Evaluation of artificial neural networks for fine particulate pollution (PM10 and PM2.5) forecasting

Multi-layer perceptron (MLP) artificial neural network (ANN) models are compared with traditional multiple regression (MLR) models for daily maximum and average O3 and particulate matter (PM10 and PM2.5) forecasting. MLP particulate forecasting models show little if any improvement over MLR models and exhibit less skill than do O3 forecasting models. Meteorological variables (precipitation, wind, and temperature), persistence, and co-pollutant data are shown to be useful PM predictors. If MLP approaches are adopted for PM forecasting, training methods that improve extreme value prediction are recommended.     

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.     

Modelling of dissolved oxygen content using artificial neural networks: Danube River, North Serbia, case study

The aims of this study are to create an artificial neural network (ANN) model using non-specific water quality parameters and to examine the accuracy of three different ANN architectures: General Regression Neural Network (GRNN), Backpropagation Neural Network (BPNN) and Recurrent Neural Network (RNN), for prediction of dissolved oxygen (DO) concentration in the Danube River. The neural network model has been developed using measured data collected from the Bezdan monitoring station on the Danube River. The input variables used for the ANN model are water flow, temperature, pH and electrical conductivity. The model was trained and validated using available data from 2004 to 2008 and tested using the data from 2009. The order of performance for the created architectures based on their comparison with the test data is RNN > GRNN > BPNN. The ANN results are compared with multiple linear regression (MLR) model using multiple statistical indicators. The comparison of the RNN model with the MLR model indicates that the RNN model performs much better, since all predictions of the RNN model for the test data were within the error of less than ±10 %. In case of the MLR, only 55 % of predictions were within the error of less than ±10 %. The developed RNN model can be used as a tool for the prediction of DO in river waters.     

Neural Network and Multiple Regression Models for PM10 Prediction in Athens: A Comparative Assessment

Abstract

Particulate atmospheric pollution in urban areas is considered to have significant impact on human health. Therefore, the ability to make accurate predictions of particulate ambient concentrations is important to improve public awareness and air quality management. This study examines the possibility of using neural network methods as tools for daily average particulate matter with aerodynamic diameter <10 µm (PM₁₀) concentration forecasting, providing an alternative to statistical models widely used up to this day. Based on a data inventory, in a fixed central site in Athens, Greece, ranging over a two-year period, and using mainly meteorological variables as inputs, neural network models and multiple linear regression models were developed and evaluated. Comparison statistics used indicate that the neural network approach has an edge over regression models, expressed both in terms of prediction error (root mean square error values lower by 8.2–9.4%) and of episodic prediction ability (false alarm rate values lower by 7–13%). The results demonstrate that artificial neural networks (ANNs), if properly trained and formed, can provide adequate solutions to particulate pollution prognostic demands.     

Systematic Development of an Artificial Neural Network Model for Real-Time Prediction of Ground-Level Ozone in Edmonton,Alberta, Canada

Abstract

Ground-level ozone is a secondary pollutant that has recently gained notoriety for its detrimental effects on human and vegetation health. In this paper, a systematic approach is applied to develop artificial neural network (ANN) models for ground-level ozone (O₃) prediction in Edmonton, Alberta, Canada, using ambient monitoring data for input. The intent of these models is to provide regulatory agencies with a tool for addressing data gaps in ambient monitoring information and predicting O₃ events. The models are used to determine the meteorological conditions and precursors that most affect O₃ concentrations. O₃ time-series effects and the efficacy of the systematic approach are also assessed. The developed models showed good predictive success, with coefficient of multiple determination values ranging from 0.75 to 0.94 for forecasts up to 2 hr in advance. The inputs most important for O₃ prediction were temperature and concentrations of nitric oxide, total hydrocarbons, sulfur dioxide, and nitrogen dioxide.     

A new hybrid system of QSAR models for predicting bioconcentration factors (BCF)

The aim was to develop a reliable and practical quantitative structure–activity relationship (QSAR) model validated by strict conditions for predicting bioconcentration factors (BCF). We built up several QSAR models starting from a large data set of 473 heterogeneous chemicals, based on multiple linear regression (MLR), radial basis function neural network (RBFNN) and support vector machine (SVM) methods. To improve the results, we also applied a hybrid model, which gave better prediction than single models. All models were statistically analysed using strict criteria, including an external test set. The outliers were also examined to understand better in which cases large errors were to be expected and to improve the predictive models. The models offer more robust tools for regulatory purposes, on the basis of the statistical results and the quality check on the input data.     

