Modeling Air Pollution from Urban Transport and Strategies for Transitioning to Eco-Friendly Mobility in Urban Environments

Abstract

Urban air pollution caused by vehicular emissions remains one of the most pressing environmental challenges, negatively affecting both public health and climate processes. In Kokshetau, Kazakhstan, where electric vehicle (EV) adoption accounts for only 0.019% of the total fleet and charging infrastructure is nearly absent, reducing transport-related emissions requires short-term and cost-effective solutions. This study proposes an integrated approach combining urban ecology principles with computational modeling to optimize traffic signal control for emission reduction. An artificial neural network (ANN) was trained using intersection-specific traffic data to predict emissions of carbon monoxide (CO), nitrogen oxides (NO_x), sulfur dioxide (SO₂), and particulate matter (PM_2.5). The ANN was incorporated into a nonlinear optimization framework to determine traffic signal timings that minimize total emissions without increasing traffic delays. The results demonstrate reductions in emissions of CO by 12.4%, NO_x by 9.8%, SO₂ by 7.6%, and PM_2.5 by 10.3% at major congestion hotspots. These findings highlight the potential of the proposed framework to improve urban air quality, reduce ecological risks, and support sustainable transport planning. The method is scalable and adaptable to other cities with similar urban and environmental characteristics, facilitating the transition toward eco-friendly mobility and integrating data-driven traffic management into broader climate and public health policies.

1. Introduction

One of the most acute and multifaceted environmental problems facing modern megacities and urban agglomerations remains air pollution, largely caused by the transport sector. Apte et al. note that high levels of PM_2.5 reduce average life expectancy, especially in densely populated regions [1]. According to the European Environment Agency [2], cars are the main source of pollutants. The most dangerous of these are nitrogen oxides (NO_x), particulate matter (PM), and carbon dioxide (CO₂), which contribute to air quality deterioration, climate change, and increased morbidity. Grange et al. [3] noted that light-duty diesel vehicles, especially in low temperatures, lead to increased NO_x emissions, which creates serious problems for regions with colder continental climates.

However, in recent years, researchers have begun to focus not only on chemical but also on biological pollution of the atmosphere. The study “Composition Analysis of Airborne Microbiota in Outdoor and Indoor Based on Microbial Size” [4] shows that solid particles contain microorganisms and their extracellular vesicles, which can be transported over long distances and affect human health and environmental processes. The study found that nanoscale bio-bubbles (n-MEV) demonstrate higher mobility and stability compared to microbial cells (m-MB), and their concentration in urban air increases with increasing pollution and traffic flow. This indicates that atmospheric aerosols in megacities are a complex mixture of chemical and biological components that exacerbate the effects of pollution on the human body [4].

The adverse health effects associated with transport-related air pollution, including increased mortality, further highlight the need for cleaner urban mobility [5]. Data from urban areas such as Barcelona and Madrid illustrate how switching to electric vehicles (EVs) and reducing dependence on conventional transport can significantly reduce NO_x and PM levels while also reducing the concentration of biologically active aerosol particles [4,6]. While EVs curb tailpipe emissions in cities, system-wide outcomes depend on life-cycle factors (battery production/end-of-life and grid carbon intensity); hence, EVs are treated here as contextual background, whereas our quantitative results focus on the current fleet.

Globally, EV adoption has been spurred by various policy tools. Financial instruments such as subsidies and tax incentives, coupled with the expansion of charging infrastructure, have proven effective in markets across the United States, China, and the European Union [7]. In London, for example, the introduction of the Ultra Low Emission Zone (ULEZ) resulted in a measurable decline in NO₂ concentrations [8]. Norway’s experience demonstrates that charging infrastructure availability may play a more significant role in driving adoption than temporary fiscal incentives [9], whereas in the United States, direct subsidies have shown greater impact than tax credits in encouraging hybrid vehicle purchases [10].

Nonetheless, the path to widespread EV adoption is fraught with obstacles. High initial costs, limited access to charging points, and persistent range anxiety remain key deterrents—particularly in large urban areas [11]. Technological innovations like inductive wireless charging have been proposed to mitigate some of these challenges and improve EV usability in dense urban environments [12]. Beyond technological barriers, human behavior and perception play a vital role in sustainable transport choices. Steg and Gifford [13] argue that perceived quality of life significantly influences the willingness to switch to greener alternatives, and Bobeth and Kastner [14] emphasize that moral or ethical motivations can shape consumer decisions toward purchasing EVs.

Social equity is another essential consideration in the transition to clean mobility. Socioeconomic and cultural differences affect access to and acceptance of low-carbon technologies. Strambo et al. highlight social inequalities resulting from the low-carbon transition: cultural differences may hinder the adoption of new transport solutions [15]. Cai et al. and Varghese et al. found that charging infrastructure is often inaccessible to residents of socially disadvantaged areas [16,17]. Trust in institutions, coupled with opportunities for public participation, has also emerged as a critical factor in shaping public attitudes toward new technologies [18].

Planning EV infrastructure requires alignment with demographic and spatial realities. Optimization models that factor in population density and projected demand are increasingly utilized to determine ideal locations for charging stations [19,20]. Broader decision-making frameworks now incorporate economic, technical, and spatial criteria to ensure investments in charging infrastructure are both effective and equitable [21].

Efforts to build public trust are equally important. Educational campaigns, clear policy communication, and financial incentives have all been shown to positively shape public opinion and encourage EV adoption [22]. These activities often take place under the banner of smart city initiatives, which aim to integrate technology, sustainability, and social equity. Albino et al. presented the concept of smart city as a synthesis of technology, ecology, and social inclusion [23]. Caragliu et al. emphasized the importance of ICT in traffic management [24].

Urban form and spatial design also influence emission levels and transportation behavior. In Krakow, for instance, Bielińska-Dusza et al. [25] show that intelligent infrastructure has successfully encouraged shifts toward more sustainable mobility patterns. Research on urban sprawl highlights the environmental burden associated with automobile-dependent development models [26].

Recognizing the interplay between transportation, climate policy, and sustainable development, many scholars now call for cross-sectoral coordination and interdisciplinary strategies. These ensure that mobility transitions are not only technically viable but also socially and economically balanced [27]. In emerging economies like Kazakhstan, initiatives are underway to power EV infrastructure with renewable energy sources [28] and localize vehicle production as a means of reducing costs and increasing resilience [29]. Garmash et al. highlight how electric trucks are gaining attention as a way to reduce emissions in the urban logistics sector [30].

The transition to clean transportation is a complex, multi-level process. It requires technological innovation, social inclusion, competent infrastructure planning, and active public participation. Only through a comprehensive approach can real environmental and social benefits be achieved in cities.

Real-world examples underscore the value of comprehensive strategies. The complete electrification of Shenzhen’s bus fleet, made possible by centralized management and strategic investment, has led to tangible improvements in both environmental protection and transport efficiency [31]. In Europe, electric vehicle sharing services are being actively promoted as a solution to reducing private vehicle ownership; however, as the experience of France shows, insufficient regulation and inadequate urban planning can exacerbate inequality and create additional strain on infrastructure [32].

Analysis of various scenarios based on models confirms an important point: technological innovation alone does not guarantee sustainability.

To ensure truly sustainable urban mobility, it is necessary not only to develop the electrification of transport, but also to reduce dependence on private cars, invest in high-quality public transport, rethink spatial design, and eliminate socio-economic disparities. Otherwise, sustainable transport can only serve as a modernized continuation of highly unsustainable systems.

