Abstract
This paper aims to introduce a novel maximum power point tracking (MPPT) strategy called transfer reinforcement learning (TRL), associated with space decomposition for Photovoltaic (PV) systems under partial shading conditions (PSC). The space decomposition is used for constructing a hierarchical searching space of the control variable, thus the ability of the global search of TRL can be effectively increased. In order to satisfy a real-time MPPT with an ultra-short control cycle, the knowledge transfer is introduced to dramatically accelerate the searching speed of TRL through transferring the optimal knowledge matrices of the previous optimization tasks to a new optimization task. Four case studies are conducted to investigate the advantages of TRL compared with those of traditional incremental conductance (INC) and five other conventional meta-heuristic algorithms. The case studies include a start-up test, step change in solar irradiation with constant temperature, stepwise change in both temperature and solar irradiation, and a daily site profile of temperature and solar irradiation in Hong Kong.
Highlights
In the past decade, a continuous decline in the overall price of photovoltaic (PV) modules can be witnessed around the world, thanks to the advancement of new materials and manufacturing, as well as the ever-growing attention to greenhouse gas emissions [1,2]
Capability of knowledge transfer: Through a positive knowledge transfer from past optimization tasks, the optimal knowledge matrices of the new optimization task can be approximated by transfer reinforcement learning (TRL), this method can efficiently harvest an optimum of high quality; Capability of online learning: TRL can continuously learn new knowledge from interactions with the environment based on reinforcement learning (RL), which can rapidly adapt to maximum power point tracking (MPPT) under different solar irradiation, temperatures, and partial shading conditions (PSC)
Due to the beneficial guidance by knowledge transfer, TRL can significantly alleviate the power fluctuations without a blind/random search. This reveals that, for real-time MPPT, TRL is capable of speedily seeking an optimum of high quality through the space decomposition on the basis of RL and beneficial knowledge transfer
Summary
A continuous decline in the overall price of photovoltaic (PV) modules can be witnessed around the world, thanks to the advancement of new materials and manufacturing, as well as the ever-growing attention to greenhouse gas emissions [1,2]. In reference [21], an ant colony optimization (ACO) combined with a novel strategy of pheromone updating was developed for MPPT, which can effectively improve the speed of tracking, accuracy, stability, and robustness under various weather conditions and different partial shading patterns. All of these meta-heuristic algorithms have two main deficiencies as they are independently utilized for MPPT under various scenarios, as follows:. In comparison with the aforementioned meta-heuristic algorithms, TRL has the following two advantages: Capability of knowledge transfer: Through a positive knowledge transfer from past optimization tasks, the optimal knowledge matrices of the new optimization task can be approximated by TRL, this method can efficiently harvest an optimum of high quality; Capability of online learning: TRL can continuously learn new knowledge from interactions with the environment based on RL, which can rapidly adapt to MPPT under different solar irradiation, temperatures, and PSC
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
