Environmental Science and Pollution Research An hourly PM10 diagnosis model for the Bilbao metropolitan area using a linear regression methodology --Manuscript Draft-Manuscript Number:
ESPR-D-12-01028R1
Full Title:
An hourly PM10 diagnosis model for the Bilbao metropolitan area using a linear regression methodology
Article Type:
Research Article
Corresponding Author:
iratxe gonzalez-aparicio, MSc Engineer TECNALIA Derio, Vizcaya SPAIN
Corresponding Author Secondary Information: Corresponding Author's Institution:
TECNALIA
Corresponding Author's Secondary Institution: First Author:
iratxe gonzalez-aparicio, MSc Engineer
First Author Secondary Information: Order of Authors:
iratxe gonzalez-aparicio, MSc Engineer Julia Hidalgo, PhD Alexander Baklanov, Proff. Ales Padro, Eng Oscar Santa-Coloma, Eng
Order of Authors Secondary Information: Abstract:
There is extensive evidence of the negative impacts on health linked to the rise of the regional background of PM10 levels. These levels are often increased over urban areas becoming one of the main air pollution concerns. This is the case on the Bilbao metropolitan area, Spain. This study describes a data-driven model to diagnose PM10 levels in Bilbao at hourly intervals. The model is built with a training period of sevenyear historical data covering different urban environments (inland, city centre and coastal sites). The explanatory variables are quantitative - log [NO2], temperature, short-wave incoming radiation, wind speed and direction, specific humidity, hour and vehicle intensity - and qualitative - working days/weekends, season (winter/summer), the hour (from 00 to 23 UTC) and precipitation/no precipitation. Three different linear regression models are compared: simple linear regression (LIN); linear regression with interaction terms (INT) and linear regression with interaction terms following the Sawa's Bayesian Information Criteria (INT-BIC). Each type of model is calculated selecting two different periods: the training (it consists on 6 years) and the testing dataset (it consists on 1 year). The results of each type of model shows that the INTBIC based model (R2 = 0.42) is the best. Results were R of 0.65, 0.63 and 0.60 for the city centre, inland and coastal sites, respectively, a level of confidence similar to the state-of-the art in the methodology. The error related calculated for longer time intervals (monthly or seasonal means) diminished significantly (R of 0.75-0.80 for monthly means and R of 0.80 to 0.98 at seasonally means) with respect to shorter periods
Response to Reviewers:
It is a document attached
Powered by Editorial Manager® and Preprint Manager® from Aries Systems Corporation
Manuscript Click here to download Manuscript: GonzalezAparicioetal.doc Click here to view linked References
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
An hourly PM10 diagnosis model for the Bilbao metropolitan area using a linear regression methodology González-Aparicio I.1, Hidalgo J.2, Baklanov A.3, Padró A.1 and Santa-Coloma O.1 (1)
Energy and Environmental Division, TECNALIA Research and Innovation, Edificio 700 C/ Geldo 48160, Derio, Spain (2) GAME/CNRM, Météo-France & CNRS, 42 Av G. Coriolis 31057 Toulouse, CEDEX, France (3) Research Department, Danish Meteorological Institute, DMI, Lynbyvej 100, DK-2100 Copenhagen Ø, Denmark
(Corresponding Author): Iratxe González-Aparicio. Tel:+ 34 667178995. Fax: +34 946 460 900. E-mail:
[email protected] Postal Address: C/Geldo - Parque Tecnológico de Bizkaia. Ed. 700 48160 Derio (Bizkaia)
Abstract There is extensive evidence of the negative impacts on health linked to the rise of the regional background of PM10 levels. These levels are often increased over urban areas becoming one of the main air pollution concerns. This is the case on the Bilbao metropolitan area, Spain. This study describes a data-driven model to diagnose PM10 levels in Bilbao at hourly intervals. The model is built with a training period of seven-year historical data covering different urban environments (inland, city centre and coastal sites). The explanatory variables are quantitative – log [NO2], temperature, short-wave incoming radiation, wind speed and direction, specific humidity, hour and vehicle intensity - and qualitative – working days/weekends, season (winter/summer), the hour (from 00 to 23 UTC) and precipitation/no precipitation. Three different linear regression models are compared: simple linear regression (LIN); linear regression with interaction terms (INT) and linear regression with interaction terms following the Sawa‟s Bayesian Information Criteria (INT-BIC). Each type of model is calculated selecting two different periods: the training (it consists on 6 years) and the testing dataset (it