Application of Traffic Load-Balancing Algorithm—Case of Vigo

Received: 10 August 2025 Revised: 24 September 2025 Accepted: 25 September 2025 Published: 9 October 2025 Citation: Dündar, S.; Alp, S.; Ulu, ˙ I.M.; Dursun, O. Application of Traffic Load-Balancing Algorithm—Case of Vigo. Sustainability 2025,17, 8948. https://doi.org/10.3390/su17198948 Copyright: © 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/ licenses/by/4.0/). Article Application of Traffic Load-Balancing Algorithm—Case of Vigo Selim Dündar * , Sina Alp, ˙ Irem Merve Ulu and Onur Dursun Faculty of Engineering and Natural Sciences, Istanbul Okan University, 34959 Istanbul, Turkey; [email protected] (S.A.); [email protected] (˙ I.M.U.); [email protected] (O.D.) *Correspondence: [email protected]; Tel.: +90-216-677-1630 Abstract Urban traffic congestion is a significant challenge faced by cities globally, resulting in delays, increased emissions, and diminished quality of life. This study introduces an innovative traffic load-balancing algorithm developed as part of the IN2CCAM Horizon 2020 project, which was specifically tested in the city of Vigo, Spain. The proposed method incorporates short-term traffic forecasting through machine learning models—primarily Long ShortTerm Memory (LSTM) networks—alongside a dynamic routing algorithm designed to equalize travel times across alternative routes. Historical speed and volume data collected from Bluetooth sensors were analyzed and modeled to predict traffic conditions 15 min ahead. The algorithm was implemented within the PTV Vissim microsimulation environment to assess its effectiveness. Results from 20 distinct traffic scenarios demonstrated significant improvements: an increase in average speed of up to 3%, an 8% reduction in delays, and a 10% decrease in total standstill time during peak weekday hours. Furthermore, average emissions of CO 2 , NO x , HC, and CO were reduced by 4% to 11% across the scenarios. These findings highlight the potential of integrating predictive analytics with real-time load balancing to enhance traffic efficiency and promote environmental sustainability in urban areas. The proposed approach can further support policymakers and traffic operators in designing more sustainable mobility strategies and optimizing future urban traffic management systems. Keywords: traffic load balancing; urban traffic management; short-term traffic prediction; long short-term memory; PTV Vissim 1. Introduction Traffic congestion is a significant social issue faced by nearly all major cities today. It arises when the demand for road usage exceeds the network’s capacity, but it can be alleviated through effective traffic management strategies. Various techniques, including demand management, optimization of traffic signal timings, and traffic balancing, have been developed and implemented to address this challenge. Travelers typically opt for the least congested routes, selecting paths that minimize travel time to lessen the effects of traffic congestion, often relying on mobile applications to identify the quickest options. Both traffic management methods and route planning software fundamentally depend on accurately predicting near-future traffic conditions. In addition to these general causes, it is essential to note that road traffic volumes exhibit specific temporal distribution patterns. On weekdays, the traffic volume across particular road types follows regular hourly trends; however, during public holidays, these patterns are notably disrupted, resulting in significant variations in daily traffic distribution. Such temporal characteristics of traffic Sustainability 2025,17, 8948 https://doi.org/10.3390/su17198948 Sustainability 2025,17, 8948 2 of 38 flow play a crucial role in analyzing and understanding congestion dynamics. Previous research has examined this issue in detail, highlighting the differences between public holidays and workday traffic distributions across selected road network elements [ 1 ]. Incorporating these temporal variations into congestion analysis enhances the accuracy of traffic prediction and provides a more comprehensive perspective for developing effective management strategies. The IN2CCAM project (Enhancing Integration and Interoperability of the CCAM Eco-System) is conducted within the framework of the Horizon 2020 program. Istanbul Okan University participates in the project as one of the partner institutions. The primary objective is to design, implement, and validate innovative technologies and services that support both infrastructure and end-users. This includes the integration of connected and autonomous vehicles (Cooperative, Connected, and Automated Mobility—CCAM) for both passenger and freight transportation. The overarching goal of IN2CCAM is to benefit all citizens by fully embedding CCAM services into the transportation system. With driverless cars, buses, and trucks operated by Artificial Intelligence (AI)-based vehicle control systems, the project aims to significantly reduce road accidents caused by human error, thereby enhancing road safety. Additionally, this initiative is expected to have positive environmental impacts, such as reducing emissions and optimizing traffic flow to prevent unnecessary travel. It also promotes inclusiveness by ensuring that the elderly and individuals with disabilities can access and benefit from transportation services. This comprehensive approach is centered on the implementation and integration of advanced physical, digital, and operational infrastructures to enhance CCAM services and improve both safety and traffic efficiency. IN2CCAM comprises 21 partners from nine European countries. The project aims to implement digital and operational CCAM solutions across six living labs. The pioneering living labs where the proposed driverless vehicles will be tested in real traffic conditions are located in Tampere (Finland), Trikala (Greece), Turin (Italy), and Vigo (Spain). Additionally, Bari (Italy) and Quadrilatero (Portugal) will serve as follower living labs, evaluating CCAM services in large simulation environments and conducting impact assessments. The IN2CCAM consortium will share the results and findings from the CCAM integration process in both urban and peri-urban traffic with the public. Vigo, one of the pioneering living labs within IN2CCAM, is located in northwest Spain and has a population of 292,817. It is a significant urban center renowned for its intense industrial activity, particularly in the automotive and shipbuilding sectors. The city features a bustling fresh fish port -one of the leading ports in Europealong with a container terminal and various commercial enterprises. However, the city’s dense activity, challenging geography, and irregular growth give rise to a range of transportation challenges. This study developed a load-balancing algorithm to address traffic demand in the city center of Vigo. The algorithm redistributes traffic across multiple route alternatives to optimize network performance, with the primary goal of equalizing travel times on all alternative routes between designated points. The effectiveness of this algorithm was assessed using the PTV Vissim 2023 microsimulation software. In the grand scheme of sustainable transportation, reducing traffic and improving the efficient use of existing infrastructure are crucial for cutting energy demand and emissions. As urban regions expand, methodologies that enhance traffic flow with a lean reliance on large-scale new construction promote sustainability by reducing the environmental burden of mobility systems. Dynamic traffic regulation strategies, such as the load-balancing algorithm, are beneficial for enhancing mobility and optimizing resources in the context of Sustainability 2025,17, 8948 3 of 38 sustainable urban development. This makes them efficient tools in building a robust and environmentally friendly transport system. A longstanding theoretical tension in traffic assignment arises between user equilibrium (UE) and system optimum (SO), as articulated in Wardrop’s principles [ 2 ]. Under UE, each driver minimizes their individual travel time, leading to a stable yet often suboptimal outcome for the entire system. By contrast, SO seeks to minimize total travel time across the network, which may require deviations from individually optimal routes. Most existing studies on dynamic routing and load balancing implicitly align with the UE framework, since drivers are modeled as rational agents optimizing their own trips. However, this emphasis on individual rationality overlooks the gap between user-based and system-wide efficiency. In this study, while the simulation framework reproduces driver behavior consistent with UE, the proposed load-balancing algorithm functions as a corrective mechanism that guides routing decisions toward more balanced traffic distributions. In this sense, the method can be regarded as a step toward reconciling UE with SO. It demonstrates potential both as a practical congestion management tool and as a contribution to the theoretical discussion on the trade-off between individual choice and collective efficiency. The main contributions and innovative aspects of this study can be summarized as follows: • This paper presents the first combination of 15-min Bluetooth-based short-term traffic forecasts and a real-time load-balancing algorithm. This link effectively connects predictive modeling to dynamic routing decisions. • It introduces a solid data processing framework that combines Monte Carlo-based methods for filling in missing data with a decade’s worth of high-resolution Bluetooth data. This ensures the reliability and completeness of the input. • The proposed approach has been thoroughly tested under 20 realistic peak-hour scenarios in PTV Vissim, showing its practical use and strength beyond theoretical development. Additionally, the simulation workflow is designed to be replicable and focused on supporting decision making. • It provides valuable insights for both local and central governments, as well as mobility service providers working on data-driven congestion reduction. All the codes used in this study are provided in the GitHub link: https://github.com/ OnurPSR/IN2CCAM/ which exists since 1 August 2025. 