Abstract: To address the low optimization efficiency of the Non-dominated Sorting Genetic Algorithm-II (NSGA-II) when applied to multi-objective operation of reservoir systems, an improved NSGA-II for multi-objective optimal operation of reservoir groups is proposed by incorporating a Sobol sequence–random hybrid initialization, a dynamic-parameter differential evolution operator, and an external archive guidance mechanism. A multi-objective joint optimal operation model for the dry season is established and applied to a case study of the six cascade reservoirs on the mainstem of the Ganjiang River. The results show that, compared with the NSGA-II algorithm under normal, dry, and extremely dry year scenarios, the Pareto front obtained by the improved algorithm exhibits better convergence and uniformity, and demonstrates significant stability in extremely dry years; Compared with conventional operation, the scheme maximizing hydropower generation of the improved algorithm increases the power output by 255 million kW·h(10.7%), 189 million kW·h(12.2%), and 92 million kW·h(7.6%), respectively; The scheme maximizing navigation benefits increases the minimum release flow of the last reservoir during the dry season by 253.54 m3/s(42.9%), 247.45 m3/s(57.0%), and 220.22 m3/s(99.2%), respectively; The scheme maximizing ecological benefits reduces the river flow deviation by 0.737(49.1%) in normal years and by 0.384(20.7%) in dry years, whereas in extremely dry years, the river flow deviation of this scheme is slightly higher than that of the conventional operation scheme. The research results can provide a scientific decision-making basis for the multi-objective joint optimal operation of the cascade reservoirs on the mainstem of the Ganjiang River.