysis is higher than that out of cluster analysis.