Photovoltaic Energy Generation Forecasting: Attention-based Network Using Sky Images

Meeting energy demand with renewable energy sources is rapidly increasing globally. However, renewable energy generation can be affected by external factors. In particular, photovoltaic energy generation is an unstable process due to reasons such as weather conditions, and variability in sunlight. Therefore, it is important to forecast it in advance. This study offers a system that will support energy management by forecasting photovoltaic energy generation. Using historical sky images and photovoltaic energy generation data, energy to be generated in the future was estimated. In the study, historical sky images were processed with the VGG-16 model. Attention mechanisms were used to increase forecasting success. By combining the features obtained from sky images with historical photovoltaic energy generation data processed by LSTM, the generation value after 15 minutes is forecasted. Experimental results show that adding the attention mechanism increases success. In addition, adding features extracted from photovoltaic energy generation data to the forecasting model positively affected the success.