Abstract: Traditional annual runoff classification is often divided into wet years,normal years and dry years by formulating quantitative indicators of annual runoff. However,it is difficult to reflect the situation of annual runoff only by dividing the results of annual runoff. This paper introduces the SOM neural network,and divides the annual runoff sequence according to the monthly runoff of the runoff sequence,the average monthly runoff during the year,the uneven coefficient of the year and other indicators,and obtains the results of the annual runoff division with the characteristics of the year. Taking Miyun Reservoir as an example,the annual runoff of Mi-Yun Reservoir is applied to research,and the annual runoff sequence of Miyun Reservoir is divided into 3 categories:wet and normal and dry,and then the results are subdivided into 9 categories. The results indicate that different classification results have different properties in runoff,intra-year distribution,unevenness,concentration,and intra-year change range. The representative years of the 9 sub-categories are similar in runoff in the same major category,and in the runoff distribution,the runoff in the dry season has its own characteristics and different distributions. It shows that the SOM neural network considering multiple indicators can divide the annual runoff sequences in detail.