Within this broader transition, a substantial body of work has explored machine-learning approaches, especially Artificial Neural Networks (ANNs), for predicting and managing air pollution. The key insights from this literature are summarised below. In recent years ANN have become an integral tool in forecasting and analyzing air pollution. The nonlinear nature of the relationships between meteorological parameters, anthropogenic impacts, and the dynamics of pollutants makes traditional statistical methods such as Multiple Linear Regression (MLR) less effective in solving environmental modeling problems. Modern research demonstrates that neural networks are able not only to identify hidden dependencies in data, but also to provide much more accurate prediction of pollutant concentrations in a wide variety of environments—from urban atmospheric layers to indoor areas.

The study by Ozdemir and Taner was one of the first to demonstrate the advantages of neural networks in modeling PM10 concentrations taking into account meteorological factors. A comparison of ANN and MLR revealed that the neural network is much better at detecting nonlinear relationships between temperature, humidity, wind speed, and suspended particle content, providing more accurate predictions [33]. Similar results were obtained by other authors. Thus, Shams and co-authors in the journal Environmental Pollution confirmed that ANNs are superior to linear models in predicting air quality indices, especially in conditions of complex urban factors, including traffic activity, building density and green areas. However, the researchers emphasize that the interpretability of such models is lower than that of traditional statistical approaches, which remains one of the key problems of using neural networks in ecology [34].

With the development of computing technologies, special attention has been paid to optimizing ANN architectures and learning algorithms. Gulati and coauthors compared several neural network variants—LM-ANN, BR-ANN, and SCG-ANN—when estimating PM_2.5 concentrations and concluded that the use of regularization methods such as Bayesian Regularization significantly increases the model’s resistance to data noise and seasonal fluctuations. These results confirm that the choice of learning algorithm and network structure plays a critical role in ensuring the quality of forecasts [35]. The latest research, including the work of Mohapatra and colleagues, indicates that further improvement of ANN should go towards creating hybrid systems where classical neural networks are combined with Bayesian methods and regularization algorithms that increase the reliability of the model [36].

One of the significant areas of development is the integration of neural networks with optimization methods. In the work of Zhang and co-authors, a hybrid system based on ANN and a genetic algorithm (GA) was proposed for managing the local indoor air quality index. This combination allows not only to predict the level of pollution, but also to control ventilation parameters, minimizing the concentration of harmful substances in the respiratory zone. This approach is especially relevant for modern energy-efficient buildings and smart climate management systems, where it is important to maintain optimal conditions with minimal energy consumption [37].

The use of neural networks in transport and urbanized systems also demonstrates high efficiency. Park and co-authors used ANN to predict the concentration of PM10 in Seoul subway stations, taking into account external weather conditions and train schedules. The model showed the ability to accurately determine the dynamics of pollution and helped to develop recommendations for the management of ventilation systems [38]. Maleki and colleagues conducted similar studies in one of the most polluted cities in the world—Ahvaz (Iran). Based on an annual series of data, ANN was able to successfully predict hourly pollutant concentrations and the AQI index even in extreme dust storms where traditional methods were losing accuracy [39].

The use of ANN in various climatic and urban conditions—from Turkey and Malaysia to India and Iran—shows the universality of the approach. However, most authors emphasize that each model requires adaptation to local data, since the composition of pollutants, the structure of emissions and climatic characteristics vary significantly. For example, the results obtained in the arid climate of Ahvaz cannot be directly used for temperate latitudes without recalibrating the input parameters.

Comparative studies, such as the work of Kamal and co-authors, where neural networks were first used to predict air quality in a tropical region, laid the foundation for the further development of ANN in ecology. Since then, the approach has evolved significantly, from simple single-layer models to multi-layer architectures with adaptive learning, regularization, and the ability to integrate with optimization methods. Today, neural networks are considered not only as a forecasting tool, but also as an element of active management of environmental processes [40].

An analysis of all the publications reviewed shows that artificial neural networks consistently demonstrate higher accuracy than linear models, especially when modeling complex nonlinear interactions between factors. However, the effectiveness of an ANN directly depends on the quality of the source data, correct preprocessing, and feature selection. Among the limitations are difficulties in interpretation, the risk of overfitting, and the need for significant computing resources. Nevertheless, the potential of ANNs in environmental forecasting is huge, especially when they are integrated with modern Explicable AI methods that allow us to reveal the contribution of individual factors to the final result.

Thus, a generalized analysis of works for the period from 2014 to 2025 [33,34,35,36,37,38,39,40] attests to the high efficiency and wide application possibilities of artificial neural networks in forecasting and managing atmospheric air quality. ANN has established itself as a flexible and reliable tool that is able to adapt to a variety of conditions and provide accurate forecasts with a minimum number of assumptions about the data structure. In the future, explicable and hybrid models are expected to be developed that combine the advantages of neural networks with physico-chemical modeling methods and optimization algorithms, which will make it possible to move from simple forecasting to active management of atmospheric and indoor air quality in the interests of sustainable development and public health. Building on these advances, the present study leverages an ANN-based predictor integrated with traffic-signal optimisation to quantify and mitigate transport emissions under Kokshetau’s local conditions.

In this context, empirical analysis of urban traffic is particularly important, as it allows us to assess the real efficiency of transport infrastructure and identify the main sources of pollution. This study includes six key intersections located in the central and busiest part of the city of Kokshetau. These intersections were selected based on high traffic intensity, diversity of traffic flows, and functional significance for the urban transport network. This approach ensures the representativeness of the sample and allows the data obtained to be considered a reflection of typical urban traffic conditions.

In this regard, it becomes evident that a sustainable transition to clean urban mobility is only possible under a set of critically important conditions:

• comprehensive analysis and accurate modeling of pollution sources;
• equitable and strategically justified distribution of charging infrastructure;
• integration of behavioral, cultural, and social factors into transport planning;
• active and long-term government support through economic and regulatory mechanisms;
• synchronization of transport development with energy and urban planning systems.

A comprehensive, structured approach is essential not only to reduce urban air pollution but also to lay the groundwork for a sustainable and inclusive urban future. Integrated frameworks that jointly model traffic-related emissions and implement dynamic mitigation remain scarce—particularly in mid-sized cities of developing regions such as Kokshetau, Kazakhstan.

To address this gap, a methodology is proposed that couples an artificial neural network (ANN) for urban transport-emission prediction with an optimized traffic-signal control strategy tailored to Kokshetau’s conditions. The six intersections analyzed were selected to reflect the diversity of the city’s network central high-volume nodes, residential areas with moderate flows, and peripheral links capturing both commercial and transit routes and enabling comparable emission assessments.

The novelty lies in integrating intelligent emission modeling with adaptive signal control, which enables quantitative assessment of traffic impacts on air quality and direct mitigation through optimized cycle lengths and delays. The aim is to evaluate how this integrated ANN plus optimization framework can effectively reduce motor-vehicle emissions. The results provide evidence-based guidance for transitioning cities like Kokshetau toward more sustainable, lower-carbon mobility.

The article is organized as follows: Section 1 reviews the relevant literature and outlines the context of Kokshetau, Section 2 details the methodology (including the ANN model development and the traffic signal optimization procedure), Section 3 presents the results, Section 4 discusses the findings and their implications for sustainable mobility transitions, and Section 5 concludes the study.

2. Materials and Methods

2.1. Study Area and Object of Research

The present study was conducted in the city of Kokshetau, a regional administrative and economic center located in northern Kazakhstan. The city of Kokshetau is the administrative center of the Akmola Region of the Republic of Kazakhstan, with a population of more than 170 thousand people (2024–2025). The main sources of air pollution in the city are industrial enterprises, smoke emissions from private households, and motor vehicles. The industrial enterprises are mainly concentrated in the northern and eastern industrial zones of the city. Among the industrial sources are KAMAZ-Engineering JSC (vehicle assembly), ENKI LLP (construction materials production), polymer packaging manufacturers—Novopek LLP and Asyl Arman LLP, as well as State Communal Enterprise Kokshetau Zhylu (heat and power generation), Kokshetau Zholdary LLP, and Kokshetauavtodor LLP (road construction).

