Abstract: In the prediction of the taxi destination, the traditional Markov prediction method relies only on the first 2 to 3 GPS points, and does not apply to trajectories that have very long dependencies. In order to solve the long-term dependencies, this paper uses Recurrent Neural Network（RNN） to predict the taxi destination, this is because the multiple hidden layers of RNN can store this dependencies. However, with the increasing amount of data, the hidden layers of RNN is very sensitive to small perturbations and the perturbations will be exponentially enlarge in the latter part of prediction, reducing the prediction accuracy. In order to improve the prediction accuracy of taxi destination and reduce the training time, this paper applies SDZ to RNN, and proposes a new taxi destination prediction method based on SDZ-RNN（SRTDP）. SDZ can not only improve the robustness of SRTDP, but also reduce the training time by adopting partial update instead of full update. Experiments show that SRTDP is superior to RNN prediction method in accuracy and speed, the prediction accuracy is improved by 12%, and the training completion time is reduced by 7%.

Key words: taxi destination prediction, recurrent neural networks, SRTDP method, prediction accuracy

摘要： 在预测出租车目的地时，传统的马尔科夫预测方法仅仅依赖于前面2到3个GPS点，对于那种具有很长依赖关系的轨迹并不适用。为了解决这种长期依赖关系，采用循环神经网络（RNN）进行出租车目的地预测，因为RNN的多个隐藏层能够存储这种依赖关系。但是随着数据量的增大，RNN的隐藏层对较小的扰动变得十分敏感，较小的扰动就会被指数级放大，最终导致预测准确率降低。为了提高预测准确率，同时缩短训练时间，将SDZ应用到RNN中，提出一种基于SDZ-RNN的出租车目的地预测方法（SRTDP）。SDZ不但能够提高SRTDP的鲁棒性，而且SDZ采用局部更新而不是全部更新的方式，降低了训练时间。实验表明，SRTDP在精度和速度上都优于RNN预测方法，预测准确率提高了12%，训练完成时间降低了7%。

关键词: 出租车目的地预测, 循环神经网络, SRTDP方法, 预测准确率