Location-Specific Optimization of Photovoltaic Forecasting Models Using Fine-Tuning Techniques

Abstract

The global energy landscape is increasingly shaped by renewable energy sources, particularly photovoltaic systems, which are influenced by economic, environmental and geopolitical factors. Technological advances and political incentives have reduced the cost of photovoltaic systems and driven their integration into the energy grid. This trend has been accelerated by energy crises and geopolitical events, underlining the importance of energy security and diversification of energy sources. However, the inherent variability of solar energy, which depends on climatic and natural cycles, poses a major challenge for the management of energy production. In this paper, short-term forecasting models for photovoltaic power plants are developed and applied using a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) system. These models improve grid management by predicting photovoltaic output more accurately, enabling better integration and reducing dependence on balancing power plants, thereby optimising cost efficiency and reliability. A novel approach for fine-tuning the models at different geographical locations is presented, where the pre-trained models are adapted to local datasets to maintain forecast accuracy despite varying climatic conditions. The evaluation results show that the fine-tuned CNN-LSTM model achieves a mean absolute error (MAE) of 242.46 W, a root mean square error (RMSE) of 406.63 W and an R2 value of 0.78, indicating a significant improvement in forecast accuracy over the baseline model. This robust framework demonstrates the practical applicability of the model for real-world photovoltaic power generation forecasting, highlighting the importance of fine-tuning to local conditions.

Keywords

• Deep learning
• energy security
• grid stability
• model fine-tuning
• photovoltaic systems
• renewable energy sources