Predicting hourly air pollutant levels using artificial neural networks coupled with uncertainty analysis by Monte Carlo simulations

Recent progress in developing artificial neural network (ANN) metamodels has paved the way for reliable use of these models in the prediction of air pollutant concentrations in urban atmosphere. However, improvement of prediction performance, proper selection of input parameters and model architecture, and quantification of model uncertainties remain key challenges to their practical use. This study has three main objectives: to select an ensemble of input parameters for ANN metamodels consisting of meteorological variables that are predictable by conventional weather forecast models and variables that properly describe the complex nature of pollutant source conditions in a major city, to optimize the ANN models to achieve the most accurate hourly prediction for a case study (city of Tehran), and to examine a methodology to analyze uncertainties based on ANN and Monte Carlo simulations (MCS). In the current study, the ANNs were constructed to predict criteria pollutants of nitrogen oxides (NO_x), nitrogen dioxide (NO₂), nitrogen monoxide (NO), ozone (O₃), carbon monoxide (CO), and particulate matter with aerodynamic diameter of less than 10 μm (PM₁₀) in Tehran based on the data collected at a monitoring station in the densely populated central area of the city. The best combination of input variables was comprehensively investigated taking into account the predictability of meteorological input variables and the study of model performance, correlation coefficients, and spectral analysis. Among numerous meteorological variables, wind speed, air temperature, relative humidity and wind direction were chosen as input variables for the ANN models. The complex nature of pollutant source conditions was reflected through the use of hour of the day and month of the year as input variables and the development of different models for each day of the week. After that, ANN models were constructed and validated, and a methodology of computing prediction intervals (PI) and probability of exceeding air quality thresholds was developed by combining ANNs and MCSs based on Latin Hypercube Sampling (LHS). The results showed that proper ANN models can be used as reliable metamodels for the prediction of hourly air pollutants in urban environments. High correlations were obtained with R ² of more than 0.82 between modeled and observed hourly pollutant levels for CO, NO_x, NO₂, NO, and PM₁₀. However, predicted O₃ levels were less accurate. The combined use of ANNs and MCSs seems very promising in analyzing air pollution prediction uncertainties. Replacing deterministic predictions with probabilistic PIs can enhance the reliability of ANN models and provide a means of quantifying prediction uncertainties.     

Controls on hourly variations in urban background air pollutant concentrations

Average 21st century concentrations of urban air pollutants linked to cardiorespiratory disease are not declining, and commonly exceed legal limits. Even below such limits, health effects are being observed and may be related to transient daytime peaks in pollutant concentrations. With this in mind, we analyse >52,000 hourly urban background readings of PM₁₀ and pollutant gases throughout 2007 at a European town with legal annual average concentrations of common pollutants, but with a documented air pollution-related cardiorespiratory health problem, and demonstrate the hourly variations in PM₁₀, SO₂, NO_x, CO and O₃. Back-trajectory analysis was applied to track the arrival of exotic PM₁₀ intrusions, the main controls on air pollutants were identified, and the typical hourly pattern on ambient concentrations during 2007 was profiled. Emphasis was placed on “worst case” data (>90th percentile), when health effects are likely to be greatest. The data show marked daytime variations in pollutants result from rush-hour traffic-related pollution spikes, midday industrial SO₂ maxima, and afternoon O₃ peaks. African dust intrusions enhance PM₁₀ levels at whatever hour, whereas European PM incursions produce pronounced evening peaks due to their transport direction (across an industrial traffic corridor). Transient peak profiling moves us closer to the reality of personal outdoor exposure to inhalable pollutants in a given urban area. We argue that such an approach to monitoring data potentially offers more to air pollution health effect studies than using only 24 h or annual averages.     

Prediction of sulfur dioxide concentrations at a site near downtown Santiago,Chile

We have analyzed the possibility to predict hourly averages of sulfur dioxide concentrations in the atmosphere at a site not far from the downtown area in the city of Santiago, Chile. We have compared the forecasts produced assuming persistence, linear regressions and feed forward neural networks. The effect of meteorological conditions is included by using forecasted values of temperature, relative humidity and wind speed at the time of the intended prediction as inputs to the different models. The best predictions for hourly averages are obtained with a three-layer neural network that has hourly averages of sulfur dioxide concentrations every 6 h on the previous day plus the actual values of the meteorological variables as input. Training the network with 1995 data, error in 8 h in advance prediction for 1996 data is of the order of 30%.     

