Transfer Learning for Solar Energy Forecasting: A Comparative Study of BGRU and T-BGRU Models

Achieving accurate predictions of solar energy is a critical requirement for reliable operation of modern power grids. In the current work, we introduce a comparative analysis of Bidirectional Gated Recurrent Unit (BGRU) and Transfer Learning-BGRU (T-BGRU) models for short-term solar energy forecasting over the Maharashtra-Mumbai region, with Global Horizontal Irradiance (GHI) as the main predicted variable. By transferring the knowledge, significant improvement was observed concerning Normalized Root Mean Squared Error (nRMSE) and Mean Absolute Error (MAE), Standard Deviation with the T-BGRU model as compared with the standard BGRU model. Perfect MAE of 0.001444 and nRMSE of 0.0155 was achieved using the T-BGRU model. This infers that the T-BGRU offers more accuracy and predictability. Most importantly, the generalization capability of the T-BGRU model proves it to be more than ideal for real-time applications in forecasting solar resources. Thus, these results demonstrate the feasibility of applying transfer learning for enhanced performance of models, especially when training data is limited to certain localities or conditions.