The use of models capable of forecasting the production of photovoltaic (PV) energy is essential to guarantee the best possible integration of this energy source into traditional distribution grids. Long Short-Term Memory networks (LSTMs) are commonly used for this purpose, but their use may not be the better option due to their great computational complexity and slower inference and training time. Thus, in this work, we seek to evaluate the use of neural networks MLPs (Multilayer Perceptron), Recurrent Neural Networks (RNNs), and LSTMs, for the forecast of 5 min of photovoltaic energy production. Each iteration of the predictions uses the last 120 min of data collected from the PV system (power, irradiation, and PV cell temperature), measured from 2019 to mid-2022 in Maceió (Brazil). In addition, Bayesian hyperparameters optimization was used to obtain the best of each model and compare them on an equal footing. Results showed that the MLP performs satisfactorily, requiring much less time to train and forecast, indicating that they can be a better option when dealing with a very short-term forecast in specific contexts, for example, in systems with little computational resources.

Research on alternative energy sources has been increasing for the past years due to environmental concerns and the depletion of fossil fuels. Since photovoltaic generation is intermittent, one needs to predict solar incidence to alleviate problems due to demand surges in conventional distribution systems.Many works have used Long Short-Term Memory (LSTMs) to predict generation. However, to minimize computational costs related to retraining and inference, LSTMs might not be optimal. Therefore, in this work, we compare the performance of MLP (Multilayer Perceptron), Recurrent Neural Networks (RNNs), and LSTMs for the task mentioned above. We used the solar radiance measured throughout 2020 in the city of Maceió (Brazil), taking into account periods of 2 hours for training to predict the next 5-minutes. Hyperparameters were fine-tuned using an optimization approach based on Bayesian inference to promote a fair comparison. Results showed that the MLP has satisfactory performance, requiring much less time to train and forecast. Such results can contribute, for example, to improving response time in embedded systems.