Identifying urban trafic patterns is critical for reducing CO 2 emissions, yet existing research lacks standardized benchmarks for objectively evaluating clustering algorithms. This fundamental gap prevents accurate assessment because real-world trafic data typically lacks ground truth labels, making the validation of clustering quality impossible. In this work, we propose a methodology for controlled comparison of clustering algorithms using expert-verified ground truth labels derived from SUMO simulations of real urban scenarios. We systematically evaluate six clustering algorithms (HDBSCAN, K-Means, MeanShift, AfinityPropagation, BayesianGMM, AgglomerativeClustering) on both aggregated and concatenated vector representations of trafic data. Our experiments reveal that HDBSCAN achieves the highest accuracy in recovering ground truth scenarios (ARI=0.73, V-measure=0.79) on aggregated data, outperforming K-Means by 0.03 in ARI. Furthermore, aggregated representations systematically outperformed detailed temporal data for all algorithms with an average ARI improvement of 0.15. The study provides a validated benchmarking methodology enabling objective algorithm selection for trafic management systems aimed at emission reduction.
The growth of urbanization and urban trafic intensity creates serious challenges for sustainable city
development, especially in the context of combating climate change. The transport sector accounts
for over one-third of CO2 emissions from final energy consumption in cities, making trafic flow
optimization critically important for achieving climate goals [
Modern urban transport systems face a critical problem of lacking reliable methods for objectively assessing the quality of clustering algorithms when analyzing trafic flows. Most existing research is based on real GPS data or trafic detector readings, which by their nature lack ground truth labels, making accurate assessment of clustering quality impossible. This fundamental problem creates a significant barrier to developing trafic management systems aimed at reducing CO 2 emissions.
The previously unsolved part of the general problem of trafic flow optimization lies in the absence of standardized methodologies for controlled comparison of clustering algorithms under conditions where the true structure of trafic patterns is known. This gap is especially critical in the context of cities’ climate commitments, where accurate identification of trafic patterns can significantly impact emission reduction.
Manifestations of this problem include the inability to determine which clustering algorithm best identifies real trafic patterns under diferent urban conditions, lack of consensus on optimal metrics for evaluating trafic data clustering quality, and shortage of controlled experimental conditions for validating research results in this field.
The main contribution of this research is a proposed methodology for controlled comparison of clustering algorithms for trafic data using expert-verified ground truth labels created from real urban trafic scenarios, which allows objective evaluation of diferent clustering approaches under conditions maximally approximating real transport systems. The research also contributes to understanding the impact of diferent trafic data aggregation approaches on clustering quality, which has direct practical significance for developing trafic management systems oriented toward reducing CO 2 emissions through trafic flow optimization.
The structure of the paper is as follows. The “Literature Review” section analyzes existing approaches to trafic data clustering. The “Materials and Methods” section describes the experimental methodology and algorithms. The “Results” section presents quantitative algorithm indicators. The “Discussion” section interprets the obtained results and compares them with existing approaches.
This section provides an overview of current research on trafic flow clustering, simulation approaches
using SUMO for CO2 emission assessment, and intelligent trafic management systems. Trafic data
clustering represents an actively developing research area evolving under the influence of growing
needs to reduce CO2 emissions from transport. The transport sector accounts for over one-third of
CO2 emissions from final consumption [
Analysis of current research reveals two dominant approaches: centroid-based and density-based
clustering methods. Systematic analysis of K-means application for zone classification by congestion
level showed its quality in identifying delay patterns associated with diferent types of trafic flows [
Mathematical models of urban mobility optimization are developing in parallel [
Comparative algorithm analysis reveals HDBSCAN’s advantage due to its ability to automatically
determine the number of clusters and handle noise. The integration of visual analytics approaches with
machine learning algorithms [
Further methodology development led to creation of emission-sensitive clustering algorithms [
The transition from real data to controlled experiments determines the growing role of simulation
tools. SUMO became the standard thanks to integration capabilities with real sensor data [
General environmental emission reduction trends drive integration of trafic pattern analysis with
emission assessment. Systematization of carbon emission reduction technologies [
Predictive models for intersections using portable measurement systems and density clustering
algorithms [
Modeling approaches for intelligent transport systems [20] emphasize the need for environmental considerations in urban mobility optimization, aligning with the emission reduction focus of this research. Multi-agent deep reinforcement learning approaches [21] and connected vehicle coordination systems [22] demonstrate evolution from centralized to distributed control. Cooperative trafic light control methods with deep learning [23] ensure coordination between multiple intersections, creating an adaptive control network.
Methodological analysis shows transition from one-dimensional to multi-level approaches. Comprehensive reviews emphasize the importance of quality simulation data [24], implemented through two-level machine learning architectures [25]. Integration of spatiotemporal data with real-time route optimization [26] and multi-scale models for medium-term forecasting [27] demonstrate growing complexity of predictive systems. Network approach to mobility analysis through cluster detection methods [28] is complemented by high-resolution cellular network data analysis [29]. Systematization of pattern identification methods using smart card data and deep learning application for spatiotemporal analysis [30] demonstrate evolution from descriptive to predictive urban mobility modeling.
Critical analysis of traditional methods reveals their limitations in determining optimal cluster numbers. Metaheuristic approaches [31] and evolutionary K-means methods [32] ofer solutions through automatic parameter optimization, especially important for trafic data with unknown cluster structure. Diversity-based approaches to clustering [33] demonstrate that ensemble methods leveraging multiple clustering perspectives can improve robustness and accuracy, particularly relevant for heterogeneous trafic pattern identification. Selection criteria for ensemble models [ 34] provide theoretical basis for comparing multiple clustering algorithms systematically, which motivates the multi-algorithm evaluation approach adopted in this study.
Thus, analysis revealed that despite significant progress in the considered field, critical gaps remain. Absence of benchmark data complicates objective algorithm comparison. Fuzzy clustering application emphasizes the need for controlled experimental conditions and ground truth labels for reliable result validation. The conducted analysis reveals a fundamental contradiction: despite the diversity of available clustering algorithms and their theoretical advantages, absence of labeled benchmark datasets makes objective quality comparison impossible for trafic data. This contradiction determines the research goal: improving trafic pattern identification quality through developing an approach that ensures objective clustering algorithm comparison on controlled simulation data with known ground truth structure.
To achieve this goal, the following tasks are formulated: 1. Create a trafic flow simulation model verified by experts and based on real urban scenarios. 2. Conduct systematic comparison of six representative clustering algorithms using standardized metrics.
The general schema of the proposed approach is shown in Figure 1. The proposed schema implements the research hypothesis that using a city map and empirical knowledge about existing trafic flow behavior, SUMO can simulate trafic movement that corresponds to real trafic and allows objective evaluation of clustering algorithm quality.
The approach consists of five main stages: (1) creating a simulation model based on real urban scenarios; (2) generating trafic data in time window format; (3) converting data into vector representation for clustering; (4) applying clustering algorithms; (5) evaluating result quality using standardized metrics.
Trafic flow is formalized as a sequence of trafic light intersection states at discrete time moments. Each trafic light state at time is characterized by a vector of vehicle counts on each trafic lane. Two types of vector data representation are used for analysis: concatenated and averaged values.
Concatenated values represent detailed temporal representation of trafic flow with preservation of complete information about vehicle count changes over time. For each 30-minute time window , a vector of dimensionality = 70 × 180 = 12, 600 is formed, where 70 is the total number of trafic lanes at all trafic lights, and 180 is the number of time slices (data is recorded every 10 seconds during 30 minutes): = [111, 112, . . . , 117, 121, . . . , ], (1) where is the number of vehicles on the -th lane of the -th trafic light at time moment , ∈ {1, 2, . . . , 180}, ∈ {1, 2, . . . , 10}, ∈ {1, 2, . . . , }, where is the number of lanes at the -th trafic light. This approach preserves temporal trafic dynamics but creates a high-dimensional feature space, which may lead to the curse of dimensionality problem in clustering.
Averaged values represent aggregated temporal representation, where one summary value is computed for each trafic lane over the entire window period. A vector of dimensionality = 70 is formed: = [11, 12, . . . , 17, 21, . . . , ], where each component is computed as arithmetic mean: (2) (3) = 1 ∑︁ ,
=1 where = 180 is the number of time slices in the window. Thus, is the average number of vehicles on the -th lane of the -th trafic light over the entire window period.
This approach sacrifices temporal detail for reducing feature space dimensionality by 180 times, providing better conditions for clustering algorithms and increasing resistance to short-term data lfuctuations. The trade-of between information completeness and clustering quality is investigated experimentally by comparing results on both representation types.
Trafic pattern identification is implemented through a sequential process integrating expert knowledge with controlled simulation to ensure objective clustering algorithm evaluation. The complete algorithmic procedure is presented in Algorithm 1 and Figure 2.
Detailed schema of trafic pattern determination method is shown in Figure 2. At the first stage, four base trafic scenarios are formed based on surveillance camera data analysis and expert knowledge from municipal trafic specialists. The morning scenario is characterized by intensive trafic to the city center and market zone, reflecting typical commuting migrations on working days. The evening scenario represents reverse flow from center and market to residential areas. The random scenario models uniformly distributed trafic without clearly expressed dominant direction. The special scenario reflects characteristic trafic from the peripheral Hrechany district, which difers from general city patterns due to its location specifics and transport infrastructure. Each scenario is verified by experts to ensure correspondence with real city trafic flows.
At the second stage, created scenarios are implemented in SUMO simulation environment version 1.15.0 using a real city map. The simulation covers an 11-hour period with recording states of 10 key trafic light intersections. Data is collected at 10-second intervals, providing suficient temporal resolution for capturing trafic flow dynamics. In total, 4,080 trafic light state records are generated. The simulation is configured considering real urban road network parameters.
At the third stage, generated data is segmented into 30-minute time windows to ensure suficient information volume for statistical pattern analysis. An overlapping window method with 10-minute shift step is used, allowing increase of observation count from 22 to 66 and ensuring temporal result stability.
At the fourth stage, each time window is converted into vector representation according to the formalization described in Section 3.2. Obtained vectors are standardized using z-score normalization.
At the fifth stage, six representative clustering algorithms with optimized parameters determined through preliminary validation on pilot dataset are applied to vectorized data. AfinityPropagation is used with damping parameter equal to 0.8. MeanShift is applied with automatic bandwidth deAlgorithm 1 Trafic Pattern Determination and Evaluation Method termination. BayesianGMM is configured with _ = 20 and full covariance type. AgglomerativeClustering uses distance threshold 0.15. HDBSCAN is applied with cosine metric and __ = 4. K-Means is tested with 5 and 7 clusters.
Ground truth labels are formed based on the simulation time schedule, where each of 66 time windows receives the label of the corresponding trafic scenario according to its activity period. Scenario time boundaries are determined considering window overlap and the need to ensure suficient observation count for each pattern type. ◁ 11 hours × 3,600 sec/hour
◁ Set of time windows ◁ Ground truth labels for windows ◁ Step 10 minutes
Clustering quality was evaluated by two metric categories. Internal metrics (Silhouette Score, DaviesBouldin Index, Calinski-Harabasz Index) characterize geometric properties of formed clusters without using external information. External metrics (V-measure, Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), Fowlkes-Mallows Score) compare clustering results with ground truth labels created by experts.
Experiments were conducted on simulation data generated in SUMO 1.15.0 using a real city map. Clustering was performed using scikit-learn 1.3.0 in Python 3.9 environment. Ground truth labels were created based on simulation time schedule, where each of 66 windows received the label of corresponding scenario according to activity period.
Table 2 shows comparison results with ground truth labels of expert-verified scenarios, allowing evaluation of accuracy in recovering true trafic pattern structure. Figure 3 presents comparison of Silhouette Score and Adjusted Rand Index indicators for all algorithms. The graph demonstrates the advantage of averaged data (circles) over concatenated (triangles), and positioning of HDBSCAN and K-Means in the upper right part indicates their optimal balance between geometric cluster quality and accuracy of ground truth scenario recovery.
Figure 4 presents a heatmap of normalized values of five quality metrics for averaged data. HDBSCAN demonstrates the most balanced high indicators across all external metrics, while BayesianGMM shows critically low values across practically all criteria.
Internal metrics analysis revealed a clear pattern: all algorithms demonstrate better results on aggregated (averaged) data compared to detailed (concatenated) values. K-Means with 7 clusters showed the highest Silhouette Score (0.56) and Calinski-Harabasz Index (279.58) for averaged values, indicating best geometric cluster quality. MeanShift demonstrated the lowest Davies-Bouldin Index (0.64) on averaged data, indicating optimal ratio of intra-cluster compactness and inter-cluster separation. Critical quality deterioration is observed for concatenated data: average Silhouette Score decrease is 0.25 points, and Calinski-Harabasz Index decreases on average by 8.7 times. AgglomerativeClustering showed the most dramatic quality drop on concatenated data.
External metrics demonstrate HDBSCAN’s advantage for accurate recovery of expert-verified trafic scenarios. HDBSCAN achieved the highest ARI (0.73) and V-measure (0.79) on averaged data, meaning 73% consistency with ground truth labels and balance between completeness and cluster homogeneity. K-Means (5 clusters) showed second place in accuracy (ARI = 0.70). MeanShift, despite high internal metrics, showed somewhat lower ground truth scenario recovery accuracy (ARI = 0.63). The worst results were demonstrated by AgglomerativeClustering on concatenated data (ARI = 0.03), practically corresponding to random point distribution across clusters.
Results demonstrate clear advantage of aggregated (averaged) data over detailed (concatenated) values for all studied algorithms. HDBSCAN showed highest results by external metrics (ARI = 0.73, V-measure = 0.79), confirming its quality for trafic pattern identification. K-Means with 7 clusters achieved the highest Silhouette Score (0.56) but showed lower results in ground truth scenario recovery accuracy.
Significant quality deterioration on concatenated data (for example, HDBSCAN ARI decreases from 0.73 to 0.61) indicates that high dimensionality and temporal detail complicate stable pattern detection. AgglomerativeClustering showed critically low ARI (0.03) on concatenated data due to creating excessive numbers of small clusters.
Unlike existing research focusing on GPS trajectory analysis, our approach uses aggregated data from trafic light intersections, which better corresponds to practical urban trafic management needs. Results align with previous work conclusions regarding HDBSCAN advantages for trafic data but first demonstrate quantitative comparison on a controlled dataset.
Main limitations include: (1) using simulation data that may not fully reflect real trafic complexity; (2) limited number of scenarios (4 types) that may not cover all urban trafic pattern diversity; (3) focus on one city, limiting result generalizability; (4) absence of considering external factors (weather, events, accidents).
The research demonstrated HDBSCAN’s capability for trafic pattern identification based on expertverified simulation data, achieving the highest ground truth scenario recovery accuracy (ARI = 0.73, V-measure = 0.79) on aggregated data. Key numerical results show a significant advantage of using averaged values over detailed time series, with an average ARI improvement of 0.15 for all algorithms. HDBSCAN outperformed the baseline K-Means by 0.03 in ARI and 0.06 in V-measure, proving the efectiveness of density-based clustering in this domain. A critical finding is the susceptibility of concatenated, high-dimensional data to the curse of dimensionality, which drastically reduced the performance of algorithms like AgglomerativeClustering. The main limitation of this study lies in using simulation data from a single city with a limited set of four scenarios, which may restrict the generalizability of results to other urbanized territories with diferent topological complexity. To address this, future research expansion is planned through integrating real camera data to validate simulation findings, testing the methodology on multiple cities to ensure robustness, and developing hybrid approaches to improve mixed trafic scenario identification. These steps will further refine the benchmark methodology, enabling more efective trafic management systems capable of significant CO2 emission reductions.
This research was funded by the European Union’s Horizon Europe Framework Programme under grant agreement No. 101148374, project “U_CAN: Ukraine towards Carbon Neutrality.” The views and opinions expressed are the authors’ own and do not necessarily reflect those of the European Union or the funding agency, the European Climate, Infrastructure and Environment Executive Agency.
The authors would like to express their gratitude to the European Union’s Horizon Europe Framework Programme for the financial support that made this research possible. Wealso extend our sincere appreciation to the developers and open-source communities behind the essential software tools used in this study, including SUMO, scikit-learn, pandas, and NumPy, whose contributions were invaluable to our work.
The authors have not employed any Generative AI tools.