According to the Department of Ecology for the Akmola Region and the Department of Natural Resources and Environmental Management of the Akmola Region, the enterprises that have the greatest impact on air quality are State Communal Enterprise Kokshetau Zhylu, Kokshetau Zholdary LLP, and Kokshetauavtodor LLP. The activities of these enterprises make a significant contribution to the background stationary emissions of pollutants such as particulate matter (PM), nitrogen oxides (NO_x), sulfur dioxide (SO₂), and volatile organic compounds (VOC).

The established emission limits of pollutants into the atmosphere for these enterprises are as follows: 8.6 thousand tons per year for Kokshetau Zhylu, 27.48 thousand tons per year for Kokshetau Zholdary LLP, and for Kokshetauavtodor LLP—124.04 thousand tons per year at the asphalt-concrete plant and 135.45 thousand tons per year at the crushing and screening facility. The combined impact of these enterprises, along with emissions from the private sector and motor vehicles, constitutes the main share of air pollution in the city.

The map above (Figure 1) shows the main industrial zones of the city—the Northern and Eastern Industrial Zones.

Table 1. Major Industrial Enterprises of Kokshetau and Their Geographic Coordinates.

No.	Enterprise Name	Type of Activity	Coordinates
1	State Communal Enterprise Kokshetau Zhylu	Heat and power generation	53°16′54″ N 69°23′37″ E
2	ENKI LLP	Construction materials production	53°18′38″ N 69°24′50″ E
3	Novopek LLP	Polymer packaging manufacturing	53°16′07″ N 69°26′35″ E
4	Asyl Arman LLP	Polymer packaging manufacturing	53°19′53″ N 69°24′47″ E
5	Kokshetauavtodor LLP	Road construction	53°22′27″ N 69°26′21″ E
6	Kokshetau Zholdary LLP	Road construction	53°16′27″ N 69°26′41″ E
7	KAMAZ-Engineering JSC	Vehicle assembly	53°16′50″ N 69°25′05″ E

presents major industrial enterprises operating within the city of Kokshetau, classified according to their primary production activities. The listed coordinates indicate the approximate geographic locations of these facilities, many of which are concentrated in the northern and eastern industrial zones of the city. These enterprises represent the main contributors to local industrial output and are relevant to environmental and urban development assessments.

Situated in northern Kazakhstan, Kokshetau a growing regional administrative and economic center illustrates transport and environmental challenges typical of medium-sized post-Soviet cities, including heavy reliance on internal-combustion engine (ICE) vehicles, limited access to sustainable mobility options, and insufficient mechanisms to mitigate traffic-related air pollution.

According to data from the Akmola Region Traffic Police Department, at the beginning of 2025, the total number of registered vehicles in the city of Kokshetau was 68,214, and according to the latest registration data, 68,866. Analysis shows that the structure of the city’s vehicle fleet corresponds to the general trends in the region: it consists mainly of private passenger cars, while the share of buses, minibuses, trucks, and specialized vehicles is significantly lower.

Gasoline-powered cars remain predominant, with more than 190,000 throughout the Akmola region, while diesel vehicles are mainly represented by trucks and buses used by utilities and construction companies.

At the same time, electric vehicles (EVs) are still virtually absent from the city’s vehicle fleet: only 13 units are registered in Kokshetau, accounting for just 0.019% of the total number of vehicles (see

Table 2. Number of Registered Vehicles in the City of Kokshetau (31 December 2024).

Type of Vehicles	Legal Entities	Private Individuals
A (Motorcycles)	15	611
B (Passenger cars)	4836	52,861
C (Heavy trucks, over 3.5 tons)	2045	246
D (Buses)	525	160
E (Trailers)	553	1079
Electric vehicles	0	13
BE (Passenger cars with trailers)	34	241
CE (Heavy trucks with trailers)	651	1111
DE (Buses with trailers)	0	1
A1 (Light motorcycles)	0	7
B1 (Tricycles and quad bikes (ATVs))	1	36
C1 (Medium trucks, 3.5–7.5 tons)	589	686
D1 (Minibuses)	238	78
C1E (Medium trucks with trailers)	8	34
D1E (Minibuses with trailers)	0	0
TOTAL	9495	59,371

). The situation is similar across the region, where electric buses are not yet in operation.

Thus, despite the measures being implemented in Kazakhstan to promote sustainable transport and reduce emissions, Kokshetau is in the earliest stages of transition to low-carbon transport technologies. This underscores the need for further development of charging infrastructure, incentives for the purchase of electric vehicles, and improvement of regional transport policy.

Data on the number of registered vehicles in Kokshetau, both owned by legal entities and individuals, was obtained from the Traffic Police Department of the Akmola Region. The analysis showed that the structure of the city’s vehicle fleet reflects the general trends characteristic of the region as a whole. A similar picture can be observed on a broader scale throughout the Akmola region.

As shown in Figure 2, gasoline-powered cars predominate: more than 190,000 passenger cars run exclusively on gasoline. Diesel-powered cars are mainly represented by freight and passenger transport—in particular, buses and special equipment operated by enterprises in the utilities and construction sectors. In contrast, electric vehicles still account for an extremely small share of the vehicle fleet: to date, only 68 units have been registered, of which 66 are passenger cars and only 2 are trucks. Electric buses are not currently in operation in the Akmola region.

These Table 2 and Figure 2 highlight the structural dominance of fossil fuels in both the city and the region, underscoring the need for coordinated infrastructure development and policy support to enable a broader transition to sustainable transport solutions.

In response to this limited adoption, some progress has been made in developing the necessary infrastructure to support EV use. The city is currently served by three publicly accessible charging stations installed at two sites: two chargers at the Qazaq Energy Charge location on Ualikhanov Street and one charger at Auelbekov Street. Monthly data from the first quarter of 2025 show a modest but increasing use of these facilities, with the number of charging sessions growing from 119 in January to 217 in March. The cumulative energy throughput also increased accordingly, from 6267 kWh to 6328 kWh. Nevertheless, this infrastructure remains insufficient to support meaningful EV expansion. The spatial distribution of the available stations is illustrated in Figure 3, which highlights the narrow coverage area and emphasizes the infrastructural bottlenecks faced by early adopters.

In addition to infrastructure and fleet composition, field studies were carried out to assess the environmental footprint of traffic in key areas of the city. Six major signalized intersections were selected as monitoring sites based on their high traffic volumes and centrality within the urban road network. Spatial distribution of the six research sites selected in Kokshetau is shown in Figure 4. The selected locations include: (1) Kenesary Kasymuly Street and M. Auezov Street; (2) M. Auezov Street and N. Nazarbayev Avenue; (3) Suleymenov Street and Abylai Khan Avenue; (4) Zh. Tashenov Street and Sh. Ualikhanov Street; (5) Nauryzbai Batyr Street and K. Satpaev Street; and (6) Sh. Ualikhanov Street and Kenesary Street (Figure 5).

At each study site, data were collected during three characteristic time intervals reflecting the typical daily rhythm of transport activity: morning peak (08:00–08:30), afternoon period (15:00–15:30), and evening peak (18:30–19:00). This approach made it possible to cover the main time phases of traffic intensity and identify differences in the structure of traffic flows depending on the time of day.

During each observation, detailed parameters characterizing the traffic situation were recorded: intersection ID, specific direction of travel (north, east, south, or west), type of vehicle (passenger car, truck, bus, minibus, or specialized transport), number of cars waiting for a red light, and the duration of the green, yellow, and red phases of the traffic light (in seconds). Together, these indicators provided a comprehensive picture of the dynamics of traffic flows in space and time, which formed the basis for subsequent quantitative analysis.

