Atmospheric urban pollution: applications of an artificial neural network (ANN) to the city of Perugia

doi:10.1016/S0304-3800(01)00434-3

Ecological Modelling

Volume 148, Issue 1, 1 February 2002, Pages 27-46

https://doi.org/10.1016/S0304-3800(01)00434-3 Get rights and content

Abstract

Urban air pollution is a growing problem. Large cities in particular show, at some point of time, high concentrations of substances dangerous for human health. The difficulty in forecasting pollutant's concentration trends with a reasonable error is still an open problem. In this paper, a new approach is proposed. An artificial neural network (ANN) is used to forecast short and middle long-term concentration levels for some of the well-known pollutants. The results seem to be in good accord with the monitored data and allow its use as the forecasting model on a 24–48 h basis requiring only the meteorological conditions and the traffic level.

Introduction

In recent times, urban air pollution has been represented as a growing problem for communities. The danger represented by air pollutants has been largely demonstrated from toxicological studies both for short and long exposition times.

Some compounds like sulphur dioxide, atmospheric particulate, nitrogen compounds, carbon monoxide and ozone are considered as typical indicators of the air quality. This situation appears to be relevant in urban areas where particular atmospheric and orographic conditions can cause an accumulation of the pollutants.

Recently some community actions together with the renewal of the vehicles have brought about lower pollutant levels in urban areas, but new pollutants are becoming important (IPA, benzene, etc.); therefore, at the moment one of the best ways of controlling the pollutants' concentration seems to be the correct management of both urban development and traffic. The best management actually seems to be the possibility to forecast the pollutants' concentration trends 1 or 2 days earlier according to the meteorological conditions and the traffic expectation in defined areas, so that the authorities have an opportunity to act. In fact the concentration of the pollutants in atmosphere in the urban areas are mainly due to the traffic emissions, connected to the traffic jam, the vehicular status, the geographic and local site characteristics, and meteorological conditions. The use of deterministic models or lagrangian ones to forecast the pollution levels in urban areas is made rather difficult for the complex fluidodynamic field which arise from the complex orography and for the heat phenomena which characterise the urban contest. One of the advantages of the artificial neural network (ANN) lies in the fact that the deterministic model needs a lot of information, whereas the neural network acts like a black-box model. The drawback of the neural approach is that no deep understanding on the physical phenomena is gained by using a neural network, since it resembles the behaviour of a black-box method. Moreover, the ANN, once trained, is fast at predicting the desired values. The last and the most convincing advantage is that the accuracy of the neural prediction is generally higher than the other kind of models.

In the present work, an ANN is used to forecast the pollutant concentration levels in the urban area of Perugia.

Section snippets

Model description

ANNs are a branch of artificial intelligence developed in the 1950s aiming at imitating the biological brain architecture. They are parallel-distributed systems made of many interconnected non-linear processing elements, called neurons (Hecht-Nielsen, 1990). A renewal of interest has grown exponentially in the last decade, mainly for the availability of suitable hardware (e.g. parallel computers, analogue/digital neural card for personal computers) that has made them convenient for fast data

Applications

The application of an ANN to the urban context in Perugia, particularly for the area of Fontivegge, near the F.S. Railway Station, is presented. The experimental data were obtained from the monitoring unit of the Presidio Multizonale di Prevenzione, located in Fontivegge and operative since 1997. The variables monitored were: sulphur dioxide, nitrogen oxides (NO, NO₂, NO_x), total suspended particulate and PM10, benzene, carbon monoxide, ozone, horizontal wind speed, moisture, pressure,

Short-term forecasts

The methodology adopted for the short-term forecasts is the following: a matrix has been given to the ANN as configured in Table 1, where A, B, C… represent pollutants' concentrations, weather parameters, and traffic; X is the pollutant concentration to be forecasted; n_A, n_B,…,n_X represent the number of hours preceding the present hour t and h is the number of hours subsequent to the present hour t (forecast at h hour).

The correct choice of the number of hours preceding the present hour t for

Discussion

Although the monitored period considered (2 years, i.e. only two time series) seems to be quite short and the lack of continuous hourly data occurred to many parameters, from the present study, it has appeared that the ANN with a single hidden layer based on the standard backpropagation algorithm described above, using eventually only the simple sigmoid (Eq. (5)) as activation function, resulted as a very efficient model to forecast both short and middle long-term air pollutant concentrations

Conclusions

The perceptron with backpropagation algorithm model have shown very good performances for the 1 h forecasts. It is necessary to mention that in order to use the model for forecasting aims (both short and middle long-term forecasts) single pollutant ANNs have to be built. For the middle (24 h) and long-term forecasts, ANNs can be used introducing hypothesis about the values of the meteorological and traffic parameters. In this case, although the ANN forecasts appear to be worse than the 1 h

References (21)

