[go: up one dir, main page]
Next Article in Journal
A CIA Triad-Based Taxonomy of Prompt Attacks on Large Language Models
Previous Article in Journal
Frame Aggregation with Simple Block Acknowledgement Mechanism to Provide Strict Quality of Service Guarantee to Emergency Traffic in Wireless Networks
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Future Internet
Volume 17
Issue 3
10.3390/fi17030112
futureinternet-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

A Regional Multi-Agent Air Monitoring Platform

by
Stanimir Stoyanov
1,2,*,
Emil Doychev
1,
Asya Stoyanova-Doycheva
1,
Veneta Tabakova-Komsalova
1,2,*,
Ivan Stoyanov
2 and
Iliya Nedelchev
1
1
Department of Computer Systems, University of Plovdiv “Paisii Hilendarski”, 4000 Plovdiv, Bulgaria
2
Institute of Information and Communication Technologies, Bulgarian Academy of Sciences, 1113 Sofia, Bulgaria
*
Authors to whom correspondence should be addressed.
Future Internet 2025, 17(3), 112; https://doi.org/10.3390/fi17030112
Submission received: 16 January 2025 / Revised: 25 February 2025 / Accepted: 26 February 2025 / Published: 3 March 2025
(This article belongs to the Special Issue Intelligent Agents and Their Application)
Browse Figures
Figure 1
<p>A map of data sources for the city of Plovdiv (visualization of our sensor network).</p> "> Figure 2
<p>A chart showing the concentration of fine particulate matter for Plovdiv-Kamenitsa.</p> "> Figure 3
<p>A chart showing the concentration of fine particulate matter for Plovdiv-Tsentar.</p> "> Figure 4
<p>A chart showing the concentration of fine particulate matter for Plovdiv-Thrace/Trakia.</p> "> Figure 5
<p>A chart showing the concentration of fine particulate matter for Plovdiv-Sahat Tepe.</p> "> Figure 6
<p>A chart showing the daily average concentrations of PM10 and PM2.5 and temperature for 2022 and 2023.</p> "> Figure 6 Cont.
<p>A chart showing the daily average concentrations of PM10 and PM2.5 and temperature for 2022 and 2023.</p> "> Figure 7
<p>A chart showing the daily average concentrations of PM10 and humidity for 2022 and 2023.</p> "> Figure 8
<p>ACreM platform architecture.</p> "> Figure 9
<p>AM general architecture.</p> "> Figure 10
<p>Segment of the AM’s Beliefs Base (initial state).</p> "> Figure 11
<p>AM’s reasoning cycle state chart diagram.</p> "> Figure 12
<p>Segment of the AM’s Plan Library and reasoning cycle.</p> "> Figure 13
<p>Sample agent test session.</p> "> Figure 14
<p>Air Pollution Ontology taxonomy.</p> "> Figure 15
<p>Annotation properties of the Carbon_Monoxide class in the Air Pollution Ontology.</p> "> Figure 16
<p>SubClass Of axiom for the Nitrogen_Dioxid class.</p> "> Figure 17
<p>Individual Plovdiv and data properties for different pollutions.</p> "> Figure 18
<p>Attributes that account for local factors of the Plovdiv region.</p> "> Figure 19
<p>Component architecture.</p> "> Figure 20
<p>Interface to the AM’s environment.</p> "> Figure 21
<p>Flow chart of the AMCo’s reasoning cycle.</p> "> Figure 22
<p>Segment of the AMCo’s code.</p> "> Figure 23
<p>Sample AMCo agent test session.</p> "> Figure 24
<p>Comparative characteristics of the two agents.</p> ">
Versions Notes

Abstract

:
Plovdiv faces significant air pollution challenges due to geographic, climatic, and industrial factors, making accurate air quality assessment critical. This study presents a hybrid multi-agent platform that integrates symbolic and sub-symbolic artificial intelligence to improve the reliability of air quality monitoring. The platform features a BDI agent, developed using JaCaMo, for processing real-time sensor measurements and a ReAct agent, implemented with LangChain, to incorporate external data sources and perform advanced analytics. By combining these AI approaches, the platform enhances data integration, detects anomalies, and resolves discrepancies between conflicting air quality reports. Furthermore, its scalable and adaptable architecture lays the foundation for future advancements in environmental monitoring. This research represents the first stage in developing an AI-powered system that supports more objective and data-driven decision-making for air quality management in Plovdiv.

1. Introduction

Air is essential for life, and people are often deeply concerned about its quality because with each breath we inhale not only the oxygen we need but also potentially harmful gases and fine particles. Chronic exposure to polluted air has been shown to have negative health effects on humans and is one of the main factors in reduced life expectancy. Although air monitoring is a heavily exploited topic, it is not losing its relevance for various reasons—on the contrary, it is becoming increasingly topical. The reasons for this are numerous, including climate change, increasing frequency of fires and droughts, and industrial and agricultural production. Thus, monitoring air conditions and air conditioning maintenance have become important in many industrialized, environmental, and urban areas. Besides pollutants, air quality is negatively affected by various emissions from transportation, electricity, and fuel use. The accumulation of harmful gases poses a life risk even in smart cities [1]. According to the IHME Institute, although deaths from air pollution decreased overall by 46% from 1990 to 2021, the total number of deaths caused by particulate matter increased by 93% from 1990 to 2021. This now represents one of the deadliest risks to human health globally [2].
In many of the threatened regions, effective air quality research and assessment systems are already in place or being developed to collect data on pollutant intensity. In addition, air quality is also influenced by other factors such as location, weather, and climate, and, therefore, air quality assessment has become an important research area. The air quality control procedure is a cycle of interrelated components [3]. Government and regional bodies, as well as non-governmental organizations, research institutes, and universities, are involved in ensuring this procedure. To protect human health, standards for acceptable levels of air pollutants are established and adopted. Air quality administrators use a variety of air monitoring devices, air property modeling, and other assessment tools to fully recognize air quality problems. The air monitoring cycle is a dynamic process involving continuous review and evaluation of goals and strategies based on their effectiveness. All parts of this process must be based on scientific research that provides air quality managers with a basic understanding of how pollutants are emitted, transported, and transformed in the air as well as their impact on human health and the environment. Recently, technologies involving smart components and big data have been increasingly used for air monitoring.
Bulgaria is no exception, with several regions with poor air quality, including Plovdiv. Due to its geographical location, Plovdiv is surrounded by multiple pollution sources such as a lead-zinc plant, landfills, and agricultural production. With fires common during summer, the Plovdiv region suffers from heavy particulate pollution. At the same time, Plovdiv is one of the oldest continuously inhabited cities in the world and a major tourist destination, so the topic of air cleanliness is directly related to its prestige and economy. Currently, data and information are obtained from various sources, including government and local authorities, non-governmental organizations, universities, and private initiatives. However, differences between datasets are often the cause of disputes about their reliability and the conditions under which data were collected. There are often various conflicting measurements, assumptions, opinions, and publications about the reliability of the measured data and the location of the measuring instruments. Discrepancies are the cause of ineffective decision-making. For these reasons, air monitoring in the Plovdiv region is a highly sensitive topic.
To establish an objective representation of air quality in the region of Plovdiv, this study aims to develop a prototype software platform that delivers the most reliable information by doing the following:
  • Collecting and analyzing measurement data from a previously established sensor network in Plovdiv [4];
  • Incorporating additional information from external sources;
  • Continuously comparing data to identify deviations and discrepancies;
  • Ensuring reliability and accuracy in the measured and obtained data to support objective conclusions and informed decision-making.
To achieve these objectives, we adopt a structured research methodology that integrates both structured and unstructured data sources. Our approach begins with the collection and analysis of measurement data from a previously established sensor network in Plovdiv, ensuring accurate and reliable air quality assessment. In addition to this structured data, we retrieve supplementary information from external sources, such as the World Health Organization’s system, enriching the dataset with broader environmental insights.
As part of our methodology, we have developed prototypes of two intelligent agents: a BDI agent and a ReAct agent. These agents are designed to extract, process, and integrate data from various sources, providing a comprehensive understanding of air quality conditions. Each agent independently interacts with users by periodically providing updates on air quality status, thereby improving real-time awareness and decision-making.
By combining these methodologies within a hybrid multi-agent platform, we enhance the accuracy, reliability, and objectivity of air quality monitoring in the Plovdiv region.
This paper summarizes the results of the first stage of developing a prototype hybrid multi-agent regional air monitoring platform with the working acronym ACreM (Air Credible Monitoring). The platform is hybrid in that it combines methods from symbolic and sub-symbolic artificial intelligence. Increasingly, their integration is referred to as integrated artificial intelligence, which is defined as the next generation of artificial intelligence [5,6]. Because the data measured and obtained from external sources, as well as the documents used, are very diverse and of different formats, we are convinced that the mutual use of components from both fields of artificial intelligence is appropriate. Furthermore, we find this problem suitable for demonstrating the capabilities of integrated artificial intelligence. In the first stage, our goal was to develop the architecture of the platform and to verify its operability by developing prototypes of its main components and conducting experiments with them. After developing and stabilizing the basic architecture, we intend to use it to solve various tasks related to conducting objective monitoring of the air condition in the Plovdiv region.