Key variables such as average speed, acceleration, and idle time were obtained from actual measurements collected by road sensors at major intersections. Vehicle speeds were determined using induction loops, and to increase reliability, the results were verified by comparing them with video recordings and manual traffic counts.

Acceleration and deceleration indicators were calculated based on sequential speed observations with time stamps, which made it possible to identify the characteristics of traffic dynamics depending on the traffic light phase. Downtime was determined by analyzing the time intervals of vehicle stops at regulated intersections.

To confirm the accuracy of the data, additional short-term field verification was carried out, covering morning and evening rush hours. The results obtained formed the basis for subsequent analysis of traffic intensity, emissions, and operational efficiency of traffic flows.

2.2. Estimation of Traffic-Related Emissions at Intersections

As a preparatory step for the subsequent modeling phase, pollutant emissions generated by traffic at selected intersections were estimated based on standardized calculation procedures. Emission levels were calculated based on traffic observations and emission coefficients derived from the official methodology RND 211.2.02.11-2004: Methodology for determining vehicle emissions for conducting summary calculations of urban air pollution [41]. Since more than 45% of the city’s vehicle fleet has been in operation for over ten years. The age of vehicles, irregular maintenance, and the widespread use of imported used cars are additional factors contributing to the increase in pollutant emissions. This circumstance was recognized by the researchers and discussed in detail in the section on uncertainty analysis.

The estimation method assumes that the greatest volume of pollutants is released during periods of deceleration, idling, and acceleration—especially at signal-controlled intersections. Emissions were therefore calculated for each vehicle category and direction based on the number of cars waiting at red lights, the duration of the red signal (including the yellow phase), and the number of signal cycles within the observed time interval.

The estimation covered four key pollutants: carbon monoxide (CO), nitrogen oxides (NO_x), sulfur dioxide (SO₂) and soot. Vehicles were grouped into five main categories: passenger cars, minibuses, trucks, special-purpose vehicles, and buses. Each group was assigned specific emission rates (in grams per minute) based on national norms that reflect vehicle behavior at urban intersections. These emission rates (

Table 3. Average specific emission rates (g/min) for different vehicle categories and pollutants under urban intersection conditions, based on RND 211.2.02.11-2004 methodology.

Vehicle Category	Emission [g/min]
	CO	NOx	Soot	SO2
passenger cars	0.3	0.01	0.01	0.3 × 10−2
minibuses	1.8	0.03	0.03	0.6 × 10−2
trucks	1.3	0.10	0.22	1.0 × 10−2
special transport	2.5	0.13	0.45	1.2 × 10−2
buses	1.1	0.10	0.20	1.0 × 10−2

) capture the average output under typical stop-and-go conditions and were multiplied by the observed number of vehicles and signal cycle data to estimate the total emissions per observation point.

These coefficients (see Table 3) characterize the average emissions during cyclic movement with stops and accelerations and were used to calculate the total emissions for each observation point. The final values were obtained by multiplying the specific emission rates by the recorded number of vehicles and traffic light cycle parameters, which made it possible to quantitatively estimate the total emissions of pollutants in real urban traffic conditions.

The specific emission rates $M_{{\Pi }_i}^3$ in Table 3 are modal time-weighted factors for urban intersections (braking, idling, acceleration; per RND-211.2.02.11-2004) and were used, together with the observed vehicle counts and signal-cycle data, to compute emissions at each observation point.

The general form of the calculation is expressed as:

$$M_{{\Pi }_i}^3={\displaystyle \frac{R_c}{60}}\sum _1^{N_c}\sum _1^{N_{gr}}{M}_{Pic}{G}_k\left[{\displaystyle \frac{\mathrm{g}}{\mathrm{m}\mathrm{i}\mathrm{n}}}\right]$$

(1)

where

• $M_{{\Pi }_i}^3$—estimated emission of pollutant i (in g/min);
• $R_{\mathrm{c}}$—duration of one red light cycle (including yellow), in minutes;
• $M_{Pic}$—specific emission rate of pollutant i for vehicle group k (g/min);
• $N_c$—number of red signal cycles (including yellow) during a 20 min interval;
• $N_{gr}$—the number of groups of cars;
• $G_k$—number of vehicles in group k observed in the queue at red signal (including yellow).

These coefficients capture average emissions under cyclic stop-and-go conditions. The stop/idling duration is incorporated explicitly via the measured signal phases (green/yellow/red) and the observed queues $G_k$ in Equation (1), while the number of cycles $N_c$ follows the measured cycle length $R_c$. In the analyzed data, cycle length ranged from 48 to 108 s (median 67 s), with median phase shares green 0.417, yellow 0.045, red 0.528. Queues at red (median vehicles per approach) were 6 (passenger cars), 0 (trucks), 0.5 (buses), 1 (minibuses), 0 (special), and these values are used in Equation (1) via $N_c$ and $G_k$.

This formula was applied independently for each combination of intersection, direction, vehicle category, and time period. The result was a set of pollutant estimates disaggregated by spatial and temporal conditions, which served as inputs for the modeling stage presented in Section 3.

2.3. Modelling Methodology

The aim of this part of the study is to develop and apply an integrated model for analyzing the impact of urban transport on air pollution, followed by the justification of measures for optimizing traffic flow and transitioning to environmentally friendly modes of transportation, using the city of Kokshetau as a case study.

The modeling process was carried out in two sequential stages. In the first stage, an artificial neural network (ANN) was developed to predict traffic-related air pollutant emissions, including carbon monoxide (CO), nitrogen oxides (NO_x), sulfur dioxide (SO₂), and particulate matter (soot). The prediction was based on a set of input variables characterizing traffic and signaling conditions. These included the intersection number (ranging from 1 to 6, corresponding to key traffic junctions in the city of Kokshetau), the specific point or direction at the intersection (coded 0 to 3, representing north, east, south, and west legs, respectively), the vehicle type (encoded from 1 to 5, covering categories such as passenger cars, trucks, buses, minibuses, and special), and the number of waiting cars at the signal. Additionally, the time of day was considered as a categorical input (0 representing 8:00 AM, 1 for 3:00 PM, and 2 for 6:30 PM) to capture variations in traffic volume and driving behavior. Signal parameters included the duration of the green phase and the red phase, both expressed in seconds.

The time factor was also included in the model as a categorical variable, which made it possible to record differences in traffic intensity and the nature of traffic flows in the morning, afternoon, and evening hours. This approach provided a comprehensive description of traffic dynamics and allowed the neural network to identify nonlinear relationships between traffic characteristics and pollutant emission levels.

Particular attention was paid to the difference between soot and particulate matter (PM). Soot is the carbonaceous fraction of the total particle mass formed as a result of incomplete fuel combustion. Its concentration varies significantly depending on the type of engine, the age of the vehicle, and the quality of the fuel. In this study, different emission factors were applied for gasoline and diesel engines, which made it possible to improve the accuracy of forecasts and take into account the actual specifics of the urban vehicle fleet.

This model was trained on a dataset collected under real-world traffic conditions and emissions estimations, capturing variations across intersections, vehicle flows, and signal timings. The neural network approach was selected due to its ability to capture complex, nonlinear relationships between input parameters and emission outcomes. Furthermore, the ANN allowed for the simultaneous prediction of multiple emission indicators, which is essential for comprehensive environmental assessments.

To develop the emission prediction model, a feedforward artificial neural network was implemented using MATLAB’s 2024b Neural Fitting Toolbox. The model was trained to solve a multi-output regression problem, where the objective was to estimate the values of four air pollution indicators (CO, NO_x, SO₂, and soot) based on seven explanatory variables: intersection, point, type of car, number of waiting cars, time of day, green light duration, and red light duration.

