Solar Energy Forecasting Using Neural Networks: Insights From Real Systems in the Amazon Region
Abstract
This work presents a neural network-based methodology to predict the daily energy output of photovoltaic (PV) systems by using meteorological and geographical information. The study relies on a dataset of 1,444 real operational records collected from two different clients in the state of Pará, Brazil, a region with high solar potential and significant seasonal variability. The input variables include solar radiation, ambient temperature, wind speed, latitude, longitude, and day of the year, which are used to train a multilayer perceptron (MLP) with two hidden layers containing 100 and 50 neurons, respectively. The model was trained and validated separately for each PV system, allowing for a comparison of performance under different local conditions. Results show that the neural network achieved satisfactory predictive accuracy for Client 1 (R² = 0.68), successfully capturing seasonal trends and the influence of environmental parameters. However, for Client 2, the performance was lower (R² = 0.49), likely due to system anomalies such as inverter malfunctions or operational irregularities, which introduced unexpected deviations in the dataset. These findings highlight both the potential and the limitations of data-driven models when dealing with realworld distributed generation systems. Despite discrepancies, the proposed approach demonstrates that ANNs can provide valuable insights for short-term solar energy forecasting in tropical regions, supporting the integration of PV systems into the grid.
Keywords
photovoltaic systems, energy prediction, neural networks, machine learning, solar forecasting