2. Related Works

There are various global air monitoring platforms that offer different analyses, usually using classical statistical models. In the European area, one of the best-known platforms is DELTA v7.0 [7]. DELTA Assessment & Planning software was developed to assess models and visualizations from complex numerical data in accordance with the FAIRMODE benchmarking initiative [8]. Two basic tools are integrated in the platform, the first for assessment and planning and the second for the benchmarking of emission inventories. The European Air Quality Index application [9] provides access to up-to-date air quality monitoring data from stations located in European cities. This platform provides air quality forecasts and health advisories in 24 European languages depending on the users’ location. The World Health Organization provides technical support to Member States in developing regulatory guidance and tools and providing authoritative advice on health issues related to air pollution and its sources [10]. The organization prepares reports on global trends and changes in health outcomes related to actions taken to address air pollution at global, national, and regional scales.
We will briefly present the solutions that were announced as some of the best air quality monitoring solutions for 2024. SafetyCulture [11] offers one of the best air quality management software platforms due to its included feature set that allows various industries to remotely track and analyze air quality in facilities, commercial sites, and vehicles. Similar features are offered by the Aeroqual platform [12], established in 2001 and based in New Zealand. Kaiterra [13] can search for air quality monitoring solutions through advanced software analysis and a team of experienced experts. Ref. [14] supports companies in Indonesia, India, and Lagos with analyses using parametric, location, and time comparisons. One of the world’s leaders in environmental management solutions, Envirosuite [15], offers high-quality air monitoring software for mining, manufacturing, and transportation companies. Rated as one of the top environmental software providers, ERA Environmental [16] has always had a comprehensive approach to environmental management, with advanced emissions tracking capabilities and custom reporting features. Building on its many years of experience in air quality monitoring, ENVEA [17] provides a wide range of digital solutions, hardware, and services for industrial facilities, government agencies, and research institutions.
The assessment of air quality is complicated by the fact that pollution is transported from one location to another by wind in the atmosphere, thus transforming the problem from a local one into a global one. Therefore, ground-based and satellite observations of larger areas can provide valuable data on air quality. Copernicus Atmosphere Monitoring Service [18] provides computational models of the atmosphere that can be applied to make daily forecasts of air quality using satellite and ground-based observations. The combination of the huge number of daily observations and the accuracy and efficiency of the models used is the great advantage of this system. In addition, the system is able to prepare daily analyses and forecasts of the global movement of atmospheric gases over the whole of Europe.
Understanding the importance of the issue, MDPI has published many special issues on air monitoring, and some of the articles will be referenced in this section. Increasingly, analyses applying artificial intelligence approaches and models, such as machine learning using large datasets, transform models, and similar techniques, are being published. Such examples can be found in many of MDPI’s special issues. In [19], a framework is proposed for global air quality forecasting incorporating Busan, South Korea, as a model city using an attention-based convolutional BiLSTM autoencoder model trained to forecast PM2.5 and PM10 particulate matter pollution intensity. A new version, extended with an artificial intelligence algorithm, of a previously proposed platform for air quality monitoring based on carbon dioxide concentration measurements is presented in [20]. In [21], the authors claim that a comparative analysis in the USA, conducted with monitors using two methods (Federal Equivalent Method and Federal Reference Method), shows that there is a problem with the accuracy and reliability of data when measuring fine particulate matter concentrations. Comparability between the two types of monitors can have significant implications for maintaining compliance with National Ambient Air Quality Standards (NAAQS). Furthermore, this study examined the effectiveness of the performance of both monitors used to conduct PM2.5 measurements collected from 276 monitoring stations. In the paper [22], various methods such as support vector regressions, gradient boosting decision trees, neural networks, and others are presented to analyze the impact of air quality index in Xi’an city, China, starting from 1 October 2022. Comparative tests were also conducted. In [23], it is argued that traditional air monitoring methods are often limited by spatial coverage and accessibility. Therefore, the authors propose using methods whose mobility and advanced sensing capabilities overcome these limitations. In [24], a multi-scale fusion model is proposed to study both single-step and multistep prediction. In [25], the results of developing an adaptive hardware-software platform are summarized, following the IoT paradigm, demonstrating a high level of accuracy in predicting carbon dioxide trends by analyzing only a limited window of recent data.
Air quality is influenced by multidimensional factors, including location, weather, and uncertain variables. Currently, many scientists use big data for their analyses thanks to advances in this field and the availability of a numerous environmental monitoring and sensor networks. In [1], various big data and machine learning-based techniques are explored for air quality prediction. In addition, on the paper highlights some of the challenges and future research needs.