During the modeling process, a series of preliminary experiments were conducted to identify the optimal network architecture and training configuration. Networks with a single hidden layer and a varying number of neurons—ranging from 5 to 15—were tested. In addition, multiple training algorithms were evaluated, including Levenberg–Marquardt backpropagation (trainlm), Bayesian regularization (trainbr), and scaled conjugate gradient (trainscg). Default Levenberg–Marquardt control parameters were used unless noted: initial damping μ₀ = 0.001, decrease factor mu_dec = 0.1, increase factor mu_inc = 10, upper bound mu_max = 1 × 10¹⁰, maximum epochs = 1000, early stopping with max_fail = 6, and minimum gradient min_grad = 1 × 10⁻⁷. Because LM is second-order, no fixed learning rate is applied, convergence is governed by μ updates. To ensure reproducibility, a fixed random seed was set prior to data division and initialization. Hidden units used a hyperbolic tangent activation (tansig), while the output layer used linear activation (purelin) to allow unconstrained real-valued predictions for the four pollutants. The dataset was randomly partitioned into training, validation, and testing subsets in proportions of 70%, 15%, and 15%, respectively. This split allowed for monitoring of generalization performance and early stopping during training. Model accuracy was assessed using the mean squared error (MSE) and root mean squared error (RMSE) metric, along with residual plots and regression diagnostics, which confirmed the suitability of the trained model for multi-output prediction of emissions.

In the second stage of the study, the previously trained neural network model was embedded within an optimization framework designed to identify traffic signal timings that would minimize pollutant emissions at urban intersections. The aim of the optimization was to determine the most favorable duration of the green signal phase for a selected traffic direction (specifically, directions A and C at a given intersection), while also considering the resulting red phase duration for the opposing directions (B and D).

The optimization of traffic light control parameters was formulated as a constrained nonlinear programming problem aimed at minimizing total pollutant emissions while maintaining acceptable intersection capacity conditions. The controlled variables were the duration of the green phase for directions A and C (green_AC) and the total duration of the traffic light cycle (cycle_length), from which the duration of the red phase for directions B and D (red_BD) was automatically determined as the difference between the total cycle length and the green phase.

To ensure realistic control, operational constraints were introduced: the green phase could not be shorter than 20 s, and the total cycle duration was limited to a range of 60 to 120 s. These conditions ensured that the developed signal plans remained practical and met standard traffic management requirements.

The objective function was based on the deviation of the predicted emission values (CO, NO_x, SO₂, and soot) from the specified environmental targets corresponding to the permissible air quality thresholds. The optimization aimed to minimize the total deviation for all four pollutants for the priority traffic direction (A and C), while taking into account the impact of increasing the duration of the red phase on the opposite approaches (B and D). This comparative analysis demonstrated that an imbalance in the distribution of traffic light phases can cause secondary effects of increased emissions in other directions, even with a reduction in emissions in the main flow axis.

Additionally, a sensitivity analysis was performed, in which the cycle duration range was extended to 150 s. The results showed that when the cycle length is increased, total emissions change by less than 2% compared to the base range of 60–120 s. This demonstrates the stability and reliability of the proposed optimization model, which remains effective with different settings of the time parameters of traffic light control.

The optimization procedure was implemented using MATLAB’s fmincon solver, enabling the simultaneous adjustment of signal timing parameters under nonlinear constraints. The resulting integrated predictive-optimization model provided a data-driven decision support tool, capable of identifying locally optimal signal strategies that contribute to lower emissions at specific intersections. This tool can be further extended to simulate various traffic scenarios and support the transition to more sustainable urban transport systems.

3. Results

3.1. Results of Traffic-Related Emission Estimation

To evaluate the applicability of the proposed model, comparative analysis was conducted using transport and emission data from several medium-sized cities in Kazakhstan with demographic and functional characteristics similar to Kokshetau—Petropavl, Pavlodar, Ust-Kamenogorsk, Karaganda, and Taldykorgan.

These cities are characterized by a high dependency on conventional vehicles, limited adoption of electric transport (less than 0.05% of the fleet), and comparable road network configurations.

The modeling results indicate that the proposed traffic signal optimization framework can achieve reductions in CO, NO_x, and PM emissions by 9–12% on average, which aligns with findings reported for Pavlodar and Karaganda (8–13%) under similar optimization initiatives.

This comparison suggests that the developed approach can be effectively scaled and adapted to other medium-sized cities in Kazakhstan to enhance the environmental performance of transport systems and mitigate urban air pollution.

Based on the input traffic data and the methodology described in Section 2.2, emission levels for four key pollutants—CO, NO_x, SO₂, and soot—were calculated for each intersection, vehicle group, and observation period. The results provide a spatial and temporal profile of traffic-related air pollution in Kokshetau, which subsequently informed the structure and training of the predictive modeling stage.

Given the significant amount of data collected, this section presents detailed results for only two representative intersections. These sections were selected to reflect the upper and lower limits of emission intensity recorded within the network under study, allowing the analysis to focus on the most representative and contrasting scenarios. This approach not only provides clarity, but also allows for a deeper analysis of the key patterns of emissions formation and an assessment of their compliance with Euro 5 standards.

During the field studies, six observation points located at key transport hubs in the city were initially considered. However, only two of them were used for subsequent calibration and validation of the model. The remaining four intersections were excluded from the analysis due to incomplete data and the lack of stable time coverage, which could lead to distortion of the neural network training results and reduce the reliability of the forecast.

The two intersections selected provided continuous, high-quality observation series containing the entire set of parameters necessary for building, training, and testing the developed emission prediction model. This made it possible to conduct a detailed analysis of emission dynamics in real urban conditions and to assess the applicability of the proposed approach to modeling air pollution from transport.

Among the six monitored sites, the selected intersections are:

• Zh. Tashenov St & Sh. Ualikhanov St (

  Table 4. Calculated emissions (g/min) (Zh. Tashenov St & Sh. Ualikhanov St).

  Vehicle Category	Zh. Tashenov St & Sh. Ualikhanov St				Nauryzbai Batyr St & K. Satpaev St
  	Emission [g/min]				Emission [g/min]
  	CO	NOx	Soot	SO2	CO	NOx	Soot	SO2
  passenger cars	32.760	1.092	0.3276	1.092	21.660	0.722	0.2166	0.722
  trucks	60.840	4.680	0.4680	10.296	15.64333	1.203333	0.120333	2.647333
  buses	29.55333	2.686667	0.268667	5.373333	0.000	0.000	0.000000	0.000
  minibuses	72.540	1.209	0.2418	1.209	21.660	0.361	0.0722	0.361

  )—the location with the highest overall pollutant loads, characterized by extreme soot and NO_x emissions from heavy-duty trucks and very high CO levels from minibuses and passenger cars.
• Nauryzbai Batyr St & K. Satpaev St (Table 4)—the site with the lowest total emissions in the dataset, including a zero bus reading during the sampling window, yet still recording values far exceeding Euro 5 thresholds for LDVs.

These two intersections serve as contrasting case studies, providing insight into both worst-case pollution hotspots and comparatively less intensive sites. This comparative approach supports the identification of priority locations for short-term mitigation measures and offers a robust empirical foundation for targeted traffic management strategies in the context of urban air quality improvement.

It should be noted that the Euro 5 framework applies to light-duty vehicles (LDV) and expresses emission limits in grams per kilometer (g/km); for heavy-duty vehicles (HDV), including buses, type-approval values are reported in grams per kilowatt-hour (g/kWh), so the comparisons for this category are indicative only [42,43]. For SO₂, no direct tailpipe limit is set under Euro 5; however, emissions are constrained indirectly by the fuel sulfur content limit of 10 ppm, as specified in Directive 2009/30/EC [44].

To ensure a fair comparison with the Euro 5 standard values, all calculated emissions were converted from grams per minute (g/min) to grams per kilometer (g/km), based on the assumption of a constant reference speed of 20 km/h. This approach reflects typical traffic conditions at regulated intersections, where vehicles travel at reduced speeds, often slowing down or coming to a complete stop at traffic lights, and then accelerating again after the signal changes.