consists on 1 year). The results of each type of model shows that the INT-BIC based model (R2 = 0.42) is the best. Results were R of 0.65, 0.63 and 0.60 for the city centre, inland and coastal sites, respectively, a level of confidence similar to the state-of-the art in the methodology. The error related calculated for longer time intervals (monthly or seasonal means) diminished significantly (R of 0.75-0.80 for monthly means and R of 0.80 to 0.98 at seasonally means) with respect to shorter periods. Keywords: Diagnosis model, air quality, PM10, anthropogenic aerosols 1. Introduction Anthropogenic aerosols are one of the main air pollution concerns in the Bilbao metropolitan area. Ample endowment with natural resources and an appropriate geographical location favoured a significant industrialization of Bilbao since 1800‟s. The city runs along a 16 km long estuary aligned in a southeast-northwest direction. Two mountain ranges lie parallel to the waterway: the highest with 700 m height and the other 300 m (Figure 1). Based on a 2009analysis of the Basque Statistical Institute (EUSTAT, 2011), the population density is 3732.1 inhabitants km-2 over an area of 360 km2. The main local sources of air pollution are traffic and industrial activities: oil and coal-fired thermal generating stations, iron and steel works, a large refinery, sulphuric and nitric acid manufacturing plants, fertilisers (Inza-Aguirre 2010; Gangoiti et al. 2002, Millán et al. 1987; Zorraquino 1970; Allende-Landa 1982). The most remarkable air quality related-incidents took place in 1968, when sulphate emissions reached 2000 µg m-3. The frequent population strikes and demonstrations ended with casualties (Zorraquino 1970; Zorraquino 1971; Allende-Landa 1982). The legal gaps of air quality control were addressed by the 38/1972 Act and 833/1975 Decree on Atmospheric Environment Protection. The regional authorities led the first efforts to implement air quality management policies creating the Department of Environment (DoE) of the Basque Country through the Basque Air Quality
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Surveillance and Control Network1 (hereinafter, AQSN). The DoE and the Environmental Management Public Society (Ihobe)2 are in charge of developing the emission inventories since the 90‟s (heavy metals and VOC‟s (Labein 1995-2003); traffic emissions (Labein 1993, Labein 1996); Green House Gases (Labein 1996-2010; Tecnalia and Limia-Martin 2011)) and the particulate matter, acidifying and ozone troposphere precursors inventories of the Basque Autonomous Region (Labein 2004-2008; Tecnalia and Limia-Martin 2011). The initiatives undertaken during the last three decades were translated into remarkable improvements of the air quality indicators throughout the Bilbao metropolitan area. In this period, SO2 emissions were reduced by 95% up to daily means around 15 µg m-3 (AQSN, 2011). However, concentrations of particulate matter (PM) did not follow this tendency and, for example, PM10 levels registered in 2003 exceeded the daily means limit values of 50 µg m-3 (the legal limit fixed by the 1999/30/EC directive (updated as 2008/50/EC)) more than 35 times per year. In 2006, the Air Quality Action Plan (Labein 2006) was developed on the emissions recorded in 2004. The plan was updated in 2009 showing a better compliance of pollutant thresholds (Labein 2009). Additionally, some studies have established the PM10 regional sources over the area (Viana et al. 2006 and Inza-Aguirre 2010) and identified the types of events influencing PM levels in Bilbao as: (1) a reduction in the concentration due to intensive Atlantic advection episodes associated with the atmospheric scavenging of PM10 by precipitation; (2) a PM10 level increase due to Southern African dust transport (Inza-Aguirre 2010; Viana et al. 2003; Viana et al. 2006) and (3) high pressure systems over Northern Europe or low pressure systems over south-eastern Europe which produce episodes of high stability in the North of the Iberian Peninsula leading to higher PM10 levels. Viana et al. (2003) analysed the main urban sources of PM10 in Bilbao city. Up to 70% of the total PM10 is related to traffic and industrial activities (Zn/Pb/steel works). The rest is considered as regional sulphate, marine and crustal (Querol et al. 2004; Viana et al. 2006 and Inza-Aguirre (2010)). Consequences of high PM10 concentrations on human health have been pointed out in different studies of particular cases (California, Torrelavega (North of Spain), Bordeaux, Marseille, Paris, etc.) which conclude that brief episodes rising the regional background concentration (typically for Bilbao 20-30 µg m-3) are more damaging than lower daily means (Moreno et al. 2009; Delfino et al. 