Reinforcement learning-based photovoltaic cleaning optimization under site-specific soiling conditions: a data-driven simulation study

Soiling of photovoltaic modules is a significant challenge in regions where dust accumulation has a considerable effect on the performance of the modules and causes the necessity of frequent cleaning. This paper discusses how a reinforcement learning method could be applied to find the optimal moments of time when cleaning should occur considering site-specific soiling conditions. The investigation is based on open-source data collected at a clean reference and a naturally soiled photovoltaic modules during six months in Laghouat, Algeria, and consists of 4,368 hourly observations. A Proximal Policy Optimization (PPO) algorithm was tested and compared to the traditional threshold strategy considering the same environmental conditions. During the 28-day test period, the threshold strategy resulted in 4.6% residual soiling loss and five cleaning episodes, while PPO algorithm decreased the residual soiling loss to 1.9% with just three cleaning episodes. Therefore, the residual soiling loss decreased by 58.7%, while the number of cleaning episodes decreased by 40%. Additionally, the economic sensitivity analysis showed that the same advantage of PPO strategy persists for different combinations of electricity tariff and cleaning cost (favorable, base, unfavorable). Moreover, PPO strategy showed lower residual soiling losses for different soiling severities (low-, base-, and high-severity soiling). Besides, 1,000 block-bootstrap Monte Carlo simulations were performed which gave non-overlapping 95% confidence intervals of the net economic benefit for both strategies. These findings suggest that PPO algorithm could be used for making better cleaning decisions of PV modules considering changing soiling conditions.

Hali tarjima qilinmagan