On the use of sky images for intra-hour solar forecasting benchmarking: Comparison of indirect and direct approaches

The transient stability of the grid is challenged by short-term photovoltaic output fluctuations, which are mainly caused by local clouds. To address this issue, intra-hour solar forecasting has been widely adopted. Sky images have been proved as promising sources to produce intra-hour solar forecasts. To incorporate with cloud dynamics, sky images are typically embedded into solar forecasting models either indirectly or directly. While the performance of these methods varies across different forecasting environments, a detailed analysis on indirect and direct approaches have not been investigated yet. In this research, we conduct a comprehensive study on the performance of 7 commonly-used sky image-based solar forecasting approaches, including four indirect and three direct models. A total of 72 forecasting settings are established to evaluate the performance of these models. Three critical parameters are specially considered, namely image resolution, image sequence length, and forecast horizon. Results show that among these forecasting models, the stacking ensemble learning and the convolutional neural network + long short-term memory network model typically show the best forecasting performance for indirect and direct workflows, respectively. Compared with the direct approaches, the indirect approaches advance at detecting ramp events with an average the ramp score of 21.65 W/(m ²×min). The direct approaches, on the other hand, outperform the indirect approaches on forecasting accuracy with an average forecast skill of 24.62%. The results of this work can be used as a general guideline for intra-hour solar forecasting benchmark selection.