Neural networks and periodic components used in air quality forecasting

Forecasting of air quality parameters is one topic of air quality research today due to the health effects caused by airborne pollutants in urban areas. The work presented here aims at comparing two principally different neural network methods that have been considered as potential tools in that area and assessing them in relation to regression with periodic components. Self-organizing maps (SOM) represent a form of competitive learning in which a neural network learns the structure of the data. Multi-layer perceptrons (MLPs) have been shown to be able to learn complex relationships between input and output variables. In addition, the effect of removing periodic components is evaluated with respect to neural networks. The methods were evaluated using hourly time series of NO₂ and basic meteorological variables collected in the city of Stockholm in 1994–1998. The estimated values for forecasting were calculated in three ways: using the periodic components alone, applying neural network methods to the residual values after removing the periodic components, and applying only neural networks to the original data. The results showed that the best forecast estimates can be achieved by directly applying a MLP network to the original data, and thus, that a combination of the periodic regression method and neural algorithms does not give any advantage over a direct application of neural algorithms.     

An enhanced PM2.5 air quality forecast model based on nonlinear regression and back-trajectory concentrations

An enhanced PM_2.5 air quality forecast model based on nonlinear regression (NLR) and back-trajectory concentrations has been developed for use in the Louisville, Kentucky metropolitan area. The PM_2.5 air quality forecast model is designed for use in the warm season, from May through September, when PM_2.5 air quality is more likely to be critical for human health. The enhanced PM_2.5 model consists of a basic NLR model, developed for use with an automated air quality forecast system, and an additional parameter based on upwind PM_2.5 concentration, called PM24. The PM24 parameter is designed to be determined manually, by synthesizing backward air trajectory and regional air quality information to compute 24-h back-trajectory concentrations. The PM24 parameter may be used by air quality forecasters to adjust the forecast provided by the automated forecast system. In this study of the 2007 and 2008 forecast seasons, the enhanced model performed well using forecasted meteorological data and PM24 as input. The enhanced PM_2.5 model was compared with three alternative models, including the basic NLR model, the basic NLR model with a persistence parameter added, and the NLR model with persistence and PM24. The two models that included PM24 were of comparable accuracy. The two models incorporating back-trajectory concentrations had lower mean absolute errors and higher rates of detecting unhealthy PM2.5 concentrations compared to the other models.     

Development of a Speciated,Hourly, and Gridded Air Pollutants Emission Modeling System—A Case Study on the Precursors of Photochemical Smog in the Seoul Metropolitan Area,Korea

ABSTRACT

A speciated, hourly, and gridded air pollutants emission modeling system (SHEMS) was developed and applied in predicting hourly nitrogen dioxide (NO₂) and ozone (O₃) levels in the Seoul Metropolitan Area (SMA). The primary goal of the SHEMS was to produce a systemized emission inventory for air pollutants including ozone precursors for modeling air quality in urban areas.

The SHEMS is principally composed of three parts: (1) a pre-processor to process emission factors, activity levels, and spatial and temporal information using a geographical information system; (2) an emission model for each source type; and (3) a post-processor to produce report and input data for air quality models through database modeling. The source categories in SHEMS are point, area, mobile, natural, and other sources such as fugitive emissions. The emission database produced by SHEMS contains 22 inventoried compounds: sulfur dioxide, NO₂, carbon monoxide, and 19 speciated volatile organic compounds. To validate SHEMS, the emission data were tested with the Urban Airshed Model to predict NO₂ and O₃ concentrations in the SMA during selected episode days in 1994. The results turned out to be reliable in describing temporal variation and spatial distribution of those pollutants.     

