In photovoltaic(PV) systems, PV arrays are often affected by shadows which results in a reduction of generated power. The power-voltage ($P-V$) characteristics of PV array will contain multiple peaks under partial shading condition (PSC), while conventional maximum power tracking (MPPT) techniques are easy to fall into a local optimum. Therefore, this paper designs a novel MPPT algorithm based on transfer reinforcement learning (TRL). The algorithm decomposes the original large-scale search space into several small-scale sub-search spaces, which can effectively improve the global search ability of TRL. Meanwhile, the knowledge transfer is adopted to transfer the optimal knowledge matrix of the old task to the new task, such that the convergence rate of TRL could be improved significantly. Three cases are carried out, e.g., constant temperature and varying solar irradiation, varying temperature and varying solar irradiation, as well as HongKong field test. Simulation results demonstrate that TRL can achieve the fastest MPPT under PSC and lowest power fluctuations in comparison to incremental conductance (INC), genetic algorithm (GA), particle swarm optimization (PSO), artificial bee colony (ABC), cuckoo search algorithm (CSA), teaching-learning based optimization (TLBO), and Q-learning. Finally, dSpace based hardware-in-loop (HIL) test verifies the implementation feasibility of TRL.