The trajectory of bus vehicles from the upstream to the downstream of road in a relatively long period still demonstrated strong uncertainty even on condition of exclusive bus lanes, due to the influence of ahead vehicles, platform capacity, pedestrian crossing and other factors. Time series models which only operate on individual targets in simple scenarios were not efficiently utilized for the impact of uncertainty on the prediction of bus trajectory. The future global trajectory sequence was suggested to be transformed into a high-dimensional time series composed of multiple simple short-term local trajectory sequences, aiming at the above problem. Subsequently, a simple and hierarchical recurrent encoder-decoder architecture was established based on the recurrent neural network to describe the mapping relationship between the local sequence and the globe sequence in the high-dimensional time series. Starting from the current period trajectory sequence, each local sequence of the high-dimensional time series was cyclically predicted until the final future global sequence was obtained. Finally, the two types of architecture were trained and tested by using GPS data of bus trajectory measured on Wensan Road in Hangzhou, China. Results show that the above method outperforms multi-step sequence to sequence method and the hierarchical recurrent encoder-decoder architecture obtains the best prediction results and generalization performance to complex environment.