Thus, the calculated pollutant indicators (in g/min) for different intersections and vehicle categories were converted to g/km values based on a fixed speed of 20 km/h. This value was chosen based on the results of an analysis of average urban traffic speeds in Kokshetau, which were determined using data obtained from Google Maps and Yandex (2025). Traffic. According to these sources, the average speed of traffic flows during rush hour (morning and evening) is about 20 km/h, which makes this indicator representative of real operating conditions.

The use of empirically verified speeds made it possible to reliably reproduce the nature of urban traffic and increase the accuracy of emission calculations in conditions of high road congestion. This methodological approach ensures that the model is consistent with actual traffic conditions and makes the results comparable with the emission limits set by the Euro 5 standard.

The rationale for selecting 20 km/h is that all measurements were taken at intersections, where vehicles naturally decelerate when approaching traffic lights or preparing to turn, often coming to a complete stop. The converted results are compared to the Euro 5 emission standards defined by Regulation (EC) No 715/2007 and related guidance [42,45], while SO₂ is discussed in the context of fuel sulfur limits [44]. This approach is standard in real-driving emission (RDE) assessments and is documented in the Handbook Emission Factors for Road Transport (HBEFA), version 3.3 [46].

Building on the above conversion step, detailed per-kilometer results are presented only for two representative intersections that bracket the observed range of emission intensity.

Table 5. Calculated emissions at 20 km/h (g/km) with Euro 5 limits (Zh. Tashenov St & Sh. Ualikhanov St and Nauryzbai Batyr St & K. Satpaev St).

Vehicle Category	Emission [g/km]
	Zh. Tashenov St & Sh. Ualikhanov St
	CO (g/km)	Euro 5 Limit (CO)	NOx (g/km)	Euro 5 Limit (NOx)	SO2 (g/km)	PM (g/km)	Euro 5 Limit (PM)
passenger cars	98.280	≤1.0/0.5	3.276	≤0.06/0.18	0.983	3.276	≤0.005
trucks	182.520	≤0.5	14.040	≤0.18	1.404	30.888	≤0.005
buses	88.660	HDV norm (g/kWh)	8.060	HDV norm (g/kWh)	0.806	16.120	HDV norm (g/kWh)
minibuses	217.620	≤0.5	3.627	≤0.18	0.725	3.627	≤0.005
	Nauryzbai Batyr St & K. Satpaev St
passenger cars	64.980	≤1.0/0.5	2.166	≤0.06/0.18	0.650	2.166	≤0.005
trucks	46.930	≤0.5	3.610	≤0.18	0.361	7.942	≤0.005
buses	0.000	HDV norm (g/kWh)	0.000	HDV norm (g/kWh)	0.000	0.000	HDV norm (g/kWh)
minibuses	64.980	≤0.5	1.083	≤0.18	0.217	1.083	≤0.005

reports the 20 km/h per-kilometer emission rates for Zh. Tashenov St & Sh. Ualikhanov St and Table 5 provides the corresponding results for Nauryzbai Batyr St & K. Satpaev St under identical assumptions (20 km/h), enabling a like-for-like contrast.

Under the 20 km/h normalization introduced above, Zh. Tashenov St & Sh. Ualikhanov St (Table 5) is the upper-bound case. Trucks dominate the local pollution load: per-kilometer soot (benchmarked against the LDV particulate limit) equals 30.888 g/km, i.e., approximately 6200 times the Euro 5 LDV PM limit of 0.005 g/km; NO_x = 14.040 g/km (≈78 times the 0.18 g/km diesel limit) and CO = 182.520 g/km (≈365 times the 0.5 g/km diesel limit). Minibuses record the highest CO of all categories (217.620 g/km; >430 times the diesel limit), while passenger cars also show extreme exceedances (e.g., CO 98.280 g/km; ≈197 times, soot 3.276 g/km; ≈655 times, NO_x 3.276 g/km; ≈18 times). Bus values in g/km are large as well; however, HDV type-approval is defined in g/kWh, so such comparisons remain indicative.

By contrast, Nauryzbai Batyr St & K. Satpaev St (Table 5) represents the lower-bound case in absolute terms; nevertheless, exceedances remain substantial under intersection conditions. Trucks reach soot = 7.942 g/km (≈1590 times the LDV PM limit), NO_x = 3.610 g/km (≈20 times the diesel limit) and CO = 46.930 g/km (≈94 times the diesel limit). Passenger cars and minibuses, although lower than at the hotspot, remain far above Euro 5 benchmarks (e.g., passenger cars: CO ≈ 65 g/km, NO_x ≈ 2.17 g/km, soot ≈ 2.17 g/km; minibuses: CO ≈ 65 g/km, NO_x ≈ 1.08 g/km, soot ≈ 1.08 g/km). Zero bus readings reflect the absence of bus movements during the sampling window rather than compliance.

Taken together, Table 5 show that even the least polluted monitored intersection has emissions per km that are orders of magnitude above the Euro 5 LDV thresholds, with trucks and vans driving the soot and NO_x exceedances and passenger cars contributing disproportionately to CO. This pattern motivates targeted, intersection-specific mitigation measures (e.g., signal timing optimization, inspection and enforcement) alongside longer-term fleet upgrades.

To complement the site-specific contrasts,

Table 6. Maximum calculated emissions by Vehicle Category (g/min).

Vehicle Category	Emission [g/min]
	CO	NOx	SO2	Soot
passenger cars	47.200	1.482	0.7227	1.417
minibuses	61.580	1.286	0.2178	1.163
trucks	14.982	1.384	0.1210	2.621
buses	17.111	1.473	0.4015	3.088

aggregates the maximum per-minute emission rates recorded across the monitoring network for each vehicle category, while

Table 7. Maximum Calculated Emissions by Vehicle Category at 20 km/h (g/km) with Euro 5 Limits.

Vehicle Category	Emission [g/min]
	CO	NOx	SO2	Soot
passenger cars	141.6	4.45	2.168	4.25
minibuses	184.7	3.86	0.653	3.49
trucks	44.95	4.15	0.363	7.86
buses	51.33	4.42	1.205	9.26
Euro 5 Limit	≤1.0/0.5	≤0.06/0.18	—	≤0.005

reports the corresponding per-kilometer values at 20 km/h along with the Euro 5 reference limits for LDVs. This perspective isolates category-specific “worst-case” behavior under intersection conditions and clarifies which pollutants dominate within each group.

For passenger cars, the network maxima reach CO = 141.6 g/km, NO_x = 4.45 g/km, and soot = 4.25 g/km. Relative to LDV diesel references, these values correspond to ≈280× (CO vs. 0.5 g/km), ≈25× (NO_x vs. 0.18 g/km), and ≈850× (soot vs. 0.005 g/km). Such magnitudes are consistent with widespread malfunction or absence of after-treatment, notably DPF for particulates and SCR/EGR for NO_x.

Minibuses exhibit the highest CO among all categories (184.7 g/km), equivalent to ≈370× the diesel limit, with NO_x = 3.86 g/km (≈21×) and soot = 3.49 g/km (≈700×). This pattern indicates that, under stop-and-go operation, minibuses are a disproportionate source of incomplete-combustion products (CO) while also contributing substantially to soot and NO_x.

For trucks, the maxima are CO = 44.95 g/km (≈90×), NO_x = 4.15 g/km (≈23×), and soot = 7.86 g/km (≈1570×). Although CO remains below passenger-car and minibus maxima, the combination of very high soot and NO_x confirms heavy-duty diesel traffic as a primary driver of local air-quality exceedances during intersection cycling.

