Redes neurais recorrentes e XGBoost aplicados à previsão de radiação solar no horizonte de curto prazo

• Autor(a) principal: Heinen, Eduarth Dapper
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/11828
• topic: Redes neurais artificiais; Rede Neural Recorrente; Radiação solar; Previsão; Seleção de atributos; Modelos de Persistência; Índice de Céu Claro; Índice de Claridade; Artificial neural networks; Recurrent Neural Networks; Solar radiation; Forecasts; Feature selection; Persistence models; Clear Sky Index; Clearness Index; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO

Precise estimates of energy production and consumption are essential to promote the integration of renewable energy sources to the electrical grid, guiding power supply balance through cycles and fluctuations inherent to these resources. In this scenario, many experiments have developed solar radiation forecasting methods, but not so many of those apply Recurrent Neural Networks and their potential of modeling time series. This dissertation reviews some of the studies and outlines an experiment comparing Recurrent Neural Networks, XGBoost and persistence of Clear Sky and Clearness indexes in solar irradiance forecasts. Model precision is verified in minute observations from Denver and Las Vegas, re-sampled in resolutions of 5 minutes, 30 minutes, 1 hour and daily mean. XGBoost gives the best results in all forecast horizons, with nRMSE between 12.9% at 5 minute resolutions, and 21.2% at daily mean. Persistence models for Clear Sky and Clearness indexes show comparable precision, with nRMSE from 14.2% to 22.5% in forecasts up to 1 hour ahead. RNNs outperform persistence in forecasts of daily mean irradiance, reaching nRMSE 21.3%.