A new hybrid system of QSAR models for predicting bioconcentration factors (BCF)

The aim was to develop a reliable and practical quantitative structure-activity relationship (QSAR) model validated by strict conditions for predicting bioconcentration factors (BCF). We built up several QSAR models starting from a large data set of 473 heterogeneous chemicals, based on multiple linear regression (MLR), radial basis function neural network (RBFNN) and support vector machine (SVM) methods. To improve the results, we also applied a hybrid model, which gave better prediction than single models. All models were statistically analysed using strict criteria, including an external test set. The outliers were also examined to understand better in which cases large errors were to be expected and to improve the predictive models. The models offer more robust tools for regulatory purposes, on the basis of the statistical results and the quality check on the input data.     

Variation of surface ozone in Campo Grande,Brazil: meteorological effect analysis and prediction

The effect of meteorological variables on surface ozone (O₃) concentrations was analysed based on temporal variation of linear correlation and artificial neural network (ANN) models defined by genetic algorithms (GAs). ANN models were also used to predict the daily average concentration of this air pollutant in Campo Grande, Brazil. Three methodologies were applied using GAs, two of them considering threshold models. In these models, the variables selected to define different regimes were daily average O₃ concentration, relative humidity and solar radiation. The threshold model that considers two O₃ regimes was the one that correctly describes the effect of important meteorological variables in O₃ behaviour, presenting also a good predictive performance. Solar radiation, relative humidity and rainfall were considered significant for both O₃ regimes; however, wind speed (dispersion effect) was only significant for high concentrations. According to this model, high O₃ concentrations corresponded to high solar radiation, low relative humidity and wind speed. This model showed to be a powerful tool to interpret the O₃ behaviour, being useful to define policy strategies for human health protection regarding air pollution.     

Formation of Disinfection Byproducts (DBPs) in Surface Water Sources: Differential Ultraviolet (UV) Absorbance Approach

Chlorination for drinking water forms various disinfection byproducts (DBPs) of trihalomethanes (THMs) and haloacetic acids (HAAs). Chlorination has been attributed to the destruction of activated aromatic sites of the natural organic matter (NOM) predominantly at electron rich sites. Experiments with Istanbul surface waters showed that the magnitude of the decrease in the ultraviolet (UV) absorbance at 272 nm (UV₂₇₂) was an excellent indicator of destruction of these sites by chlorine. The main objective of the present study is to develop the differential UV₂₇₂ absorbance (ΔUV₂₇₂) related models for the prediction of the formation of THM, HAA, and their species in raw water samples from Terkos, Buyukcekmece, and Omerli lakes under different chlorination conditions. Significant factors affecting DBP formation in the raw waters were identified through numerical and graphical techniques. The R² values of the models varied between 0.94 and 0.97, indicating excellent predictive ability for THM₄ and HAA₉ in the raw waters. The models were validated using additional data. The results of this study concluded that addition of ΔUV₂₇₂ parameter into THM₄ and HAA₉ models make the prediction of these DBP compounds more precisely than those of DBP models developed in the past. A better understanding of these modeling systems will help the water treatment plant operators to minimize the DBP formation, providing a healthier and better drinking water quality as well as identifying strategies to improve water treatment and disinfection processes.     

Particulate air quality model predictions using prognostic vs. diagnostic meteorology in central California

Comparisons were made between three sets of meteorological fields used to support air quality predictions for the California Regional Particulate Air Quality Study (CRPAQS) winter episode from December 15, 2000 to January 6, 2001. The first set of fields was interpolated from observations using an objective analysis method. The second set of fields was generated using the WRF prognostic model without data assimilation. The third set of fields was generated using the WRF prognostic model with the four-dimensional data assimilation (FDDA) technique. The UCD/CIT air quality model was applied with each set of meteorological fields to predict the concentrations of airborne particulate matter and gaseous species in central California. The results show that the WRF model without data assimilation over-predicts surface wind speed by ～30% on average and consequently yields under-predictions for all PM and gaseous species except sulfate (S(VI)) and ozone(O₃). The WRF model with FDDA improves the agreement between predicted and observed wind and temperature values and consequently yields improved predictions for all PM and gaseous species. Overall, diagnostic meteorological fields produced more accurate air quality predictions than either version of the WRF prognostic fields during this episode. Population-weighted average PM_2.5 exposure is 40% higher using diagnostic meteorological fields compared to prognostic meteorological fields created without data assimilation. These results suggest diagnostic meteorological fields based on a dense measurement network are the preferred choice for air quality model studies during stagnant periods in locations with complex topography.     

Neural network and multiple regression models for PM10 prediction in Athens: a comparative assessment

Particulate atmospheric pollution in urban areas is considered to have significant impact on human health. Therefore, the ability to make accurate predictions of particulate ambient concentrations is important to improve public awareness and air quality management. This study examines the possibility of using neural network methods as tools for daily average particulate matter with aerodynamic diameter <10 microm (PM10) concentration forecasting, providing an alternative to statistical models widely used up to this day. Based on a data inventory, in a fixed central site in Athens, Greece, ranging over a two-year period, and using mainly meteorological variables as inputs, neural network models and multiple linear regression models were developed and evaluated. Comparison statistics used indicate that the neural network approach has an edge over regression models, expressed both in terms of prediction error (root mean square error values lower by 8.2-9.4%) and of episodic prediction ability (false alarm rate values lower by 7-13%). The results demonstrate that artificial neural networks (ANNs), if properly trained and formed, can provide adequate solutions to particulate pollution prognostic demands.     

Source Apportionment of Particulate Matter (PM10) and Indoor Dust in a University Building

A study on source apportionment of indoor dust and particulate matter (PM₁₀) composition was conducted in a university building by using chemometrics. The objective of this study was to investigate the potential sources of selected heavy metals and ionic species in PM₁₀ and indoor dust. PM₁₀ samples were collected using a low-volume sampler (LVS) and indoor dust was collected using a soft brush. Inductively coupled plasma spectrometry (ICP-MS) was used to determine the concentration of heavy metals, while the concentration of cations and anions was determined by atomic absorption spectrometer (AAS) and ion chromatography (IC), respectively. The concentration of PM₁₀ recorded in the building throughout the sampling period ranged from 20 ± 10 μgm^?3 to 80 ± 33 μgm^?3. The composition of heavy metals in PM₁₀ and indoor dust were dominated by zinc (Zn), followed by lead (Pb), copper (Cu), and cadmium (Cd). Principle component analysis (PCA) and multiple linear regression (MLR) showed that the main sources of pollutants in PM₁₀ came from indoor renovations (73.83%), vehicle emissions (16.38%), earth crust sources (9.68%), and other outdoor sources (0.11%). For indoor dust, the pollutant source was mainly earth crust. This study suggests that chemometrics can be used for forensic investigation to determine the possible sources of indoor contaminants within a public building.     

Particulates and Metallic Elements Monitoring at Two Sampling Sites (Harbor,Airport) in Taiwan

This study characterized the dry deposition flux and dry deposition velocity (Vd) of metallic elements attached on particulate matter. Specifically, large particles (>10 μm), coarse particles (10 μm～2.5 μm), and fine particles (<2.5 μm) were studied at the Gong Ming Junior High School (Taichung Airport) and Taichung Harbor sampling sites in central Taiwan. Ambient air samples were collected to determine total suspended particulate matter (TSP), dry deposition plate (DDP), Vd, coarse particulate matter (PM_2.5–10) and fine particulate matter (PM_2.5), and metallic elements concentrations at the Airport and Taichung Harbor sites between June 17, 2013, and November 14, 2013. The results revealed that the average TSP, DDP, Vd, PM_2.5–10, and PM_2.5 particulate at the Airport were 54.55 (μg/m³), 902.25 (μg/m²-min), 17.11 (m/sec), 0.003 (μg/m³), and 0.010 (μg/m³), respectively; while these values at Taichung Harbor were 63.66 (μg/m³), 539.69 (μg/m²-min), 9.94 (m/sec), 0.003 (μg/m³), and 0.014 (μg/m³), respectively. In addition, the results showed that the average Cu and Pb concentrations were higher than Cr, Ni, and Cd for both the airport and harbor sampling sites. Furthermore, Cr, N, Cu, Cd, and Pb had the highest average concentrations versus those reported for other study areas, with one exception: The results obtained in Kacanik, Kosovo, during 2005. The average metallic elements concentrations order was Cu > Pb > Cr > Ni > Cd.     

Ensemble and bias-correction techniques for air quality model forecasts of surface O3 and PM2.5 during the TEXAQS-II experiment of 2006

Several air quality forecasting ensembles were created from seven models, running in real-time during the 2006 Texas Air Quality (TEXAQS-II) experiment. These multi-model ensembles incorporated a diverse set of meteorological models, chemical mechanisms, and emission inventories. Evaluation of individual model and ensemble forecasts of surface ozone and particulate matter (PM) was performed using data from 119 EPA AIRNow ozone sites and 38 PM sites during a 50-day period in August and September of 2006. From the original set of models, two new bias-corrected model data sets were built, either by applying a simple running mean average to the past 7 days of data or by a Kalman-Filter approach. From the original and two bias-corrected data sets, three ensembles were created by a simple averaging of the seven models. For further improvements three additional weighted model ensembles were created, where individual model weights were calculated using the singular value decomposition method. All six of the ensembles are compared to the individual models and to each other in terms of root mean square error, correlation, and contingency and probabilistic statistics. In most cases, each of the ensembles show improved skill compared to the best of the individual models. The over all best ensemble technique was found to be the combination of Kalman-Filtering and weighted averaging. PM_2.5 aerosol ensembles demonstrated significant improvement gains, mostly because the original model's skill was very low.     

Near-road air pollutant concentrations of CO and PM2.5: A comparison of MOBILE6.2/CALINE4 and generalized additive models

The contribution of vehicular traffic to air pollutant concentrations is often difficult to establish. This paper utilizes both time-series and simulation models to estimate vehicle contributions to pollutant levels near roadways. The time-series model used generalized additive models (GAMs) and fitted pollutant observations to traffic counts and meteorological variables. A one year period (2004) was analyzed on a seasonal basis using hourly measurements of carbon monoxide (CO) and particulate matter less than 2.5 μm in diameter (PM_2.5) monitored near a major highway in Detroit, Michigan, along with hourly traffic counts and local meteorological data. Traffic counts showed statistically significant and approximately linear relationships with CO concentrations in fall, and piecewise linear relationships in spring, summer and winter. The same period was simulated using emission and dispersion models (Motor Vehicle Emissions Factor Model/MOBILE6.2; California Line Source Dispersion Model/CALINE4). CO emissions derived from the GAM were similar, on average, to those estimated by MOBILE6.2. The same analyses for PM_2.5 showed that GAM emission estimates were much higher (by 4–5 times) than the dispersion model results, and that the traffic-PM_2.5 relationship varied seasonally. This analysis suggests that the simulation model performed reasonably well for CO, but it significantly underestimated PM_2.5 concentrations, a likely result of underestimating PM_2.5 emission factors. Comparisons between statistical and simulation models can help identify model deficiencies and improve estimates of vehicle emissions and near-road air quality.     

Artificial neural network for the identification of unknown air pollution sources

Artificial neural networks (ANN), whose performances to deal with pattern recognition problems is well known, are proposed to identify air pollution sources. The problem that is addressed is the apportionment of a small number of sources from a data set of ambient concentrations of a given pollutant. Three layers feed-forward ANN trained with a back-propagation algorithm are selected. A test case is built, based on a Gaussian dispersion model. A subset of hourly meteorological conditions and measured concentrations constitute the input patterns to the network that is wired to recover relevant emission parameters of unknown sources as outputs. The rest of the model data are corrupted adding noise to some meteorological parameters and we test the effectiveness of the method to recover the correct answer. The ANN is applied to a realistic case where 24 h SO₂ concentrations were previously measured. Some of the limitations of the ANN approach, together with its capabilities, are discussed in this paper.     

Correlations between fine particulate matter (PM2.5) and meteorological variables in the United States: Implications for the sensitivity of PM2.5 to climate change

We applied a multiple linear regression (MLR) model to study the correlations of total PM_2.5 and its components with meteorological variables using an 11-year (1998–2008) observational record over the contiguous US. The data were deseasonalized and detrended to focus on synoptic-scale correlations. We find that daily variation in meteorology as described by the MLR can explain up to 50% of PM_2.5 variability with temperature, relative humidity (RH), precipitation, and circulation all being important predictors. Temperature is positively correlated with sulfate, organic carbon (OC) and elemental carbon (EC) almost everywhere. The correlation of nitrate with temperature is negative in the Southeast but positive in California and the Great Plains. RH is positively correlated with sulfate and nitrate, but negatively with OC and EC. Precipitation is strongly negatively correlated with all PM_2.5 components. We find that PM_2.5 concentrations are on average 2.6 μg m^?3 higher on stagnant vs. non-stagnant days. Our observed correlations provide a test for chemical transport models used to simulate the sensitivity of PM_2.5 to climate change. They point to the importance of adequately representing the temperature dependence of agricultural, biogenic and wildfire emissions in these models.     

Characterizing the spatiotemporal variability of PM2.5 in Cusco,Peru using kriging with external drift

Advancing the understanding of spatiotemporal aspects of air pollution in the urban environment is an area where improved methods can be of great benefit to exposure assessment and policy support. This paper explores the potential of a technique known as kriging with external drift (KED) to provide high resolution maps of fine particulate matter for a downtown region of Cusco, Peru. There were three stages in this research. The first was to conduct a pilot level monitoring campaign to investigate ambient, regional, and street-level air pollutant concentrations for particulate matter (PM_2.5, PM₁₀) and carbon monoxide (CO) in the Province of Cusco. The second was to compile observations within a geographic information system (GIS) in order to characterize the proximal effect of the local transportation network, elevation, and land use classifications on PM_2.5. Third, regression, ordinary kriging and kriging with external drift were used to model PM_2.5 for three select time periods during a 24-h day. Statistical evaluations indicate kriging with external drift resulted in the strongest models explaining 64% of variability seen with morning particle concentrations, 25% for afternoon particles, and 53% in evening particles. These models capture spatial and temporal variability for air pollution in Cusco. These variations seem to be influenced, to varying degrees, by elevation, meteorological conditions, spatial location, and transportation characteristics. In conclusion, combining GIS, meteorological data and geostatistics proved to be a complementary suite of tools for incorporating spatiotemporal analysis into the air quality assessment.     

Systematic development of an artificial neural network model for real-time prediction of ground-level ozone in Edmonton, Alberta, Canada

Ground-level ozone is a secondary pollutant that has recently gained notoriety for its detrimental effects on human and vegetation health. In this paper, a systematic approach is applied to develop artificial neural network (ANN) models for ground-level ozone (O3) prediction in Edmonton, Alberta, Canada, using ambient monitoring data for input. The intent of these models is to provide regulatory agencies with a tool for addressing data gaps in ambient monitoring information and predicting O3 events. The models are used to determine the meteorological conditions and precursors that most affect O3 concentrations. O3 time-series effects and the efficacy of the systematic approach are also assessed. The developed models showed good predictive success, with coefficient of multiple determination values ranging from 0.75 to 0.94 for forecasts up to 2 hr in advance. The inputs most important for O3 prediction were temperature and concentrations of nitric oxide, total hydrocarbons, sulfur dioxide, and nitrogen dioxide.     

Prediction of the adsorption capability onto activated carbon of a large data set of chemicals by local lazy regression method

Accurate quantitative structure–property relationship (QSPR) models based on a large data set containing a total of 3483 organic compounds were developed to predict chemicals’ adsorption capability onto activated carbon in gas phrase. Both global multiple linear regression (MLR) method and local lazy regression (LLR) method were used to develop QSPR models. The results proved that LLR has prediction accuracy 10% higher than that of MLR model. By applying LLR method we can predict the test set (787 compounds) with Q²_ext of 0.900 and root mean square error (RMSE) of 0.129. The accurate model based on this large data set could be useful to predict adsorption property of new compounds since such model covers a highly diverse structural space.