Taxi destination prediction can grasp the flow direction of taxis and facilitate the taxi scheduling. Most existing prediction methods only use the original features of the trajectory sequence as the input of the prediction model and ignore the spatiotemporal data behind the original features, resulting in the lack of spatiotemporal information. To address the above problems, a Deep Learning method for taxi Destination Prediction (DLDP) is proposed. Firstly, the method uses a sliding window to calculate the high-level features of the trajectory base on speed and turning rate. Secondly, the auto-encoder is used to convert the high-level features into a fixed-length potential spatial representation, so as to obtain the depth features of the trajectory. Finally, the depth features are combined with the original features and used together as an input of the LSTM (Long Short-Term Memory network) for prediction. Experiments show that this method improves the accuracy by 9% and reduces the average distance error by 1km in comparison to the traditional RNN prediction model.