基于深度学习的短时交通流预测研究

针对交通流时间序列，在深度学习的理论框架下，构建基于LSTM-RNN的城市快 速路短时交通流预测模型.根据交通流的时空相关性完成时间序列的重构，依靠模型训练对时 空关联特性进行识别和强化，兼顾精度和时效性确定神经网络深度，完成短时交通流预测模 型搭建.基于TensorFlow 的Keras 完成LSTM-RNN的逐层构建和精细化调参，利用路网实测数 据样本完成算法验证，实现模型本地保存并根据预测精度进行自适应更新.结果表明，本文所 采用的预测算法精度高，受训练样本量的限制较小，实时性、扩展性和实用性均得到有效提高.

Short-term Traffic Flow Prediction Based on Deep Learning

This paper proposes a traffic flow time series prediction model for urban expressway based on LSTMRNN under deep learning framework. First, we refactor the traffic time series with integrated spatial and temporal correlation of traffic flow, making LSTM-RNN obtain and strengthen the ability of data mining. Next, network depth is determined by both precision and timeliness during model designing. And then, we take use of Keras based on TensorFlow to implement LSTM- RNN with building model layer by layer and regulating all the parameters subtly. We validate the model utilizing the measured data from real express way, and implement local model saving and updating regularly according to the prediction accuracy. It is proved that the proposed model performs an accurate prediction for short-term traffic flow which is not restricted by the training sample size to a large extent. Meanwhile, the extensibility and practicability of the model is improved significantly.