Short-term solar forecasting based on sky images to enable higher PV generation in remote electricity networks

• Main Authors: Schmidt Thomas, Calais Martina, Roy Eric, Burton Ashton, Heinemann Detlev, Kilper Thilo, Carter Craig
• Format: Article
• Language: English
• Published: EDP Sciences 2017-01-01
• Series: Renewable Energy and Environmental Sustainability
• Online Access: https://www.rees-journal.org/articles/rees/full_html/2017/01/rees170028s/rees170028s.html

The integration of a high share of photovoltaic (PV) power generation in remote electricity networks is often limited by the networks’ capabilities to accommodate PV power fluctuations caused by passing clouds. Increasing the share of PV penetration in such networks is accompanied by an increased effort to achieve integration. In the absence of solar forecasting, sufficient spinning reserve must always be provided to cover unforeseen reductions. The expected ramp rates are magnified in small and centralised PV systems and can be in the order of a few seconds. In this study, we investigate the use of a low-cost sky camera for very short-term solar forecasting. Almost 2 months of sky camera data have been recorded in Perth, Western Australia and processed for to provide high-resolution irradiance forecasts based on visible sky images. For performance validation, the capability to provide reliable forecasts under constant clear sky conditions is investigated. During these times, PV generation is expected to be high and reliable, which provides an opportunity to reduce the online spinning reserve often enabling power station operation with one less operating diesel generation. For networks with disconnected diesel generators, we assume that clouds that could reduce the PV generation output have to be predicted at least 2 min before their arrival to have enough time for a diesel generator to start and synchronize with the grid. Therefore, we define an irradiance threshold discriminating between the persistent state of constant clear sky (stays clear) and the non-persistent state (cloud shading event) based on a 2–5 min time horizon. In a binary evaluation, we achieve an overall accuracy of 97% correct forecasts and low 3% false alarms of cloud events indicating a high potential for fuel savings. Focusing on the rare (2% of the time) but more critical non-persistent conditions, we found 8 out of 84 cloud events have not been predicted in advance. Reasons for erroneous forecasts and suggestions for model improvements are provided.