Accurate short-term traffic flow prediction plays an important role in intelligent transportation systems. Considering the daily fluctuations of traffic flow patterns and the excellent performance of recurrent neural networks in time series analysis, a deep learning algorithm named time slot recurrent neural networks (TS-RNN) is proposed to predict future traffic flow by learning multiple sub-patterns instead of a complete traffic flow pattern. The historical series data were break into groups according to their time slot, and the optimal grouping size was determined by grid searching algorithm. Removing irrelevant historical time series data can effectively reduce noise and interference during learning, so that each independent sub-mode only needs to be trained on highly correlated data and learns fewer but more accurate feature representations. To the best of our knowledge, this is the first time that a piecewise pattern has been used in time series forecasting. The comparison results with six traffic flow prediction models suggest that the performance of proposed model is far superior than those of comparison models due to the more powerful and accurate time series pattern learning and representing ability. Furthermore, small-scale models can also make full use of computing resources, improve the parallel computing capabilities and achieve real-time online prediction.