TEMPERATURE FORECAST AND SOLAR IRRADIANCY THROUGH ARTIFICIAL NEURAL NETWORKS OF THE LSTM TYPE
Artificial Neural Networks; Computational modeling; Forecast.
Accuracy in the prediction of temperature and solar radiation is a very important factor for the generation of solar photovoltaic energy, especially in short-term predictions. The deep neural networks known as Long Short-Term Memory (LSTM) are models that have shown excellent performance in the work with these environmental variables. Through the application of different architectures based on the LSTM model, this study aims to make the prediction of temperature and solar radiation in the short term, through the use of data obtained by equipment installed on the Tapajós campus of UFOPA, close to the city center, when compared with station installed by INMET, in the countryside, four architectures were executed, with differences in the number of neurons, and ten repetitions each for a time series of surface air temperature with 12053 records, 10299 for training and 1754 for testing the LSTM network . The acquired data were processed using Python language software, where after the data treatment, the network training was performed with later prediction of the time series. The results obtained had their performance evaluated using the MAE metric (Mean Absolut Error), the average of the values found ranged from 0.0420 in the architecture with the lowest number of neurons to 0.0213 in the architecture with the highest number of neurons, where they were performed statistical tests in order to identify the best applied architecture. According to the Kruskall Wallis test, which showed a statistically significant difference between the architectures, and the Dunn test, which indicates which pairs have a statistically significant difference, the architecture with the highest number of neurons per layer showed the best performance. It is expected that for the solar radiation data that will be acquired later, the results will be similar, where the LSTM network tended to follow well the variations of the temperature prediction.