Previsão de geração fotovoltaica utilizando o método Cascade Forward Back Propagation de redes neurais artificiais

Abstract

The work presents the study of a Artificial Neural Network (ANN) called CFBP applied for the implementation of predicting 1 hour ahead output power of PhotoVoltaic (PV) system. The intention of the research was the implementation of a model capable of improve reliable forecasts due to the growing scenario of PV source, that presents high flashing in the production level, making this kind of information essential for a great management of the energy system and for growth of this source in the country’s electric matrix. The proposed RNA features: CFBP architecture, which number of inputs, hidden layers, neurons and learning rate were adjusted according to the data provided for ANN’s training and validation, the learning mechanism applied is backpropagation and the sigmoid activation function. The training and validation method applied is accomplished using the output power data from a residential PV plant, which were collected through the monitoring system of the PV inverter and sized for the ANN implementation. For the purpose of comparing the accuracy of the applied model, an MLP ANN was implemented and had its performance analyzed. The both models have been applied in two distinct conditions in relation to the processing of data, the fist considering the 24 hours a day and the second considering only the system PV operation time course. Metrics were implemented at the end of the work to analyze the error presented by the models. The CFBP architecture implemented for the data scenario without the zeros, for example, obtained the following results: RMSE= 54.14W, R2= 99.97% and MAE= 43.41 W, becoming the ANN with the best performance presented in the study.