3. State of the Air in Plovdiv

Although there are standards for acceptable air quality, pollution in urban settings is caused by a variety of factors. The overall pollution picture in a city results from a complex overlay of different backgrounds—natural, regional, and urban. Depending on the specific conditions, these backgrounds may amplify or neutralize each other. Urban air pollution is a complex issue that requires comprehensive and systematic analysis. Only by understanding all the factors—natural, regional, and urban—and how they interact can an accurate picture of the situation be developed and effective measures taken to improve air quality. This integrated approach, combining different levels of analysis, is key in developing sustainable solutions to tackle urban air pollution. Air quality depends not only on pollutant emissions from different sources such as industry, transport, and tertiary and domestic sources, but also on weather conditions that control the dispersion of pollutants or, conversely, their accumulation. For these reasons, it is necessary to have an objective understanding of the air quality in the region of interest (in this case, the Plovdiv region). Plovdiv is the second largest and most important city in the Republic of Bulgaria, located on both sides of the Maritsa River. The natural environment of Plovdiv is characterized by a specific relief, dominated by the six syenite hills, which are not only an emblematic element of the landscape but also protected areas with high biodiversity. Together with the Maritza River valley, they create a microclimate characterized by mild winters and hot, humid summers. The urban environment is urbanized, with high building densities, intensive car traffic, and industrial activity. Some analyses from external sources regarding the Plovdiv region are published in [26,27]. In this section, we summarize the results of our measurement analysis.
Our network includes 11 sensor complexes located at key points in the city of Plovdiv and multiple sensors in the surrounding towns. The sensors measure the concentration of various pollutants, including fine particulate matter (PM10 and PM2.5), nitrogen oxides, ozone, and other gases that affect air quality. In addition to major air pollutants, sensors in our network also report data on temperature, atmospheric pressure, humidity, gamma radiation, and more. The data, available at [4], is collected in real time every 5 s and analyzed using specialized software platforms. The aim of the network is to provide the public and local authorities with accurate and timely information on the air quality in the measured residential areas. The system is part of an effort to raise awareness of environmental issues and create a healthier and more sustainable urban environment. In Figure 1, we have marked in red the areas of the city that have sensors from our network and external sources, and in green, we have marked some of the sensors we monitor in areas of the city where data from other sources are not available.
With the following charts, we visualize the comparison of data from our sources and those of the European Environment Agency. The chart (Figure 2) shows the concentrations of fine particulate matter (PM) PM2.5 and PM10 in Plovdiv-Kamenitsa, comparing them with the current European Union standards and the European Environment Agency data for 2022 and 2023.
According to EU standards, the annual limits are as follows: PM2.5–5 µg/m3; PM10–15 µg/m3. Our sensor data for this area for 2022 are as follows: PM2.5–15 µg/m3 and PM10–21 µg/m3. It can be seen that in 2023, the levels increased to 19 µg/m3 for PM2.5 and 26 µg/m3 for PM10. Data from an external source, the European Environment Agency, show that the levels for PM2.5 are 17 µg/m3 and for PM10 33 µg/m3. In 2023, their data report levels are 16 µg/m3 for PM2.5 and 37 µg/m3 for PM10. We can conclude that the levels of PM pollution in this area are significantly above EU air quality standards, with values remaining above the permissible limits. Compared to the European Environment Agency data, the area has higher levels of PM, indicating serious air quality problems in this area.
The following chart (Figure 3) shows the concentrations of fine particulate matter (PM) PM 2.5 and PM 10 in Plovdiv-Tsentar, comparing our data with external sources.
In 2022 and 2023, PM2.5 and PM10 levels in TSENTAR significantly exceed EU standards, indicating serious air quality problems. For 2023, PM10 levels (33 µg/m3) are particularly high and significantly above the acceptable standards.
In 2024, there is a significant decrease in PM concentrations in the TSENTAR. PM2.5 levels drop to 11 µg/m3 and PM10 levels drop to 15 µg/m3, both of which are within the EU standards. Data from the European Environment Agency for 2022 and 2023 show that PM levels in TSENTAR were above the EU average. In 2024, data from the Agency are not available. Despite the apparent reduction in pollution in 2024, efforts must continue to maintain and further reduce PM pollution, especially PM2.5. Regular monitoring and implementation of pollution control measures are critical to improving air quality in Plovdiv. With the current actions and pollution control measures, there is a positive trend in the improvement of air quality in Plovdiv. However, long-term efforts will still be needed to reach and maintain EU clean air standards.
The image (Figure 4) is a chart comparing annual levels of fine particulate matter (PM2.5 and PM10) in Plovdiv-TRACIA, Bulgaria, for different sources and years. The chart is interesting and relevant because it shows how air pollution levels have changed over the years and how they compare to current EU air quality standards. This is important for assessing air quality and taking measures to improve the environment.
The annual data from our and external sources for the Plovdiv-Trakia region for PM2.5 and PM10 by year compared with data from the European Environment Agency and EU standards show the following:
  • In 2022, the levels of PM2.5 (18 µg/m3) and PM10 (24 µg/m3) were above the EU standard, and they increased in 2023 to PM2.5 (21 µg/m3), PM10 (29 µg/m3), leaving the values above the limit.
  • Data from the European Environment Agency in 2022 show PM2.5 levels of 17 µg/m3 and in 2023, the levels are 16 µg/m3. These levels remain stable but still above the EU standard. Data in 2022 for PM10 (33 µg/m3) levels are also above the EU standard and increased to 37 µg/m3 in 2023.
In the period from 2022 to 2023, the levels of PM2.5 and PM10 in Plovdiv-Trakia remain above the EU standards, and an increase is observed, especially for PM2.5. When compared with data from the European Environment Agency for the same period, PM2.5 levels remain relatively stable. PM10 levels are high, exceeding the EU standard more than twice in 2023. The data show that the air quality in Plovdiv-Trakia is not within the permissible limits according to both EU and European Environment Agency standards.
The chart (Figure 5) shows the annual levels of fine particulate matter (PM2.5 and PM10) for the Sahat Tepe area of Plovdiv and it also compares the current European Union (EU) standards and European Environment Agency data for 2022 and 2023. Our measurement data in 2022 and 2023 show that the levels of PM2.5 and PM10 in the Sahat Tepe area exceed European air quality standards but are still at levels that can be improved. Comparison with European Environment Agency data for 2022 and 2023 shows levels of PMF above EU standards, but with significant variation by year and particle type. In 2022, the European Environment Agency PM10 levels were higher (33 µg/m3) compared to our data (23 µg/m3), and in 2023, the EPA PM10 levels (37 µg/m3) were higher than our data (19 µg/m3).
For a better understanding of the pollution in the city of Plovdiv, we consider our other sources located in different areas of the city (Plovdiv-Western District; Plovdiv-Southern District; Plovdiv-Marasha; Plovdiv-Gagarin; Plovdiv-Garata).
Using mathematical-statistical and analytical methods, we establish spatio-temporal patterns in the distribution of the concentration of PM10 in the air of Plovdiv and determine to what extent the meteorological elements temperature, atmospheric pressure, humidity, wind direction, and wind speed influence this distribution. In order to reveal the intra-annual distribution of FPH10 content, the daily average concentrations of FPH10 in the air of Plovdiv were calculated for each month of the year of the study period 2022–2023. From the data in this table, we compiled statistics for 730 days of both years which can be viewed in Table 1: Statistics.
In the statistical analysis, an inverse relationship was observed between temperature and PM10 and PM2.5 content (Figure 6). This correlation is often explained by meteorological factors and human activities, such as the following:
  • Inversions in the atmosphere that keep pollutants close to the ground at low temperatures;
  • Increase in heating fuels in winter (especially solid fuel or wood);
  • Reduced ventilation in urban areas due to stagnant air.
It is important to note that the relationship between temperature and PM can vary depending on local conditions and pollution sources.
The relationship between relative humidity (RH) and PM10 (Figure 7) content can be described as a positive correlation under certain conditions, for example,
  • Condensation of water molecules on dust particles. This increases their size and mass, resulting in higher measured PM10 concentrations.
  • Atmospheric stability: High humidity is often associated with stagnant air, which makes it difficult for dust particles to disperse.
  • Formation of secondary aerosols: Humidity facilitates chemical reactions in the atmosphere that lead to the formation of secondary particles, increasing the total PM10.
However, this relationship may not be linear. At very high humidity (>80–90%), e.g., during fog, PM10 concentrations can increase significantly. Conversely, at low humidity (<30%), particulate matter is drier and does not increase by condensation but may remain in the atmosphere for longer periods if precipitation is absent.
The correlation between RH and PM10 depends on the specific pollution sources and local meteorological conditions. For example, in industrial areas or with high use of heating systems the correlation may be more pronounced.
Air temperature, relative humidity, wind speed, and wind direction are key climatic parameters that determine the characteristics of a region’s atmosphere. Modern automated weather stations provide continuous measurements of these parameters, providing data at short intervals (e.g., every 5 s) during the day.
The climate in the Plovdiv region is transient-continental with moderate rainfall and prolonged summer droughts. The region is dominated by westerly and easterly winds with relatively low speed. The autumn–winter period (October 2022–March 2023) is characterized by insignificant rainfall. During this period, local heating systems are widely used as temperatures drop. In the domestic sector, mainly solid fuels (coal, wood, household waste) are used. Due to the large number of sources and the low quality of fuels, the concentration of controlled pollutants increases in this period. The specific meteorological conditions in the area also contribute to the high levels of fine particulate matter—a large number of days with calm weather (wind speeds below 1.5 m/s), temperature inversions (occurring on about 81% of days in the year), and fogs along the Maritza River all lead to the retention and accumulation of pollutants.
The analysis of the data (Figure 6 and Figure 7) for the winter period shows that the values vary above the average daily and annual standards for the protection of human health, specified in [9,10]. The comparison of the values during the winter season with those registered during the summer period leads to the conclusion that the levels of this pollutant are directly related to domestic heating (increased consumption of solid fuels for heating) and car traffic. In combination with unfavorable climatic conditions and characteristic topographic features that disrupt the dispersion of emitted pollutants, high concentrations are registered. These factors, combined with Plovdiv’s specific topographical and climatic characteristics, hinder pollutant dispersion, leading to elevated concentrations of fine particulate matter. The urban landscape, featuring high-density construction and syenite hills, further amplifies pollution retention.
Monitoring air quality in Plovdiv requires data integration from multiple sources, including the European Environment Agency and local sensor networks. A comprehensive approach is necessary to track pollution trends, detect anomalies, and assess environmental and public health impacts accurately. The current data from Plovdiv’s sensor network, which captures real-time measurements every five seconds, reveals that particulate matter (PM2.5 and PM10) concentrations frequently exceed EU air quality standards, particularly in densely populated areas. Despite some improvements in 2024, persistent efforts are needed to maintain long-term reductions in pollution.
Regular air quality monitoring is essential to mitigating health risks, improving control measures, and understanding the relationship between pollution and climatic factors. Given the complexities of urban air pollution, an integrated assessment approach is crucial for implementing effective strategies to improve air quality. In response, our team is expanding sensor coverage to include data-deficient areas and incorporating automated real-time analysis to enable prompt pollution detection. These initiatives will support sustainable urban development and contribute to a healthier environment in Plovdiv.

