DATE2020-10-29 16:10-17:00


SPEAKER黃楓南 教授(國立中央大學數學系

TITLE Data Assimilation Technique with Long Short Term Memory Networks for Highway Traffic Flow Prediction

ABSTRACT The development of accurate and reliable computational traffic flow prediction tools has always been an active research topic in transportation engineering and planning. Generally, the available predictive tools are divided into three categories, namely parametric methods, nonparametric methods, and PDE-based simulations. In particular, the machine learning methods (such as the k-nearest neighbor (K-NN) method and the long short term memory networks (LSTM)) are the nonparametric techniques, and the autoregressive integrated moving average (ARIMA) and its variants are one of the most representative parametric methods. In this work, we propose the data assimilation technique with the long short term memory networks (LSTM) for predicting the highway traffic flows. The proposed method is developed based on the framework of the Karman filtering (KF) algorithm, which consists of two key components: the prediction step and the correction step. The predicted value is obtained by performing numerical simulation and then corrected by Karman filtering with real data. As the numerical simulator, which is a kernel component of the predictive tool, we use an explicit Godunov’s method to discretize the Lighthill-Whitham-Richards model, where the MacNicholas formulation is used as the fundamental relation between the velocity and density. Since the data at the upstream boundary point in the future period is not available. The pseudo-predicted values obtained by using LSTM are used for setting boundary conditions. In this study, we use Seasonal ARIMA (SARIMA), LSTM methods as baseline methods and compare them with our proposed method. The numerical results based on the real traffic for the Hsuehshan Tunnel extracted from the Freeway Bureau Ministry of Transportation and Communications (MOTC) database in Taiwan show that our method outperforms SARIMA and LSTM as well as the classical KF method.