TØIRoads: A Road Data Model Generation Tool

Modelling and predicting traffic is fundamental for the administration of a city and its surroundings as it directly impacts the economy and everyday lives of the population [11, 2, 9, 1]. While works on traffic prediction tend to focus on short-term prediction (e.g., 12 steps of 5 minutes in the future) [14, 18, 5], a high-level view of the main properties of a road network is crucial for long-term city planning. For instance, one may want to plan for the increase of road capacity by enlarging road segments or to plan for the development of new links in the road network if a systematic congestion scenario is foreseen due to an increase of the population or significant changes in traffic flow resulting from the construction of new points of interest in the city.

The solutions to the high-level setting are classically explored in the context of agent-based traffic simulators [16]. However, agent-based traffic simulators, such as MATSIM [10], are co-evolutionary tools which inherently cannot be parallelized. The complete configuration of features and parameters used in the real world traffic simulators include the specification of a number of agent-based traffic routines. This complexity makes their format hard to use in machine learning approaches, which are parallelizable. Machine learning approaches, in particular those based on graph neural networks, are useful for predicting what would happen in these scenarios [12, 20, 6], however, they need datasets with the relevant features.

In this work we present a road data model generation tool, which we call “TØIRoads”, that can create datasets synthetically, based on user defined parameters. We describe road data models which can represent high level features of a road network such as population, points of interest, and road length/cost and capacity. Such abstraction allows for a simplified traffic usage and congestion analysis that focus on the high level features. We provide theoretical results regarding mass conservation and sufficient conditions for avoiding congestion within the model.

A main benefit of generating road data synthetically is that the data is independent of any particular city structure. Datasets of varying sizes can be created, simulating the fact that road networks also vary in size. Moreover our road data generator has the interesting feature that the user can give a particular structure (which could realistically represent the information from a real world road network) and use our road data model generation tool to create variants (“mutations”) of this network, e.g., with increasing population, which could indicate possible future congestion issues if the population grows without an increase of capacity in the road network.

In Section 3, we describe our road data models and establish some theoretical properties related to these data models. In Section 4, we describe several parameters that can be specified by a TØIRoads user to create graph data that can serve as input for training graph neural networks (or another learning approach that receives graph data as input) for predicting congestion within the model. In Section 5 we present some statistics associated with the datasets generated by TØIRoads. Finally, we conclude in Section 6.

Traffic prediction is a broad research field that has been heavily studied in the literature [3, 11, 13]. Most works focus on short term or even real time traffic prediction, with the goal of supporting drivers to find an optimal path. Indeed, as many cars and drivers today are sending their information to the internet, some researchers have been very successful in predicting traffic based on real time driver resolution data [7].

Jiang and Luo (2022) describe some datasets used for traffic prediction with graph neural networks as well as what variants of graph neural networks are used for the prediction tasks [11]. They categorize studies by scale and domain, and include passenger flow for bus and metro systems, as well as various studies focusing on road traffic flow, speed, congestion and other variables. What is common among the road traffic datasets is the usage of counting stations in city road systems. Counting stations can have different time resolution as well as various measurements in addition to simple counting of vehicles. Olug et al. (2024)  [17] show for a dataset like this that adding features, such as population density and districts, to the counting station data points can greatly improve predictions. These datasets related to various modes of transportation are based on available data and as such are quite relevant to traffic prediction research.

Bui et al writes about using spatial temporal graph neural networks for traffic forecasting [3], and points to how to construct the adjacency matrix of a graph based on traffic counting station nodes is an open problem, in terms of effectively capturing information streams. For instance, in the urban vehicle detection system dataset [4] that they use, there is no single way to make the adjacency matrix, as there are many road links between each counting station. The TØIRoads dataset generator can take as input a graph structure based on real world data and make datasets with variants of the original data, and so, it might be used as a tool for studying this graph construction problem.

Here we consider traffic from the point of view of road administrators, also called road planners, who have the goal of predicting the need e.g., of improving road capacity in certain road segments to avoid congestion in the city. Our model is made to work as an aggregate of traffic data, representing high level features of a road network, such as population, points of interest, and road length/cost and capacity, while representing the graph structure in detail, at a road segment level. Our model is not restricted to a particular graph structure and it can generate graph structures synthetically. There are various works in the literature on generating graph data synthetically [8, 15], however, here we use features and structural constraints that are meaningful for road traffic prediction. We prove theoretical results related to traffic conservation and congestion for our road data models.

