In terms of the instability of inflow and the year-end water level setting in cross-year long term optimal scheduling of large-scale hydropower cascade stations, the essay proposes the conception of the inflow at risk. In order to obtain the maximum generation benefit, the author establishes a mid-long term optimal scheduling model. The power generated, together with the water level of the reservoir in the first year, is regarded as basic statistics. The multi-scene scheduling scheme of the second year is regarded as a compensation for the relevant decisions of the first year. Therefore, the existing scheduling can be optimized. The conclusions are as follows: considering the benefit and risk of power generation, the model is designed as a unified scheduling model. The risk preference coefficient is introduced into the model. With Benders decomposition algorithm, the efficiency of the calculation is improved. The model can adjust the risk preference coefficient according to the expectation of the decision-maker to obtain balanced combinations of various returns and risks. At the same time, the model can reduce the risk resulted by the price fluctuation of water and electricity and the optimized model can adapt to different scenarios. The calculation results show that the proposed model can significantly improve the benefit of power generation of hydropower stations as well as reduce the wasted water of reservoirs.