2002 and Lanki et al. 2008). Research efforts are being carried out in order to predict temporal disaggregated air quality levels in the Bilbao metropolitan area. For example, the existing prediction tools focus on the short term are to the 48h-forecasting of the SO2, CO, NO2, NO and O3 pollutants (Agirre-Basurko et al. 2006, Ibarra-Berastegi et al. 2008, IbarraBerastegi et al. 2001a, Ibarra-Berastegi et al. 2001b). However, there are no methodologies to predict the PM10 pollutant and then, to predict PM10 in a long-term basis. The objective of the study is to obtain a diagnosis PM10 data-driven model for the Bilbao metropolitan area that depends on climatic and urban parameters. This model is intended to be used in a near-future to study PM10 levels based on future climatic evolution and urban growth scenarios for Bilbao to evaluate the climatic and air quality shifts under a context of climate change. There is a large series of methods available in the literature (e.g. Neural Networks for gaseous pollutants (Voukantsis et al. 2011; Ibarra-Berastegi et al. 2008; Agirre et al. 2006); Models for gaseous and aerosols: Multiple Linear Regression (Henry and Hidy 1979); Principal Component Regression (Abdul-Wahab et al. 2003); Independent Component Regression (Pires et al. 2008) and finally Cluster-wise methods (such us Poggi and Portier 2011)). Here three main considerations drove the choice of the method: a method which not requires expensive computational resources is needed to be used in long-term simulations; not built with black 1
http://www.ingurumena.ejgv.euskadi.net/r493614/es/contenidos/informacion/red_calida_aire_capv/es_975/red.html 2 http://www.ihobe.net/
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
boxes to modify the independent variables for the future projections generation (as is the case for Neuronal Networks behaviour) and proved robustness in the atmospheric sciences. In this sense Linear Regression Models, used extensively in air quality studies, is the statistical method chosen to obtain the PM10 diagnostic model for Bilbao. The article is structured as follows: section 2 introduces the historical records and its analysis to construct the PM10 model; section 3 discussed the consistency and validation of the model within a new period. The main conclusions are discussed in section 4. 2. Data and Methods 2.1 Observations Hourly observations from 2005 to 2011 of PM10 and NO2 concentrations, near surface air temperature, specific and relative humidity, surface pressure, wind speed and direction, short wave incoming radiation, precipitation and traffic intensity have been used as initial variables for the analysis. Hourly PM10 and NO2 concentrations are obtained from the three monitoring stations (Figure 1) selected from the air quality network of the DoE which cover a heterogeneous range of urban environments defined for Bilbao in González-Aparicio et al. (2010). (A) COASTAL-Residential (Getxo: 43º 21‟ 22”N; 03º 00‟ 52”W), a sub-urban area close to the coast surrounded by a residential low density building district. (B) URBAN-Centre (Erandio: 43º 17' 28" N; 2º 56' 53”W) at the city centre, close to the river. It is characterized by high density of buildings with moderate density of traffic and industrial activities. (C) URBAN-Inland (Basauri: 43º 14‟ 10”N; 02º 53‟ 43”W) in the outer city, placed at the end of the valley in a narrow channel intersecting with parallel valleys. It is characterized as a high density of traffic and industrial activities. Finally, a RURAL station (Mundaka: 43º 23' 57”N; 02º 49'W) in coastal surroundings at a distance of 39 km from the city centre is considered as reference for PM10 background levels. Original time series are 99% complete. A reconstruction has been carried out through a multivariate linear regression and empirical orthogonal function analysis following Beckers and Rixen (2003). In order to keep only the urban peaks of PM10 concentration regional influences (mainly the Saharan intrusions) have been removed from the data following the methodology developed by Escudero et al. (2007). The Mundaka station is defined as a regional background site in order to identify the Saharan episodes. Those stations are equipped with automatic and continuous equipments which measure the PM10 through beta radiation. They are not reference equipments for measuring the PM10 in filters (under the EN 12341 directive), so a correction factor (CA), defined by the Royal Decree 102/2011 as the ratio between the beta radiation and filter radiation, is applied. CA for each station is 1.35, 1.0, 0.88 and 1.08 for inland, centre, coastal and rural stations, respectively. Hourly meteorological variables, which have been provided by the Basque meteorological agency (EUSKALMET), have been extracted from the closest mast (< 1km) to each air quality station. The traffic data have been provided by the Biscay Autonomous Council (DFB) extracted from the same streets where the stations are located (see Figure 1). The original traffic data are given as daily mean intensities per year. A temporal disaggregation into hourly mean intensity is performed using the methodology explained in the Basque Traffic Evolution Document (2010). 