Data assimilation with Long Short-Term Memory Networks based on Attention for Highway Traffic Flow Prediction

• Main Authors: Chia-Ming Chang, 張家銘
• Other Authors: Feng-Nan Hwang
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/qcd9w4

碩士 === 國立中央大學 === 數學系 === 107 === Traffic flow prediction an active research topic in transportation engineering. In general, the traffic flow prediction model can be divided into three categories, one is PDE-based simulation, another one is parametric approaches, and the other is non-parametric approaches. There are further hybrid approaches to parametric approaches and non-parametric approaches. In this work, we propose combining the data-driven simulation technique with machine learning tools to decrease prediction error, and use the Kalman Filter (KF) on this basis to achieve the effect of data assimilation. The KF consists of two steps: prediction and correction. In the prediction step, we use the EX method to discretize the LWR model where the MacNicholas model is used as the fundamental relation between the velocity and density. Since the data at the boundary points in the future period are not available. The predicted values obtained by using LSTM with the attention mechanism are used for setting the boundary condition. In the correction step, we use the predicted value obtained by the LSTM with attention mechanism as the observation value, which is used to weight our predicted value and get the correction predicted value. In this study, we use SARIMA and the LSTM Attention as the baseline methods. Autoregressive Integrated Moving Average (ARIMA) is one of the most widely used methods of prediction for university time series data prediction. SARIMA is an extension of ARIMA with seasonal components. Long Short Term Memory networks (LSTMs) is a special kind of RNN that can learn long-term dependencies better than RNN. In addition, adding attention mechanisms can help us better predict the future. We compare them with our proposed method. The experimental results demonstrate that our method outperforms SARIMA and LSTM Attention.