Resource-Efficient PV Energy Yield Nowcasting with Sky Images: a Hybrid Global Annealing Schedule

The increasing integration of small-scale dispersed photovoltaic (PV) systems into the grid necessitates highly accurate forecasting models that offer fine-grained insight on future PV generation. However, such forecasting models can be difficult to fine-tune under dynamic cloudy conditions. Shallow data-driven models trained with gradient descent optimization methods are prone to local optima entrapment, while deeper architectures require excessive computational resources, often prohibitive for small-scale local PV systems. In this paper, we propose a novel, hybrid optimization algorithm that consists of an adaptive hill-climbing annealing schedule, coupled with a fast local search component, to enhance global search without compromising the rate and speed of convergence. The proposed algorithm is employed for the optimal calibration of a data-driven PV energy yield model, part of a nowcasting framework based on sky images. The primary objective of this paper is to introduce a cost-effective and resource efficient nowcasting framework to facilitate the integration of small-scale local PV systems. Experiments on a 180W PV module prove that the proposed PV energy yield nowcasting framework results in highly accurate and detailed predictions, maintaining the average forecasting error below 9%, regardless of the prevailing sky conditions and without compromising computational efficiency.