Traffic time series prediction is becoming increasingly important in many real applications, ranging from resource scheduling of company, make decisions of government, abnormal data detection and so on. History time series describes the overall variation of data, such as upward, downward or steady trend. In reality, due to the many factors affecting the value of data in traffic fields, so it is difficult to consider the comprehensive factors and use an appropriate model to correctly represent it. In this paper, We use a novel hybrid network called TreNet to predict the time series value. TreNet is target data drive and does not require any prior knowledge in practice. It uses convolutional neural networks (CNN) to extract local salience features from adjacent raw data and uses long-short term memory networks (LSTM) to capture long-range memory. On the other hand, it is the first time to apply TreNet on traffic field and use it to predict future value. Experiments demonstrate that this method can capture useful information to enhance the prediction performance. In seven traffic datasets, TreNet achieves the best performance in six datasets compared with other methods. The experimental results show that this kind of hybrid neural networks architecture is suitable for many different types of traffic time series prediction problems.