Power Generation Forecasting of a Solar Photovoltaic Power Plant by a Novel Transfer Learning Technique with Small Solar Radiation and Power Generation Training Data Sets

Deep learning (DL) which is widely used for photovoltaic generation forecasting requires large data to be trained properly, which is not available initially for new power plants. Transfer learning is a powerful technique used in computer vision problems with capability to achieve high performance even when small training data are available. This study applies transfer learning to utilize solar radiation data to train a deep neural network which is fine-tuned on power forecasting using a small dataset for a 400 kWp PV power plant in India. The transfer learning technique can only be used if two problems are similar, therefore, a novel comparison between solar radiation and power generation datasets based on Pearson’s correlation coefficient, is carried out which shows that these two problems are similar. A comparison of the model forecasts before and after applying transfer learning is carried out. The results are validated with real data for the power plant which shows that, the proposed model achieved high performance in forecasting generation with percentage error improved by 1.3% and R-value increased by 3.2% after applying transfer learning. Therefore, this methodology can be used for all types of time series problems as it guarantees a high forecasting accuracy