Air quality significantly impacts public health, with nitrogen dioxide (NO2) being a key pollutant linked to respiratory and cardiovascular diseases. In this study, we developed a machine learning model to accurately predict hourly NO2 concentrations in Ternopil, Ukraine, using readily available meteorological and temporal data. The model was trained on a large dataset and tested using data from the Ecocity monitoring station, known for recording NO2 levels exceeding legal limits. By employing neural networks, the model demonstrated high accuracy in predicting NO2 concentrations, with the error of 3.9% and 1.4%, respectively, in the test samples. Our findings underscore the potential of machine learning techniques to enhance air quality monitoring and forecasting, particularly in urban areas with limited resources. This approach offers a valuable tool for real-time pollution management and public health protection.
a variety of gases, droplets, and particles that degrade air quality; therefore, their exposure to
humans is believed to lead to serious health problems, especially in urban areas where pollution
levels are high [
Air quality is considered good when it meets a certain level of purity and transparency, and
there are no gaseous pollutants such as smoke, dust, smog, and other impurities in the
atmosphere. The Air Quality Index (AQI) is a numerical indicator used by government agencies
to inform the public about the current state of the air or projected levels of air pollution. When
the AQI rises due to an increase in air pollutants (e.g., during peak traffic hours or when there
are downwind wildfires), a growing proportion of the population can quickly experience serious
negative health effects [
The issue of constant monitoring of air quality in real time is also relevant, as the authors
says in paper [
The purpose of this study is to assess and analyse the impact of atmospheric pollution on air quality, as well as to establish its dependence on atmospheric factors.
Many studies have established a link between cardiovascular and pulmonary diseases and
long-term exposure to pollutants, including nitrogen dioxide (NO₂) and particulate matter
(PM2.5 and PM10). According to the European Environment Agency [
Natural NO₂ emissions have very low background concentrations. Emissions associated with
anthropogenic activities are the most significant factor affecting human health. The main source
of NO₂ emissions is human-caused combustion processes, such as heat, electricity, or internal
combustion engines. In particular, motor vehicles are the main sources of nitrogen oxides [
Such an analysis requires models that can accurately reflect air quality in order to identify
periods of increased pollution and respond quickly to short-term fluctuations, especially in
urban areas. Prediction of pollutants by deterministic methods is accompanied by significant
uncertainties due to the complexity of the physical and chemical processes that determine the
formation and transport of pollutants in the urban atmosphere [
In this regard, sophisticated machine learning methods are becoming increasingly common
in air quality modelling, outperforming traditional statistical approaches. In the review by
Cabañeros et al. [
ANNs are capable of detecting complex, nonlinear relationships between meteorological
variables and pollutant concentrations, and generalizing information from training datasets to
form functional relationships between variables, even if the nature of these relationships is
unknown. Unlike regression analysis, ANNs work effectively in the presence of significant noise
in the data [
The first successful applications of ANNs for modelling NO₂ concentrations in urban areas
were presented in the works of Gardner and Dorling [
A key aspect in developing machine learning-based air quality models is the selection of
appropriate input parameters. The concentration of NO₂ in urban air is influenced by many
variables that reflect meteorological conditions and pollution sources. Several studies have
identified meteorological variables such as temperature, humidity, and wind speed as important
predictors, as well as concentrations of other pollutants [
An alternative approach is to use previously measured concentrations of the target pollutant
as predictors, relying on temporal autocorrelation between successive values of the same
variable. This is especially effective for forecasting several hours in advance [
Other studies have used traffic data obtained by vehicle counts or other models [
In our study, we set out to develop a model that would accurately estimate hourly NO₂ concentrations in Ternopil, Ukraine, using only available standard meteorological and temporal data as input variables.
As in many cities, air quality in Ternopil is monitored by several separate measurement stations. In previous years, several of them have recorded high levels of NO₂ that exceeded the regulatory limits. For this study, the Ecocity station in the central part of Ternopil, known as a pollution hotspot, was chosen, where NO₂ concentrations often exceed the legal thresholds. This location is particularly important due to the high density of residential development in the immediate vicinity of the station.
Ternopil is located in western Ukraine, near the Seret River, on the Ternopil Plateau of the
Podillia Upland of the Eastern European Plain. The city is located in the temperate climate zone
of the broadleaf forest zone. Ternopil has a moderately continental climate with warm and
humid summers and mild winters [
The street where the air quality monitoring station used in this study is located is in the central part of the city. It is a heavily trafficked main transportation artery connecting the highway with the centre of Ternopil. At the same time, there is a very high density of residential buildings along this street. For several months, both the concentrations of pollutants measured at this station and the number of days with peak pollution levels exceeded the established permissible limits.
The correct choice of input parameters, is crucial for predicting pollutant concentrations.