O Antonic et al.
Modelling groundwater regime acceptable for the forest survival after the building of the hydro-electric power plant
Ecol. Model.
(2001)
J.J Heymans et al.
A carbon flow model and network analysis of the northern Benguela upwelling system, Namibia
Ecol. Model.
(2000)
C Karul et al.
Case studies on the use of neural networks in eutrophication modelling
Ecol. Model.
(2000)
P Werbos
Generalization of Backpropagation with application to a recurrent gas market model
Neural Networks
(1988)
M.J.A Berry et al.
Data Mining Techniques
(1997)
D.P Bertsekas
Non-linear Programming
(1995)
D.P Bertsekas et al.
Neuro-Dynamic Programming
(1996)
A Bishop
Neural Networks for Pattern Recognition
(1995)
A Blum
Neural Network in C++
(1992)
Boger, Z., Guterman, H., 1997. Knowledge extraction from artificial neural network models. IEEE Systems, Man and...

There are more references available in the full text version of this article.

Cited by (142)

Investigating the relationship between air temperature and the intensity of urban development using on-site measurement, satellite imagery and machine learning
2024, Sustainable Cities and Society
Given the seriousness of the urban heat problem, the relationship between urbanization and air temperature has become a critical concern worldwide. In this study, common urban planning indicators, including the building coverage ratio (BCR), floor area ratio (FAR), and fractional vegetation cover (FVC), were extracted from satellite images to determine the intensity of urban development. On-site measurements and machine learning (ML) were used to observe and analyze the relationship between the intensity of urban development and air temperature. From the on-site measurement results, the air temperature in downtown Taipei decreased by an average of approximately 0.32 °C with every 10 % increase in the FVC. However, it increased by an average of approximately 0.28 °C and 0.03 °C with every 10 % increase in the BCR and FAR, respectively. The results obtained from the ML models demonstrated the same trend, with minor differences from the on-site measurement results, which were regarded as reasonable and acceptable. In this study, a more convenient method was proposed to extract urban planning indicators, describe the intensity of urban development within an area, and help estimate air temperature in areas without measuring instruments. The relationship determined herein may aid in the decision-making process of the balance of urbanization and vegetation.
ANFIS and ANN models to predict heliostat tracking errors
2023, Heliyon
The efficiency and performance of solar tower power are greatly influenced by the heliostats field. To ensure accurate tracking of reflectors often requires an evaluation of the beam reflected positions. This operation is costly time-consuming due to the number of heliostats. It is also necessary to set up a fast and less expensive method able to evaluate tracking heliostat. In this paper, prediction models based on the Adaptive Neuro-Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN) were applied to estimate rapidly and accurately heliostat error tracking. The modeling is based on the experimental data of seven different days. The input parameters are time and day number and the output is the beam reflected position following the altitude and azimuth axes. Both techniques have been able to predict the beam reflected position. A comparison of results showed that intelligent methods recorded better performance than conventional model based on geometric errors. For ANFIS model, coefficients of correlation (R²) of 0.97 is obtained compared to that of the ANN, 0.96 and 0.92 for altitude and azimuth axes respectively. The intelligent methods may be a promising alternative for predicting heliostat beam reflected the position.
Comparison Of Anfis And Ann Modeling For Predicting The Water Absorption Behavior Of Polyurethane Treated Polyester Fabric
2021, Heliyon
Citation Excerpt :
Artificial neural networks (ANNs) are highly parallel computing systems inspired by biological neural networks [47]. ANNs were first established in the 1950s to emulate the architecture of the biological brain of humankind [48]. The ANN can develop an internal representation of a signal pattern introduced to the network as an input.
Nowadays, the polyurethane and its derivatives are highly applied as a surface modification material onto the textile substrates in different forms to enhance the functional properties of the textile materials. The primary purpose of this study is to develop prediction models to model the absorption property of the textile substrate using the Adaptive Neuro-Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN) methods. In this study, polyurethane (PU) along with acrylic binder was applied on the dyed polyester knitted fabric to develop and validate the prediction models. Through the morphological study, it was evident that the solution prepared with the polyurethane and the acrylic binder was effectively coated onto the fabric surface. The ANFIS model was constructed by considering binder (ml) and PU (%) as input parameters, whereas absorbency (%) was the only output parameter. On the other hand, the system was trained with 70% data for constructing the ANN model whereas testing and validation were done with 15% data, respectively. To train the network, feed-forward backpropagation with Levenberg–Marquardt learning algorithm was used. The coefficient of determination (R²) was found to be 0.98 and 0.93 for ANFIS and ANN model, respectively. Both prediction models exhibited an excellent mean absolute error percentage (0.76% for the ANFIS model and 1.18% for the ANN model). Furthermore, an outstanding root-mean-square error (RMSE) of 0.61% and 1.28% for ANFIS and ANN models was observed. These results suggested an excellent performance of the developed models to predict the absorption property of the polyurethane and acrylic binder treated fabric. Besides, these models can be taken as a basis to develop prediction models for specific types of functional applications of the textile materials to eliminate heaps of trial and error efforts of the textile industries, which eventually be helpful in the scalable production of functional textiles.
