Long-term traffic flow estimation: a hybrid approach using location-based traffic characteristic

Abstract

Traffic speed estimation plays a key role in various situations, ranging from individual's trip planning to urban traffic management. Despite many studies on short-term prediction, there is only a limited number of studies focusing on long-term prediction and only a couple of them does go beyond 24 h. On the contrary, this study presents a novel hybrid architecture using location-based traffic characteristic for traffic speed estimation up to 7 days. In this architecture, the introduced mean filtering estimation (MFE) model and long short-term memory (LSTM) neural network are jointly utilized for minimizing the error for traffic flow estimation. Both MFE and LSTM utilizes the speed data, collected from roadside sensors in Istanbul, of previous weeks that have the same weekday and the same time with target time to be predicted. Results in this study indicate that the use of MFE gives lower error rates for locations with low traffic complexity while LSTM outperforms MFE model for locations with high traffic complexity. Thanks to the introduced MFE and the proposed hybrid architecture, we are able to predict the speed data of a given location with an error of lower than +/- 10 km/h.