Relating meteorological variables and trends in motor vehicle emissions to monthly urban carbon monoxide concentrations

A model which quantifies the relationship between the monthly time series for CO emissions, the monthly time series in ambient CO concentration, and meteorologically driven dispersion was developed. Fifteen cities representing a wide range of geographical and climatic conditions were selected. An eight-year time series (1984–1991 inclusive) of monthly averaged data were examined in each city. A new method of handling missing ambient concentration values which is designed to calculate city-wide average concentrations that follow the trend seen at individual monitor sites is presented. This method is general and can be used in other applications involving missing data. The model uses emissions estimates along with two meteorological variables (wind speed and mixing height) to estimate monthly averages of ambient air pollution concentrations. The model is shown to have a wide range of applicability; it works equally well for a wide range of cities that have very different temporal CO distributions. The model is suited for assessing long-term trends in ambient air pollutants and can also be used for estimating seasonal variations in concentration, estimation of trends in emissions, and for filling in gaps in the ambient concentration record.     

Air Pollution Problems of the Steel Industry

Urban air pollution has traditionally been modeled using annual, or at best, seasonal emissions inventories and climatology. These averaging techniques may introduce uncertainty into the analysis, if specific emissions (e.g. SO₂) are correlated with dispersion factors on a short-term basis. This may well be the case for space heating emissions. An analysis of this problem, using hourly climatological and residential emission estimates for six U.S. cities, indicates that the errors introduced using such averages are modest (~ ± 12%) for annual average concentrations. Maximum hourly concentrations vary considerably more, since maximum heat demand and worst case dispersion are in general not coincident. The paper thus provides a basis for estimating more realistic air pollution Impacts due to residential space heating.     

Forecasting air quality time series using deep learning

This paper presents one of the first applications of deep learning (DL) techniques to predict air pollution time series. Air quality management relies extensively on time series data captured at air monitoring stations as the basis of identifying population exposure to airborne pollutants and determining compliance with local ambient air standards. In this paper, 8 hr averaged surface ozone (O₃) concentrations were predicted using deep learning consisting of a recurrent neural network (RNN) with long short-term memory (LSTM). Hourly air quality and meteorological data were used to train and forecast values up to 72 hours with low error rates. The LSTM was able to forecast the duration of continuous O₃ exceedances as well. Prior to training the network, the dataset was reviewed for missing data and outliers. Missing data were imputed using a novel technique that averaged gaps less than eight time steps with incremental steps based on first-order differences of neighboring time periods. Data were then used to train decision trees to evaluate input feature importance over different time prediction horizons. The number of features used to train the LSTM model was reduced from 25 features to 5 features, resulting in improved accuracy as measured by Mean Absolute Error (MAE). Parameter sensitivity analysis identified look-back nodes associated with the RNN proved to be a significant source of error if not aligned with the prediction horizon. Overall, MAE's less than 2 were calculated for predictions out to 72 hours.

Implications: Novel deep learning techniques were used to train an 8-hour averaged ozone forecast model. Missing data and outliers within the captured data set were replaced using a new imputation method that generated calculated values closer to the expected value based on the time and season. Decision trees were used to identify input variables with the greatest importance. The methods presented in this paper allow air managers to forecast long range air pollution concentration while only monitoring key parameters and without transforming the data set in its entirety, thus allowing real time inputs and continuous prediction.     

Prediction of maximum of 24-h average of PM10 concentrations 30 h in advance in Santiago,Chile

We have developed a neural network based model that uses values of PM10 concentrations measured until 6 p.m. on the present day plus measured and forecasted values of meteorological variables as input in order to predict the level reached by the maximum of the 24-h moving average (24MA) of PM10 concentration on the next day. We have adjusted the parameters of the model using 1998 data to predict 1999 conditions and 1999 data to forecast 2000 maximum concentrations. We have found that among the relevant meteorological input variables, the forecasted difference between maximum and minimum temperature is the most important. Due to the fact that local authorities impose restrictions to emissions on days when the maximum of 24MA of PM10 concentration is expected to exceed 240 μg/m³, we have corrected the measured concentrations on these days before evaluating the efficacy of the forecasting model. Percent errors in forecasting the numerical value are of the order of 20%. The performance of the neural network is better than that of a linear model with the same inputs, but the difference being not great is an indication that the right choice of input variables may be more important than the decision to use a linear or a nonlinear model.     

