Previsão de Geração de Usina Solar Fotovoltaica Utilizando Rede Neural Artificial e Algoritmo PSO

• Ano de defesa: 2023
• Autor(a) principal: Silva, André Wagner de Barros
• Orientador(a): Leão, Ruth Pastôra Saraiva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Link de acesso: http://repositorio.ufc.br/handle/riufc/74946
• topic: Energia solar; Aprendizado de Máquina; Geração de energia; Rede neural artificial; Algoritmos de Enxame; Usinas Fotovoltaica; Forecast; Machine Learning; Regression; Neural Network; Swarm Algorithms; PV Generation

Having the intermittent character and the increasing insertion of solar photovoltaic generation (PV) in the global power plant in recent years, it is imperative to develop even more accurate forecasting models for generation, allowing better planning of the PV plant operation and the entire electrical system. Artificial neural networks have become very popular for presenting promising assertive results in predicting photovoltaic generation and robust model performance. The main contribution of this work is the implementation and comparison of photovoltaic generation hourly forecasting models for a 164 MWp power plant, using types of Focused Time-Delay Neural Networks (FTDNN). Backpropagation, Adam Optimization, Particle Swarm Optimization (PSO), Chaotic PSO (CPSO), and PSO with Aging and Weakening Factor (PSO-AWF) were tested during the network training, although PSO-AWF was also used in the optimization of the architecture parameters for the FTDNN network. For performance comparison purposes, the following reference models were used: multilayer perceptron regression, linear regression, decision tree regression and persistence. Based on different statistical performance metrics, the FTDNN model with PSO-AWF training technique obtained the best result between the algorithms with manual parameter adjustment, with Root Mean Square Error (RMSE) 18.354 MW, Mean Absolute Error (MAE) 13.784 MW, Pearson Correlation Coefficient (R) 80.042 %, Normalized Root Mean Square Error (NRMSE) 14.155%, and Normalized Mean Absolute Error (NMAE) 10.631%. Among the models with automatic adjustment of parameters and forecast for 1h ahead, the FTDNN network that applies PSO-AWF for structuring and Adam for training performed better, with RMSE 18,542 MW, MAE 13,565 MW, R 79,631%, NRMSE 14,300% e NMAE 10,462%. In addition to the forecast for 1h ahead, models for forecasting 3h and 6h ahead (with and without hourly resolution) were also implemented, besides the analysis of the effect of changing the amount of input data and the cross-validation technique on a given automatic model. A small improvement in the result was observed, for the models that provide the forecast for 3h and 6h ahead with hourly resolution, with no improvement being identified for the other models tested with automatic adjustment of parameters.