Deep Learning Models to Estimate and Predict the Solar Irradiation in Brazil
摘要
Solar irradiation is the backbone of photovoltaic power technologies and its quantization allows to optimize energy generation. However, solar irradiation can be difficult to detect, mostly due to the design and disposition of sensors, as well as their high cost. To address this limitation, this paper proposes a deep neural network-based model to estimate global solar irradiation by only relying on weather data, focusing on applications targeting the Brazilian territory. The model uses a deep neural network trained with data from the Brazilian National Institute of Meteorology (INMET), which includes 606 nationwide weather stations and over 39 million hourly records of meteorological variables cataloged from years 2010 to 2022. Thus, in this paper i) a deep neural network is used to estimate irradiation, and ii) a long short-term memory is used to predict solar irradiation considering different time granularities: 5 min, 30 min, 6 h, and 1 day. The results show a small error between the measured irradiation data and the calculated results with regard to the following six meteorological variables: time, temperature, relative humidity, wind speed, precipitation, and atmospheric pressure. Moreover, experimental validations conducted using a weather station set up by the authors demonstrate that the proposed models can accurately predict solar irradiation. Thus, the developed model stands as a promising approach for applications within the Brazilian perspective, improving the efficiency and reliability of solar energy generation.