Modeling short-term variability of semivolatile organic chemicals in air at a local scale: an integrated modeling approach

Monitoring campaigns from different locations have recently shown how air concentrations of persistent semivolatile contaminants such as polychlorinated biphenyls (PCBs) often exhibit short-term (less than 24 h) variations. The observed patterns have been ascribed to different factors, such as temperature-mediated air-surface exchange and variability of planetary boundary layer (PBL) height and dynamics. Here, we present a new modeling approach developed in order to investigate the short-term variability in air concentrations of organic pollutants at a local scale. A new dynamic multimedia box model is supplied by a meteorological preprocessor (AERMET) with hourly values of air compartment height and wind speed. The resulting model is tested against an existing dataset of PCB air concentrations measured in Zurich, Switzerland. Results show the importance of such modeling approach in elucidating the short- and long-term behavior of semivolatile contaminants in the air/soil system.     

Development of a speciated, hourly, and gridded air pollutants emission modeling system--a case study on the precursors of photochemical smog in the Seoul metropolitan area, Korea

A speciated, hourly, and gridded air pollutants emission modeling system (SHEMS) was developed and applied in predicting hourly nitrogen dioxide (NO2) and ozone (O3) levels in the Seoul Metropolitan Area (SMA). The primary goal of the SHEMS was to produce a systemized emission inventory for air pollutants including ozone precursors for modeling air quality in urban areas. The SHEMS is principally composed of three parts: (1) a pre-processor to process emission factors, activity levels, and spatial and temporal information using a geographical information system; (2) an emission model for each source type; and (3) a post-processor to produce report and input data for air quality models through database modeling. The source categories in SHEMS are point, area, mobile, natural, and other sources such as fugitive emissions. The emission database produced by SHEMS contains 22 inventoried compounds: sulfur dioxide, NO2, carbon monoxide, and 19 speciated volatile organic compounds. To validate SHEMS, the emission data were tested with the Urban Airshed Model to predict NO2 and O3 concentrations in the SMA during selected episode days in 1994. The results turned out to be reliable in describing temporal variation and spatial distribution of those pollutants.     

Short-term prediction of means of artificial neural urban NO2 pollution by networks

A neural network model for the short-term prediction of concentrations of urban pollutants was developed and applied to the Turin (Northern Italy) air quality network. In particular, the study was focused on NO₂ concentrations measured at five stations; t + 3 and t + 24 hour NO₂ concentration forecasting based on hourly meteorological and concentration data gave good agreement with observed concentrations. This is particularly true for the mean concentration values and concentration distribution. The time of occurrence of peak values was correctly forecast but the amounts were generally underestimated. To reduce this underestimation, an empirical step function was applied in the t + 24 case. This allowed an accurate estimate to be obtained of the few cases in which 50% of the air quality monitoring stations exceeded the attention level (200 μg m^-3) during the following day for at least one hour.     

The New England Air Quality Forecasting Pilot Program: Development of an Evaluation Protocol and Performance Benchmark

Abstract

The National Oceanic and Atmospheric Administration recently sponsored the New England Forecasting Pilot Program to serve as a “test bed” for chemical forecasting by providing all of the elements of a National Air Quality Forecasting System, including the development and implementation of an evaluation protocol. This Pilot Program enlisted three regional-scale air quality models, serving as prototypes, to forecast ozone (O₃) concentrations across the northeastern United States during the summer of 2002. A suite of statistical metrics was identified as part of the protocol that facilitated evaluation of both discrete forecasts (observed versus modeled concentrations) and categorical forecasts (observed versus modeled exceedances/nonexceedances) for both the maximum 1-hr (125 ppb) and 8-hr (85 ppb) forecasts produced by each of the models. Implementation of the evaluation protocol took place during a 25-day period (August 5–29), utilizing hourly O₃ concentration data obtained from over 450 monitors from the U.S. Environment Protection Agency’s Air Quality System network.     

