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

Speaker: Feng-Nan Hwang (Department of Mathematics, National Central University, Taiwan)

Abstract

Developing an accurate and reliable computational tool for traffic flow prediction has always been an active research topic in transportation engineering and planning. In general, the available predictive tools are falls into three categories, i.e., 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) belong to the nonparametric methods, while the Autoregressive integrated moving average (ARIMA) and its variants are 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 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 show that our method outperforms SARIMA and LSTM. This is joint work with Chia-Ming Chang (NCU, Math) and Chia-Hui Chang (NCU, CSIE).