Generation degradation due to soiling represents a critical barrier to the efficiency of solar photovoltaic systems and results in significant revenue losses for utilities. In this regard, image-regression techniques may enable precise quantification of power output reduction, which is critical for real-time monitoring and automated control in operational photovoltaic systems. However, constructing effective convolutional neural network (CNN) models poses a major challenge for neural architecture search (NAS) methods due to the vast and complex search space. In this context, we present a Rainbow Deep Q-Network-based NAS (RainbowDQN-NAS) framework that integrates Double Deep Q-Network, prioritized experience replay, multi-step learning, distributional value learning, task-specific reward formulations, and Polyak-averaged target updates to achieve superior predictive accuracy, stable training, fast convergence, and satisfactory computation time. We evaluated our framework on the large-scale DeepSolarEye dataset, which encompasses diverse practical soiling conditions. The numerical findings demonstrate that our proposed methodology surpasses contemporary NAS approaches in terms of prediction accuracy with satisfactory computational effectiveness. Thus, our proposed approach facilitates the acquisition of direct image-based solar panel performance information via sensor-free diagnostics to support the efficient operation of photovoltaic power plants. • Sensor-free prediction of soiling-induced power loss using solar panel images. • Reinforcement learning–based neural architecture search. • Image regression framework for scalable solar panel performance monitoring.
Abid et al. (Tue,) studied this question.