Recovery of the histogram of hourly ozone distribution from weekly average concentrations

A simple method is presented for estimating hourly distribution of air pollutants, based on data collected by passive sensors on a weekly or bi-weekly basis with no need for previous measurements at a site. In order for this method to be applied to locations where no hourly records are available, reference data from other sites are required to generate calibration histograms. The proposed procedure allows one to obtain the histogram of hourly ozone values during a given week with an error of about 30%, which is good considering the simplicity of this approach. This method can be a valuable tool for sites that lack previous hourly records of pollutant ambient concentrations, where it can be used to verify compliance with regulations or to estimate the AOT40 index with an acceptable degree of exactitude.     

Forecasting summertime surface-level ozone concentrations in the Lower Fraser Valley of British Columbia: an ensemble neural network approach

Empirical models for predicting daily maximum hourly average ozone concentrations were developed for 10 monitoring stations in the Lower Fraser Valley (LFV) of British Columbia. According to data from 1991 to 1996, ensemble neural network models increased explained variance an average of 7% over multiple linear regression models using the same input variables. Without modification, all models performed poorly on days when the observed peak ozone concentration exceeded 82 parts per billion, the National Ambient Air Quality Objective. When numbers of extreme events in training data were increased using a histogram equalization process, models were able to forecast exceedances with improved accuracy. Modified generalized additive model (GAM) plots and associated measures of input variable importance and interaction were generated for a subset of the trained models and used to investigate relationships between input variables and ozone levels. The neural network models displayed a high degree of interaction among inputs, and it is likely the ability of these model types to account for interactions, rather than the nonlinearity of individual input variables, that explains their improved forecast skill. Inspection of GAM-style plots indicated that the relative importance of input variables in the ensemble neural network models varied with geographic location within the LFV. Four distinct groups of stations were identified, and rankings of inputs within the groups were generally consistent with physical intuition and results of prior studies.     

Forecasting human exposure to atmospheric pollutants in Portugal – A modelling approach

Air pollution has become one main environmental concern because of its known impact on human health. Aiming to inform the population about the air they are breathing, several air quality modelling systems have been developed and tested allowing the assessment and forecast of air pollution ambient levels in many countries. However, every day, an individual is exposed to different concentrations of atmospheric pollutants as he/she moves from and to different outdoor and indoor places (the so-called microenvironments). Therefore, a more efficient way to prevent the population from the health risks caused by air pollution should be based on exposure rather than air concentrations estimations. The objective of the present study is to develop a methodology to forecast the human exposure of the Portuguese population based on the air quality forecasting system available and validated for Portugal since 2005. Besides that, a long-term evaluation of human exposure estimates aims to be obtained using one-year of this forecasting system application. Additionally, a hypothetical 50% emission reduction scenario has been designed and studied as a contribution to study emission reduction strategies impact on human exposure.To estimate the population exposure the forecasting results of the air quality modelling system MM5-CHIMERE have been combined with the population spatial distribution over Portugal and their time-activity patterns, i.e. the fraction of the day time spent in specific indoor and outdoor places. The population characterization concerning age, work, type of occupation and related time spent was obtained from national census and available enquiries performed by the National Institute of Statistics. A daily exposure estimation module has been developed gathering all these data and considering empirical indoor/outdoor relations from literature to calculate the indoor concentrations in each one of the microenvironments considered, namely home, office/school, and other indoors (leisure activities like shopping areas, gym, theatre/cinema and restaurants). The results show how this developed modelling system can be useful to anticipate air pollution episodes and to estimate their effects on human health on a long-term basis. The two metropolitan areas of Porto and Lisbon are identified as the most critical ones in terms of air pollution effects on human health over Portugal in a long-term as well as in a short-term perspective. The coexistence of high concentration values and high population density is the key factor for these stressed areas. Regarding the 50% emission reduction scenario, the model results are significantly different for both pollutants: there is a small overall reduction in the individual exposure values of PM₁₀ (<10 μg m^?3 h), but for O₃, in contrast, there is an extended area where exposure values increase with emission reduction. This detailed knowledge is a prerequisite for the development of effective policies to reduce the foreseen adverse impact of air pollution on human health and to act on time.     