4. The ACreM Platform

The ACreM (Air Credible Monitoring) platform is agent-centric, hybrid, and regional. It is agent-centric because the two main components of the platform are implemented as agents. In modern artificial intelligence, the emergence of the agent-oriented concept is traced back to the mid-1990s. An intelligent agent is a computer system that can operate autonomously, reactively, proactively, and socially within some environments to achieve designated goals [28]. The platform is hybrid because an agent can be classified as symbolic AI while another can be classified as sub-symbolic AI. In the evolution of artificial intelligence, the two main directions have been competitive and often mutually exclusive. Recently, integrated artificial intelligence systems that combine components from both areas have gained growing attention. We believe that hybridity is very appropriate in our case for a variety of reasons. One of the most important ones is that for its intended use, the platform needs to acquire and analyze data with different formats from a variety of sources. The data can be grouped as follows:
  • Internal—hot data extracted from our sources (mainly our sensor network) and stored in a relational database. In addition, specialized expertise for the specific application domain is stored in an appropriate repository (ontology). In general, the internal data and knowledge are stored in structured formats.
  • External—the data are obtained from various sources such as publications, statistics, and foreign measurements. This information is usually unstructured and may include a variety of different, including scientifically unproven, content.
The platform is mainly intended for use in the Plovdiv region. For this purpose, parameters specific to the region are incorporated (as detailed in the previous section) in the platform’s operation. The architecture of the ACreM platform is given in Figure 8.