Air pollution in Germany: Spatio-temporal variations and their driving factors based on continuous data from 2008 to 2018
2021, Environmental Pollution
This study analyzed long-term observational data of particulate matter (PM_2.5, PM₁₀) variability, gaseous pollutants (CO, NO₂, NO_X, SO₂, and O₃), and meteorological factors in 412 fixed monitoring stations from January 2008 to December 2018 in Germany. Based on Hurst index analysis, the trend of atmospheric pollutants in Germany was stable during the research period. The relative correlations of gaseous pollutants and meteorological factors on PM_2.5 and PM₁₀ concentrations were analyzed by Back Propagation Neural Network model, showing that CO and temperature had the greater correlations with PM_2.5 and PM₁₀. Following that, PM_2.5 and PM₁₀ show a strong positive correlation (R² = 0.96, p < 0.01), suggesting that the reduction of PM_2.5 is essential for reducing PM pollution and enhancing air quality in Germany. Based on typical PM₁₀/CO ratios obtained under ideal weather conditions, it is conducive to roughly estimate the contribution of natural sources. In winter, the earth’s crust contributed about 20.1% to PM₁₀. Taken together, exploring the prediction methods and analyzing the characteristic variation of pollutants will contribute an essential implication for air quality control in Germany.
Green blasting policy: Simultaneous forecast of vertical and horizontal distribution of dust emissions using artificial causality-weighted neural network
2021, Journal of Cleaner Production
The environmental impacts of the mining industry as well as reduction strategies are important issues that have been highlighted as a result of the environmentally-friendly policy of green mining. As part of that policy, this study developed an approach based on a multi-layer artificial neural network and fuzzy cognitive map to predict the vertical and horizontal distribution of blast-induced dust emissions, simultaneously. Hence, a fuzzy cognitive map based on the cause-and-effect analysis concept was first designed to extract inputs’ weights. An optimal network with two hidden layers was then implemented to predict the vertical and horizontal dust distributions in a mine close to residential and agricultural areas. The performance evaluation of the approach indicated good results by the R² of 0.9933 and 0.9267 between the measured and predicted values for the horizontal and vertical distributions, respectively. Furthermore, the predicted and measured outputs over every blasting round were in accordance due to the slight mean absolute and root mean square errors of 0.009 and 0.018, respectively. A sensitivity analysis revealed that all inputs, excluding air humidity on the vertical distribution, had acceptable effects on the dust distribution outputs by the strength of higher than 0.7. Based on a reduction solution that we suggested, one round was blasted by taking water capsules together with stemming material in 40% of the whole blast holes. The results indicated that the maximum horizontal distribution of blast-induced dust was decreased about six times. The approach can be straightforwardly updated and used for other mining cases.
Comparison of maritime transport influence of SO2 levels in Algeciras and Alcornocales Park (Spain)
2021, Transportation Research Procedia
The main aim of this work was to measure the influence of the volume of shipping over the Sulphur dioxide (SO₂) concentration in the air pollution in two monitoring stations located at Algeciras city and Alcornocales Park developing the same analysis in these two locations. The target is to demonstrate the assumption that Algeciras is more affected by SO₂ than Alcornocales Park which is 30 km far away from Algeciras Port. A multiple regression approach has been applied using wind data: wind direction (degrees) and wind speed (km/h) recorded in two weather stations, together with the volume of the gross tonnage per hour (GT/h) of vessels in the Bay of Algeciras to estimate SO₂ concentration values in the two stations Algeciras and Alcornocales. The database contains records of hourly samples of these variables during the year 2019. Different artificial neural networks (ANNs) models were compared and the results showed that SO₂ in Algeciras station could be better explained than the same pollutant in Alcornocales station. On the other hand, ANNs produced better results than linear models which means that nonlinear models fit best the data. A cross-validation procedure has been applied in order to assure the generalisation capabilities of the tested models. The results showed that in Algeciras a more reliable estimation could be done reaching a correlation estimation between the model and the target (real) values of SO₂. This fact highlights the major influence of maritime transport in the Bay of Algeciras.

View all citing articles on Scopus

View full text

Atmospheric urban pollution: applications of an artificial neural network (ANN) to the city of Perugia

Abstract

Introduction

Section snippets

Model description

Applications

Short-term forecasts

Discussion

Conclusions

Ecol. Model.

Ecol. Model.

Ecol. Model.

Neural Networks

Data Mining Techniques

Non-linear Programming

Neuro-Dynamic Programming

Neural Networks for Pattern Recognition

Neural Network in C++