2.2 Historical 2005 to 2011 PM10 records analysis In a yearly basis, it is observed that there are two main periods where the PM10 levels recorded at the different stations are more likely to exceed the legal limits: winter (from December to February, hereinafter DJF) and summer (from June to August, JJA in the following). During spring (March, April and May) and autumn (September, October and November) the profiles of PM10 levels are smoothed, with daily peaks below 25 µg m-3. This result is in accordance with
3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
others studies over the same area (Inza-Aguirre 2010 and Labein 2009). At the city centre station, for winter time, the seasonal mean PM10 concentrations range between 20 µg m-3 in 2010 and 39 µg m-3 in 2006. It is observed a negative tendency from 2008 to 2010, where the mean seasonal concentrations decrease from 37 µg m-3 to 24 µg m-3 and a slight increase of 10 % - in 2011 with respect to 2010. During summer time, concentrations ranged from 33.5 µg m-3 in 2006 to 22.6 µg m-3 in 2009. The same trend is found in the other stations. The reduction is probably associated with a decrease in traffic intensity from 2008. Analysing together winter and summer time, the year with more concentrations above legal limits is 2006, with 31, 45 and 22 days with averaged daily means of 82, 137 and 91 µg m-3 at inland, centre and coastal stations, respectively. On the contrary, the legal limits are respected during all summer 2009 and all winter 2010. A monthly analysis (for both winter and summer seasons) shows low levels of PM10 during August, with the exception of the coastal site (characterized by crowded beaches), and during the end of the year. This monthly trend is also widely observed in other cities (Querol et al. 2004 and Voukantis et al. 2011). 7-day seasonal mean week variability is also reported and shows a differentiation between working days, Saturdays and Sundays as usual reported in the literature (Kapprara et al. 2003, Hörmann et al. 2005, Gong et al. 2007, Yusof et al. 2008, Basque Traffic Evolution Document (2010). At the inland site, during working days, the seasonal mean PM10 levels ranges between 33-36 µg m-3 and 40-45 µg m-3 for summer and winter, respectively. At the centre site, for working days the seasonal mean PM10 range between 26-28 µg m-3 and 35-38 µg m-3. For Saturdays there is a reduction of 20% which increase again up to 35% on Sundays. At the coastal site the mean is similar for working days and Saturdays (23-26 and 20-27 µg m-3 for summer and winter) but there is a 10% reduction on Sundays. For winter time, the distribution of the PM10 concentration is 58% and 42% for working days and weekends, respectively. During working days, 52% is associated to peak hours (5-9 and 17-20 LT) and 48% to the remaining hours. For weekends, the distribution is practically homogeneous. For summer time, pollutants are diluted in the deep mixing layer present at the afternoon and only one maximum is visible during the morning hours (6-9 Local Time, hereinafter, LT). The distribution of the PM10 concentration is 61% during morning hours, and 39% during the rest of the day. The distribution is different as well on working days and weekends, 56% and 44% respectively. During weekends, the PM10 contribution at peak hours is almost evenly distributed during daytime. On a daily basis, taking the centre station as a representative example, the daily means (hereinafter, reference means) obtained are 31 and 26 µg m-3 for winter and summer, respectively. The daily mean 90th percentile (i.e. the score below which 90 percent of the observations are found, associating to a threshold of maximum PM10 levels) correspond to 66 and 48 µg m-3 (hereinafter, reference 90th percentile). However, pollutant concentrations varied greatly during the day, with peaks far above the legal 50 µg m-3. In winter, the PM10 peaks reach 118-238, 111-530 and 136-344 µg m-3 at inland, centre and coastal sites, respectively. During summer, PM10 levels reach 129-320, 139-242 and 183-397 µg m-3. In general, the types of days with high peaks of PM10 were the days which exceeded the reference mean and the reference 90th percentile. That can be observed, for example, during two of the