4.1. Air Monitor Agent

The Air Monitor (AM) is a personal assistant that is designed to identify and locate various anomalies related to the state of the air based on our measurements. AM was developed as a BDI (Beliefs-Desires-Intention) agent in the JaCaMo development environment [29]. The BDI model has its roots in the philosophical tradition for understanding practical reasoning in humans. Practical reasoning is action-oriented—a process of seeking what to do. A practical deliberation is a matter of weighing conflicting considerations for and against competing options dependent on the desires, concerns, or value judgments of the agent [30]. As a process, practical reasoning involves two distinct activities. The first one, known as deliberation, is deciding what state of affairs (goal) we want to achieve. The second, called planning, is deciding how to achieve that state of affairs (that goal). One of the advantages of this architecture is the ability to flexibly represent and work with the agent’s environment, which is essential for our goals. Another advantage is that AM works with verified expertise for our application domain of interest (in this case air pollution). This circumstance is at the same time a disadvantage of the ReAct agent, which also works with free text created by non-professionals such as journalists, politicians, and interested citizens. Thus, we are of the opinion that combining the two approaches is an adequate way to achieve the aim of the study, namely the search for reliable information about the air situation in Plovdiv. At the same time, we want to demonstrate and explore the power of the integration of methods from the two main areas of artificial intelligence. Along with this, we want to show a concrete example that demonstrates the advantages of integrated artificial intelligence.
Figure 9 shows the general architecture of AM situated in its working environment. The diagram presents the main components of the AM and its environment.
The Beliefs Base is a discrete structure that models an agent’s perception of its environment. Individual beliefs are represented as predicates (in the style of the Prolog logic programming language). The structure consists of two base components, static and dynamic (Figure 9). The static component contains parameters with immutable values (usually threshold and limit values) used to estimate the air condition and the type of pollutants. This component is initialized by interacting with the Air Pollution Ontology, from which the necessary parameters are extracted. The dynamic component models the current air status. The values of the dynamic parameters are extracted from the Sensor DB, which contains hot data from the actual measurements. Any change to the Beliefs Base (deleting or adding a belief) triggers an actual event and updates at a predefined time interval. A sample Beliefs Base segment is given in Figure 10.
The Plan Library (Figure 11) stores AM plans, each following a three-component structure: triggering_event: context <- body, where: and <-serve as separators. Updates to the AM’s beliefs or intentions—either additions (+) or deletions (−)—are modeled through events.
Belief-changing events occur when the AM updates its beliefs based on new perceptions of the environment. Intention-changing events (i.e., adding intentions) typically result from executing other plans. Contexts determine when a plan is applicable to the agent. Since the agent operates in a dynamic environment, contexts help delay commitment to a specific action until the latest possible moment. Thus, when selecting a plan to achieve a goal, the context ensures that a plan with the highest probability of success is chosen. The body of the plan contains a series of executable actions, separated by “;”. A plan may include subgoals, which are achieved using separate plans. Plan continuation requires that corresponding subgoals are met (i.e., the associated subplans are completed). We distinguish two types of subgoals:
  • Achievement goals (denoted by “!”)
  • Test subgoals (denoted by “?”)
Test subgoals are commonly used to extract data from various data structures or repositories.
AM’s reasoning cycle is based on the concept of practical reasoning by rational agents [31] and includes the following steps (Figure 11):
Initialization. During initialization, the agent perceives the environment (Air Pollution Ontology and Sensor DB) to initialize the static Beliefs Base. Limiting and threshold values of observed air pollutants are usually obtained from the Air Pollution Ontology. This phase also specifies the initial intention !start_air_monitoring causing the monitoring process to start.
Reasoning. In this phase, the dynamic Beliefs Base is periodically updated by reading the Sensor DB. This database stores the current values of the observed sensors for measuring various air pollutants. Practical BDI agents operate by continuously handling events, which represent either perceived changes in the environment or changes in the agent’s own goals. The AM agent continuously monitors its environment for changes. This could be new sensor data (interacting with the Sensor DB also located in the environment), updates to the Air Pollution Ontology, or external events. Thus, changes in the Beliefs Base cause the generation of a current event, which is used in the next phase to select an applicable plan.
Plan Selection. The plan selection is done in two steps. First, a set of possible plans is determined through the successful unification of the triggering event of a plan with the current event. Second, from the set of possible plans, one applicable plan is selected through the successful unification of the context of the plan with the Beliefs Base.
Plan Execution. The AM agent executes the selected applicable plan. The plan consists of a series of actions, which may include subgoals that need to be achieved. Typical actions are comparisons between the current values of the monitored pollutants and the permissible threshold values to determine whether the air condition is normal or anomalies are observed. In both cases, AM prepares informative messages for users. The actions could also involve querying the ontology, updating the Beliefs Base, or triggering external processes. If the plan includes subgoals, the agent will recursively apply the reasoning cycle to achieve these subgoals. Subgoals can be either achievement goals (denoted by “!”) or test subgoals (denoted by “?”).
Termination. The reasoning cycle continues until the AM agent’s intentions are achieved or the agent is terminated. For example, if the agent’s intention is to monitor air quality continuously, the cycle will repeat indefinitely. In each cycle, the agent informs the user whether any deviations have been detected during air monitoring. If the agent is shut down, it will terminate its reasoning cycle and stop monitoring the environment.
The execution of the reasoning cycle is supported by JaCaMo run-time interpreter [30].
Part of the code that implements the library plan and reasoning cycle is given in Figure 12.
Figure 13 shows a sample test session for working with the AM.
AM Environment. Two main repositories are located in the AM environment. The core knowledge for understanding and monitoring air quality is stored in an ontology called Air Pollution Ontology, developed under the World Health Organization’s global air quality guidelines of [32]. The Executive Environmental Agency of Bulgaria [33] and Environmental Monitoring System of Plovdiv Municipality [34] offer comprehensive information on air quality in Bulgaria, including measurements of key pollutants such as PM10, PM2.5, SO2, O3, CO, and NO2, available on their official website. This foundational knowledge serves as the basis for evaluating air quality by comparing it with dynamically entered data in the databases. Inferences can be made about the occurrence of certain pollution events, the impact of specific pollutants in a region, or even the prediction of potential air quality incidents to which local authorities and citizens should respond.
This section introduces the basic ontologies for storing the underlying knowledge for air quality management and explains the concept of knowledge processing in the ACreM platform. The Air Pollution Ontology captures structured knowledge about air pollutants, their sources, and their environmental and health impacts. This ontology forms the foundation for understanding and analyzing air quality data, providing a semantic framework for integrating dynamic data from IoT sensors and other real-time inputs. In the ACreM platform, air quality parameters, including pollutant types and regulatory thresholds, are formally structured within the ontology to ensure consistency and accurate reasoning. The ontology defines key pollutants such as PM10, PM2.5, NO2, SO2, CO, and O3 as distinct classes, each associated with attributes specifying their emission sources, health impacts, and chemical properties. Additionally, threshold values are integrated based on internationally recognized standards, such as WHO and EU air quality guidelines, using properties like hasLimitValue for concentration thresholds and hasDuration for applicable timeframes. This structured representation allows the system to dynamically compare real-time sensor data with regulatory limits, identify anomalies, and support decision-making processes for air quality management in the Plovdiv region.
The ontology represents a comprehensive taxonomy of air pollutants, organized hierarchically to reflect their chemical and physical characteristics. The taxonomy includes classes for the following (Figure 14):
  • Particulate Matter (PM)—differentiated into PM10 and PM2.5 based on particle size, with attributes for their sources, effects, and regulatory limits.
  • Gaseous Pollutants—includes NO2, SO2, CO, and O3, each described in terms of their chemical properties, emission sources, and atmospheric behaviors.
For each pollutant class, the ontology includes annotation properties (Figure 15):
  • rdfs:comment provides detailed descriptions of the pollutant’s characteristics, sources, and health impacts.
  • dcterms:description aligns with the Dublin Core Metadata Initiative’s guidelines. It is commonly used to provide concise, descriptive information about ontology elements.
To describe the standard norms according to the WHO Global Air Quality Guidelines of the World Health Organization, axioms have been created for each pollutant. For example, the axiom defining the maximum values for nitrogen dioxide is presented in Figure 16.
The SubClass Of axiom for the Nitrogen_Dioxid class models the regulatory limits for nitrogen dioxide (NO2) based on its concentration over two distinct timeframes: annual and daily. These constraints are represented using specific properties that associate the pollutant with its permissible limits and their durations:
  • The annual average concentration of nitrogen dioxide is constrained to a maximum of 10 µg/m3, represented by the property hasLimitValue. This limit is further specified as applying over an “annual” duration using the hasDuration property.
  • The daily average concentration is constrained to a maximum of 25 µg/m3. Similarly, the hasDuration property specifies that this limit applies to a “daily” timeframe.
  • Such axioms are developed for all pollutant classes in the ontology. These axioms enable the ontology to precisely capture and differentiate between the temporal applicability of air quality standards. Using properties such as hasLimitValue and hasDuration ensures semantic clarity and allows for automated reasoning about compliance. For example, the ontology can determine whether observed pollutant levels adhere to these thresholds and flag non-compliance for further action.
This structured representation facilitates integration with real-time sensor data, allowing applications to dynamically monitor, analyze, and report air quality violations. By incorporating such temporal and quantitative constraints, the ontology provides a robust foundation for reasoning about air quality and supporting decision-making in environmental management.
The axioms created in the ontology are associated with threshold values for various air pollutants as defined by the World Health Organization (WHO) standards. These values, measured in milligrams per cubic meter (mg/m3), typically represent average daily concentrations, although some thresholds may also include annual averages or shorter-term values. The threshold values may vary across different regions depending on their specific characteristics, such as geographic location, the presence of industrial zones, heating methods, traffic levels, and other environmental or socioeconomic factors.
To address this variability, we introduced individuals within the class RegionPollutants_(City). Each individual in this class represents a specific city or region and is defined using data properties that link the individual to its corresponding threshold values. For example, the thresholds for air pollutants in the region represented by the individual Plovdiv are detailed in Figure 17.
The ontology facilitates the modeling of region-specific pollutant thresholds by integrating both universal guidelines (e.g., WHO standards) and context-sensitive data tailored to local environments. The RegionPollutants_(City) class provides a framework to capture the unique air quality characteristics of specific regions, while maintaining consistency with global benchmarks.
For each region (e.g., Plovdiv), the ontology defines the following:
  • Thresholds for pollutants such as PM10, PM2.5, SO2, NO2, CO, and O3.
  • All values are expressed in mg/m3.
  • Thresholds correspond to specific durations, such as daily averages (e.g., 24 h periods) or annual means.
  • Attributes that account for local factors, such as dominant industrial activities, population density, and prevalent weather patterns, are presented as annotation properties for the concrete region (Figure 18).
The Air Pollution Ontology provides a robust framework for understanding and monitoring air quality. This ontology-based approach enables data-driven decision-making, helping to identify hotspots, predict pollution events, and develop tailored interventions to improve air quality in Plovdiv.
The ontology incorporates the standard threshold values for air pollutants defined by the World Health Organization (WHO), ensuring a globally recognized baseline for air quality assessment. Simultaneously, its architecture is designed to be extensible, allowing for the addition of specific data for individual regions. For example, the region of Plovdiv is included with detailed pollutant thresholds based on its unique geographic, industrial, and climatic characteristics.
Furthermore, in cases where specific regional data is unavailable, the ontology’s flexibility ensures that the WHO standard thresholds can be applied universally, serving as default values for any region. This dual capability makes the ontology both adaptable to local contexts and reliable for broader applications. By facilitating the integration of region-specific attributes alongside global standards, the ontology supports effective air quality management, compliance monitoring, and sustainable urban development across diverse geographic areas.
The second repository is a relational database that stores our measurements. The databases contain the dynamic data received from various IoT nodes or from real-time data entered by the platform’s users. The data flow commences with its inception—measurements conducted by devices within the sensor group (see Figure 19). The recorded values are subsequently read and consolidated into a singular data package by a dedicated controller device. This controller functions as an intermediary, adapting between the diverse sensor devices, each with its specific interface, and the broader system.
At regular intervals (typically every 1 to 5 min), the controller transmits the data package to the Data Endpoint of the software system. Upon arrival, the data package is directed to a Data Transformation module. Its primary function is to decompose the data package into a series of individual measurements, each associated with its respective value and metric. The implementation of the Data Transformation module leverages the Node Red’s Flows [35]. For every measurement, the Data Transformation module initiates a discrete REST request to the OGC’s SensorThings API, which, in turn, stores the measurement in the Data Storage. The Fraunhofer IOSB implementation [36] serves as the backbone for the SensorThings API. PostgreSQL, enhanced with the TimescaleDB extension [37], serves as the designated Data Storage. Once the data are securely stored in the Data Storage, retrieval becomes possible through a variety of SensorThings API read methods. These methods accommodate a broad spectrum of filters, allowing for meticulous customization of the retrieved data. Applications, such as Grafana [38], can leverage SensorThings API read methods for monitoring, visualization, or data export. Grafana, for instance, features a convenient preconfigured data source for seamless integration with the SensorThings API.
Interaction between AM and its environment. The interaction of the AM with its environment is based on the concept of A&A (Artifacts & Agents). The A&A meta-model and first ideas for its use in agent-oriented applications are discussed in [39,40]. One major challenge in developing agent-oriented applications is modeling the agents’ environment. Usually, a suitable representation is devised for each specific case. For this reason, the creation of A&A as a formalism for representing and modeling the environment, regardless of any particular application, is a significant advancement. Accordingly, an agent-oriented system is modeled as a triple (Ags, Bdg, Env) containing the following three sets: the set of agents (Ags), the environment (Env), and the bridge (Bdg) between them [41]. Considering the abstract nature of the metamodel, various agent-oriented languages and infrastructures can offer concrete implementations.
To interact with its environment, AM uses two interfaces implemented in Java and embedded as external routines in the corresponding plans. Integration is relatively easy since JaCaMo itself is also implemented in Java. The interfaces to the Air Pollution Ontology and Sensor DB are presented in Figure 20.

