To reveal the response characteristics of reservoir group operation to the non-stationary inflow runoff and achieve adaptive reservoir operation, the reservoir group in the Xijiang River Basin was used as a research object. An evolutionary multi-objective direct policy search algorithm was employed to construct a multi-objective optimization operation model for the reservoir group, and the optimal solutions that balance multiple objectives of power generation, ecology, and water level were obtained. Under both stationary and non-stationary runoff conditions, the sensitivity analysis method was used to re-evaluate the optimal solutions, and further to reveal the sensitivity response characteristics of operation objectives to inflow runoff. The results indicate that there are significant trade-off relationships between the rates of ecological water shortage and water level risk at the Wuzhou station and the average output of the reservoir group in the Xijiang River Basin. The sensitivity of operation objectives to inflow runoff does not show significant differences among different optimal solutions, but it is more responsive to non-stationary runoff. The wet-dry variations in runoff under non-stationary conditions are more pronounced, which to some extent enhances the average output of the cascade reservoir group, while also leading to higher rates of ecological water shortage and water level risks at the Wuzhou station.