Buses present the highest absolute soot among the four categories (9.26 g/km). While HDV type-approval is expressed in g/kWh rather than g/km (hence direct regulatory comparison is indicative), benchmarking against the LDV particulate limit implies an order-of-magnitude exceedance of ≈1850×; NO_x likewise attains 4.42 g/km at the maximum. These values reinforce the dominant contribution of heavy-duty diesel engines under low-speed, transient operation.

At the representative intersection speed of 20 km/h, all vehicle categories exceed the Euro 5 LDV thresholds for CO, NO_x and soot by one to three orders of magnitude. Given the proximity of receptors (pedestrians and cyclists) at signalized intersections, these maxima imply high short-term exposure potential and help explain the prominence of intersection hotspots in urban pollution maps.

The evidence supports a two-track response. In the short term, signal timing should be optimized to reduce queuing and harsh transients. Technical inspections should prioritize DPF integrity and NO_x control operability. Access restrictions or low emission zones are warranted at the worst nodes. Heavy-duty flows should be diverted away from dense pedestrian corridors during peak hours. Fuel sulfur compliance should be enforced to limit SO₂ formation. In the long term, sustainable reductions will require fleet modernization and electrification of high-emission duty cycles. Charging infrastructure and retirement of high emitting vehicles are essential. The reported maxima provide high-contrast targets for the predictive ANN and define targets for emission-aware signal optimization. These results motivate the next step of the study. A predictive model of traffic-related emissions using artificial neural networks will be developed.

3.2. Development of a Predictive Model for Traffic-Related Emissions Using Artificial Neural Networks

The most effective network configuration identified through experimental testing consisted of a single hidden layer with 8 neurons, trained using the Levenberg–Marquardt algorithm. This configuration consistently yielded the lowest prediction errors across training, validation, and test subsets, providing a good balance between model complexity and generalization ability.

Table 8. Training results of ANN for best network.

Model	ANN
Training algorithms	Levenberg–Marquardt
Neurons in hidden layer	8
Epoch	68
Performance	1.29
Gradient	0.722
Best validation performance (at epoch)	2.2008 (62)
MSE	1.29
RMSE	1.21

shows the results of the training process for ANN.

Based on the training results, the ANN demonstrated high predictive accuracy and stability across all data subsets. The network trained for 68 epochs, achieving a final performance (mean squared error) of 1.29, RMSE—1.21 and a gradient magnitude of 0.722, which indicates smooth convergence of the optimization process. The best validation performance, marking the minimum validation error, was recorded at epoch 62, with a value of 2.2008 (Figure 6). These values confirm that the training process avoided overfitting and maintained generalization.

The training state plot (Figure 7) further supports convergence, showing a stable gradient and learning rate (mu), along with minimal validation failures until late epochs. The error histogram (Figure 8) reveals that the vast majority of prediction errors are centered around zero, with symmetrical distribution, indicating no significant prediction bias.

Finally, the regression plots (Figure 9) present a near-perfect linear fit between the predicted and actual target values for all subsets. The regression coefficients (R-values) approached 0.99, with best-fit lines almost overlapping the ideal Y = T line, which strongly support the robustness of the model for emission forecasting.

The regression and residual diagnostics should be read with two sources of uncertainty in mind. Aleatoric variability reflects short-term fluctuations in demand and signal operations and observation noise, epistemic uncertainty arises from limited spatio-temporal coverage and model specification/initialization. The reported test-set RMSE quantifies the expected prediction error under operating conditions comparable to those observed at the study intersections. The transformation from g/min to g/km uses a constant reference speed (20 km/h) and therefore scales all per-kilometer values proportionally, leaving relative comparisons and conclusions unchanged. Our results are most reliable for peak periods and signalized approaches considered here, performance may differ under substantially different control plans or traffic regimes.

Together, these results validate the choice of an 8-neuron architecture with the Levenberg–Marquardt training algorithm as optimal for the emission prediction task. The model is thus well-suited for integration into the optimization framework described in the second phase of the study.

3.3. Optimization of Traffic Signals for Emission Reduction

In the second stage of the study, the trained neural network model was embedded in a computational optimization framework. The primary objective of this stage was to determine the optimal duration of the green signal phase and total signal cycle at selected intersections, such that the predicted emissions of key four pollutants are minimized or kept as close as possible to regulatory target values. The approach preserved traffic flow symmetry by accounting for both directions of movement.

The optimization algorithm explored various combinations of green time allocations and cycle lengths within predefined constraints. The green time allocated to one direction could not fall below 20 s to maintain basic traffic throughput. For each candidate configuration, the emission levels were forecast using the neural model, and average predicted pollutant values were computed separately for both primary (A&C) and opposing (B&D) traffic directions. A composite objective function was formulated as the total deviation of these emission levels from their respective permissible thresholds. Minimizing this deviation ensured that the proposed configuration remained environmentally acceptable from a multi-pollutant perspective.

The optimization relied on a gradient-based solver (Sequential Quadratic Programming algorithm), and the model evaluated performance dynamically for each time of day and intersection separately. A key feature of this procedure was that it simultaneously considered the adverse effect of longer red signals on the opposing traffic streams, helping avoid displacement of emissions rather than their actual reduction.

To make the methodology accessible to planners and engineers, a dedicated graphical user interface (GUI) application was also developed. This user-friendly tool allows the user to select a specific intersection and time of day, execute the optimization with a single click, and immediately view the optimal timing alongside predicted emission levels for both directions. The output is visualized via dynamic bar charts and numerical summaries, offering an intuitive understanding of trade-offs and benefits from the proposed signal timing strategy.

The interface (Figure 10) consists of two dropdown menus allowing the user to select the intersection ID (from 1 to 6) and the time of day (0—Morning 8.00 am, 1—Afternoon 3.00 pm, 2—Evening 6.30 pm). Upon pressing the “Optimize” button, the application runs the optimization algorithm described earlier, which searches for the optimal duration of the green signal for directions A and C while respecting constraints on minimum green time (≥20 s) and total cycle time (bounded between 50 and 120 s).

Once the optimization is completed, the results are displayed both textually and graphically. The textual panel (bottom area) provides numerical values for the optimized green duration and cycle time, followed by predicted average emission levels in g/km (CO, NO_x, SO₂, and soot) for both direction groups: A&C (directions optimized for lower emissions) and B&D (which experience the corresponding red phase). The accompanying bar chart visually compares emission values across the four pollutants and direction groups, allowing users to easily assess trade-offs. Figure 11 shows an example for Intersection 2 during the morning period and Figure 12—for Intersection 1 in the afternoon.

This visualization enables informed decision-making regarding traffic control strategies by transparently presenting both the benefits and trade-offs in air pollution for different signal settings.

4. Discussion

The main objective of this study was to identify realistic ways to reduce air pollution caused by urban transport in Kokshetau [47]. Although Kazakhstan’s long-term transport development strategy envisages a gradual transition to environmentally friendly modes of transport (EFT), including electric vehicles (EVs), the results of the analysis presented in this paper show that significant improvements in air quality are possible at the current stage, even before a large-scale transformation of the vehicle fleet.

Short-term measures, such as optimizing traffic light cycles, are particularly important as they are a practical and quickly implementable tool for reducing harmful emissions at the busiest intersections. Field measurements conducted in Kokshetau confirmed that peak traffic loads directly coincide with maximum pollutant concentrations, which underscores the effectiveness of traffic flow regulation in areas with high traffic intensity. A similar correlation between traffic density and emission levels has been observed in other cities around the world [38], further confirming the potential of interventions such as traffic light optimization in combating local air pollution.

As shown in the sections on experimental results and modeling, the artificial neural network (ANN) developed in the first stage demonstrated high reliability and accuracy in predicting emissions of four key pollutants associated with motor vehicles. After successful validation, the model was integrated into the second stage of the study, where it was used to optimize the timing parameters of traffic light control.

The integration of ANN into the structure of a nonlinear optimization model with constraints made it possible to determine the locally optimal values for the duration of the green phase and the overall cycle, at which emissions are minimized without disrupting the symmetry of traffic flows.

The simulation results showed that the proposed measures could reduce the impact of PM_2.5 and NO_x pollutants on the population by 6–8%, which is consistent with the World Health Organization (WHO) estimates based on dose–response functions. This effect can significantly reduce the number of respiratory and cardiovascular diseases, thereby highlighting the social and medical importance of managing transport emissions as part of urban environmental policy.

Importantly, this method accounted not only for the direct effect of signal settings on emissions at the prioritized directions (A&C), but also for the secondary effects on opposing flows (B&D), which would otherwise experience increased red light times. This comparative approach helped ensure that the optimization did not merely displace pollution from one stream to another, but instead sought balanced, system-wide emission reductions.

At the policy level, these findings hold critical implications. According to the broader research project on the potential transition to EFT in Kokshetau, the current share of electric vehicles is exceedingly low—only 13 electric cars out of nearly 69,000 registered vehicles. While the development of charging infrastructure is ongoing, including the installation of three EV charging stations and the creation of a dedicated training center, widespread adoption remains a long-term goal, heavily dependent on public readiness, affordability, and infrastructure scaling.

In this context, optimizing traffic light regulation is seen as a realistic interim solution that can significantly reduce emissions without the need for capital investments in upgrading transport infrastructure. This approach is cost-effective and technically feasible using existing computing resources and software tools.

An additional advantage is the introduction of an interactive software application (see Figure 10), which increases the availability of analysis tools for urban planners and transportation engineers. The application allows you to evaluate specific intersections and quickly test different traffic control scenarios depending on the time of day, which ensures flexibility and adaptability of decisions. Thus, local governments are able to make decisions based on data, and the methodology itself becomes a complementary tool to long-term initiatives such as electrification of transport and the development of public transport. Overall, the integration of neural network forecasting with emission-based optimization forms a scalable and reproducible platform for managing the quality of the urban environment. Although this approach cannot replace the systemic effect of a complete transition to electric transport, it serves as a practical bridge between current constraints and future goals, providing tangible improvements in air quality today.

The developed tool with a graphical interface (GUI) is aimed at practical application by municipal structures. It is capable of operating on standard office computers, does not require highly qualified personnel, and can be integrated with existing municipal GIS platforms. Short-term employee training is sufficient for its implementation, which makes this solution operational and accessible for widespread implementation in the framework of environmental management of urban transport.

In a broader context, the presented approach is consistent with the global practice of reducing transport emissions in cities implementing integrated sustainable mobility strategies. For example, in London, the introduction of low-emission zones, culminating in the Ultra Low Emission Zone (ULEZ), led to a marked improvement in air quality: the concentration of nitrogen dioxide (NO₂) in the central areas of the city decreased by about 18% in the first year of operation of the system [8].

Similar results are observed in Barcelona, where the “Superblocks” program, which provides for the restriction of automobile traffic and the redistribution of street space in favor of pedestrians, provided a decrease in NO₂ levels by about 25%, as well as a significant decrease in concentrations of particulate matter (PM) [48]. Additional simulations conducted in Barcelona and Madrid show that a massive switch to electric vehicles can significantly reduce NO_x and PM emissions within urban areas.

The example of the Chinese city of Shenzhen deserves special attention, which in 2017 became the first megacity in the world to fully electrify its public transport fleet. This step not only reduced the level of urban pollution, but also increased the efficiency of the transport system [31].

All these examples confirm that the integration of policy measures, infrastructure solutions and technological innovations provides significant environmental benefits. They also demonstrate that temporary measures such as optimizing traffic management can play an important role as part of a comprehensive transition to low-carbon urban mobility.

A comparative analysis shows that Kokshetau’s achievements in reducing emissions as a result of the introduction of optimized traffic light regulation (at the level of 9–12%) are comparable to the results recorded in London (≈11%) and Barcelona (≈10%) after the implementation of similar measures [6,8]. This indicates that the proposed approach is capable of providing comparable efficiency even in conditions of limited resources and infrastructure typical of medium-sized cities.

5. Conclusions

The primary objective of this study was to develop and validate an integrated methodological framework for reducing urban traffic-related air pollution in Kokshetau, with a focus on short- to medium-term measures that can be implemented prior to large-scale adoption of environmentally friendly transportation (EFT), such as electric vehicles. By combining an artificial neural network (ANN) model for emission forecasting with constrained nonlinear optimization of traffic signal timing, the research aimed to identify operational strategies that minimize emissions of CO, NO_x, SO₂, and particulate matter without compromising traffic flow symmetry.

The results of the study confirmed that optimizing traffic light regulation for individual intersections can significantly reduce the level of pollutant emissions, especially in areas with high traffic congestion. At the same time, such reductions can be achieved using existing infrastructure and available computing tools, without the need for large-scale capital investments. The developed artificial neural network (ANN) demonstrated high accuracy in predicting multicomponent emissions (R ≈ 0.99), and the implemented optimization procedure provided balanced system effects, taking into account secondary emissions in oncoming traffic directions.

The results obtained are especially relevant for the city of Kokshetau, where the share of electric vehicles remains extremely low (only 0.019% of the fleet), the infrastructure of charging stations is limited to three points, and rush hour is accompanied by the highest concentrations of pollutants in the air. In such conditions, optimization of traffic light cycles becomes a realistic and economically justified tool for improving the quality of the urban environment.

However, the study has a number of limitations that should be considered when interpreting the results. Firstly, the emissions assessment was based on the regulatory coefficients of the RND 211.2.02.11-2004 methodology used in domestic practice. Although this methodology is suitable for basic calculations, it does not reflect individual differences in vehicles, including maintenance conditions, fuel quality, and the effectiveness of exhaust control systems. As a result, it is possible to both underestimate and overestimate the actual emissions, especially given that Kazakhstan’s fleet is older than the European one and in many cases is not equipped with catalysts or particulate filters (DPF).

Secondly, the optimization model and algorithm focused exclusively on traffic light intersections, not covering other sources of pollution such as highway sections, roundabouts, and cargo terminals. Thirdly, the analysis considered static traffic control scenarios for fixed time intervals, without the use of adaptive or intelligent systems capable of dynamically responding to fluctuations in traffic flows in real time. Fourthly, the pollutant scope in this study was limited to CO, NO_x, SO₂ and PM_2.5 to maintain comparability with available city datasets and routine monitoring. Other toxic exhaust constituents especially carbonyls such as formaldehyde and acetaldehyde were not quantified due to limited routine measurements and the lack of locally validated emission factors at the intersection level. Given their co-emission under idling and stop-and-go conditions, future deployments should include targeted carbonyl monitoring to complement the present analysis.

Future research should address these gaps by integrating real-time traffic and emissions sensing into the optimization loop, enabling dynamic control strategies that can adapt to sudden changes in traffic patterns. Expanding the spatial scope beyond intersections to entire corridors would provide a more comprehensive emission reduction strategy. In addition, coupling the methodology with dispersion modeling could help translate emission reductions into direct estimates of exposure and health impacts. In the context of the broader EFT transition project for Kokshetau, further studies should also assess the combined effect of short-term traffic management measures and medium- to long-term measures such as expanding the EV fleet, electrifying public transport, and introducing low emission zones. Such multi-level strategies, supported by robust modeling, offer the greatest potential for achieving sustainable improvements in urban air quality and public health while supporting the city’s environmental and climate goals. To strengthen external validity, a before–after street measurement campaign is planned at treated and control intersections covering criteria pollutants and selected carbonyls to verify the modeled reductions under real operating conditions.