4.2. Air Monitor Copilot Agent

The Air Monitor Copilot (AMCo) is a copilot agent based on the ReAct architecture that we use to examine and evaluate air pollution data in the Plovdiv region from external sources. The ReAct framework [42] uses a combination of task decomposition, reasoning loops, and multiple issue resolution tools. The ReAct agents of the LangChain library can support a complete query processing process. Using these agents, we can break complex queries into manageable steps and execute them systematically. In the context of LangChain, dialog agents can exhibit dynamic behavior, integrating different data sources and services. LangChain agents can use a language model as a reasoning engine. Unlike chains where the sequence of actions is predefined, agents use language models to determine the sequence of actions based on the given context. LangChain allows us to easily integrate tools and APIs to improve the functionality of language models. This includes connecting models to external data sources and services, allowing for more dynamic and intelligent applications. By giving language models access to APIs and custom tools, developers can create more flexible and context-aware applications ranging from data mining to performing specific actions based on model output.
The current prototype version of AMCo consists of the following three components: the AMCo Agent, AirQualityApi, and AirQualityMobile. The AMCo Agent is a Python 3.9 project using the LangChain library and the OpenAI Large Language Model. The agent is able to analyze particle levels of unstructured data obtained from different sources of information. The agent first reads the resources (in the experiment presented, these are the World Health Organization guidelines for monitoring particulate matter pollution [32]), which it uses to initialize its internal vector database. For this purpose, it does an internal transformation of the text (converting it from .pdf to .txt) before generating the representation in the vector database. The next action of the agent is to connect OpenAI to the local vector database. This completes the preparation phase, and AMCo goes into operational mode. In the test experiment, every minute, the agent generates a query to an external source (in this case [27]) to retrieve information about the current state of particulate matter in the city center of Plovdiv. If the result is different from the previous state, a query is generated and sent to OpenAI to analyze the results, and then the result is forwarded as a query to AirQualityApi. AirQualityApi is a .NET API using MongoDB as the internal database. This AMCo component operates as a service used to store the results and provide access to them to the mobile application. In our case, it stores the current air dust status along with the analysis results from OpenAI. The AirQualityMobile application, used by AMCo, visualizes these analysis results. This app makes a request to AirQualityApi for visualization data every minute, ensuring that the analysis is refreshed whenever pollution levels change.
The AMCo agent’s reasoning cycle is presented in Figure 21.
A segment of program code implementing the AMCo behavior described above is given in Figure 22. The individual steps of the agent’s operational cycle are commented on in the code.
Results from a sample test session with the AMCo agent are presented in Figure 23.
A comparative characteristic of the two agents is given in Figure 24.

5. Conclusions and Future Work

The paper summarizes the results of the first stage of developing a platform supporting air monitoring in the Plovdiv region. By introducing the platform, we aim to contribute to the objective assessment of air quality, facilitating more effective decision-making. Currently, conflicting data is published by state and local authorities, as well as by NGOs, scientific institutes, and universities. Our goal is to start using the data that we have been collecting and storing for the last three years. Additionally, due to the layering of the causes of air pollution, we strive to identify the specific factors at any given time and thus contribute to timely and targeted interventions.
In the second stage of development, we intend to work in three directions. The first one is to refine the current prototype platform. Interaction between the two agents is one of the main tasks for refining the prototype. To this end, we intend to extend the AM operational cycle with a new step to receive messages from other agents and to prepare an appropriate response. The challenge here is that agents are developed in different environments with different technologies. In the current version, the AMCo presents the results in an unstructured format (text) that is easy for a human user to read. In future interactions with the AM, data exchange should also be in a structured or semi-structured (e.g., JSON) format. Another task is customization towards different risk groups of users—by profession, age, and diseases. To enhance platform personalization, we are going to extend the existing ontology and develop a new one for the different types of diseases caused by air pollution. The second direction is to implement a dedicated air monitoring chain of thought (CoT). CoT is a method for guiding LLMs through a series of steps or logical connections to reach a conclusion or solve a problem. This approach is particularly useful for tasks that require deeper understanding of the context or considering multiple factors. CoT asks LLMs to think about complex problems by breaking them down into smaller, more manageable components. This CoT will also be hybrid, i.e., it will include both ReAct agents and BDI agents. The third direction is to develop our own small but specialized language model. In a very short time, large language models have made an impact with their natural language processing capabilities. However, their resource requirements somewhat limit their applicability. Small language models are a compact and efficient alternative, applicable to various applications. For the purposes of our study, these models are valuable for integrating very specific, limited state and air monitoring information. A small language model is easier to fine-tune for solving air monitoring tasks and will simultaneously improve the performance and accuracy of the platform. Furthermore, a notable feature of such a model is its ability to process data locally, which makes it particularly useful for IoT edge devices such as our sensor network.

Author Contributions

Conceptualization, S.S.; Methodology, S.S.; Software, E.D., I.S. and I.N.; Formal analysis, A.S.-D. and V.T.-K.; Investigation, I.S. and I.N.; Resources, E.D.; Data curation, A.S.-D. and V.T.-K.; Writing—original draft, S.S., E.D. and A.S.-D.; Writing—review & editing, V.T.-K.; Visualization, I.S. and I.N. All authors have read and agreed to the published version of the manuscript.

Funding

This study is financed by the European Union-NextGenerationEU, through the National Recovery and Resilience Plan of the Republic of Bulgaria, project No. BG-RRP-2.004-0001-C01.