most recent episodes of high levels, one in winter time, on 20th -29th of January 2010 (episode I) and other in summer time on 13th -18th July 2009 (episode II). The highest hourly peaks are reached on 23 rd and 26th of January (from 60 to 120 µg m-3) and on 15th and 16th of July (60-161 µg m-3). During those days the daily average exceeds in 50% the 2010-winter means (22 µg m-3) and in 38% the 2009-summer means (20 µg m-3). Moreover, it exceeds by 15% the reference average and by 20% (winter) and 10% (summer) the reference 90th percentile for the period 2005-2011. Local meteorological parameters also influence the PM10 concentrations. The effect of the temperature is greater in days with weak winds, when the atmospheric circulation is dominated
4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
by land-to-sea breezes, with clear skies and thermal inversion that favours stagnation of the pollutants within the mixing boundary layer. For Bilbao, the strongest dependency was found between hourly PM10 levels and the wind direction patterns. Having in consideration the wind roses of the Basque country region (see Figure 2- extracted from the Spanish Meteorological Agency - AEMET3) it is observed that the main wind direction comes from the south-west and North, north-west. If the wind direction analysis is bracketed into four lagged quarters (from 291º to 20º; from 21º to 111º; from 112º to 201º and from 201º to 290º); in general it is observed up to 30 µg m-3 of difference among the quarters. The greatest difference appears at the inland station, since pollutant dispersion is prevented by the narrow channel of the river in this area of the city. In this station, PM10 levels differences about 45 µg m-3 for southerly winds, 35 µg m-3 for westerlies and 30 µg m-3 for the other quarters. At the centre the difference between the four quarters is insignificant; the mean is between the 20-30 µg m-3. By the coast, it is observed a difference between southerly winds (with the mean of 30 µg m-3) and the rest quarters (the mean is 20 µg m-3). In addition, regarding the spatial distribution of the observed PM10 levels, there is a difference between the tree stations due to the surrounding characteristics (meteorological, morphological and in traffic). The observed PM10 concentrations are 10% (in summer) and 18% (in winter) lower than at the urban station. The observed PM10 levels at the inland sub-urban site are 35% (in summer) and 20% (in winter) higher than in the city centre 2.3 Methodology The goal of this study is to build a data-driven model to diagnose the PM10 hourly levels as a function of meteorological fields and traffic information available from past observations but also potentially available from regional climatic scenarios and local traffic intensity scenarios. The resulting tool is intended to be used in a near-future within the framework of the KEGOKITZEN project4 in order to explore changes in PM10 concentrations under a context of climatic change. A linear regression (LR) has been applied to model the relationship between the explanatory variables (quantitative and qualitative) and a response variable (PM10 in this case). Three different methodologies have been explored (Robinson 2011): Multiple Linear Regressions (LIN): the response variable as a linear function of several explanatory variables independently between each other. Linear regression with interactions (INT); i.e. produces interaction variables that are uncorrelated with their component variables and with any lower-order interaction variables. Advantages of the method include clarity of tests of regression coefficients, and efficiency of winnowing out uninformative predictors. Linear stepwise regression with interaction terms following the Sawa‟s Bayesian Information Criteria (INT-BIC). For each of the independent variables, the F-statistic is calculated to determine each variable‟s contribution to the model; the stepwise regression evaluates all of the variables already included in the model and removes any variable that has an insignificant F. That estimates a measure of the difference between a given model and the “true” underlying model. The model with the smallest BIC among all competing models is deemed the best model. The BIC is a function of the number of observations n, the sum of square errors (SSE), the pure error variance fitting the full model (σ2), and the number of independent variables k ≤ p + 1 where k includes the intercept
3 4
http://www.aemet.es/es/portada http://info.labein.es/k-egokitzen/
5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Each type of model is built selecting two different periods: the training (it consists on 6 years) and the testing dataset (it consists on 1 year). This process has been carried out 7 times (each time considers one different year for the testing period) and finally it is selected the best model of the iterations. A good linear model has small root mean square error (RMSE) and a high adjusted coefficient of determination (R2) close to 1. These coefficients evaluate the degree of agreement between modelled vs. observed PM10 concentrations. R2 is the square of the correlation coefficient (R), where n is the number of observations, Xi and Yi are the observed and calculated values at each time step (i). and are the mean of the observations and calculated sample, respectively.
The best model gives the highest R2, with the lowest RMSE both in the training and testing periods. The Figure 3 presents the box-plots of the RMSE of each type of model at the urban station. It can be observed that comparing the testing and training period of each type of model, the LIN and INT-BIC based-models has the smallest difference between them. Thus, the INTbased model is ruled out. In addition, the INT-BIC based-model has lower RMSE both in the testing and training period than the LIN-based model. Moreover, the INT-BIC (R2 = 0.42) fits better than the INT (R2 = 0.36) and than the LIN- based model (R2 = 0.29). The selection of the explanatory variables (i.e. the input variables to build the model) is also a key issue under the LR approach since irrelevant or noisy variables have negative effects on the training process. In addition, these methods are limited due to the failure to capture the nonlinear behaviour of air pollutants, or the incomplete understanding of the physical and chemical processed involved. Thus, an analysis is carried out showing that the input variables selected are linear (most of them follow a normal distribution) and independent between each other (see Figure 4). Therefore, the input explanatory dataset (Table 1) is treated quantitatively and qualitatively depending on the observed patterns, and independently for each location (inland, centre and coastal). Categorical variables are defined as a function of the season (winter – December, January and February - or summer – June, July and August -), the type of day (week days and weekends), the hour (from 00 to 23 UTC), if there is (not) precipitation and the wind direction. The wind direction variable is transformed into lagged variables taking into account the four quarters presented in the previous section according to the wind rose from AEMET. The quantitative variables are based on the meteorological (temperature, wind speed, specific humidity and radiation), chemical (log [NO2] because it is closer to a normal distribution than the NO2) and traffic data. As a result, an algorithm is obtained with constant coefficients for the quantitative variables and different factors for each categorical variable and for each interaction for each site. As a representative example, the Equation (1) defines the model for the city centre. The Table 2 presents the coefficients of the equations; the “residuals” gives the difference between the experimental and predicted signals; the estimates for the model‟s coefficients are provided along with their standard deviations („Std Error‟) and a t-value and probability for a null hypothesis that the coefficients have values of zero. At the bottom of the table it is found
6
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
the standard deviation about the regression (residual standard error), the correlation coefficient and an F-test result on the null hypothesis. It must be noted that in the model selected (INT-BIC), if the hours of the day (HH) are included as input variables; the traffic data (TRAF) and the working days/non working days (WD) are excluded and vice-versa; the results are very similar in both cases. This is explained by the fact there is close dependence between the hour of the day, working days and traffic intensity. This gives an advantage of the method when the model is used for climatic scenarios because the future traffic data are often not available. Additional sensitivity tests are performed directly with the INT-BIC model: it has been demonstrated that the inclusion of other variables: i.e. the relative humidity, pressure, month and year, do not improve the results of the predictions. Equation 1:
Where „Cte‟ is the constant value in the equation 3. Results and discussions Validation of the model performance: The INT-BIC model suitability is firstly visually examined analysing the residuals. If the model is appropriate, then the residual errors are random and normally distributed. In addition, removing one case should not be significantly impact the model‟s suitability. The Figure 5, in the upper left, shows the residual errors plotted versus their fitted values. The residuals are randomly distributed around the horizontal line representing a residual error of zero; that is, there is not a distinct trend in the distribution of points (Robinson 2011). The Figure 5 in the lower left is the standard Q-Q plot, which is a graphical method for comparing two probability distributions by plotting their quantiles against each other. It can be observed that data from the distribution follow the normal curve (y=x) fairly closely until the second theoretical quantile that is light tile, that means most of the observations piles near the upper boundary, being skewed to the right. These points are the high PM10 observed levels which are not able to reproduce by the model. It counts for less than the 15% of the total amount of observations, thus, it can be said that it is normally distributed. The scale-location of the Figure 5b plot in the upper right shows the square root of the standardized residuals (sort of a square root of relative error) as a function of the fitted values. Again, there is no obvious trend in this plot, fulfilling the condition to be an appropriate model. Finally, the Figure 5d plot in the lower right presents each points leverage, which is a measure of its importance in determining the regression results. Distances larger than 1 are suspicious and suggest the presence of a possible outlier or a poor model, which is not the case here. Further, the reproducibility of the model is tested. On an hourly basis, the correlations for each algorithm are 0.62, 0.65 and 0.60 at the coast, centre and inland locations, respectively. The best correlations are found in the cases and at the locations in which the PM10 levels are higher (see Table 3). For example, the best correlation are found at the city centre, on working days and in winter, when there is no precipitation and the wind flows from the south west. The highest discrepancies produced by the model occur in the hours when the PM10 levels observed between consecutive hours are very high (i.e. short peaks with an increase of up to 200% differences). The Figure 6 shows (first quadrant of the scatter plot) isolated observed concentrations greater that 225 µg m-3 at the urban station which are underestimated by the model. Similar results are obtained by other studies as well, and the discrepancies are often
7
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
explained by unexpected traffic congestions that the disaggregation method of the vehicle intensity, from annual to hourly values, is not able to reproduce. Similar performance has been obtained by Yusof et al. 2008 that calculated hourly PM10 predicting equations with a correlation coefficient (R) of 0.60 for Seberang Perai (Penang). Voukantis et al. (2011) obtained R of 0.65-0.80 at different stations in Thessaloniki, (Greece) of an hourly based frequency model. At a daily basis, Pires et al. (2008) obtained in the application of several PM10 forecasting methodologies R of 0.60-0.79 in Oporto (Portugal) and Hörmann et al. (2005) obtained an R 0.81 of the regression equation for Graz (Austria). Reproducibility of the physical mechanisms: For both periods, the model performance shows a dependency on the location and on the meteorological conditions. The models fit better winter than summer observations, and at the centre and inland stations (see Table 3). In winter the mean concentration levels are higher due to the increase of anthropogenic emissions and the atmospheric conditions favouring the accumulation of pollutants around the emission sources: development of intensive thermal inversions, lower convective dynamics and low dispersive local circulations. At the centre station, the wind speed is reduced due to the roughness and the effect of the city. This reduction leads to more stable and predictable dispersion patterns, compared to the stations out of the city. At the inland station, in the narrowest part of the valley surrounded by the higher mountains (600 m height), the thermal inversions are more intense and the stagnation leads to the highest mean of PM10. The highest discrepancies are found at the coastal station, when the PM10 levels are the lowest. That is associated to the variability of wind conditions, which greatly affect pollutant dispersion. The gradient of the temperature and eddies formed due to the cliffs cause thermal and mechanical turbulences in the area being much more difficult to accurately predict the air quality levels. Diurnal cycle, daily and seasonal mean predictions: The diagnosed and observed PM10 mean distribution during the diurnal cycle follow the same behaviour at the three stations. As in the observations, the predictions are not homogenously distributed