Overcoming the Data Gap of Floating Photovoltaic Plants: Synthetic Data Generation and Transfer Learning Techniques
Abstract
The scarcity of data in newly-built photovoltaic (PV) systems, particularly floating PV (FPV) plants, presents challenges for performance estimation and forecasting. This study implements a physics-based model for synthetic data generation and a deep learning (DL)-based transfer learning technique to mitigate the problem of limited data in a new FPV plant. First, a transient physical model is calibrated using limited data from an FPV plant (7, 15, and 30 days) to generate synthetic data used to train a Long Short-Term Memory (LSTM) model. Second, a LSTM model is pre-trained with data from a conventional PV plant and fine-tuned using limited data from the FPV plant. Both approaches are compared and the LSTM model fine-tuned outperforms the model trained with synthetic data. However, the physical model provides acceptable performance even with minimal real data, showing a maximum MAE difference of 0.58 °C. This scale of difference in error translates into less than 0.1 % error in PV efficiency, demonstrating that synthetic data can serve as a reliable alternative when data from PV plants are not available.
Keywords
Synthetic data, transfer learning, floating PV, knowledge transfer, fine-tuning