Data Availability Statement

The data presented in this study are openly available in Meter, https://meter.ac/gs/nodes/. World Health Organization’s system: https://www.c40knowledgehub.org/s/article/WHO-Air-Quality-Guidelines?language=en_US#:~:text=The%20WHO%20air%20quality%20guideline,3%20%2D%204%20days%20per%20year (accessed on 15 January 2025).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Kang, G.K.; Gao, J.Z.; Chiao, S.; Lu, S.; Xie, G. Air quality prediction: Big data and machine learning approaches. Int. J. Environ. Sci. Dev. 2018, 9, 8–16. [Google Scholar] [CrossRef]
  2. Institute for Health Metrics and Evaluation. Available online: https://www.healthdata.org/research-analysis/health-risks-issues/air-pollution (accessed on 16 December 2024).
  3. EPA. Air Quality Management Process Cycle. Available online: https://www.epa.gov/air-quality-management-process/air-quality-management-process-cycle (accessed on 16 December 2024).
  4. Meter. Available online: https://meter.ac/gs/nodes/ (accessed on 16 December 2024).
  5. Maruyama, Y. Symbolic and Statistical Theories of Cognition: Towards Inte-grated Artificial Intelligence. In Software Engineering and Formal Methods, SEFM 2020 Collocated Workshops, Proceedings of the SEFM 2020 Lecture Notes in Computer Science, Amsterdam, The Netherlands, 14–15 September 2020; Cleophas, L., Massink, M., Eds.; Springer: Cham, Switzerland, 2021; Volume 12524. [Google Scholar] [CrossRef]
  6. Ilkou, E.; Koutraki, M. Symbolic Vs Sub-symbolic AI Methods: Friends or Enemies? In Proceedings of the International Conference on Information and Knowledge Management, Galway, Ireland, 19–20 October 2020; Available online: https://ceur-ws.org/Vol-2699/paper06.pdf (accessed on 15 January 2025).
  7. An Official Website of the European Union. Air Quality Modeling. Available online: https://aqm.jrc.ec.europa.eu/Section/Assessment/Background (accessed on 16 December 2024).
  8. An Official Website of the European Union. FAIRMODE. Available online: https://fairmode.jrc.ec.europa.eu/ (accessed on 16 December 2024).
  9. European Environment Agency. European Air Quality Index App Now Available in all EU Languages. Available online: https://www.eea.europa.eu/en/newsroom/news/european-air-quality-index-app (accessed on 16 December 2024).
  10. WHO. Available online: https://www.who.int/teams/environment-climate-change-and-health/air-quality-energy-and-health (accessed on 16 December 2024).
  11. SafetyCulture. Available online: https://safetyculture.com/app/air-quality-monitoring-software/ (accessed on 16 December 2024).
  12. Real-Time Air Quality Monitoring Solutions. Available online: https://www.aeroqual.com/ (accessed on 16 December 2024).
  13. Air Quality Monitors Created by Industry Experts. Available online: https://www.kaiterra.com/ (accessed on 16 December 2024).
  14. Envisom—Advanced Air Monitoring Software for Professionals. Available online: https://oizom.com/envizom-air-quality-monitoring-software/ (accessed on 16 December 2024).
  15. Real-Time Impact on the Community and the Planet. Available online: https://envirosuite.com/ (accessed on 16 December 2024).
  16. Air Emissions Management Software. Available online: https://www.era-environmental.com/solutions/environmental/air (accessed on 16 December 2024).
  17. Envea. Available online: https://www.envea.global/ (accessed on 16 December 2024).
  18. ECMWF as Part of The Copernicus Programme, Air quality. Available online: https://atmosphere.copernicus.eu/air-quality (accessed on 16 December 2024).
  19. Mengara Mengara, A.G.; Park, E.; Jang, J.; Yoo, Y. Attention-Based Distributed Deep Learning Model for Air Quality Forecasting. Sustainability 2022, 14, 3269. [Google Scholar] [CrossRef]
  20. Soares, P.H.; Monteiro, J.P.; Gaioto, F.J.; Ogiboski, L.; Andrade, C.M.G. Use of Association Algorithms in Air QualityMonitoring. Atmosphere 2023, 14, 648. [Google Scholar] [CrossRef]
  21. Khan, T.R.; Emerson, Z.I.; Mentz, K.H. Evaluation of Fine Particulate Matter (PM2.5) Concentrations Measured by Collocated Federal Reference Method and Federal Equivalent Method Monitors in the U.S. Atmosphere 2024, 15, 978. [Google Scholar] [CrossRef]
  22. Yang, J.; Tian, Y.; Wu, C.H. Air Quality Prediction and Ranking Assessment Based on Bootstrap-XGBoost Algorithm and Ordinal Classification Models. Atmosphere 2024, 15, 925. [Google Scholar] [CrossRef]
  23. Nycz, B.; Pietrucha-Urbanik, K. Advances in Air Quality Monitoring: The Role of Drone Technology. Proceedings 2024, 105, 19. [Google Scholar] [CrossRef]
  24. Zhang, J.; Xia, W. Prediction of PM2.5 Concentration on the Basis of Multi-Time Scale Fusion. Processes 2022, 10, 171. [Google Scholar] [CrossRef]
  25. Rescio, G.; Manni, A.; Caroppo, A.; Carluccio, A.M.; Siciliano, P.; Leone, A. Multi-Sensor Platform for Predictive Air Quality Monitoring. Sensors 2023, 23, 5139. [Google Scholar] [CrossRef] [PubMed]
  26. Gocheva-Ilieva, S.G.; Voynikova, D.S.; Stoimenova, M.P.; Ivanov, A.V.; Iliev, I.P. Regression trees modeling of time series for air pollution analysis and forecasting. Neural Comput. Appl. 2019, 31, 9023–9039. [Google Scholar] [CrossRef]
  27. Gocheva-Ilieva, S.G.; Ivanov, A.V.; Voynikova, D.S.; Stoimenova, M.P. Modeling of PM10 Air Pollution in Urban Environment Using MARS. In Large-Scale Scientific Computing. LSSC 2019, Proceedings of the Lecture Notes in Computer Science, Sozopol, Bulgaria, 10–14 June 2019; Lirkov, I., Margenov, S., Eds.; Springer: Cham, Switzerland, 2020; Volume 11958. [Google Scholar] [CrossRef]
  28. Wooldridge, M. An Introduction to MultiAgent Systems; Wiley: Hoboken, NJ, USA, 2009; ISBN 978-0-470-51946-2. Available online: https://www.wiley.com/en-be/An+Introduction+to+MultiAgent+Systems%2C+2nd+Edition-p-9780470519462 (accessed on 15 January 2025).
  29. Boissier, O.; Bordini, R.H.; Hubner, J.; Ricci, A. Multi-Agent Oriented Programming Programming Multi-Agent Systems Using JaCaMo; The MIT Press: Cambridge, MA, USA, 2020; ISBN 9780262044578. [Google Scholar]
  30. Bratman, M.E. What is intention? In Intentions in Communication; Cohen, P.R., Morgan, J.L., Pollack, M.E., Eds.; The MIT Press: Cambridge, MA, USA, 1990; pp. 15–32. [Google Scholar] [CrossRef]
  31. Wooldridge, M. Reasoning About Rational Agents; The MIT Press: Cambridge, MA, USA, 2003; Available online: https://mitpress.mit.edu/9780262515566/reasoning-about-rational-agents/ (accessed on 16 December 2024)ISBN 9780262515566.
  32. World Health Organization. WHO Global Air Quality Guidelines: Particulate Matter (PM2.5 and PM10), Ozone, Nitrogen Dioxide, Sulfur Dioxide, and Carbon Monoxide; WHO: Geneva, Switzerland, 2021; Available online: https://www.who.int/publications/i/item/9789240034228 (accessed on 16 December 2024).
  33. Executive Environmental Agency of Bulgaria. Air Quality Bulletin: 2023 Report on Atmospheric Pollutants in Bulgaria; World Meteorological Organization: Geneva, Switzerland, 2023. Available online: https://www.eea.government.bg (accessed on 16 December 2024).
  34. Plovdiv Municipality. Environmental Monitoring System of Plovdiv Municipality. Available online: https://ecomonitoring.plovdiv.bg (accessed on 16 December 2024).
  35. Node-RED. Low-Code Programming for Event-Driven Applications. Available online: https://nodered.org/ (accessed on 16 December 2024).
  36. FROST-Server. Fraunhofer IOSB’s Server Implementation of the OGC SensorThings API. Available online: https://fraunhoferiosb.github.io/FROST-Server/ (accessed on 16 December 2024).
  37. TimescaleDB. Time Series Database. Available online: https://www.timescale.com/ (accessed on 16 December 2024).
  38. Grafana. The Open Observability Platform. Available online: https://grafana.com/ (accessed on 16 December 2024).
  39. Weyns, D.; Omicini, A.; Odell, J. Environment as a first class abstraction in multiagent systems. Auton. Agents Multi-Agent Syst. 2007, 14, 5–30. [Google Scholar] [CrossRef]
  40. Ricci, A.; Piunti, M.; Viroli, M. Environment programming in multi-agent systems: An artifact-based perspective. Auton. Agents Multi-Agent Syst. 2011, 23, 158–192. [Google Scholar] [CrossRef]
  41. Ortiz-Hernández, G.; Guerra-Hernández, A.; Hübner, J.F.; Luna-Ramírez, W.A. Modularization in Belief-Desire-Intention agent programming and artifact-based environments. PeerJ Comput. Sci. 2022, 8, e1162. [Google Scholar] [CrossRef] [PubMed]
  42. Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K.; Cao, Y. REACT: Synergizing Reasoning and Acting in Language Models. In Proceedings of the International Conference on Learning Representations, Kigali, Rwanda, 1–5 May 2023. [Google Scholar] [CrossRef]
Futureinternet 17 00112 g001
Figure 1. A map of data sources for the city of Plovdiv (visualization of our sensor network).
Figure 1. A map of data sources for the city of Plovdiv (visualization of our sensor network).
Futureinternet 17 00112 g001
Futureinternet 17 00112 g002
Figure 2. A chart showing the concentration of fine particulate matter for Plovdiv-Kamenitsa.
Figure 2. A chart showing the concentration of fine particulate matter for Plovdiv-Kamenitsa.
Futureinternet 17 00112 g002
Futureinternet 17 00112 g003
Figure 3. A chart showing the concentration of fine particulate matter for Plovdiv-Tsentar.
Figure 3. A chart showing the concentration of fine particulate matter for Plovdiv-Tsentar.
Futureinternet 17 00112 g003
Futureinternet 17 00112 g004
Figure 4. A chart showing the concentration of fine particulate matter for Plovdiv-Thrace/Trakia.
Figure 4. A chart showing the concentration of fine particulate matter for Plovdiv-Thrace/Trakia.
Futureinternet 17 00112 g004
Futureinternet 17 00112 g005
Figure 5. A chart showing the concentration of fine particulate matter for Plovdiv-Sahat Tepe.
Figure 5. A chart showing the concentration of fine particulate matter for Plovdiv-Sahat Tepe.
Futureinternet 17 00112 g005
Futureinternet 17 00112 g006aFutureinternet 17 00112 g006b
Figure 6. A chart showing the daily average concentrations of PM10 and PM2.5 and temperature for 2022 and 2023.
Figure 6. A chart showing the daily average concentrations of PM10 and PM2.5 and temperature for 2022 and 2023.
Futureinternet 17 00112 g006aFutureinternet 17 00112 g006b
Futureinternet 17 00112 g007
Figure 7. A chart showing the daily average concentrations of PM10 and humidity for 2022 and 2023.
Figure 7. A chart showing the daily average concentrations of PM10 and humidity for 2022 and 2023.
Futureinternet 17 00112 g007
Futureinternet 17 00112 g008
Figure 8. ACreM platform architecture.
Figure 8. ACreM platform architecture.
Futureinternet 17 00112 g008
Futureinternet 17 00112 g009
Figure 9. AM general architecture.
Figure 9. AM general architecture.
Futureinternet 17 00112 g009
Futureinternet 17 00112 g010
Figure 10. Segment of the AM’s Beliefs Base (initial state).
Figure 10. Segment of the AM’s Beliefs Base (initial state).
Futureinternet 17 00112 g010
Futureinternet 17 00112 g011
Figure 11. AM’s reasoning cycle state chart diagram.
Figure 11. AM’s reasoning cycle state chart diagram.
Futureinternet 17 00112 g011
Futureinternet 17 00112 g012
Figure 12. Segment of the AM’s Plan Library and reasoning cycle.
Figure 12. Segment of the AM’s Plan Library and reasoning cycle.
Futureinternet 17 00112 g012
Futureinternet 17 00112 g013
Figure 13. Sample agent test session.
Figure 13. Sample agent test session.
Futureinternet 17 00112 g013
Futureinternet 17 00112 g014
Figure 14. Air Pollution Ontology taxonomy.
Figure 14. Air Pollution Ontology taxonomy.
Futureinternet 17 00112 g014
Futureinternet 17 00112 g015
Figure 15. Annotation properties of the Carbon_Monoxide class in the Air Pollution Ontology.
Figure 15. Annotation properties of the Carbon_Monoxide class in the Air Pollution Ontology.
Futureinternet 17 00112 g015
Futureinternet 17 00112 g016
Figure 16. SubClass Of axiom for the Nitrogen_Dioxid class.
Figure 16. SubClass Of axiom for the Nitrogen_Dioxid class.
Futureinternet 17 00112 g016
Futureinternet 17 00112 g017
Figure 17. Individual Plovdiv and data properties for different pollutions.
Figure 17. Individual Plovdiv and data properties for different pollutions.
Futureinternet 17 00112 g017
Futureinternet 17 00112 g018
Figure 18. Attributes that account for local factors of the Plovdiv region.
Figure 18. Attributes that account for local factors of the Plovdiv region.
Futureinternet 17 00112 g018
Futureinternet 17 00112 g019
Figure 19. Component architecture.
Figure 19. Component architecture.
Futureinternet 17 00112 g019
Futureinternet 17 00112 g020
Figure 20. Interface to the AM’s environment.
Figure 20. Interface to the AM’s environment.
Futureinternet 17 00112 g020
Futureinternet 17 00112 g021
Figure 21. Flow chart of the AMCo’s reasoning cycle.
Figure 21. Flow chart of the AMCo’s reasoning cycle.
Futureinternet 17 00112 g021
Futureinternet 17 00112 g022
Figure 22. Segment of the AMCo’s code.
Figure 22. Segment of the AMCo’s code.
Futureinternet 17 00112 g022
Futureinternet 17 00112 g023
Figure 23. Sample AMCo agent test session.
Figure 23. Sample AMCo agent test session.
Futureinternet 17 00112 g023
Futureinternet 17 00112 g024
Figure 24. Comparative characteristics of the two agents.
Figure 24. Comparative characteristics of the two agents.
Futureinternet 17 00112 g024
Table 1. Statistics.
Table 1. Statistics.
PM2.5PM10PTRHGamma
Radiation
[µg/m3][µg/m3][hPa][deg_C][%][CPM_5_min]
Num.Valid730730730730730730
Missing000000
Mean15.7427111621.38849419999.854164715.8205340845.0972966169.377924
Median10.3043859613.87990748999.402859515.5439236147.04336673169.2143799
Mode5.571428.82926922.33720−3.981080170
Std. Deviation13.4692963318.129996426.6964054118.79926468426.080944071.808297018
Variance181.422328.69744.84277.427680.2163.27
Skewness2.022.0290.3040.024−0.2640.696
Std. Error of Skewness0.090.090.090.090.090.09
Kurtosis4.264.270.88−1.093−0.3190.997
Std. Error of Kurtosis0.1810.1810.1810.1810.1810.181
Minimum2.3169014083.637323944922.3372037−3.9810810810164.8576512
Maximum88.1013986118.7272727922.337203733.69819005100176.6041667
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.

