The prediction of the short-term passenger flow on a bus route plays a key role in the daily dynamic bus dispatching system． A combined forecasting model based on K-means clustering algorithm was constructed through analyzing passenger flow characteristics． The historical short-term passenger demand data was divided into different clusters by using K-means algorithm according to the similarity of time-varying demand． Each cluster of the passen- ger demand was predicted individually by using Least Squares Support Vector Machine (LSSVM)，Back Propaga- tion Neural Network (BPNN) and Auto-regressive Moving Average (ARMA)． The parameters of LSSVM and BPNN were optimized by the genetic algorithms． The weather effects on passenger flow were also considered． The combination models were formed by a combination of the best prediction models for each cluster． The passenger flow data of Route 104 in Changsha of China was used for the case studies． The results show that the model combinations depend on the differences of the time-varying passenger flow． The K-means clustering method has the ability to classify the time-varying passenger flow data in different periods，which is conducive to improving the prediction performance． The K-means clustering combination models are a promising tool to predict the short-term passenger flow on a bus route，especially when considering the weather effects．