Estimation and prediction of the OD matrix in uncongested urban road network based on traffic flows using deep learning

Abstract

In this article, we propose a new method for OD (Origin–Destination)​ matrix prediction based on traffic data using deep learning. The input values of the developed model were determined based on data on the structure of the road network, origin and destination points of trips, as well as data on traffic intensity on road network sections recorded by video-sensing devices. The advantage of the method is that the complex process of data acquisition and processing is not required for the estimation and prediction of the matrix. Historical data and the iterative method of estimating a prior OD matrix were used only to generate training sequences for the neural network. The proposed method using deep learning neural networks with the long short-term memory (LSTM) or autoencoders layers (DLNa — deep learning networks with autoencoders) is characterized by relatively high accuracy and resistance to temporary missing data from several measurement points located in the urban road network. The case study was conducted for a network of a medium-sized city in Poland. The results show (average MAPE = 7.18% (LSTM), 6.80% (DLNa)) that the proposed method can have a practical implementation in real-time dynamic traffic assignment (DTA) systems for ITS applications. The proposed method of short-term forecasting the OD matrix does not require questionnaire research or detailed information on spatial development. Therefore, it is not as expensive and time-consuming as the methods based on these data.