In this section we describe our road data models. The model is an abstraction of the road traffic conditions found in real world scenarios, useful to estimate road usage and congestion. The road network is represented with a graph, where nodes are road segments and edges connect road segments which would be connected in the road network. High population contributes to increase the traffic of a road segment, while high cost contributes to decreasing the traffic. Road capacity is a feature that combines with the measure of traffic to determine if a road segment is congested or not. See Section 4.3 for a detailed example containing the notions presented here.

Here we present our road data model generator, named TØIRoads, for simulating traffic conditions in a simplified road network. The road data models generated by TØIRoads are instances of the model described in Section 3.2. We first describe the parameters that the user can give and then the procedure for generating road data models according to the parameters given by the user.

In this section we provide some experimental results using TØIRoads for generating datasets. The experiments were performed on an Intel(R) Xeon(R) Silver 4116 CPU @ 2.10GHz with 4 cores (x86_64 architecture) with 11 GB of memory. The code used for the experiments is available at https://github.com/gruwesen/TOIROADS. Our hypotheses are that (i) the run time is heavily affected by density and number of nodes, (ii) congestion is affected by the number of nodes and the maximum number of population per node, and (iii) congestion is affected by the ratio of nodes that have a population value. From the mathematical point of view, points of interest and population are interchangeable variables in S-RoadGNN. So the same behaviour for population would hold for points of interest, meaning that there would be no need to perform experiments varying the points of interest instead of population. For this reason, runtime and congestion assessments are only done varying population. We first present some results regarding the time needed to create the datasets.

The results in Table 1 show how increasing the number of nodes and density affects the runtime. For this experiment, the values of the remaining parameters were kept fix and set as follows: the road weight (indicating cost/length) was a number between $1$ and $10$, the population and points of interest rate was $0.4$, the maximum for population and points of interest (per node representing a road segment) was $3$, finally, road capacity was a number between $10$ and $20$.

We now present results showing how increasing the population (and not road capacity) can lead to more congested nodes in the road data model.

Table 2 describes how max population increases the number of congested nodes when considering three sets of $20$ graphs, each with $50$, $100$, and $200$ nodes. The capacity is set to a random number between $[3\sqrt{n},3n]$, where $n$ is the number of nodes, as this interval has been found to often yield a medium amount of congestion in S-RoadGNN. As can be seen from the table, increasing the max population increases congestion at every node number.

Table 3 describes how the population rate increases the congestion level in a similar manner to Table 2. Again, we see an increase in the number of congested nodes as we increase the population rate, similar to when we increase the maximum population. To reiterate, the max population is the max value of a node that has population, while the population rate describes the part of nodes that contain a population. Decreasing the population rate to 0.2 did more to lessen congestion than decreasing the max population.

To illustrate how this dataset can be used in the context of traffic prediction, we have made an experiment using a graph neural network (GNN) that predicts node congestion. In this experiment we trained a GNN model on a dataset generated by S-RoadGNN to predict whether a node is congested or not. We included in Figure 3 a graphical representation of the prediction results of the GNN. Figure 3 shows the resulting predictions for a single example graph after twenty epochs of training. This experiment was run using a Graph Attention Network (GAT) with three layers. The training dataset contained 100 graphs of 20 nodes, while the validation set contained 40 graphs of 20 nodes. The specific graph that is drawn in Figure 3 was drawn randomly from the validation set. In this figure, the model correctly predicted congestion in $13$ of the $20$ nodes.

We motivate and present road data models, some of their theoretical properties and a tool, called TØIRoads, for generating datasets within these models. While one may be tempted to assume that preventing congestion is always desirable, this may not be the case if congestion of non-environmental friendly means of transportation, such as private cars, serve as incentive for the use of public transportation. Nevertheless, being able to study the scenarios related to changes in the population and traffic flow depending on points of interest and to predict congestion (whether to prevent it or not) is important for long-term city planning. As future work, we plan to include a few more parameters such as a sensible strategy for adding location information (currently only represented in a relative way, based on paths and their weights) to our road data models. A potential improvement would be to allow different densities per region of the graph, to simulate more dense districts, as commonly happens to districts that are close to the center of a city.