In order to identify photovoltaic model parameters accurately and reliably, a Modified Backtracking Search Algorithm(MBSA) was proposed. In the algorithm, firstly, some individuals were selected to learn the current population and historical population information at the same time, and the others were made to learn from the best individual in the current population and stay away from the worst solution, so as to maintain the population diversity and improve the convergence speed. Then, the performances of individuals in the population were quantified by the probability. On this basis, the individuals were able to adaptively select different evolution strategies to balance the exploration and exploitation capabilities. Finally, an elite strategy based on chaotic local search was used to further improve the quality of the population. The proposed algorithm was tested on different photovoltaic models such as single diode, double diode, and photovoltaic module. Experimental results show that the convergence speed and parameter identification accuracy of Backtracking Search Algorithm(BSA) are significantly improved by the proposed strategies. Eight advanced algorithms such as Logistic Chaotic JAYA(LCJAYA) algorithm and Multiple Learning BSA(MLBSA) were compared with the proposed algorithm. Experimental results show that the robustness of the proposed algorithm is the best among these algorithms, and the identification accuracy of the proposed algorithm is better than those of JAYA, LCJAYA, Improved JAYA(IJAYA) and Teaching-Learning-Based Optimization(TLBO) algorithms on both single diode and double diode models, and the proposed algorithm outperforms JAYA, LCJAYA, IJAYA and TLBO algorithms on photovoltaic module model in identification accuracy. Under different illumination conditions and at different temperatures, the manufacturer real data on three photovoltaic modules:thin-film, mono-crystalline and multi-crystalline were used for the actual measurement test, and the results predicted by the proposed algorithm were consistent with the actual situations in the test. Simulation results show that the proposed algorithm is accurate and stable on photovoltaic model parameter identification.