Aprendizado de máquina aplicado à predição de potência de geração distribuída na rede de distribuição de média tensão

Abstract

This paper presents a model, which uses Machine Learning techniques, capable of estimating the available power capacity for insertion of Distributed Generation (DG) in any point of the electric power distribution network in Ceará. Algorithms were developed based on four different machine learning models, Random Forest, Support Vector Machine, Artificial Neural Networks and XGBoost. The algorithms were tested and from the application of performance metrics the XGBoost model was selected. The input data of the developed algorithm were obtained from information provided by the company Enel Distribuição Ceará regarding technical studies carried out by the HV / LV Network Planning area and operational data from the distributor's concession electric network. Given the great demand for DG connections in recent years, the contribution of the work is to provide the system user with transparent information about the available capacity of the electrical network for DG connection without violating operational restrictions. In addition to presenting the predictive diagnosis of the network's capacity for connecting DGs, the work contributes to the better handling of network access requests for both accessor and distributor. The accuracy rate of the prediction given by the algorithm was evaluated for different classes of customers predominant in feeder: residential, commercial, industrial and rural. The highest hit rate was found in industrial feeders and the lowest in rural feeders.