Divide and Conquer Q-Learning (DCQL) algorithm based Photovoltaic (PV) array reconfiguration scheme for alleviating the partial shading influence. (1st January 2023)

• Record Type: Journal Article
• Authors: Gao, Xin; Deng, Fang; Wu, Guoqiang; Pan, Qiufeng; Zheng, Chengyu; Wang, Wei; Cai, Tianwang; Jiang, Liyuan
• Is Part Of: Solar energy. Volume 249(2022)
• Page Start: 21
• Page End: 39
• Publication Date: 2023-01-01
• Subjects: Photovoltaic array reconfiguration -- Reinforcement learning -- Q-learning algorithm -- Divide and conquer principle -- Partial shadings; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2022.09.005
• Languages: English
• ISSNs: 0038-092X

Abstract

Abstract: Nonuniform irradiance on the Photovoltaic (PV) array would cause mismatch problem and decrease the generated power. To relieve the negative effects of partial shading, reconfiguring the interconnection of the PV array is a favorable solution. However, the current reconfiguration techniques are usually limited by different corresponding factors, e.g., not suit for various instantaneous shade cases, the shortage of experts' knowledge or large amounts of data, easy coverage to local optima, and the strict assumptions for application. Therefore, in this paper a unsupervised adaptive PV array reconfiguration technique is proposed with a Divide and Conquer Q-Learning (DCQL) based array reconfiguration scheme designed to decrease the influence of partial shades. To testify the performance of the proposed PV array reconfiguration scheme, mathematic theoretical evaluation and electric characteristic analyses are made and compared with those obtained by the conventional Total-Cross-Tied (TCT) arrangement, SuDoKu connection, Genetic Algorithm (GA) based scheme, as well as some lately proposed Grasshopper Optimization Algorithm (GOA), Harris Hawks Optimizer (HHO), and Lo Shu (LS) technique based reconfiguration methods. Besides, additional estimation indices, including mismatch loss, fill factor, percentage power loss and percentage power enhancement are calculated with substantial analyses and comparisons. Subsequently, the application of the proposed scheme in real-time conditions is evaluated, with energy saving and income analysis estimated in comparison with other eminent challenger reconfiguration schemes. At last, a comprehensive qualitative comparison assessment, with respect to dispersion ability under partial shading conditions (PSC), sensor requirement, wiring complexity and capital investment is carried out and proves the versatility of proposed methodology. Highlights: States, actions and rewards are designed in the Q-Learning based reconfiguration. The combination of divide and conquer principle fastens the decision making process. The proposed reconfiguration is estimated better performance through variety contrasts.