**1. Introduction**

Traffic congestion has been an increasing problem in the busiest urban cities. The main obstacles of traffic congestion are illegal parking, road maintenance, lane closure due to utility work, narrowing roads, accidents, and weather conditions. These incidences lead to traffic bottlenecks, which cause several adverse effects on the number of crashes, the increased cost of travelers and commuters, and fuel consumption. The traffic bottlenecks can often be located at strategic locations in a network, e.g., at off-ramps, on-ramps, and lane drop areas [1]. Practically, there have been various useful definitions of bottlenecks proposed by several authors identifying traffic fluctuations along with connected roadway segments. A bottleneck implies the congestion evolution and queue formation,

15

which consequently disturb travel delay and worsen the urban traffic environment and safety [2]. In our previous work [3], a bottleneck was defined in that the demand and supply are mismatched due to the network structure such as upstream links merging to only one downstream link in each intersection area.

Traffic managemen<sup>t</sup> and control systems have become increasingly important for developing traffic operations' efficiency and safety [4]. To improve traffic management, various sources of data are available via different traffic data collection methods. A series of spot sensors such as inductive loop detectors, remote transportation microwave sensors, and video cameras can provide accurate traffic information. However, these fixed-location sensors can only incorporate the traffic state at specific locations, and these devices might include high maintenance expenditure with frequent malfunction. For urban road networks, sensors need to cover a wide range of areas, especially for bottleneck and upstream/downstream analysis. The high installation and maintenance costs of those sensors are proven to be impractical for many cities.

Compared to the traditional way of using fixed-location sensors, with further advances in technology, using vehicles equipped with Global Positioning System (GPS) sensors as probes to collect traffic data has become popular. Ubiquitous traffic data are available everywhere automatically, and this helps develop Intelligent Transportation Systems (ITSs). The emerging widespread availability of probe data significantly helps to overcome the geographic coverage and spacing restrictions of traditional loop detector data [5]. Using such vehicles can provide real-time information on the traffic conditions along with the entire road network. These GPS-equipped vehicles can collect and output mobility data periodically, including longitude, latitude, speed, vehicle headings, and timestamps.

Traffic condition prediction is one of the primary components of ITSs, and its attention has grown in transportation research. The objective is to provide individual commuters or travelers with accurate traffic information on time. Some of the most promising predictions could improve travel time reliability with the accurate prediction for the same trips compared with the same day of adjacent weeks because time series of traffic data usually have recurrent temporal patterns. For example, similar traffic congestion happens every morning and evening rush hour. Further, this pattern is likely to occur weekly, monthly, and yearly.

This research aims at identifying bottleneck at each intersection and predicting urban network gridlocks. To summarize, the primary contributions of this paper are listed as follows:


The remainder of the paper is organized as follows. Section 2 provides the related works. Section 3 presents the simulation framework. Section 4 proposes the methodology of traffic data processing, the description of bottleneck and gridlock identification, and the study of sample sizes for gridlock detection. Section 5 presents the urban gridlock prediction based on the bottleneck with LSTM. Section 6 discusses the experimental results. Finally, concluding remarks are presented in Section 7.
