Large-scale power systems contain a large number of generating units, and many factors need to be considered during their operation. Unit commitment optimization is a multi-objective and multi-constrained nonlinear large-scale optimization problem. Existing methods to solve this have many shortcomings. The artificial fish swarm algorithm has good performance in solving nonlinear optimization problems, but it has disadvantages such as low optimization efficiency and possibly falling into local extremes. In order to overcome these deficiencies, a modified artificial fish swarm algorithm is proposed. The algorithm introduces variable vision, adjusts the move strategy of artificial fish and combines the mutation operation in a genetic algorithm. A multi-objective optimization model considering both economy and environmental protection is constructed. A phased optimization method is adopted to solve the problem of overlong calculation time caused by the increase of unit scale. The modified algorithm is applied to the unit status arrangement phase, which determines the statuses of units. Then the mixed integer programming method is used to distribute the load. Simulation experiments are carried out for a large-scale power grid unit optimization with up to 1000 units. The experimental results show that the convergence and global search ability of the modified optimization algorithm are improved, the calculation time of large-scale unit commitment is greatly shortened, and ideal results under multi-objective conditions are also obtained. This verifies the effectiveness of the proposed method. This work is supported by the National Natural Science Foundation of China (No. 51677072) and the Fundamental Research Funds for the Central Universities (No. 2020MS120).