基于递归框架的高速公路交通流量实时预测方法

实时与准确的断面交通流量预测是实现高速公路智能化管理与控制的基础。高速公路交通流量预测要求对数据噪声进行有效处理，且需要满足实时性需求。然而，少有研究从实时性的角度对高速公路交通流量预测的准确性进行改善。研究了结合自适应卡尔曼滤波与长短时记忆神经网络（long short-term memory,LSTM）自编码器的高速公路交通流量递归预测框架，可以满足智能交通系统的实时性与准确性需求。收集高速公路的交通流量和速度等历史数据，应用卡尔曼滤波方法进行数据平滑，以提高原始数据的可预测性能；引入无监督机器学习算法LSTM自编码器对交通流量的时变特征进行建模，以提高模型的运算效率；考虑到高速公路交通流量预测的实时性需要，进一步提出递归预测框架，用LSTM自编码器的预测值代替卡尔曼滤波值；根据获取的实时数据，执行自适应卡尔曼滤波算法以修正当前的最佳状态值，并将该修正值输入LSTM自编码器进行迭代预测。选取美国明苏尼达双子城高速公路的实测交通数据进行案例分析，结果表明：所提出的高速公路实时交通流量递归预测框架在计算成本与预测精度2个方面具有相对竞争优势，模型预测的平均绝对百分比误差为5.0%，优于...

A Recursive Framework-based Approach for Real-time Traffic Flow Forecasting for Highways

Real-time and accurate traffic flow forecasting is a prerequisite for intelligent management and control of highways, which requires an effective approach for data processing as well as for meeting the real-time requirement. However, few studies have considered the accuracy of traffic flow forecasting for highways from a real-time perspective. Based on this consideration, a recursive framework for traffic flow forecasting is developed combining adaptive Kalman filter (KF) and long short-term memory (LSTM) autoencoder to meet the real-time and accuracy requirements of intelligent transportation systems. Historical data of traffic flow and speed are adopted, and smoothed by a KF method to enhance the prediction accuracy. An unsupervised machine learning algorithm, LSTM autoencoder, is introduced to model the time-varying characteristics of highway traffic flow efficiently. Considering the real-time requirement of traffic flow forecasting for highways, a recursive forecasting framework is proposed. The output of the KF algorithm is replaced by the predicted value of LSTM autoencoder. Based on the real-time data, the adaptive KF algorithm is conducted to correct the current optimal state value. A case study is conducted based on a real-world traffic dataset collected from the Minnesota Twin Cities, USA. Study results show that the recursive framework of forecasting the highway traffic flow proposed in this study has relatively competitive advantages in terms of both computational cost and prediction accuracy. The mean absolute percentage error of prediction is 5.0% (< 7.4% of the combined KF and LSTM autoencoder model); and total training time is 85 s, which is lower than the standard LSTM (101 s).