Share and Cite

MDPI and ACS Style

Stoyanov, S.; Doychev, E.; Stoyanova-Doycheva, A.; Tabakova-Komsalova, V.; Stoyanov, I.; Nedelchev, I. A Regional Multi-Agent Air Monitoring Platform. Future Internet 2025, 17, 112. https://doi.org/10.3390/fi17030112

AMA Style

Stoyanov S, Doychev E, Stoyanova-Doycheva A, Tabakova-Komsalova V, Stoyanov I, Nedelchev I. A Regional Multi-Agent Air Monitoring Platform. Future Internet. 2025; 17(3):112. https://doi.org/10.3390/fi17030112

Chicago/Turabian Style

Stoyanov, Stanimir, Emil Doychev, Asya Stoyanova-Doycheva, Veneta Tabakova-Komsalova, Ivan Stoyanov, and Iliya Nedelchev. 2025. "A Regional Multi-Agent Air Monitoring Platform" Future Internet 17, no. 3: 112. https://doi.org/10.3390/fi17030112

APA Style

Stoyanov, S., Doychev, E., Stoyanova-Doycheva, A., Tabakova-Komsalova, V., Stoyanov, I., & Nedelchev, I. (2025). A Regional Multi-Agent Air Monitoring Platform. Future Internet, 17(3), 112. https://doi.org/10.3390/fi17030112

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop