Forecasting PM10 in metropolitan areas: Efficacy of neural networks

Environmental Pollution 163 (2012) 62e67

Contents lists available at SciVerse ScienceDirect

Environmental Pollution journal homepage: www.elsevier.com/locate/envpol

Forecasting PM10 in metropolitan areas: Efficacy of neural networks H.J.S. Fernando a, M.C. Mammarella b, G. Grandoni b, P. Fedele b, R. Di Marco b, R. Dimitrova a, *, P. Hyde c a

University of Notre Dame, Civil Engineering & Geological Sciences & Environmental Fluid Dynamics Laboratories, Notre Dame, IN 46446, USA ENEA, Italian National Agency for New Technologies, Energy and Sustainable Economic Development, Lungotevere Thaon di Revel, 76, 00196-Roma, Italy c Arizona State University, The School for Engineering of Matter, Transport and Energy (SEMTE), Tempe, AZ 85287-9809, USA b

a r t i c l e i n f o

a b s t r a c t

Article history: Received 24 September 2011 Received in revised form 6 December 2011 Accepted 11 December 2011

Deterministic photochemical air quality models are commonly used for regulatory management and planning of urban airsheds. These models are complex, computer intensive, and hence are prohibitively expensive for routine air quality predictions. Stochastic methods are becoming increasingly popular as an alternative, which relegate decision making to artificial intelligence based on Neural Networks that are made of artificial neurons or ‘nodes’ capable of ‘learning through training’ via historic data. A Neural Network was used to predict particulate matter concentration at a regulatory monitoring site in Phoenix, Arizona; its development, efficacy as a predictive tool and performance vis-à-vis a commonly used regulatory photochemical model are described in this paper. It is concluded that Neural Networks are much easier, quicker and economical to implement without compromising the accuracy of predictions. Neural Networks can be used to develop rapid air quality warning systems based on a network of automated monitoring stations. Ó 2011 Elsevier Ltd. All rights reserved.

Keywords: Neural network Air quality prediction Human health warnings

1. Introduction The US Environmental Protection Agency (EPA) classifies primary air pollutants based on their human health impacts, and imposes upper limits on allowable ambient pollutant concentrations to protect public health. These limits define the US National Ambient Air Quality Standards, NAAQS (EPA, 2001). Exceedences of NAAQS lead to substantial governmental penalties on cities, and illnesses related to high pollutant levels lead to social and economic disruptions such as workplace slowdown, absenteeism and congested emergency rooms. Therefore, it is of best interest for local authorities to predict air quality sufficiently in advance and warn the public of pending unhealthy conditions so that direct exposure to ambient air can be avoided during such periods. The case in point here, Phoenix metropolis, has been classified as a serious non-attainment area of NAAQS for particulate matter of aerodynamic diameter less than 10 mm, PM10 (ADEQ, 2005; MC-AQD, 2001e2009), and local authorities have taken major steps to achieve compliance. For example, with the support of a USEPA Challenge Grant, the development of a health warning system with hourly updates was initiated in 2007, which is expected to replace operator based daily forecasting derived from meteorological

* Corresponding author. E-mail addresses: [email protected], (R. Dimitrova).

[email protected]

0269-7491/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.envpol.2011.12.018

predictions. Since deterministic photochemical models such as USEPA’s Community Air Quality Modelling System (CMAQ) are too computationally intense for daily forecasting and require frequent (expensive) updating of pollution inventories, an affordable yet scientifically defensible methodology needs to be adopted and tested for PM forecasting. In this regard, an attractive methodology is the use of Neural Networks (NN), which has been tested for several European cities, including Milan, Rome and Athens; and some of these systems are now routinely used for operational forecasting (Mammarella et al., 2005; Kolehmainen et al., 2001; Grivas and Chaloulakou, 2006). Nevertheless, several issues were of concern, which have not been adequately addressed before: how does the fidelity of NN predictions compare with that of photochemical models? And how do NN perform with respect to available data, especially during the periods of health-adverse high PM10 events? The present study attempts to answer these questions. Most of the severe PM10 violations of the Phoenix area have been attributed to regional natural ‘exceptional events’ or local exceptional (episodic) events associated with windblown dust emanating from area sources such as construction and agricultural sites, vacant lots and alluvial channels. During these events, the pollution concentration spikes for a short period of time, thus raising the 24-h averaged PM10 concentration anomalously. This may lead to PM10 concentrations above the currently set 24averaged NAAQS, 150 mg/m3. Even if the NAAQS is not exceeded, severe health repercussions can occur due to pulsed PM10 events.

H.J.S. Fernando et al. / Environmental Pollution 163 (2012) 62e67

For example, thunderstorm-induced asthma epidemics that swamp hospital emergency rooms within 20 min of the onset of a storm have been attributed to the suspension of micron-sized particulate matter (Venables et al., 1997). Unlike for the case of pollution sources associated with industrial and transportation networks, it is difficult to predict, using air pollution models, PM pollution arising from natural episodic events; poor coupling of dust entrainment and meteorology as well as strong non-stationarity of meteorological processes are some of the causes (Choi et al., 2006; Choi and Fernando, 2008). Neural networks are expected to be particularly helpful in such cases. Traditionally, the term neural network is analogous to a circuit of biological neurons, but current usage of the term implies networks made up of artificial neurons or “nodes”. These networks have the capability of learning inputeoutput relationships via imbedded functions and historic data. While this may appear as metaphysical, the crux of NN concept is a set of data and predictive equations, which acquires ‘knowledge’ and improved predictive ability via simple ‘acquisition and prediction’ exercises. More the availability of data, the better the neural network is trained. In this work the application was limited to the prediction of past PM10 concentrations (hindcasting), based on training using data taken before that period. An operational neural network, in this spirit, would entail prediction of air pollution a few hours to a few days ahead, which can be disseminated to the community though electronic media. This network would acquire more ‘skill’ with time as additional hourly/daily relationships between PM10 concentrations and meteorological variables become available, processed, and predicted. 2. EnviNNet development The Italian National Agency for New Technologies, Energy and Sustainable Economic Development (ENEA) has developed a prototype stochastic NN model based on neural networks, called EnviNNet, for air quality forecasting in Rome, Milan and Napoli. The development of EnviNNet required careful selection of a subset of input variables, paying attention to site-specific exceptional events, including time lag effects. The noise of data needed to be considered in order to satisfactorily adapt the non-linear dynamic interaction between meteorological and pollution related processes. A three-layer MLP (MultiLayer Perceptrons) network, with the number of hidden nodes selected to reliably rebuild data from a test data-set was used. In training the NN, characteristics of high, low and episodic air pollution events at a particular data site were taken into account. Considering that PM10 concentrations are heterogeneous throughout the year, specific input data subsets were built by combining time section data from different years to ensure a robust and adaptive response of NN. In the NNs installed in Italians cities, meteorological and pollution data were taken from selected fourto-six-month windows as well as time periods that show noteworthy patterns. Considering its wide usage in atmospheric applications (Mayora-Ibarra et al., 1998; Niska et al., 2004), a MLP type NN was selected (Gardner and Dorling, 1998), paying attention to its eventual function of predicting PM10. The MLP structure consists of an interconnected system of nodes (neurons) within a hidden layer that employs non-linear continuous transfer functions to connect input and output vectors. The topology included full connection between layers and no connection between neurons in the same layer. An exponential transfer function was used to ensure positive functions at the output node and hyperbolic tangent for hidden nodes, and information flow operated with forward feed. This is an excellent compromise for a non-linear function both globally and locally.

63

Standardized data inputs eliminated problems associated with different measurement scales of different predictors. The reconstruction of the output signal employed a linear combination and exponentialization of outputs of hidden nodes. In training, the system parameters were learnt or estimated through the conjugate-gradient method. It is preferred over the backpropagation method as it exploits both first-order (gradient) and second-order (curvature) data during optimization of the objective function. Mathematically, a three-layer NN has the following form,

y ¼ f ð4ðx; wÞÞ where x represents the input data, w the coefficients (parameters estimated by learning), f the activation functions from layers 2 to 3 and 4 the activation functions from layers 1 to 2. The choice of f and 4 determines the output. For example, if 4 is hyperbolic tangent and f is linear, an input of meteorological and pollution variables are transformed to an output with both negative and positive values, but if f is exponential, the output can have only positive values. As such, an exponential function for the output and a hyperbolic tangent activation function for the hidden neurons were selected.

2.1. The case study As the first step, PM10 and meteorological data collected at the Central Phoenix (CP) monitoring station in the metropolitan Phoenix area was selected. CP is one of the seven continuous pollution (including PM10) monitoring stations operated in the Phoenix area; see Fig. 1. EnviNNet was trained using a selected set of appropriately preprocessed historic time series data at CP, whereupon it was used to predict PM10 in the hindcasting mode. The input vector is made of the following variables determined based on physically plausible governing variables (i.e., the same approach as in conventional dimensional analysis): 2.1.1. Pollution data These constitute the endogenous component of the NN. Time lags are selected by optimizing the data reconstruction efficiency of the NN, so as to achieve the best trade-off between sensitivity and specificity (Shaltaf and Mohammad, 2009). Since these are multi-step operations, the measured pollutant levels are used in a rolling mode to generate the reconstructed values.

Fig. 1. The locations of pollution and meteorological monitoring stations in Phoenix, Arizona.

64

H.J.S. Fernando et al. / Environmental Pollution 163 (2012) 62e67

2.1.2. Meteorological measurements The meteorological data comprise of exogenous component of the NN, and have a modulating effect on PM10 concentration levels. The time lags were selected by optimizing, as before, the relationship between data reconstruction efficacy and network efficiency (Frank et al., 2001; Boulle et al., 2001). Forecasted meteorological data were not used for hindcasting. 2.1.3. Statistical and descriptive indicators They represent the behaviour of the (CP) neighbourhood via classes of input parameters that account for aggregated homogeneous spatio-temporal bands of microclimatic factors and air pollutant concentration. The latter is dependent on the wind velocity and direction (transport and diffusion of PM10) as well as emissions specified by the ’day’ (working days or holidays). A mathematical description of the procedure used for EnviNNet case study is shown in Appendix 1. 2.2. Training and processing phase The characteristics of high, low and episodic air pollution events at CP had to be accounted, and therefore the training strategy is critical for reliable predictions. To this end, specific techniques were employed for the case study. The data for training were obtained from a selected smaller window as well as from time periods of noteworthy patterns (e.g. exceptional events) outside these time windows. The input variables were thus selected to have the following attributes:  A historical series of PM10 and meteorological (hourly) data collected at CP during 2005, at least 50 days prior to the period selected for hindcasting,  Days based on analysis and correlations of historical series, representing salient PM10 trends in the time window selected (15 days),  Noteworthy patterns (e.g. significant PM10 peaks representative of exceptional events) observed outside the time window of training. In this phase, the intrinsic parameters of the NN were estimated by numeric optimization until a satisfactory convergence (i.e., the maximum of the objective function) was obtained (Dorling et al., 2003; Vidyasagar, 1998). The weights and time lags so identified (given in Appendix 1) naturally depend on the site, for example, local weather and neighbourhood characteristics. The most important aspect of training is obtaining of a non-linear dynamic relationship that connects input data with PM10 output so that the NN can be operated in hindcasting mode, with 24 h advance predictive capability. 3. Comparisons between CMAQ and EnviNNet e results and discussion Two different types of predictive systems e deterministic (Models3-CMAQ) and stochastic (EnviNNet) e were evaluated against observations at one monitor in the central Phoenix area. A one month ‘design’ period covering November 2005 was selected, considering that, in general, winter months exhibit the highest PM10 concentrations. The selected period has both high and low PM10 days. The predictive system consists of three integrated models: the Pennsylvania State University and National Centre for Atmospheric Research (NCAR) Mesoscale Meteorological Model e MM5v3.7 (Grell and Dudia, 1994) for simulating the weather, the Sparse Matrix Operator Kernel Emissions e SMOKEv2.2 model for

emissions processing (Carolina Environmental Program, 2005), and the Community Multiscale Air Quality Model e CMAQv.4.5 (Byun and Ching, 1999) for simulating pollutant concentrations. Temporally, spatially, and chemically resolved model-ready emissions were prepared using the recent Western Regional Air Partnership (WRAP) emissions inventory including the BRAVO-1999 study for the Mexico emissions inventory (WGA, 2006). The modelling domain was based on a Lambert Projection centered at (97 W, 40 N), with three nested domains having 36, 12 and 4 km grids to predict meteorology, emissions, and air quality, respectively. Vertically 29 levels were specified with the layer closest to the ground being 7 m thick to capture boundary-layer processes. The inner domain with 4  4 km grids was centered in the Salt River Valley in central Phoenix. The Models3-CMAQ air quality modelling system has been validated for the exact domains for meteorological variables and PM10 concentrations (Fernando at al., 2009; Dimitrova et al., 2009). The outcomes from the finest domain were compared with results from EnviNNet and observed data. The outer domains were exclusively used to provide boundary and initial conditions for the inner domains. The meteorological and PM10 observations from the year 2005 were used with the stochastic model EnviNNet to select a set of appropriate pre-processed historic time series at the Central Phoenix site (CP). Statistical and descriptive indicators were found to represent behaviour of the CP neighbourhood via various classes of the input parameters. The indicators account for aggregated homogeneous spatio-temporal bands of microclimatic factors and air pollutant concentration. The latter is dependent on the wind transport, diffusion of PM10 and emissions, which can be separately specified for working days and holidays. The EnviNNet output vis-à-vis the prediction of the deterministic modelling system is presented next. The regulatory PM10 enforcements are made with 24-h averages, which have also been used to estimate the relationship between pollution and asthma events for Phoenix (Dimitrova et al., 2011). Both models were evaluated against observations as shown in Fig. 2. The neural network satisfactorily captures PM10 peaks whereas CMAQ encountered problems in this regard (November 10, 17, 21e22). The model performance, based on hourly data modelled by individual methods and the observed data are shown in Fig. 3. The coefficients of determination (R2) are also shown. The R2 is approximately two times higher for EnviNNet compared to CMAQ for the one month period considered. The predicted hourly concentrations are closer to the observed values and clustered around the perfect coefficient of determination (R2 ¼ 1) while the scatter plot for CMAQ is more dispersed. It is difficult to predict hourly PM10 concentration, compared to forecasting of common gaseous air pollutants, due to complexities of processes involved in the formation, transport and removal of atmospheric aerosols. In addition, CMAQ cannot predict the hourly PM10 concentrations at

Fig. 2. 24-h PM10 concentrations predicted by CMAQ and EnviNNet in comparison with observations (error bars are 25% of predicted value).

H.J.S. Fernando et al. / Environmental Pollution 163 (2012) 62e67

65

Fig. 3. Scatter plots of PM10 concentration predicted by CMAQ (a) and EnviNNet (b) in comparison with hourly observed data.

Fig. 4. Comparison of hourly PM10 concentration predicted by CMAQ and EnviNNet with observed data for two high pollution episodes: November, 7e11, 2005 (a) and November, 21e25, 2005 (b).

a local observational site, given that deterministic models calculate only the averaged quantities over individual grid cells (4 km resolution in our case). In addition, emission inventories contribute substantial uncertainties to deterministic models. In order to investigate the efficacy of hourly predictions, two five day periods were selected, 7e11 and 21e25 (November 2005). A comparison between the two models is shown in Fig. 4. In addition, different statistics (defined in the Appendix 2) were calculated to evaluate the performance of both models, and they indicate reasonable agreement between the calculated and observed values (Table 1). The Index of Agreement (IA) for both models is more than 0.6, showing a good correspondence between calculations and observations. Generally, EnviNNet gives better IA in comparison with CMAQ, except during 21e25 of November. The Mean Absolute Error (MAE) is less than 26 mg/m3 for CMAQ and less than 19 mg/m3 for EnviNNet for the month, and MAE increases for high

pollution events. The Root Mean Square Error (RMSE) is in the range 25e40 mg/m3 for different periods, with EnviNnet yielding smaller errors than CMAQ. The model performance for the prediction of high PM10 concentration events was also evaluated, considering perceived impacts of such events on public health and hence the importance of issuing advisories on PM10 episodic events. Note that there are no regulatory or recommended criteria for such events, and hence in the present study the threshold defining high pollution events was set as one standard deviation above the averaged concentration level. The performance measures used are the probability of detection (POD e total number of successful forecasts of high-concentration events or hits as a fraction of total observed events; this is also called the hit rate), the false alarm rate (FAR e the number of false alarms divided by the total number of events observed, that is the number of false alarms and correct rejections), and the critical

66

H.J.S. Fernando et al. / Environmental Pollution 163 (2012) 62e67

Table 1 Summary of the statistical measures that compares the observations and model concentrations of PM10 for different periods. Model

MAE

RMSE

IA

POD

FAR

CSI

Period

CMAQ EnviNNet CMAQ EnviNNet CMAQ EnviNNet

26.12 19.01 32.38 23.27 25.73 24.56

34.40 25.02 40.64 31.44 36.26 33.99

0.68 0.77 0.57 0.72 0.69 0.65

0.293 0.440 0.233 0.467 0.467 0.375

0.025 0.031 0.031 0.033 0.053 0.047

0.243 0.314 0.184 0.304 0.280 0.240

November November 7e11 November 7e11 November 21e25 November 21e25 November

of grid-based models they predict pollutants at the training locations only. By having a network of monitoring stations, each armed with a neural network, and by proper spatial interpolation of predictions [e.g. using Kriging methods as in Dimitrova et al. (2011)], it is possible to make spatio-temporal predictions for a given domain, which in turn can be automated to issue health advisories for communities. The USEPA has recently approved funds to set up a pilot network of this ilk for Phoenix, Arizona. Acknowledgments

success index (CSI ethe number of correct forecasts or hits divided by the total number of events forecasts plus the number of misses, that is the addition of hits, false alarms and misses); see Wilks (1995), Chaloulakou et al. (1999), Zhou and Du (2010). The POD is in range 0e1, with a value of 1 indicating a perfect forecast. The CSI is used relatively frequently since, unlike POD and FAR, it takes into account both false alarms and missed events, and is therefore is a more balanced score. The CSI is somewhat sensitive to the climatology of event, and tends to give poorer scores for rare events. The POD range for both models was 23%e47%, with better forecasts for high-concentration events by EnviNNet. The FAR values noted were below 5% in every case. The CSI values typically exceeded 0.2 and 0.3, respectively, for CMAQ and EnviNNet. The quality of PM10 predictions by the Neural Network method for episodic events can be improved by increasing the frequency of extreme values in the training process, either by including more episodes in the training set (i.e., lengthy observations for training) or by including each episodic case multiple times (Kukkonen et al., 2003; Chaloulakou et al., 2003). 4. Conclusions In all, the EnviNNet was found to better predict moderate to high PM10 concentrations than CMAQ.The computations time for 24-h CMAQ simulation was about 5 h using 4 nodes on a High Performance Computer (HPC) Linux Cluster and for EnviNNet was about 10 min using a workstation SUN Blade 1500, 1.062 GHz UltraSPARC III, 2 GB RAM, Sun Solaris. Considering the better performance and rapid and inexpensive computations of EnviNnet, neural networks can be recommended as a superior competitor to photochemical models for practical air quality predictions, albeit the former stokes criticism of not having strong physical and dynamical bases. An additional advantage of neural networks is that they do not require an expensive emission inventory or its regular upgrading. Therein, inputeoutput relationships for a particular site are derived using large volumes of historic data, and in essence this approach is analogous to dimensional analysis based predictions, without explicitly identifying non-dimensional numbers. Strong physical, dynamical and numerical bases of photochemical models as well as their extensive spatial coverage, nevertheless, should not be underestimated. With improvement of computational power and model architecture, they are expected to be tenable for future regulatory air quality predictions and human health warnings. About 70% of ambient PM10 in Phoenix comes from fugitive emissions, which include traffic on unpaved roads, land clearings and construction activities, which are difficult to quantify accurately. Furthermore, emissions from many source categories cannot be accurately apportioned in time and space, and the ever changing land use calls for frequent and expensive upgrading of the emission inventory for photochemical models. Neural networks do not require an emission inventory, but unlike domain-wide predictions

Thismodelling and NN work was funded by the Arizona Department of Environmental Quality (Air Quality Division and Director’s Special Initiative on Children’s Health) through a challenge grant from the USEPA. Most of the work was conducted while the first and sixth authors were at Arizona State University. The authors gratefully acknowledge ENEA for facilitating collaboration between Italian and US investigators. Appendix 1. EnviNNet The input vector X combines and transfers input data to the neural network, with appropriate time lags, as follows:


 X ¼ PM10t1 ; Ttn ; Ptn ; ISDðDWÞtm ; ISDðVÞtm ; ISDðFÞt with i ¼ [1, 2, 3], n ¼ [0, 1, 2], m ¼ [0, 1], where PM10 is the concentration level at times t ¼ 1, t ¼ 2 and t ¼ 3 (t is given in hours); T e the temperature at current time t (of prediction); P e the atmospheric pressure at t; ISD(V) e the vector of statistical and descriptive indicators of wind velocity V; ISD(DW) e the vector of the statistical and descriptive indicators of wind direction DW; ISD(F) e the vector of statistical and descriptive indicators of the ‘day’ with F the vector of the day time bands. The predicted concentration Ct of PM10 at time t by the MLP three-hidden node NN is given by a formula that represent the effects of non-linear operators on the inputs,

Ct ¼ expðC0 þ D1 *H1 þ D2 *H2 þ D3 *H3 Þ where C0, D1, D2, D3 are the coefficients estimated in the learning process of the NN and H1, H2, H3 are the three-hidden nodes. The three-hidden nodes are calculated by the following formula:

  X X X Hi ¼ tanh b0;1 þ bi;n *Cn þ bi;j *Mj þ bi;v *ISDv where i ¼ [1,2,3], n ¼ [1,2,3], j ¼ [1, 2] and v ¼ [1,2,3]; Mj ¼ {Tt, Pt} is the meteorological vector made up of temperature (T) and pressure (P) at time t ¼ k with k ¼ 0, 1, 2; Cn ¼ {Ctem} e the concentration vector made up of three elements, with m ¼ 1, 2, 3 selected according to the dispersion scenario of the site concerned; ISDv ¼ {ISD(V)t, ISD(DW)t, ISD(F)t} e the vector of statistical and descriptive indicators made up of three statistical and descriptive indicator vectors, namely: ISD(V)t e indicators of wind velocity (V) at time t ¼ tep with p ¼ 0, 1; ISD(DW)t e indicators of wind direction (DW) at time t ¼ tep with p ¼ 0, 1; and ISD(F)t e indicators of working days/holidays at time t. Appendix 2. Performance measures The following indicators were used for performance evaluation (EPA, 2007; ICF, 2009). Here P is the predicted value, O e the measured value and P; O e the mean values.

MAE ¼

N 1 X jPi  Oi jðMean Absolute ErrorÞ N i¼1

H.J.S. Fernando et al. / Environmental Pollution 163 (2012) 62e67

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N uX u ðPi  Oi Þ2 u ti ¼ 1

RMSE ¼ " R2 ¼

N N X

ðRoot Mean Square ErrorÞ

#2 ðPi -PÞðOi -OÞ

i¼1 N X

ðPi -PÞ2

i¼1

N X i¼1

N X

IA ¼ 1 

ðCorrelation CoefficientÞ ðOi -OÞ2

ðPi  Oi Þ2

i¼1 N X

ðIA Index of AgreementÞ

ðjPi  Oi j þ jOi  Oi jÞ2

i¼1

The following scores are used to verify a forecast against an observation of a binary event (yes or no), where F, H, and O, respectively, refer to the numbers of forecasted events, correctly forecasted events (hits), and observed events, while N is the total number of events in a verification domain.

POD ¼

H ðProbability of detectionÞ O

FAR ¼

F H ðProbability of false detectionÞ NO

CSI ¼

H ðCritical Success IndexÞ F þOH

References ADEQ Institutional report, 2005. Revised PM10 State Implementation Plan for Salt River Area, Appendices. Arizona Department of Environmental Quality Office, Phoenix, AZ. www.azdeq.gov/environ/air/plan/pm10.html. Boulle, A., Chandramohan, D., Weller, P., 2001. A case study of using artificial neural networks for classifying cause of death from verbal autopsy. International Journal of Epidemiology 30, 515e520. Byun, D.W., Ching, J.K.S., 1999. Science Algorithms the EPA Model-3 Community Multiscale Air Quality (CMAQ) Modeling System. EPA-600/R-99/030. U.S. Environmental Protection Agency, Office of Research and Development, Washington, DC. http://www.epa.gov/AMD/CMAQ/000_cover_exec.pdf. Carolina Environmental Program, SMOKE v2.2 User’s Manual, 2005. Carolina Environmental Program, Chapel Hill, NC. http://www.smoke-model.org/documentation. cfm. Chaloulakou, A., Assimakopoulos, D., Lekkas, T., 1999. Forecasting daily maximum ozone concentrations in the Athens basin. Environmenta Monitoring and Assessment 56, 97e112. Chaloulakou, A., Saisana, M., Spyrellis, N., 2003. Comparative assessment of neural networks and regression models for forecasting summertime ozone in Athens. The Science of the Total Environment 313, 1e13. Choi, Y.-J., Fernando, H.J.S., 2008. Implementation of a windblown dust Parameterization into MODELS-3/CMAQ: application to episodic PM events in the U.S./ Mexico border. Atmospheric Environment 42, 6039e6046. Choi, Y.-J., Hyde, P., Fernando, H.J.S., 2006. Modeling of episodic particular matter events using a 3-D air quality model with fine grid: applications to a pair of cities in the U.S./Mexico border. Atmospheric Environment 40, 5181e5201. Dimitrova, R., Hedquist, B.C., Hyde, P., Fernando, H.J.S., Anderson, J., 2009. Air quality modeling in support of children’s health studies. In: Proceeded in the 89th American Meteorological Society Annual Meeting, Phoenix, Arizona, J20.1. http://ams.confex.com/ams/pdfpapers/145908.pdf.

67

Dimitrova, R., Lurponglukana, N., Hyde, P., Fernando, H.J.S., Runger, G., Hedquist, B., Anderson, J., Bannister, W., Johnson, W., 2011. Relationship between Particulate Matter and Childhood Asthma e basis of a future warning system for Central Phoenix. Atmospheric Chemistry and Physics Discussion 11, 1e35. doi:10.5194/ acpd-11-1-2011. Dorling, S.R., Foxall, R.J., Mandic, D.P., Cawley, G.C., 2003. Maximum likelihood cost function for neural network models of air quality data. Atmospheric Environment 37, 3435e3443. EPA, 2001. Technical support document for approval and promulgation of implementation plan; Arizona e Maricopa nonattainment area; PM10, notice of final rulemaking. www.epa.gov/region9/air/phoenixpm/pdf/tsd.pdf. EPA, 2007. Guidance on the Use of Models and Other Analyses for Demonstrating Attainment of Air Quality Goals for Ozone, PM2.5 and Regional Haze. EPA-454/B07e002. Office of Air Quality Planning and Standards Emissions Analysis and Monitoring Division, Air Quality Modelling Group. http://www.epa.gov/ scram001/guidance/guide/final-03-pm-rh-guidance.pdf. Fernando, H.J.S., Dimitrova, R., Runger, G., Lurponglukana, N., Hyde, P., Hedquist, B., Anderson, J., 2009. Children’s Health Project: Linking Asthma to PM10 in Central Phoenix e a report to the Arizona Department of Environmental Quality. http:// www.azdeq.gov/ceh/download/Health%20Project%20Report.pdf. Frank, R.J., Davey, N., Hunt, S.P., 2001. Time series prediction and neural networks. Journal of Intelligent & Robotic Systems 31, 91e103. Gardner, M.W., Dorling, S.R., 1998. Artificial neural networks (the multilayer perceptron) e a review of applications in the atmospheric sciences. Atmospheric Environment 32, 2627e2636. Grell, G.A., Dudia, D.R.S., 1994. A Description of the Fifth Generation Penn State/NCAR Mesoscale Model (MM5). Boulder, Colorado: National Center for Atmospheric Research, NCAR/TN-398þSTR.20. ICF International report: Evaluation of CMAQ model performance for the 812 prospective II study. Memorandum prepared for EPA office of Policy Analysis and Review (OPAR), ICF International, San Rafael, CA, 2009. www.epa.gov/air/sect812/jan10/812_AQM_mpe_memo.pdf. Grivas, G., Chaloulakou, A., 2006. Artificial neural network models for prediction of PM10 hourly concentrations, in the Greater Area of Athens, Greece. Atmospheric Environment 40, 1216e1229. ICF International, 2009. Evaluation of CMAQ Model Performance for the 812 Prospective II Study. Memorandum prepared for EPA office of Policy Analysis and Review (OPAR). www.epa.gov/air/sect812/jan10/812_AQM_mpe_memo.pdf. Kolehmainen, M., Martikainen, H., Ruuskanen, J., 2001. Neural networks and periodic components used in air quality forecasting. Atmospheric Environment 35, 815e825. Kukkonen, J., Partanen, L., Karppinen, A., Ruuskanen, J., Junninen, H., Kolehmainen, M., Niska, H., Dorling, S., Chatterton, T., Foxall, R., Cawley, G., 2003. Extensive evaluation of neural network models for the prediction of NO2 and PM10 concentrations, compared with a deterministic modeling system and measurements in central Helsinki. Atmospheric Environment 37, 4539e4550. Mammarella, M.C., Grandoni, G., Pasquale, F., Di Marco, R.A., Sanarico, M., 2005. Neural Networks For Predicting And Monitoring Urban Air Pollution: The Intelligent Automatic System A.T.M.O.S.FE.R.A. ÓÒEcosystem Conference. Accademia dei Lincei, Rome. Maricopa County Air Quality Department (MC-AQD) e Air monitoring network review., 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009. www.maricopa. gov/aq/divisions/monitoring/network.aspx. Mayora-Ibarra, O., Motta, L., Curatelli, F., 1998. Forecasting of atmospheric pollution levels with neural networks. In: Proceedings of IV International Conference on Neural Networks and theirs Applications NEURAP’98, Marseille, France. Niska, H., Hiltunen, T., Karppinen, A., Ruuskanen, J., Kolehmainen, M., 2004. Evolving the neural network model for forecasting air pollution time series. Engineering Applications of Artificial Intelligence 17, 159e167. Shaltaf, S.J., Mohammad, A.A., 2009. Neural networks based time-delay estimation using DCT coefficients. American Journal of Applied Sciences 6 (4), 703e708. Venables, K.M., Allitt, U., Collier, C.G., Emberlin, J., Grieg, J.B., Hardaker, P.J., Highham, J.H., Laing-Morton, T., Maynard, R.L., Murray, V., Strachan, D., Tee, R.D., 1997. Thunderstorm-related asthma - the epidemic of 24/25 June 1994. Clinical and Experimental Allergy 27, 725e736. Vidyasagar, M., 1998. Are analog neural networks better than binary neural networks? Circuits System Signal Processing 17 (2), 243e270. doi:10.1007/ BF01202855. Western Governors’ Association (WGA), 2006. Final Report for the Western Regional Air Partnership (WRAP) 2002 Visibility Model Performance Evaluation, WGA 30203. Wilks, D.S., 1995. Statistical Methods in the Atmospheric Sciences. Academic Press, San Diego. Zhou, B., Du, J., 2010. Fog prediction from a multimodel mesoscale ensemble prediction system. American Weather and Forecasting 25, 303e322. http:// www.emc.ncep.noaa.gov/mmb/SREF/2222289_WAF_Feb-2010.official.PDF.