Correlating Emissions with Time and Temperature to Predict Worst-Case Emissions from Open Liquid Area Sources

Abstract

The two primary factors influencing ambient air pollutant concentrations are emission rate and dispersion rate. Gaussian dispersion modeling studies for odors, and often other air pollutants, vary dispersion rates using hourly meteorological data. However, emission rates are typically held constant, based on one measured value. Using constant emission rates can be especially inaccurate for open liquid area sources, like wastewater treatment plant units, which have greater emissions during warmer weather, when volatilization and biological activity increase. If emission rates for a wastewater odor study are measured on a cooler day and input directly into a dispersion model as constant values, odor impact will likely be underestimated. Unfortunately, because of project schedules, not all emissions sampling from open liquid area sources can be conducted under worst-case summertime conditions. To address this problem, this paper presents a method of varying emission rates based on temperature and time of the day to predict worst-case emissions. Emissions are varied as a linear function of temperature, according to Henry’s law, and a tenth order polynomial function of time. Equation coefficients are developed for a specific area source using concentration and temperature measurements, captured over a multiday period using a data-logging monitor. As a test case, time/temperature concentration correlation coefficients were estimated from field measurements of hydrogen sulfide (H₂S) at the Rowlett Creek Wastewater Treatment Plant in Garland, TX. The correlations were then used to scale a flux chamber emission rate measurement according to hourly readings of time and temperature, to create an hourly emission rate file for input to the dispersion model ISCST3. ISCST3 was then used to predict hourly atmospheric concentrations of H₂S. With emission rates varying hourly, ISCST3 predicted 384 acres of odor impact, compared with 103 acres for constant emissions. Because field sampling had been conducted on relatively cool days (85–90 °F), the constant emission rate underestimated odor impact significantly (by 73%).     

A Simple Methodology for the Determination of Back Trajectories

Abstract

This study comprehensively characterizes hourly fine particulate matter (PM_2.5) concentrations measured via a tapered element oscillating microbalance (TEOM), β-gauge, and nephelometer from four different monitoring sites in U.S. Environment Protection Agency (EPA) Region 5 (in U.S. states Illinois, Michigan, and Wisconsin) and compares them to the Federal Reference Method (FRM). Hourly characterization uses time series and autocorrelation. Hourly data are compared with FRM by averaging across 24-hr sampling periods and modeling against respective daily FRM concentrations. Modeling uses traditional two-variable linear least-squares regression as well as innovative nonlinear regression involving additional meteorological variables such as temperature and humidity. The TEOM shows a relationship with season and temperature, linear correlation as low as 0.7924 and nonlinear model correlation as high as 0.9370 when modeled with temperature. The β-gauge shows no relationship with season or meteorological variables. It exhibits a linear correlation as low as 0.8505 with the FRM and a nonlinear model correlation as high as 0.9339 when modeled with humidity. The nephelometer shows no relationship with season or temperature but a strong relationship with humidity is observed. A linear correlation as low as 0.3050 and a nonlinear model correlation as high as 0.9508 is observed when modeled with humidity. Nonlinear models have higher correlation than linear models applied to the same dataset. This correlation difference is not always substantial, which may introduce a tradeoff between simplicity of model and degree of statistical association. This project shows that continuous monitor technology produces valid PM_2.5 characterization, with at least partial accounting for variations in concentration from gravimetric reference monitors once appropriate nonlinear adjustments are applied. Although only one regression technically meets new EPA National Ambient Air Quality Standards (NAAQS) Federal Equivalent Method (FEM) correlation coefficient criteria, several others are extremely close, showing optimistic potential for use of this nonlinear adjustment model in garnering EPA NAAQS FEM approval for continuous PM_2.5 sampling methods.     

Neural network modelling and prediction of hourly NOx and NO2 concentrations in urban air in London

Multilayer perceptron (MLP) neural networks were trained to model hourly NO_x and NO₂ pollutant concentrations in Central London from basic hourly meteorological data. Results have shown that the models perform well when compared to previous attempts to model the same pollutants using regression based models. This work also illustrates that MLP neural networks are capable of resolving complex patterns of source emissions without any explicit external guidance.