Ivan Puluj National Technical University in Ternopil (Ukraine), in cooperation with the
program “Clean Air for Ukraine” of the NGO Arnika (Prague, Czech Republic), NGO Free
Arduino (Ivano-Frankivsk, Ukraine) and the public monitoring network EcoCity, installed an
AirFresh air quality measurement station at the university to expand the network and conduct
research. The AirFresh air quality monitoring station is a device that enables real-time
monitoring and recording of ambient air conditions, namely temperature, humidity, and dust
concentrations of PM2.5 and PM10. AirFresh measures the concentrations of dust microparticles
(PM2.5 and PM10), carbon monoxide (CO), ammonia (NH3), ground-level ozone (O3), and
nitrogen dioxide (NO₂) as shown in Figure 1. Each station may be outfitted with radiation
background sensors or additional sensors for 16 pollutants tracked by EcoCity.[
Meteorological parameters, such as the height of the measurement site, have a significant impact on the concentration of NO₂. At the monitoring station, data on pollutant concentrations, as well as temperature and relative humidity, are measured at a height of 3 meters.
Experimental data on NO₂ concentrations, temperature and humidity were collected at the station during the period from August 6 to 13, 2024.
To ensure the ability to recognize temporal patterns in NO₂ fluctuations, the model was given the ability to detect typical hourly and daily variations of this pollutant. To this end, the analysis included time variables that reflect different frequency components in the observed data.
The development of a machine learning model consists of several stages, each of which plays an important role in creating an efficient and accurate model. The main stages of developing a machine learning model are: data collection; data preprocessing and analysis; selection of a model and machine learning algorithm; splitting the data into training and test samples; model training; and evaluation and validation. Tracking NO2 emissions, which is the most active pollutant gas, and predicting its concentration are important steps towards pollution control. Therefore, nitrogen dioxide (NO₂) was predicted using experimental data obtained from the Meteorological Station AirFresh. During learning, the dataset was divided into two unequal parts training and test samples.
The study was divided into two stages. At the first stage, the training set contained experimental dependencies of NO₂ concentration on temperature, humidity, and measurement time for six days, and a sample of a one-day dataset unknown to the system was chosen to test the quality of forecasting. And at the second stage, the test sample was randomly selected by the computer from all the experimental data for different days (Figure 2).
The dataset contained 541 elements. In particular, the training set contained 504 elements characterising temperature, humidity and measurement time over N days (six days in our study). The NO₂ concentration was predicted by neural networks and selected as the output parameter. Based on the experimental results of the NO₂ concentration for one day, a test set of 37 elements was formed to evaluate the quality of prediction. It was found that the built models can make predictions based on data that were not used in the training sample. Therefore, such results are informative for studying their quality.
0,03 0,03 0,035 0,04 0,045 0,05 0,055 0,06
a) 0,04 0,05 0,06
b) 0,07 0,06 3 0,05 m / gm0,04 / pm0,03 p , 2O0,02 N 0,01 0,06 0,05 3 m / g /mm0,04 p p , 2
NO0,03 5 10 15 20 25 30
40 August 13, 2024 a)
Exp Pred 35 Exp
Pred 0,02 0.000010.000120.000250.000380.000530.00063
August 6-13, 2024 b)
The neural networks were used to build (Figure 3 and 4) the dependence of experimental concentrations (NO2true) on the predicted ones (NO2pred), as well as NO₂-August 13, 2024 for one day. The NN method gives an error of 3.9%.
It is important that in Figure 3, the points are located quite close to the bisector of the first coordinate angle, which indicates the consistency of the predicted and experimental data.
The NO₂ concentration was predicted by temperature, humidity and time of measurement over seven days. The sample contained 541 elements, of which 80% were randomly selected for the training set and 20% were left to evaluate the quality of the prediction. The parameters of the neural network are shown in Table 1. 1 n |ytrue− y prediction| ,
∑ n i=1 |ytrue| (1)
It was found that the predicting results are in good agreement with the experimental ones. The error of the NN method is 1.4%.
In this study, a machine learning model was developed to predict NO₂ concentrations based on meteorological and temporal data in Ternopil, Ukraine.
The concentration of nitrogen dioxide (NO₂) was predicted using experimental data obtained from the Weather Station during 6-13 August 2024. It was found that regardless of the type of study (self-selected test sample or randomly selected by a computer), the forecasting results are in good agreement with the experimental data. The error of the NM method is 3.9% and 1.4%, respectively, in the test samples.
The proposed model allows for real-time forecasting of pollutant emissions, which is an important tool for monitoring air quality in urban areas with limited resources.
The use of such models can be an important step toward creating early warning systems for elevated levels of pollution and rapid response to short-term fluctuations in the concentration of harmful substances. This could have a positive impact on public health, especially in areas with heavy traffic and high building density. Future research could focus on integrating additional factors, such as traffic and other pollutants, to enable even more accurate predictions of air quality changes in modern urban ecosystems.