2. Literature Review As urban populations grow and vehicle ownership rises, transportation-related challenges have emerged as significant issues in mediumand large-sized cities. These challenges lead to traffic congestion, economic losses, accidents, wasted time during commutes, and environmental impacts associated with transportation [ 3 – 5 ]. Public transportation services often face platform failures and deteriorating service quality in high-demand situations, such as natural disasters, pandemics, and holiday periods [ 6 ]. In these circumstances, effective traffic management becomes crucial. Without a sound strategy, rising user demand can result in bottlenecks, interruptions, and delays [7]. Congestion management strategies generally fall into two categories: congestion prevention and congestion control. Congestion prevention takes a proactive approach to safeguard the network from overload, while congestion control reacts to congestion that has already occurred within the network [ 8 , 9 ]. Given the current limitations in modifying transportation infrastructure and the insufficient information available to drivers, traditional shortest path algorithms are widely employed. Algorithms such as Dijkstra and A* are commonly used to identify optimal routes and minimize travel times. Alongside these, dynamic path planning algorithms address unique situations [ 5 , 10 – 12 ]. Dynamic Sustainability 2025,17, 8948 4 of 38 path planning algorithms often utilize big data predictive analysis technologies to assess traffic conditions and forecast future scenarios, thereby enabling the estimation of the most suitable routes [13]. One alternative method, multi-agent simulation, emphasizes the individual active components within a system. This approach allows for flexibility in designing heterogeneous actors that interact with one another and their environment [ 14 ]. Conversely, reinforcement learning is a modeling technique inspired by behavioral psychology, where the system is conceptualized as agents interacting with their surroundings [15]. Another approach is optimization. In the realm of traffic engineering, optimization focuses on routing traffic across the network to enhance overall performance and efficiently utilize resources. Effective traffic management can improve service accessibility while optimizing resource use and minimizing operational costs [ 16 ]. Classical optimization methods typically involve offline routing based on a pre-estimated or predefined traffic matrix [ 17 – 19 ]. These methods function effectively as long as actual traffic patterns closely align with the predefined matrix. However, they often struggle to accommodate traffic fluctuations in unexpected scenarios, which has led to the implementation of online dynamic load-balancing techniques [ 17 ]. Load balancing enhances efficiency and mitigates the risk of overload by distributing traffic across multiple units, ensuring optimal resource utilization [ 20 ]. A notable drawback of this method is the potential for interruptions that may arise during the rebalancing of the routing protocol amidst congestion [ 17 ]. Another strategy involves robust static routing, which considers multiple potential traffic matrices to identify a routing configuration that ensures optimal performance across all scenarios [17,21,22]. In unusual road situations, traffic management centers redirect drivers to alternative routes. This approach not only enhances the efficiency and capacity of the network but also mitigates the risk of secondary accidents that can occur due to congestion [ 23 ]. In recent years, intelligent navigation systems—similar to active traffic management functions—have been developed. These systems calculate routes that minimize travel times by leveraging congestion predictions derived from the speed and location data of all devices using the application [23]. In earlier studies, Zhang et al. introduced a Dynamic Hybrid Routing framework aimed at optimizing traffic flow [ 17 ]. This approach focuses on rebalancing traffic by choosing and implementing the most suitable pre-configured routing scenario in response to fluctuations in traffic conditions. Conversely, Li et al. addressed the challenge of load balancing by formulating it as a non-linear programming problem. They then applied deep learning techniques, combined with traffic predictions, to address this optimization problem [20]. The primary objective of active traffic management is to enhance system performance by utilizing traffic assignment algorithms. In this regard, dynamic routing algorithms and traffic simulations, which help predict traffic patterns, are becoming increasingly vital for designers of active traffic systems. Active traffic management is a crucial concern, not only due to its impact on productivity loss but also because it influences fuel consumption [ 23 ]. Modern routing techniques can gather traffic congestion data using both real-time and historical information. Real-time data can be sourced from loop detectors, cameras, toll port data, or mobile phone signals. While today’s prevalent navigation systems typically direct users to the route with the shortest travel time, a significant strategic challenge persists [ 24 ]. These systems generally guide drivers based on periodically updated data, which can lead to an excessive concentration of vehicles on the “optimal” route. Consequently, this can lead to increased travel times due to congestion. Additionally, subsequent drivers often make their routing decisions before the impacts of earlier traffic flows are Sustainability 2025,17, 8948 5 of 38 reflected in the system. As a result, directing the majority of vehicles to the “best route” can paradoxically create traffic bottlenecks and lead to longer overall travel times [24]. To address the routing challenge, various algorithms have been developed that employ a cooperative routing approach, taking into account both the predicted traffic volume and estimated travel times [ 24 ]. Liu et al. introduced a participatory navigation system that utilizes data from collaborative location and route selections made by vehicles to forecast traffic speed and future traffic flow [ 24 ]. Roughgarden’s research suggests that routing without coordination yields suboptimal outcomes [ 25 ]. In this context, Wilkie et al. introduced a self-aware re-routing algorithm that integrates historical traffic data with planner-predicted routes. The algorithm explicitly considers the potential for each vehicle’s planned trajectory to exacerbate congestion on the roads; marginally, it utilizes [26]. The use of traffic simulations to assess the potential impacts of policies and strategies has become increasingly essential, particularly as transportation systems become more complex with the incorporation of new active traffic management techniques. The models applied in these analyses are typically categorized as equilibrium models (traffic assignment models) and non-equilibrium models (simulation models) [ 23 ]. While equilibrium models fall short in addressing unexpected events, non-equilibrium models are necessary to fill that gap. However, in densely populated and congested urban areas, information is often lacking, especially in situations where various stakeholders, such as smartphone applications and traffic management centers, employ different management strategies [ 23 ]. Liang and Wakahara conducted a SUMO microsimulation in a medium-sized area of London, demonstrating that proactive dynamic re-routing can significantly reduce average travel times [ 27 ]. In their study, Wang et al. tested an algorithm within SUMO that aimed to address congestion. The process began with calculating the immediate turn for a vehicle approaching a congested road segment, followed by a recalibration of the remaining route [ 7 ]. This algorithm achieved an impressive reduction of up to 51.50% in average travel time. Urban traffic management has been an active research area, encompassing traffic prediction, routing, and load balancing. Traditional parametric models, such as the Kalman filter (KF) [ 28 ], exponential smoothing (ES) [ 29 ], and the autoregressive integrated moving average (ARIMA) family [ 30 ], have been widely applied for traffic prediction. However, these models assume stationary time series, which is often not the case, especially for non-linear traffic data [ 31 ]. Non-parametric approaches, including K-nearest neighbor (KNN) [ 32 ] and support vector regression (SVR) [ 33 ], provide greater flexibility but require careful parameter tuning. Deep learning models, such as stacked autoencoders (SAEs), deep belief networks (DBNs), and artificial neural networks (ANNs), have been explored for their superior predictive capabilities [ 34 ]. Standard ANNs, however, cannot explicitly capture temporal dependencies; they often function primarily as output aggregators in deeper network architectures [ 34 ]. Recurrent neural networks (RNNs) address this limitation by learning sequential dependencies, making them suitable for time-series prediction. Vanilla RNNs suffer from vanishing gradient issues [ 35 ], which are mitigated by long short-term memory (LSTM) networks and gated recurrent units (GRUs) [ 36 ]. RNNs are often implemented in stacked layers or encoder–decoder frameworks [ 37 ], compressing input sequences into context vectors. While effective, static context vectors may create bottlenecks, limiting the model’s ability to capture complex temporal patterns. In addition to accurate traffic prediction, effective routing and load balancing are essential for minimizing congestion and improving network efficiency. Traditional shortestpath routing and uniform load distribution approaches are limited in their responsiveness to dynamic traffic conditions. Popularity-based routing methods, such as Polaris, distribute traffic according to road usage patterns, mitigating congestion and lowering CO 2 Sustainability 2025,17, 8948 6 of 38 emissions [ 38 ]. Quantum-inspired optimization algorithms provide scalable solutions to complex routing problems, outperforming classical optimization methods in specific urban networks [ 39 ]. Strength-optimized weight balancing dynamically adjusts traffic flow to prevent network congestion [ 40 ], while adaptive load balancing using Road Side Units (RSUs) redistributes traffic based on network conditions [ 41 ]. Smart Route Determining with Load Balancing (SRD-LB) integrates real-time traffic data and environmental factors to optimize routing and reduce travel time [42]. While these schemes represent tremendous advancements over the conventional approach, the most serious problems remain in properly integrating prediction into routing and load balancing. Most dynamic routing algorithms, particularly those based on Reinforcement Learning (RL), respond quickly to real-time data. This quick response can lead to delays or loss of information, resulting in poor clustering. Cars may end up following a limited number of routes, making traffic worse instead of better. Additionally, merging predictive models with load balancing brings up issues with timing and accuracy. If rerouting is too aggressive, it can disrupt traffic; if it is too cautious, it might fail to prevent congestion under changing conditions. These challenges require a broader strategy that utilizes effective traffic predictions and adjusts vehicle schedules across the network to avoid bottlenecks and enhance overall performance. These constraints drive this current work, which suggests an integrated urban traffic control model. The model combines traffic prediction optimization and load balancing, routing based on not only near-future forecasts but also more abstract network states and long-term traffic patterns. The proposed approach enhances efficiency and reliability in urban road networks by directly addressing issues such as poor clustering, prediction uncertainty, and dynamic load balancing. This is especially important in complex areas, such as Vigo. The combined approach is far superior to conventional approaches, as it provides practical guidance for traffic operators and policymakers who aim to optimize congestion control, minimize travel time, and mitigate environmental impacts. Technologies such as autonomy and connectivity can significantly enhance system performance [ 43 ]. Through vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, vehicles can exchange information without requiring extensive sensor deployment across the network, thereby improving the overall travel experience [ 43 , 44 ]. V2V communication enables vehicles to access local traffic data, thereby alleviating the computational burden on centralized systems, while automation facilitates real-time responses [44]. A study analyzing 2900 rows of vehicle data, recorded hourly from January to May 2021 at the Ba¸sak¸sehir Mahmutbey Junction in Istanbul, aimed to predict both traffic volume and speed values [ 45 ]. To achieve this, an LSTM–based prediction method was employed. Its results were compared with those obtained from linear regression, Random Forest (RF), Support Vector Machines (SVMs), Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNNs), and RNNs. The LSTM method yielded significantly superior results across all evaluation metrics, including mean squared error, root mean squared error, mean absolute error, and R 2 values. Specifically, it achieved an R 2 score of 0.897 for volume prediction and 0.883 for speed prediction. In another study, researchers sought to predict average speed using data from 15 traffic density measurement stations and 10 weather stations in Istanbul, covering the period from 1 January 2020 to 1 January 2021 [ 46 ]. Utilizing four different ANNs along with two SVM models, the General Regression Neural Network (GRNN) model demonstrated the highest performance, attaining an R2value of 0.899. Additionally, another investigation focused on predicting average flow speed under various conditions along the D100 Highway at the entrance to Tuzla District in Sustainability 2025,17, 8948 7 of 38 Istanbul [ 47 ]. This study systematically examined the meteorological factors that affect traffic flow speed, including temperature, humidity, wind speed, prevailing weather conditions, weekdays, and working hours. It further sought to predict average flow speed by analyzing different combinations of these influencing factors. The performance of the developed model was rigorously evaluated by examining the discrepancies between observed and predicted speeds. The analysis indicated that the differences ranged from a minimum of 0.09 km/h to a maximum of 5.98 km/h, demonstrating the model’s predictive accuracy. A groundbreaking study utilizing the LSTM method to predict near-future traffic speeds was conducted in Beijing, China [ 48 ]. This research utilized data from two microwave sensors positioned on a highway to forecast both traffic volume and speed. The performance of the model was evaluated against three RNN architectures, SVM, ARIMA, and Kalman Filters. The LSTM model outperformed all other methods in terms of mean absolute percentage error and mean squared error. Another significant study focused on traffic speed prediction in Chicago, USA [ 49 ]. This study employed various pre-processing techniques to support seven distinct deep learning models aimed at predicting average flow speeds along a 13-mile segment of Interstate 55 (I-55) Highway. The top-performing model achieved an R 2 value of 0.85, effectively forecasting traffic speeds during both a sports event and rainy conditions. Additionally, research utilizing data from two publicly available datasets—METR-LA and TaxiBJ—alongside data from the Chengdu Highway in Chengdu, China, aimed to predict traffic [ 50 ]. This study employed a graph-based CNN and compared its performance with that of five different methodologies, including LSTM. The evaluation used mean absolute error, root mean squared error, and mean absolute percentage error as performance measures, revealing that the graph-based CNN delivered superior results across all metrics. A recent review extensively analyzed modern deep learning-based traffic prediction methods, with a particular focus on CNN, RNN, and LSTM, while highlighting the strengths and weaknesses inherent to these deep learning approaches [51]. Traffic prediction has evolved in recent years from traditional time series and deep learning methods to graph-based methods. These new methods explicitly model spatial dependencies in road networks. GNNs have been utilized with open-government data to achieve better performance than traditional methods, such as ARIMA and historical averages. This is because they can effectively manage spatial and temporal interactions within urban traffic systems [52]. Subsequently, more advanced architectures, such as the Dynamic Graph SpatialTemporal Neural Network (DGSTN), have been proposed to overcome the limitations of static graph models. Equipped with adaptive spatial self-attention and dynamic graph generation, these models can better describe changing traffic conditions. As a result, they outperform traditional LSTM-based models in terms of prediction performance [53]. This shift from static to dynamic graph models shows the advantage of flexibility in modeling changing traffic networks. Another issue is how to retain prediction capability with new traffic data over time. In 2025, a continual learning framework based on Spatial-Temporal Graph Convolutional Recurrent Networks and regularization techniques, including the Elastic Weight Consolidation and Memory Aware Synapses, is presented. This method enables the learning of new traffic patterns without forgetting the old ones, resulting in more robust real-time prediction [54]. Alongside progress in traffic prediction methods, the development of Vehicle-toEverything (V2X) communication has enabled cooperative traffic prediction. Among the deep learning models, Bidirectional Long Short-Term Memory (BiLSTM), GRU, and LSTM, we can see that including V2X communication data significantly enhances the prediction Sustainability 2025,17, 8948 8 of 38 performance. These results indicate the usefulness of cooperative vehicle systems in traffic management [ 55 ]. Moreover, hybrid models that combine GNNs and Deep Reinforcement Learning (DRL) have been used in optimizing resource allocation for V2X communication networks, demonstrating the broader applications of graph-based learning outside the realm of traffic flow prediction [56]. Additionally, contrastive learning has been recently introduced into GNN-based traffic forecasting models. The Contrastive Learning Multi-Graph Convolution Network (CLMGCN) utilizes temporal data augmentation and contrastive representation learning to enhance feature learning. The model, supported by multiple graph structures, has been shown to outperform traditional graph convolutional models. It confirms that representation learning is effective for traffic forecasting [57]. These developments collectively indicate an interesting direction for research in graphbased, adaptive, and communication-aware prediction models. GNN-based models naturally capture network-level dependencies in a more straightforward manner than LSTMbased ones. V2X models enhance predictions by leveraging information from the cooperation of traffic states. With the addition of continuous learning and contrastive learning, adaptiveness and robustness are enhanced further. This work contributes by placing an LSTM-based prediction model within the context of a real-world decision support system, taking into account these emerging trends for future applications. 3. Materials and Methods Initially, the traffic data collected for the city of Vigo was assessed, revealing sensorrelated gaps within the time series data. These gaps were addressed through a Monte Carlo simulation, followed by short-term traffic forecasting conducted using Machine Learning techniques. Rather than relying solely on current traffic conditions, load balancing was implemented based on anticipated traffic conditions projected for the next 15 min. Ultimately, the traffic load-balancing process aimed to distribute demand in a manner that would achieve equilibrium in travel times on roads within the city center. The effectiveness of the developed load-balancing algorithm was evaluated within the PTV Vissim microsimulation environment. This section provides a comprehensive overview of the studies conducted. 3.1. Traffic Prediction In 2013, the city of Vigo launched a network of 80 Bluetooth sensors. These sensors anonymously capture the Bluetooth MAC (Medium Access Control) addresses of devices, including smartphones and vehicle systems, at various locations throughout the city. The timestamps of these detections are recorded, and an algorithm refines and filters the data to calculate the travel time between two Bluetooth sensor locations. Figure 1illustrates the Bluetooth sensor detection architecture in Vigo. The journey between each pair of sensors is referred to as a Vector, with its length and trajectory determined by a set of geographic coordinates. The city publishes all data about these vectors through an API REST (Application Programming Interface, Representational State Transfer) service. The system’s algorithm computes travel times for all vectors every three minutes. The resulting data are made available through the Vigo website, the Vigo Driving Application (APP) V1.4, and in a JSON (JavaScript Object Notation) file formatted for open data. Travel times for points of interest are displayed on physical Variable Message Panels (VMPs) and Virtual VMPs (IVI—In Vehicle Information messages) within the Vigo Driving APP (see Figure 2). Sustainability 2025,17, 8948 9 of 38 Figure 1. Bluetooth sensor detection architecture in Vigo. Figure 2. VMPs displayed on the Vigo driving app. The thresholds that determine each vector color are as follows: •Medium Speed > 30 km/h—green. •20 km/h < Medium Speed < 30 km/h—green. •15 km/h < Medium Speed < 20 km/h—yellow. •10 km/h < Medium Speed < 15 km/h—orange. •Medium Speed < 10 km/h—red. The data collected by the Bluetooth sensor network is utilized to deliver real-time information to drivers and the traffic management system. Furthermore, all records are stored in a database with a resolution of 3 min. This database contains approximately 10 years of data for each vector, detailing trip times and average speeds. Given the substantial size of the full database, the resolution has been adjusted to 15 min. The final dataset Sustainability 2025,17, 8948 16 of 38 3.2.9. Long Short-Term Memory LSTM networks are a specialized form of RNNs that effectively capture long-term dependencies in sequential data, making them particularly suited for time series prediction. LSTMs address the shortcomings of traditional RNNs by incorporating mechanisms such as input, output, and forget gates, which regulate the flow of information. This enables the network to retain relevant information over extended periods while discarding irrelevant data. As a result, LSTMs excel in various tasks, including time series forecasting, speech recognition, and natural language processing. The primary advantage of LSTMs lies in their robustness when handling sequential data, where the relationships between data points over time are critical. This characteristic makes them highly effective for applications demanding precise time series predictions, such as financial market forecasting, weather prediction, and energy load forecasting. By effectively managing the vanishing gradient problem—a common challenge in standard RNNs—LSTMs maintain performance across long sequences. Despite their strengths, LSTMs do have limitations. They are computationally intensive and require considerable training time and resources, particularly when working with large datasets. The complexity of the model can complicate the interpretation of its internal workings and the relationships it captures. Additionally, hyperparameter tuning for LSTMs can be intricate, often necessitating extensive experimentation to achieve optimal performance. Nevertheless, the advantages of LSTMs in managing complex sequential data establish them as invaluable tools within the machine learning toolkit. In summary, LSTM networks are potent methods for modeling and predicting sequential data due to their capacity to capture long-term dependencies while alleviating common issues faced by traditional RNNs. Although they are resource-intensive and require careful tuning, their applications across various fields highlight their significance and effectiveness in time series analysis. 3.2.10. Comparison of Model Performances The performance metrics (R 2 and Mean Squared Error, MSE) obtained from the Deep Learning models for each start–end pair are presented in Tables 3and 4. As indicated by the tables, the LSTM model has exhibited significantly superior performance in terms of R 2 across all origin–destination pairs when compared to the other models. However, with respect to mean squared error (MSE), it outperformed the other models only in the Martinez Garrido–Burgos and Aragon–Jenaro pairs. In the cases of Aeropuerto Hasta Ap-9 Tunnel Da Madroa, GBR, Florida–Padre Seixas A Ricardo Mella, and Praza Da Estacion A Garcia Barbon Ve Pso Alfonso XII–G Portela–Berbes, the MLP model demonstrated the best performance. Nevertheless, the performance of the LSTM model closely approached that of these models. In both tables, the best values are highlighted in bold. The hyperparameters of the LSTM model are presented in Table 5. To visualize the prediction performance of the LSTM model, the predicted speed values derived from the speed data obtained at Martinez Garrido-Burgos A Aragon-Jenaro for a single day are illustrated in Figure 4. The results indicate that the models perform well in estimating traffic volume through time-series analysis using machine learning methods, with the LSTM model demonstrating the highest levels of performance among those tested. Accurate traffic predictions are crucial for various applications, including demand management, optimizing traffic signal timings, and balancing traffic flow. Nonetheless, models trained solely on historical traffic speed data, despite achieving high accuracy, exhibit limitations in anticipating changes in traffic conditions. Additionally, they tend to provide average values during peak and trough periods of demand, including speed estimates. Sustainability 2025,17, 8948 17 of 38 Table 3. R2Value of the Models. Performance Aerop. Hasta Ap-9 Túnel Da Madroa Florida-Padre Seixas A Ricardo Mella Martinez Garrido-Burgos A Aragon-Jenaro Praza Da Estacion A Garcia Barbon Pso Alfonso XII—G Portela—Berbes MLR 0.844 0.697 0.747 0.792 0.971 SVR 0.848 0.681 0.733 0.796 0.972 RFR 0.840 0.678 0.728 0.783 0.973 MLP 0.853 0.708 0.753 0.804 0.976 ABR 0.750 0.701 0.742 0.831 0.955 GBR 0.853 0.707 0.754 0.802 0.974 KNN 0.826 0.655 0.707 0.768 0.971 ETR 0.833 0.662 0.720 0.774 0.972 LSTM 0.953 0.918 0.907 0.952 0.993 Table 4. MSE Value of the Models. Performance Aerop. Hasta Ap-9 Túnel Da Madroa Florida-Padre Seixas A Ricardo Mella Martinez Garrido-Burgos A Aragon-Jenaro Praza Da Estacion A Garcia Barbon Pso Alfonso XII—G Portela—Berbes MLR 15.282 29.082 11.341 8.483 71.137 SVR 14.861 30.558 11.958 8.314 69.796 RFR 15.677 30.922 12.186 8.837 67.333 MLP 14.377 27.982 11.039 7.985 63.403 ABR 24.526 28.803 11.438 25.235 113.920 GBR 14.369 28.122 11.026 8.071 63.936 KNN 17.039 33.126 13.091 9.477 72.928 ETR 16.325 32.394 12.547 9.207 69.935 LSTM 14.765 29.238 9.186 8.366 78.988 Table 5. Hyperparameters of the LSTM Model. Component Hyperparameter Search Space Final Value Model Input Input Sequence Length [3, 4, 5, 6, 7, 8, 9, 10] 8 Model Input Prediction Horizon [15 min] 15 min Model Input Hidden Layer Size [5, 10, 15, 20, 25, 30, 35, 40, 45, 50] 15 Model Input Number of Layers [1, 2, 3, 4] 2 Model Input Batch Size [32, 64, 128] 64 Regularization Early Stopping Patience [8] 8 Optimizer Loss Function [Adam Optimizer] Adam Optimizer Optimizer Learning Rate [1 ×10−3, 5 ×10−4] 1 ×10−3 Normalizer - - - Window Length - - - Stride - - - 0 10 20 30 40 50 Average Speed (km/h) LSTM Model Predictions Observed Speed Predicted Speed Figure 4. Predicted speed values derived from the speed data obtained at Martinez Garrido-Burgos A Aragon-Jenaro for a single day. Reported Mean Absolute Percentage Error (MAPE) = 2.87%. Sustainability 2025,17, 8948 18 of 38 3.3. Load-Balancing Algorithm Load balancing has its roots in computer technology. It refers to the process of distributing resources—such as servers, processors, and storage units—across a network or system in an equitable manner to optimize performance. This approach not only enhances the efficiency of network or system operations but also bolsters resilience against potential overloads. The primary objective of load balancing is to effectively distribute traffic or queues directed towards resources experiencing high demand. By ensuring that traffic is evenly distributed among available resources, load balancing helps prevent resource overload. It mitigates the risks associated with single points of failure, which can lead to performance degradation. This technique is particularly prevalent in large-scale networks and server clusters, where it plays a crucial role in ensuring high availability and performance. By reducing the likelihood of system crashes, load balancing improves user experiences and supports business continuity. A variety of load-balancing algorithms and technologies are employed to cater to differing requirements. In the realm of transportation engineering, load balancing can be likened to the role of a traffic police officer, who alleviates congestion by directing vehicles to optimal positions. This systematic approach not only prevents accidents but also ensures that vehicles reach their destinations as efficiently and swiftly as possible. Traffic load balancing, therefore, refers to the balanced distribution of traffic across various resources in a network, resulting in more effective resource utilization and improved network performance. Implementing traffic load balancing is crucial for increasing operational efficiency, saving time and resources, and minimizing environmental impacts in road transportation. Consequently, transportation companies and logistics operations can strategically adopt these practices to reap significant benefits, particularly in large distribution centers, logistics hubs, and warehouses, where effective traffic management is essential. This not only helps to lower costs but also improves delivery timelines. Achieving effective transportation efficiency can be accomplished through several methods: Route Optimization: This process ensures that the chosen route is the most accurate, based on time and distance parameters. To optimize routes, various factors must be taken into account, including the number and locations of necessary visit points along the route, as well as the timing of these visits. By optimizing transportation routes, vehicles can take shorter and more efficient paths, resulting in reduced fuel consumption and time savings. Effective route optimization not only enhances distribution and transportation processes but also supports business growth through cost-effectiveness. Vehicle Tracking Systems: Global Positioning System (GPS)-based vehicle tracking systems facilitate real-time monitoring of vehicles, allowing for route adjustments based on current traffic conditions and helping to balance traffic load. Capacity Planning: Efficient transportation capacity planning organizes vehicles and shipments according to demand, thereby preventing unnecessary trips made by empty vehicles. Multiple Distribution Centers: For larger companies, utilizing multiple distribution centers to distribute loads yields positive outcomes by balancing traffic and minimizing delivery times. Transportation Diversification: Employing a diverse range of interoperable transportation modes—such as trucks, trains, and ships—helps to balance traffic loads and reduce costs. Traffic Routing: Traffic routing algorithms can effectively be employed to steer clear of congested areas. Sustainability 2025,17, 8948 19 of 38 In the IN2CCAM project, the development of a load-balancing algorithm began with classifying roads in the city center of Vigo into four categories based on their respective legal speed limits. For each category, the boundary conditions of traffic flow were determined. These conditions included free flow speed (u s ), jam density (k jam ), and maximum flow (Q max ), which were calculated based on the number of lanes in one direction, ranging from one to four. Utilizing these values, the fundamental relationship between speed and flow was derived individually for each road, as illustrated in Figure 5. Figure 5. Speed–flow relationship. The speed–flow relationship predicts a decline in flow when demand surpasses Q max ; however, it fails to address how speed is affected when demand exceeds capacity. Consequently, for the saturated flow conditions depicted in Figure 6, the shift in speed resulting from demand exceeding capacity was determined by symmetrically reflecting the forced flow section of the graph around the capacity point. This approach enabled the derivation of individual curves that illustrate how vehicle demand affects flow speed on each road in the city center of Vigo. Figure 6. Speed–demand relationship [62]. Sustainability 2025,17, 8948 20 of 38 The developed load-balancing algorithm determines the optimal route between any two points by taking into account current road congestion. It subsequently guides vehicles along these routes to distribute traffic evenly and prevent bottlenecks. As each new vehicle contributes to the load on its selected path, the algorithm recalibrates the conditions and may suggest an alternative route for subsequent vehicles. Initially, a weighted and directed graph is constructed to represent all possible transactions between roads. The algorithm incorporates two distinct weights: the distances between roads and the travel durations between them (Equations (3)–(5)). G = (V, E,d,τ) (3) d:E −→ R_(>0) (4) τ:E −→ R_(>0) (5) In this context, Vrepresents the set of roads, Esignifies the set of directed connections, ddenotes the length of edge Ein kilometers, and τ indicates the expected travel time in minutes. Once the graph is created, an adjacency matrix is generated to facilitate recursive calculations and enable traversal of all nodes. Based on the structured data, all potential paths between the source and destination are determined using a depth-first algorithm with an exhaustive search method. The 15-min prediction results are used to determine how many seconds it will take for a driver to use a route, considering the car congestion on that route. With the prediction results using discrete time blocks of 1–20 s, a car congestion threshold dataset is constructed. Each column represents a car congestion threshold, which is needed at maximum to travel at that time in seconds. The weights of the duration and distance parameters are tested in different proportions. The path with the shortest distance is identified with depth-first approach, and, subsequently, all paths whose distance is less than λ times the shortest distance are filtered for evaluation, where λis a tuned threshold set by the Sobol test. The duration of the filtered paths is then calculated in relation to the road’s load. To evaluate the optimal path, durations and distances of roads are normalized with respect to their minimum and maximum values, and a weighted average approach is employed. The weight α balances time and distance and is tuned via the Sobol sensitivity test. The procedure is given as Equations (6) and (7). ∼ d=d−dmin dmax −dmin ,∼ t=t−tmin tmax −tmin (6) Weighted average = α(duration) + (1 −α/100)(distance) (7) The weighted average is calculated based on the priority assigned to the variables of duration and distance. After computing the weighted average for each candidate path, the one with the lowest value is selected. Subsequently, the travel times for the roads along that route are updated. Each time a new vehicle is routed from the source to the destination, these same steps are repeated. As traffic increases, travel times may change, which can alter the recommended path compared to the initial suggestion. Evaluations of the results are analyzed using the First-order Sobol Sensitivity Test. Based on the parameters within the specified range, 512 tests have been conducted. Tables 6and 7present the parameters of the load balancing algorithm and the Sobol sensitivity scores of the algorithm, respectively. The change in the depth range had zero variance in the Sobol Sensitivity Test, as it was observed that the algorithm’s convergence determination depends on 10 or fewer recursions. Sustainability 2025,17, 8948 21 of 38 Table 6. Parameters of the Load-Balancing Algorithm. Parameter Search Space Depth Range [10, 40] αRange [0, 1.0] γRange [2.0, 4.0] Table 7. Sobol Sensitivity Scores of the Load-Balancing Algorithm. Parameter S1 Index S1 Confidence Score Depth 0.000000 0.000000 α0.592330 0.093727 γ0.225597 0.081464 The pseudo-code of the load-balancing algorithm is given as Algorithm 2. Algorithm 2. Pseudo Code of the Load-Balancing Algorithm 1: Proc ENUMERATE_PATHS(G, s, t, Dmax): 2: Input: G = (V, E, d, τ) 3: s←source vertex 4: t←destination vertex 5: Dmax ←recursion depth limit 6: Output: P←list of tuples (π, dist(π)) 7: Procedure DFS(cur, depthleft, path, distacc): 8: If depth_left = 0: 9: Return 10: For each (cur, nxt) ∈E: 11: de←d(cur, nxt) 12: If nxt = t: 13: π←path ▷[cur, nxt] 14: Append (π, distacc + de) TO P 15: Continue 16: If nxt ∈path: 17: Continue 18: DFS(nxt, 19: depthleft −1, 20: path ▷[cur, nxt], 21: distacc + de) 22: Initialise empty list P 23: DFS(s, Dmax, [], 0) 24: Return P The description of the pseudo-code is provided in Appendix A. 4. Traffic Simulations The PTV Vissim 2023 software was employed to simulate traffic conditions. The analysis was constrained by the software’s 1000 km road drawing limit and the complexity of mapping Bluetooth data to individual road segments. Consequently, the study concentrated on a designated sample area rather than the entire city of Vigo. During the simulation phase, the geographic boundaries of this sample area were established to ensure an accurate representation of the region, as shown in Figure 7. Sustainability 2025,17, 8948 22 of 38 Figure 7. Geographic boundaries used for the Vigo traffic simulation. The road network of the selected sample area was represented in the simulation model. The representation incorporated key parameters, including the number of lanes, lane width, geometric characteristics, road classification, intersections, and turning movement data for both right and left turns. Figure 8displays the roads included in the simulation model. Figure 8. Representation of the roads in the simulation model. In the next phase, traffic control devices were integrated into the simulation model for the designated area in Vigo. This involved defining the locations and phases (red, yellow, and green lights) of each traffic signal, as depicted in Figure 9. In the simulations, the same signal plans were implemented during weekend peak hours as well. Subsequently, the routes of public transportation lines within the sample area were incorporated, along with the placement of stops and the frequency of service. In the following steps, these transportation routes, stop locations, and service frequencies were systematically added to the simulation. In the final stage, the hourly traffic flow for each road segment was established in the simulation, with average speeds reflecting the vehicle composition. Sustainability 2025,17, 8948 23 of 38 Figure 9. Visualization of the traffic light configuration in the simulation. Scenario preparation is essential for evaluating traffic performance under varying traffic conditions. In this context, multiple scenarios are developed using a calibrated and validated simulation model. By employing a load-balancing algorithm designed to distribute traffic load more evenly, simulations of load-balancing applications are conducted for historical dates and times within the modeled area. Each scenario aims to compare the algorithm’s effectiveness in balancing traffic and its responsiveness to different traffic situations. These scenarios incorporate various traffic conditions from different times, including both peak and off-peak hours. The analyses will yield insights into the algorithm’s efficacy in managing traffic flow and assessing potential changes in traffic performance. For the analysis, traffic conditions during peak hours over a ten-day period in 2023 were examined. A dataset was prepared to select these days, taking into account factors such as weekdays versus weekends, seasonal variations (summer versus winter), whether schools were in session, and differing weather conditions. The identified peak hour periods for the city are as follows: on weekdays, the morning peak spans from 07:30 to 08:30, while the evening peak occurs from 17:15 to 18:15. On weekends, the morning peak is from 11:00 to 12:00, and the evening peak is from 18:00 to 19:00. A detailed overview of the included days and their corresponding conditions can be found in Table 8. Table 8. The dates used to create the simulations. Date Day Season Special Condition School Weather Conditions 1 January 2023 Sunday Winter New Year (Holiday) Closed 13.5 ◦C, Rain 22 January 2023 Sunday Winter None Closed 12.0 ◦C, Sunny 16 January 2023 Monday Winter None Open 12.0 ◦C, Rain 10 June 2023 Saturday Summer None Closed 20.0 ◦C, Cloudy 16 June 2023 Friday Summer None Open 25.0 ◦C, Sunny 9 July 2023 Sunday Summer None Closed 21.0 ◦C, Partially Cloudy 25 July 2023 Tuesday Summer None Open 23.0 ◦C, Sunny 15 August 2023 Tuesday Summer Assumption of Mary (Holiday) Open 25.0 ◦C, Sunny 11 November 2023 Saturday Winter None Closed 18.0 ◦C, Rain 22 November 2023 Wednesday Winter None Open 15.0 ◦C, Sunny Sustainability 2025,17, 8948 24 of 38 In the Vigo study, vehicle volume and speed data were collected using Bluetooth sensors deployed at multiple locations across the network. This approach correlated detected vehicles rather than relying solely on counts at individual road sections or on origin–destination matrices derived from start and end zones. The Bluetooth sensors were assumed to operate with an accuracy of 30%, a value periodically updated by the Traffic Management Centre of Vigo through comparisons with inductive loop measurements on main roads. Accordingly, the measured volume data were scaled by a factor of 3.33 to estimate total traffic. In addition, a ± 5% normally distributed error term was added to ensure randomness. These route-based data were subsequently allocated to individual road segments, allowing for the determination of total hourly traffic volume for each segment. The O-D demand matrix used in the simulation conducted for 16 January 2023, at 17:15, is presented in Table 9, while the O-D demand matrix used in the simulation conducted for 10 June 2023, at 18:00, is presented in Table 10. In both tables, the rows represent origins and the columns represent destinations. In both tables, the diagonal cells in the OD matrix were marked with ‘–’ to indicate that trips with the exact origin and destination are not considered, since no travel demand is defined in such cases. Table 11 lists the segment represented by each letter. Table 9. The O-D matrix used in the simulation conducted for 16 January 2023—17:15. Zone A B C D E F G H I J K L M N O P R S T A - 303 303 181 209 209 73 209 73 309 309 309 309 B 163 - 829 503 14 14 112 112 126 112 112 112 C 163 600 - 619 56 56 14 112 112 126 112 112 112 D 163 600 600 - 559 256 256 303 303 303 418 174 335 174 194 194 E 97 359 359 359 - 256 256 303 303 303 363 128 256 120 194 194 F 58 150 150 89 - 312 67 46 312 65 67 67 G 600 260 260 155 285 - 493 117 100 530 169 117 117 H 97 359 239 239 239 285 426 - 303 303 363 128 218 120 139 139 I 68 239 239 239 239 218 239 - 529 443 221 216 207 241 241 J 83 239 239 239 239 270 239 635 - 547 270 270 253 294 294 K 83 329 239 239 239 170 239 635 635 - 170 170 253 1056 1056 L 270 270 229 - 270 229 229 229 M 493 493 229 229 - 229 229 229 N 83 170 170 762 170 170 - 762 762 O 186 460 460 - 859 859 492 306 P 186 239 239 239 239 239 635 635 635 460 952 - 1956 492 306 R 186 239 239 239 239 239 635 635 635 460 952 1587 - 492 306 S 186 859 859 859 - 306 T 550 550 550 550 - Table 10. The OD matrix used in the simulation conducted for 10 June 2023—18:00. Zone A B C D E F G H I J K L M N O P R S T A - 87 126 203 116 30 163 163 163 163 B 117 - 576 539 257 37 37 135 145 145 40 87 87 36 120 120 C 117 319 - 539 257 37 37 135 145 145 12 37 25 120 120 D 117 319 396 - 296 116 116 180 180 180 170 45 116 30 159 159 E 64 172 221 283 - 116 116 180 180 180 170 45 116 30 159 159 F 69 83 83 90 90 - 153 15 21 21 19 2 153 G 106 124 124 124 69 160 - 263 30 30 27 3 230 H 101 160 153 134 193 160 236 - 180 180 170 45 77 30 159 159 I 37 98 126 160 160 74 160 - 372 365 94 38 56 309 309 J 37 98 126 160 216 132 160 492 - 515 54 54 79 436 436 K 37 98 126 160 160 156 160 492 492 - 156 156 156 755 755 L 46 209 209 96 - 209 96 96 96 M 106 124 124 124 69 209 209 30 30 96 203 - 96 96 96 N 46 156 156 319 156 156 - 319 319 O 106 193 193 - 722 722 742 636 P 145 98 126 160 160 160 492 492 550 59 59 59 193 - 1477 742 636 R 145 98 126 160 160 160 492 492 550 59 59 193 193 1485 - 742 636 S 106 722 722 722 - 636 T 566 566 566 566 - Sustainability 2025,17, 8948 25 of 38 Table 11. Name of the zone represented by each letter. Letter Zone APraza de América BMonumento Á Xente do Mar CAv. de Castelao D VG-20 E Coia FAv. da Florida, 197–191 GRúa Padre Seixas HA Florida IRúa de Citroën JRúa dos Olimpícos KAv. do Alcalde Portanet LAv. do Fragoso MRúa Martín Echegaray, 1 NAv. do Fragoso OAv. de Balaídos, 2 PAv. do Alcalde Portanet, 1 RAv. de Castrelos, 236 SAv. de Antonio Palacios, 54 T Camiño Freixeiro For traffic simulations to yield reliable and accurate results, it is essential to conduct the processes of calibration and validation meticulously. Calibration involves fine-tuning the model to ensure it aligns with real-world data. During this phase, model parameters, such as vehicle speeds and signal timings, are adjusted based on observed traffic conditions. Real traffic data is collected and utilized to construct the model. Validation, on the other hand, evaluates the accuracy of the calibrated model by testing it against various datasets. Both calibration and validation are critical for ensuring the model’s realism and dependability, thus enabling decision-makers to formulate strategies based on credible and trustworthy results. During the calibration process, distinct scenarios were established for morning and evening peak hours over 10 days, with vehicle volumes and speeds allocated to various scenarios for each day. The simulation runtime was set to 3600 s, accompanied by a 900-s warm-up period. Each scenario was executed with 48 random seeds and run 10 times to minimize variability, using the average results for analysis. The model calibration was conducted with consideration for driver behaviors. Since the considered road network consists of urban roads, the calibration was performed based on the parameters of the Wiedemann 74 car-following model [63]. The Wiedemann 74 model has three parameters: the average standstill distance, the additive part of the safety distance, and the multiplicative part of the safety distance. During the calibration process, the values of these three parameters were adjusted to ensure that the traffic volumes and average speed values on each link matched the observed values. As a result of the calibration, the parameters were set as follows: average standstill distance = 2 m, additive part of safety distance = 2, and multiplicative part of safety distance = 3. Over the course of the ten days analyzed, a total of 20 primary scenarios—comprising one-hour intervals during both morning and evening peak times—were executed in the simulations for the validation process. The average traffic flow speeds for each road segment were validated using the GEH (Geoffrey E. Havers) statistic. The GEH statistic is a widely recognized non-linear metric in traffic engineering that compares volume values derived from model data with those observed in real-world data. A GEH value of less than 5 indicates a well-calibrated model, which is interpreted as a discrepancy of 0.5 vehicles per hour. Feldman provides the formula used to calculate the GEH value as Equation (8) [64]. GEH =s2(Qmodel −Qobservation)2 (Qmodel +Qobservation)(8) Sustainability 2025,17, 8948 32 of 38 predicting near-future conditions. This capability is especially relevant in urban settings that experience frequent demand fluctuations due to temporal, meteorological, or eventrelated factors. One limitation of the conducted study is that vehicle volume and speed data were obtained from Bluetooth sensors. The information that Bluetooth sensors can detect approximately 30% of the actual traffic volume was obtained from previous studies carried out by the Municipality of Vigo. In the current study, vehicle volumes were scaled based on this value by multiplying by 3.33, and a randomly varying error term of ± 5% was added. It was considered that a 30% sampling rate would represent the actual average speed with high accuracy ( ± 10.6% for a population of 200 vehicles and ± 8.7% for a population of 300 vehicles), and the values obtained from the sensors were used in the modeling. Using alternative detection methods to obtain vehicle volumes and average speeds with higher accuracy (considering the vehicle types as well) would increase the reliability of the study. Another limitation of this study is that the LSTM models were developed considering traffic conditions within a two-hour historical time window. Extending the time horizon and incorporating exogenous/spatial features such as weather, holidays, events, or neighboring link conditions could further improve the performance of the LSTM models. However, each additional parameter increases model complexity and the computational load required. In particular, under a 100% compliance rate scenario, all vehicles in the city would communicate with the TCC, and repeated calculations would be required for each route. Therefore, the number of parameters included in the model has been kept as low as possible. Consequently, a model reflecting the trade-off between accuracy and simplicity was established, with simplicity dominating. Nevertheless, the LSTM models achieve R 2 values above 0.9, demonstrating that accuracy is well beyond merely acceptable. Adding more parameters to the LSTM models could further enhance performance. In future studies, additional parameters could be incorporated into LSTM models to improve prediction performance. However, the practical applicability of the model should also be considered in such cases. A further limitation of this study is that pedestrian movements and trips made with non-motorized modes, such as micromobility systems, were not considered in the microsimulation. The traffic data obtained from Vigo Municipality does not include pedestrian activity or vehicles operating at low speeds, such as micromobility modes. Pedestrian interactions with motorized traffic primarily occur when crossing the street, typically during the red signal phase for motor vehicles. Given that the actual signal phases and timings used in the city were incorporated into the simulation, one may argue that pedestrian traffic was, to some extent, implicitly taken into account. However, including pedestrian and non-motorized vehicle traffic in the simulation would enable a much more realistic representation of traffic conditions. This, however, would proportionally increase the computational burden. Therefore, the objectives of the simulation should be carefully defined, and an appropriate balance should be struck between simplicity and accuracy. The findings are particularly encouraging for cities like Vigo, which feature irregular urban layouts and varied traffic patterns. The scalability of this approach indicates its potential applicability in a broader range of contexts. Similar frameworks could be modified to suit other urban centers with distinct traffic dynamics, infrastructure configurations, and varying levels of data availability. Future research may focus on adapting the algorithm for larger and more complex metropolitan networks, or on integrating multimodal transportation systems, including public transit, bicycles, and micromobility options. Additionally, while this study predominantly examined passenger vehicles, broadening the model to include freight vehicles, ride-sharing systems, and connected autonomous vehicles (CAVs) would yield a more thorough assessment of network-wide performance Sustainability 2025,17, 8948 33 of 38 in mixed traffic scenarios. By incorporating V2V and V2I communications into the loadbalancing framework, the responsiveness and efficiency of highly interconnected urban environments could be significantly enhanced. The results of this paper underscore the dual importance in real-life settings of the algorithm presented here: that it improves traffic network operation, as well as its sustainability. Through the redistribution of demand and the reduction in congestion, the approach avoids unnecessary fuel consumption and CO 2 emissions, thus contributing to a push for sustainable traffic systems. The approach can also synchronize with other innovative mobility modes, such as electrification and shared mobility, to create a more sustainable urban mobility ecosystem. From a policy perspective, such decision support can serve as a foundation for various planning processes. The results clearly demonstrate that traffic load balancing, when integrated with predictive analytics and simulation-based validation, presents a feasible solution for optimizing urban traffic. This approach not only helps alleviate congestion and mitigate environmental effects but also aligns with the broader goals of sustainable urban mobility and intelligent transportation systems. In conclusion, the identification of congested road segments and the application of load-balancing strategies have yielded positive outcomes regarding both traffic network performance and emission reduction. This methodology could be effectively implemented not just in the city of Vigo but also in various other cities and countries across different scales. Furthermore, the study could be expanded to encompass not only passenger vehicles but also other modes of transportation and their integration. The findings suggest that merging load balancing with intelligent transportation systems can significantly enhance urban quality of life while lessening the environmental impact associated with transportation. 7. Conclusions This study proposed and evaluated a traffic load-balancing algorithm designed for the city of Vigo, Spain, as part of the IN2CCAM Horizon 2020 project. The algorithm incorporated short-term traffic predictions, primarily generated using LSTM models, alongside a dynamic routing mechanism intended to redistribute traffic demand and optimize network performance. Results from various simulation scenarios indicated improvements in average travel speed, reductions in delays and periods of standstill, and measurable decreases in pollutant emissions. The findings confirm that integrating predictive analytics with dynamic load balancing can yield significant advantages for urban traffic management. This approach not only facilitates a more efficient utilization of existing infrastructure but also enhances environmental sustainability by reducing fuel consumption and emissions. Additionally, the methodology offers practical value for policymakers and local authorities, serving as a decision-support tool to inform the development of more adaptive and inclusive mobility strategies. Future research should concentrate on extending the application of the proposed algorithm to various cities with differing traffic and infrastructure characteristics, as well as incorporating more advanced prediction models such as GNNs. These advancements will enhance the approach’s adaptability and robustness, contributing to the establishment of smarter and more sustainable urban transportation systems. Author Contributions: Conceptualization, S.D. and S.A.; methodology, S.D. and S.A.; software, S.D., ˙ I.M.U. and O.D.; validation, S.D. and S.A.; formal analysis, S.D. and S.A.; investigation, S.D. and S.A.; resources, S.D.; data curation, S.D. and S.A.; writing—original draft preparation, S.D., ˙ I.M.U. and O.D.; writing—review and editing, S.D.; visualization, S.D.; supervision, S.D. and S.A.; project administration, S.D.; funding acquisition, S.D. All authors have read and agreed to the published version of the manuscript. Sustainability 2025,17, 8948 34 of 38 Funding: This research was funded by European Union’s Horizon CL5 2022-D6-01-04 research and innovation program, grant number 101076791. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: The datasets presented in this article are not readily available because they belong to Municipality of Vigo. Requests to access the datasets should be directed to Municipality of Vigo. Acknowledgments: During the preparation of this manuscript, the authors used Grammarly for the purposes of improving the language of this text. The authors have reviewed and edited the output and take full responsibility for the content of this publication. Conflicts of Interest: The authors declare no conflicts of interest. Abbreviations The following abbreviations are used in this manuscript: ABR AdaBoost Regressor AI Artificial Intelligence ANNs Artificial Neural Networks API REST Application Programming Interface Representational State Transfer APP Application ARIMA Autoregressive Integrated Moving Average BiLSTM Bidirectional Long Short-Term Memory CAV Connected Autonomous Vehicle CCAM Cooperative, Connected, and Automated Mobility CLMGCN Contrastive Learning Multi-Graph Convolution Network CNG Compressed Natural Gas CNNs Convolutional Neural Networks CO Carbon monoxide CO2Carbon dioxide DBNs Deep Belief Networks DGSTN Dynamic Graph Spatial–Temporal Neural Network DRL Deep Reinforcement Learning ETR Extra Tree Regressor GBs GigaBytes GBR Gradient Boosting Regressor GDPR General Data Protection Regulation GEH Geoffrey E. Havers GNN Graph Neural Network GPS Global Positioning System GRNN General Regression Neural Network GRUs Gated Recurrent Units HC Hydrocarbons I-55 Interstate 55 IN2CCAM Enhancing Integration and Interoperability of CCAM eco-system IVI In Vehicle Information JSON JavaScript Object Notation KNN K-Nearest Neighbor LPG Liquefied Petroleum Gas LSTM Long-Short Term Memory MAC Medium Access Control MAPE Mean Absolute Percentage Error MLP Multilayer Perceptron Sustainability 2025,17, 8948 35 of 38 MLR Multiple Linear Regression MSE Mean Squared Error NOxNitrogen oxide O-D Origin–Destination RF Random Forest RFR Random Forest Regressor RL Reinforced Learning RNNs Recurrent Neural Networks RSU Road Side Unit SAEs Stacked Autoencoders SD Standard Deviation SE Standard Error SO System Optimum SRD-LB Smart Route Determining with Load Balancing SSR Sum of Squared Residuals SVMs Support Vector Machines SVR Support Vector Regression TCC Traffic Control Centre UE User Equilibrium V2I Vehicle-to-Infrastructure V2V Vehicle-to-Vehicle V2X Vehicle-to-Everything VMP Variable Message Panel Appendix A This glossary defines the symbols and variables used in the path-enumeration procedure that builds candidate routes from a source node to a target node in a positively weighted graph G = (V, E, d, τ ), where d: E →R+ is edge distance, and τ : E →R+ is edge duration. The algorithm performs a depth-bounded, cycle-free DFS and returns a set P of complete paths, each annotated with its total distance (and, optionally, duration). Variable Type Definition/Role Set/Computed by Example ENUMERATE_PATHS Procedure/function Top-level routine that initializes the search and returns P; calls DFS(s, D_max, [], 0). Defined by the algorithm; orchestrates recursion and final return. Returns P = { (π1, d(π1)), (π2, d(π2)), . . . } s Vertex (node) Source node where path search starts. Given as input. s = A D_max Integer (≥0) Maximum number of edges (hop budget) allowed for any path explored. Given as input; controls search breadth. D_max = 3 depth_left Integer (0. . .D_max) Remaining hop budget at the current recursion level. Starts at D_max; decremented by 1 on each recursive step. depth_left = 2 cur Vertex (node) The node currently being expanded during DFS. Initialized as s; updated to nxt on recursion. cur = B nxt Vertex (node) A neighbor of cur reached by an outgoing edge (cur, nxt). Enumerated from outgoing edges of cur. nxt = C E Set of edges (subset of V ×V) Edge set of the graph; may be directed or represented with symmetric pairs for undirected graphs. Defined by the network/ road topology. E = {(A,B), (B,D), (A,C), (C,D)} d_e Positive real (R+)Distance (weight) of the edge currently traversed. d_e = d(cur, nxt). d_e = 3.0 (km) d Weight function d: E →R+ Maps each edge to its distance (or cost) used in accumulation and pruning. Provided with the graph; queried as d(u,v). d(A,B) = 5 t Vertex (node) Target (destination) node. When nxt == t, a complete path is recorded. Given as input. t = D π(pi) Path (ordered list of edges or vertices) A complete route from s to t discovered by DFS. π= path ⊕[(cur, nxt)] when nxt == t (⊕= concatenate). π= [(A,B), (B,D)] path Partial path The path built so far along the current recursion branch (no repeated vertices). Extended on recursion: path’ = path ⊕[(cur, nxt)]. path = [(A,B)] Sustainability 2025,17, 8948 36 of 38 Variable Type Definition/Role Set/Computed by Example dist_acc Positive real (R+)Accumulated distance along the current partial path. Updated as dist_acc’ = dist_acc + d_e. dist_acc = 5.0 P List/set of records Collection of discovered complete paths with annotations (e.g., total distance). Append (π, dist_acc + d_e) whenever nxt == t. P = [([(A,B),(B,D)], 9.0), ([(A,C),(C,D)], 5.0)] References 1. Macioszek, E.; Kurek, A. Road Traffic Distribution on Public Holidays and Workdays on Selected Road Transport Network Elements. Transp. Probl. 2021,16, 127–138. [CrossRef] 2. Wardrop, J.G. Some theoretical aspects of road traffic research. Proc. Inst. Civ. Eng. 1952,1, 325–378. [CrossRef] 3. Meghana, B.S.; Kumari, S.; Pushphavathi, T.P. Comprehensive traffic management system: Real-time traffic data analysis using RFID. In Proceedings of the International conference of Electronics, Communication and Aerospace Technology (ICECA), Coimbatore, India, 20–22 April 2017. [CrossRef] 4. Gu, Y.; Xie, L. Fast sensitivity analysis approach to assessing congestion induced wind curtailment. IEEE Trans. Power Syst. 2014, 29, 101–110. [CrossRef] 5. Sun, N.; Shi, H.; Han, G.; Wang, B.; Shu, L. Dynamic path planning algorithms with load balancing based on data prediction for smart transportation systems. IEEE Access 2020,8, 15907–15922. [CrossRef] 6. World Health Organization. Health Service Continuity Planning for Public Health Emergencies: A Handbook for Health Facilities: Interim Version for Field Testing; World Health Organization: Geneva, Switzerland, 2021. 7. Wang, S.; Djahel, S.; McManis, J. A multi-agent based vehicles re-routing system for unexpected traffic congestion avoidance. In Proceedings of the 17th International IEEE Conference on Intelligent Transportation Systems (ITSC), Qingdao, China, 8–11 October 2014. [CrossRef] 8. Chughtai, O.; Badruddin, N.; Awang, A.; Rehan, M. Congestion-aware and traffic load balancing scheme for routing in WSNs. Telecommun. Syst. 2016,63, 481–504. [CrossRef] 9. Ee, C.T.; Bajcsy, R. Congestion control and fairness for many-to-one routing in sensor networks. In Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems, Baltimore, MD, USA, 3–5 November 2004. [CrossRef] 10. Dijkstra, E.W. A note on two problems in connexion with graphs. Numer. Math. 1959,1, 269–271. [CrossRef] 11. Hart, P.; Nilsson, N.; Raphael, B. A formal basis for the heuristic determination of minimum cost paths. IEEE Trans. Syst. Sci. Cybern. 1968,4, 100–107. [CrossRef] 12. Zhao, J.; Gao, Y.; Bai, Z.; Wang, H.; Lu, S. Traffic speed prediction under non-recurrent congestion: Based on LSTM method and BeiDou navigation satellite system data. IEEE Intell. Transp. Syst. Mag. 2019,11, 70–81. [CrossRef] 13. Wang, Y.; Zhou, H.; Wang, Y. Mobile robot dynamic path planning based on improved genetic algorithm. In Proceedings of the International Conference on Green Energy and Sustainable Development (GESD 2017), Chongqing, China, 27–28 May 2017. [CrossRef] 14. Clausen, U.; Brueggenolte, M.; Kirberg, M.; Besenfelder, C.; Poeting, M.; Gueller, M. Agent-based simulation in logistics and supply chain research: Literature review and analysis. In Proceedings of the Interdisciplinary Conference on Production, Logistics and Traffic, Dortmund, Germany, 27–28 March 2019. [CrossRef] 15. Sutton, R.S.; Barto, A.G. Reinforcement Learning: An Introduction, 2nd ed.; MIT Press: Cambridge, MA, USA, 2018. 16. Akerele, J.I.; Uzoka, A.; Ojukwu, P.U.; Olamijuwon, O.J. Optimizing traffic management for public services during high-demand periods using cloud load balancers. Comput. Sci. IT Res. J. 2024,5, 2594–2608. [CrossRef] 17. Zhang, J.; Xi, K.; Luo, M.; Chao, H.J. Dynamic hybrid routing: Achieve load balancing for changing traffic demands. In Proceedings of the 2014 IEEE 22nd International Symposium of Quality of Service (IWQoS), Hong Kong, China, 26–27 May 2014. [CrossRef] 18. Fortz, B.; Thorup, M. Optimizing OSPF/IS-IS weights in a changing world. IEEE J. Sel. Areas Commun. 2002,20, 756–767. [CrossRef] 19. Zhang, J.; Xi, K.; Zhang, L.; Chao, H. Optimizing network performance using weighted multipath routing. In Proceedings of the 21st International Conference on Computer Communications and Networks (ICCCN), Munich, Germany, 30 July–2 August 2012. [CrossRef] 20. Li, J.; Luo, G.; Cheng, N.; Yuan, Q.; Wu, Z.; Gao, S.; Liu, Z. An end-to-end load balancer based on deep learning for vehicular network traffic control. IEEE Internet Things J. 2018,6, 953–966. [CrossRef] 21. Kodialam, M.; Lakshman, T.V.; Orlin, J.B.; Sengupta, S. Oblivious routing of highly variable traffic in service overlays and IP backbones. IEEE/ACM Trans. Netw. 2009,17, 459–472. [CrossRef] 22. Zhang, C.; Liu, Y.; Gong, W.; Kurose, J.; Moll, R.; Towsley, D. On optimal routing with multiple traffic matrices. In Proceedings of the 24th Annual Joint Conference of the IEEE Computer and Communications Societies INFOCOM 2005, Miami, FL, USA, 13–17 March 2005. [CrossRef] Sustainability 2025,17, 8948 37 of 38 23. Chan, C.; Kuncheria, A.; Macfarlane, J. Simulating the impact of dynamic re-routing on metropolitan-scale traffic systems. ACM Trans. Model. Comput. Simul. 2023,33, 1–29. [CrossRef] 24. Liu, R.; Liu, H.; Kwak, D.; Xiang, Y.; Borcea, C.; Nath, B.; Iftode, L. Balanced traffic routing: Design, implementation, and evaluation. Ad Hoc Netw. 2016,37, 14–28. [CrossRef] 25. Roughgarden, T. Selfish Routing and the Price of Anarchy; MIT Press: Cambridge, MA, USA, 2005. 26. Wilkie, D.; Van den Berg, J.; Lin, M.; Manocha, D. Self-aware traffic route planning. In Proceedings of the AAAI Conference on Artificial Intelligence, San Francisco, CA, USA, 7–11 August 2011. [CrossRef] 27. Liang, Z.; Wakahara, Y. Real-time urban traffic amount prediction models for dynamic route guidance systems. EURASIP J. Wirel. Commun. Netw. 2014,2014, 85. [CrossRef] 28. Guo, J.; Huang, W.; Williams, B.M. Adaptive Kalman Filter Approach for Stochastic Short-Term Traffic Flow Rate Prediction and Uncertainty Quantification. Transp. Res. Part C Emerg. Technol. 2014,43, 50–64. [CrossRef] 29. Williams, B.M.; Durvasula, P.K.; Brown, D.E. Urban Freeway Traffic Flow Prediction: Application of Seasonal Autoregressive Integrated Moving Average and Exponential Smoothing Models. Transp. Res. Rec. J. Transp. Res. Board 1998,1644, 132–141. [CrossRef] 30. Lee, S.; Fambro, D.B. Application of Subset Autoregressive Integrated Moving Average Model for Short-Term Freeway Traffic Volume Forecasting. Transp. Res. Rec. J. Transp. Res. Board 1999,1678, 179–188. [CrossRef] 31. Karlaftis, M.G.; Vlahogianni, E.I. Statistical Methods Versus Neural Networks in Transportation Research: Differences, Similarities and Some Insights. Transp. Res. Part C Emerg. Technol. 2011,19, 387–399. [CrossRef] 32. Sun, B.; Cheng, W.; Goswami, P.; Bai, G. Flow-Aware WPT K-Nearest Neighbours Regression for Short-Term Traffic Prediction. In Proceedings of the 2017 IEEE Symposium on Computers and Communications (ISCC), Heraklion, Greece, 3–6 July 2017; pp. 48–53. [CrossRef] 33. Su, H.; Zhang, L.; Yu, S. Short-Term Traffic Flow Prediction Based on Incremental Support Vector Regression. In Proceedings of the Third International Conference on Natural Computation (ICNC 2007), Haikou, Hainan, China, 24–27 August 2007; pp. 640–645 . [CrossRef] 34. Tedjopurnomo, D.A.; Bao, Z.; Zheng, B.; Choudhury, F.M.; Qin, A.K. A Survey on Modern Deep Neural Network for Traffic Prediction: Trends, Methods and Challenges. IEEE Trans. Knowl. Data Eng. 2020,34, 1544–1561. [CrossRef] 35. Hochreiter, S. The Vanishing Gradient Problem During Learning Recurrent Neural Nets and Problem Solutions. Int. J. Uncertain. Fuzziness Knowl. Based Syst. 1998,6, 107–116. [CrossRef] 36. Cho, K.; Van Merriënboer, B.; Bahdanau, D.; Bengio, Y. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. In Proceedings of the SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, Doha, Qatar, 25 October 2014; pp. 103–111. [CrossRef] 37. Cho, K.; Van Merriënboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations Using RNN Encoder-Decoder for Statistical Machine Translation. arXiv 2014, arXiv:1406.1078. [CrossRef] 38. Cornacchia, G.; Lemma, L.; Pappalardo, L. Polaris: Popularity-Based Traffic Routing Algorithm. arXiv 2024, arXiv:2406.05388. [CrossRef] 39. Bentley, C.D.B.; Marsh, S.; Carvalho, A.R.R.; Kilby, P.; Biercuk, M. J Quantum computing for transport optimization. arXiv 2022, arXiv:2206.07313. [CrossRef] 40. Rath, M.; Chatterjee, J.M. Strength Optimized Weight Balancing for Traffic Management in Vehicular Ad-hoc Networks. Int. J. Bus. Data Commun. Netw. 2025,20, 1–19. [CrossRef] 41. Haider, S.E.; Khan, M.F.; Saeed, Y. Adaptive Load Balancing Approach to Mitigate Network Congestion in VANETS. Computers 2024,13, 194. [CrossRef] 42. Upadhyay, U.; Maheshwari, S.; Sikka, G. SRD-LB: A Smart Route Determining algorithm with Load Balancing for Smart Cities in India. Nat. Volatiles Essent. Oils 2021,8, 13206–13215. Available online: https://www.nveo.org/index.php/journal/article/view/ 4269/3516 (accessed on 10 September 2025). 43. Wang, C.; Peeta, S.; Wang, J. Incentive-based decentralized routing for connected and autonomous vehicles using information propagation. Transp. Res. Part B Methodol. 2021,149, 138–161. [CrossRef] 44. Bagloee, S.A.; Sarvi, M.; Patriksson, M.; Rajabifard, A. A mixed user-equilibrium and system-optimal traffic flow for connected vehicles stated as a complementarity problem. Comput. Aided Civ. Infrastruct. Eng. 2017,32, 562–580. [CrossRef] 45. Utku, A. Derin Ö˘grenme Tabanlı Trafik Yo˘gunlu˘gu Tahmini: ˙ Istanbul ˙ Için Bir Vaka Çalı¸sması. Düzce Üniversitesi Bilim Ve Teknol. Derg. 2023,11, 1584–1598. [CrossRef] 46. Ta¸s, ˙ I.Ç.; Müngen, A.A. Yapay Sinir A˘gları ve Destek Vektör Makineleri Yöntemleri ile Bölgesel Trafik Yo˘gunluk Tahmini. Adıyaman Üniversitesi Mühendislik Bilim. Derg. 2021,8, 378–390. [CrossRef] 47. Ta¸s, N.; Sezen, B. Yapay Sinir A˘gları ile Trafik Yo˘gunlu˘gu Tahmini. Afyon Kocatepe Üniversitesi Sos. Bilim. Derg. 2020,22, 1020–1034. [CrossRef] Sustainability 2025,17, 8948 38 of 38 48. Ma, X.; Tao, Z.; Wang, Y.; Yu, H.; Wang, Y. Long short-term memory neural network for traffic speed prediction using remote microwave sensor data. Transp. Res. Part C Emerg. Technol. 2015,54, 187–197. [CrossRef] 49. Polson, N.G.; Sokolov, V.O. Deep learning for short-term traffic flow prediction. Transp. Res. Part C Emerg. Technol. 2017,79, 1–17. [CrossRef] 50. Chen, K.; Chen, F.; Lai, B.; Jin, Z.; Liu, Y.; Li, K.; Wei, L.; Wang, P.; Tang, Y. Dynamic spatiotemporal graph-based CNNs for traffic flow prediction. IEEE Access 2020,8, 185136–185145. [CrossRef] 51. Kashyap, A.A.; Raviraj, S.; Devarakonda, A.; Nayak, K.; Santosh, K.V.; Bhat, S.J. Traffic flow prediction models—A review of deep learning techniques. Cogent Eng. 2021,9, 1–24. [CrossRef] 52. Brimos, P.; Karamanou, A.; Kalampokis, E.; Tarabanis, K. Graph Neural Networks and Open-Government Data to Forecast Traffic Flow. Information 2023,14, 228. [CrossRef] 53. Jiang, M.; Liu, Z. Traffic Flow Prediction Based on Dynamic Graph Spatial-Temporal Neural Network. Mathematics 2023,11, 2528. [CrossRef] 54. Francies, M.L.; Khalil, A.T.; Amer, H.M.; Ata, M.M. A framework for continual learning in real-time traffic forecasting utilizing spatial–temporal graph convolutional recurrent networks. Complex Intell. Syst. 2025,11, 420. [CrossRef] 55. Abdellah, A.R.; Abdelmoaty, A.; Ateya, A.A.; El-Latif, A.A.A.; Muthanna, A.; Koucheryavy, A. Accurate V2X traffic prediction with deep learning architectures. Front. Artif. Intell. Sec. Mach. Learn. Artif. Intell. 2025,8, 1565287. [CrossRef] 56. Guo, K.; Tian, D.; Hu, Y.; Sun, Y.; Qian, Z.; Zhou, J.; Gao, J.; Yin, B. Contrastive learning for traffic flow forecasting based on multi graph convolution network. IET Intell. Transp. Syst. 2024,18, 290–301. [CrossRef] 57. Ji, M.; Wu, Q.; Fan, P.; Cheng, N.; Chen, W.; Wang, J.; Letaief, K.B. Graph Neural Networks and Deep Reinforcement Learning Based Resource Allocation for V2X Communications. arXiv 2024, arXiv:2407.06518. [CrossRef] 58. Blogg, M.; Semler, C.; Hingorani, D.; Troutbeck, R. Travel time and origin-destination data collection using Bluetooth MAC address readers. In Proceedings of the 33rd Australasian Transport Research Forum (ATRF), Canberra, Australia, 29 September–1 October 2010. 59. Margreiter, M.; Busch, F.; Carstensen, C. The Evolution of Bluetooth Detection Rates. In Proceedings of the 98th Annual Meeting of the Transportation Research Board, Washington, DC, USA, 13–17 January 2019. Available online: https://mediatum.ub.tum. de/1476803 (accessed on 10 September 2025). 60. Gheorghiu, R.A.; Iordache, V.; Cormo¸s, A.C. Analysis of the possibility to detect road vehicles via Bluetooth technology. Sensors 2021,21, 7281. [CrossRef] [PubMed] 61. Zinner, J. Bluetooth-Based Travel time Monitoring: An Assessment of Performance and Potential Applications. Master’s Thesis, Georgia Institute of Technology, Atlanta, Georgia, 2013. 62. Akçelik, R. Speed-Flow Models for Uninterrupted Traffic Facilities; Technical Report; Akcelik & Associates Pty Ltd.: Balwyn, VIC, Australia, 2003. Available online: https://www.sidrasolutions.com/media/119/download (accessed on 10 September 2025). 63. Wiedemann, R. Simulation des StraBenverkehrsflusses in Schriftenreihe des Tnstituts fir Verkehrswesen der Universitiit Karlsruhe; Universität Karlsruhe: Karlsruhe, Germany, 1974. 64. Feldman, O. The GEH Measure and Quality of the Highway Assignment Model; Association for European Transport and Contributors: London, UK, 2012; pp. 1–18. 65. Nouri, P.; Morency, C. Untangling the Impacts of Various Factors on Emission Levels of Light Duty Gasoline Vehicles; Technical Report n◦CIRRELT-2015-53; CIRRELT: Montréal, QC, Canada, 2015. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.