Machine learning for improving PV energy prediction to develop green and sustainable airports in Romania

Photovoltaic (PV) power generation prediction is a key factor in increasing the adoption of renewable energy solutions, especially in energy-intensive infrastructures such as airports. Given their high energy consumption and the increasing pressure to comply with European and national sustainability regulations, the integration of photovoltaics can significantly reduce both the carbon footprint and operational costs. This study explores the potential of neural models, including deep neural networks (DNNs), standard recurrent neural networks (SRNNs), long-term memory networks (LSTMs), gate recurrent units (GRUs), and convolutional neural networks (CNNs), to achieve accurate PV power generation forecasting. After evaluating 20,401 unique model configurations, CNN proved to be the most efficient architecture, obtaining a coefficient of determination (R 2 ) = 0.983 a mean absolute error (MAE) of 13.7, outperforming GRU and the other analysed architectures. The results show that accurate prediction models allow for more efficient energy consumption planning and better integration with the electrical grid within airports, facilitating the transition to green and sustainable airport infrastructures in Romania. The implementation of photovoltaic systems not only reduces dependence on fossil fuel energy, but also contributes to financial optimization by lowering energy costs and the possibility of capitalizing on surplus production. Moreover, prediction based on machine learning allows for real-time energy management, improving operational efficiency and synchronization between production and consumption. This research can be used by professionals or teachers even without an in-depth knowledge of neural networks, to build and implement prediction systems in the field of renewable energy. At the time of this study, there were no previous scientific works that provided such an applied analysis for Romania.