Pareto optical flow solutions for ensemble, satellite-based forecasts of irradiance and PV

 

The sustainable electric grid of the future will rely on comprehensive measurements and forecasts of its millions of components. For PV, such high-resolution forecasts will benefit from the near real-time data and detail provided by satellite observations. Nowcasting or forecasting into the future from detailed satellite images based on e.g. root mean squared error optimization, (e.g. machine learning), however inevitably promotes smoothing and removes detail from the forecast, calling into question the definition of forecast quality. Probabilistic forecasting in the form of ensemble solutions offers an answer, allowing for the detail from satellite images without the expectation of deterministic pixel point accuracy. (Ensemble numerical weather predictions exist on high-resolution grids, but also present smoothed predictions of clouds). 

This study creates ensemble forecasts using a new version of the optical-flow-based nowcasting system presented in past sessions that solves for the global cloud motion using Taylor-approximated streamlines. The optimized flow field is physically constrained through a combination of mass and angular momentum conservation. The errors for the motion of different structures in the image are discerned as secondary objectives to the overall optimization. The optical flow algorithm uses ant-colony, multi-objective optimization, following many solutions before arriving at a Pareto optimum. 

The experiments test the viability of using other cloud motion solutions in the Pareto front to generate ensemble forecasts of the cloud cover and subsequently of irradiance maps and regional PV power. Where regions are fully cloudy or clear, the ensemble solutions should be uniform, while moving regions of variable cloudiness aim for realistic ensemble distributions. The reliability and potential of such forecasts are evaluated and compared to numerical ensemble weather predictions using continual ranked probability scores (CRPS).