during the day: the concentrations are higher during the peak hours, showing a significant difference between working days and weekends (see Figure 7 as a representative example). Consequently, a future increase in the values of the seasonal means mainly would contribute to the peak hours during working days. As shown in the Figure 8, when means are calculated with longer time intervals, the error between the observations and calculations diminish. On a daily basis, the model fit improves by 20% while on a seasonal basis the improvement reaches 40%. For summer the R 2 are 0.81, 0.90 and 0.64 at the inland, centre and coastal site, respectively. For winter the R values are 0.72, 0.74 and 0.81. The mean seasonal errors are reduced to 0.1 µg m-3 and 0.2 µg m3 for summer and winter, respectively. Prediction of high-concentration events: This evaluation of the models‟ performance is of particular interest since successful and on-time predictions of episodic PM10 concentration can be of use in assuming activity for the protection of public health. It has been already reported that health effects are more linked with short term exposure, rather than with the 24-h averaged exposure (Gold et al. 2004). In addition, some Member States of the European Union have national legislation that requires 1-h time resolution (CAFE, 2004). Here, following Grivas and Chaloulakou (2006); Chaloulakou et al. (1999) and Wilks (1995) the performance measures used are the accuracy (Ac – percent of forecast that correctly predicted the event or no-event of high concentration levels), the probability of detection (POD – percentage of successful forecast over total high concentration events), the false alarm rate (FAR- fraction of false forecasts over total forecasts, the critical success index) and the CSI-ratio of correct forecasts over total high concentration events). The measures are applied to the entire period analysed for both seasons. Since the 1-h limit value are not legislated, the characterization of an hourly PM10 concentration as high, in the present study will be made based on the criteria selected by Grivas and Chaloulakou (2006) for each station with similar environmental characteristics: the limit is made in the case of being 2
8
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
standard deviation levels above the average measured concentration levels in each site. As a result, using the table 3, at the urban site the PM10 limit is 50 µg m-3; 65 µg m-3 at the inland and 43 µg m-3 at the coastal site. The Table 4 presents the measures with the criterion followed to assess the model. For example, the model developed for the urban site correctly forecasts the 65% of the events or non-events but the 65% of the forecasts for high events are not materialized. However, almost the 25% of the episodes are correctly predicted. Extrapolating to the rest of the sites the behaviour of each measure is in the same order. The quality of the predictions of the models developed regarding episodic concentration levels is undermined by their inability to correctly spot some distinct very high concentrations events. This verifies the general assumption explained by the unexpected traffic congestions at some locations. 4. Conclusions Anthropogenic aerosols are one of the main air pollution concerns in the Bilbao metropolitan area, Northern coast of Spain. Although different measures and policies have been implemented during the last three decades with a considerable improvement of the air quality, concentrations of particulate matter did not follow the same tendency. During the 2005-2011 period (at three different locations: inland, urban centre and coastal), the main contributions to higher daily mean concentrations were produced at peak hours during working days, reaching up to 530 µg m-3. They are mainly caused by high density of traffic (at 5-9 and 17-20 LT in winter and between 6-9 LT in summer). In general, the days with high peaks of PM10 are the days which exceed the reference mean and the reference 90th percentile. During the highest hourly peaks of the most recent episodes with high concentrations (on 13-18 Jul 2009 and 20-29 Jan 2010), PM10 levels reached up to 120 µg m-3 and 161 µg m-3, respectively. During those days the daily average exceeded by 38% the 2009-summer means (20 µg m-3) and by 50% the 2010-winter means (22 µg m-3). Quantitative (climatic and urban) and qualitative (temporal and spatial) variables have been used to build the PM10 diagnosis model. Three different types of linear models have been tested: simple linear regression techniques; linear regression with interaction terms and linear regression with interaction terms following the Sawa‟s Bayesian Information Criteria. Each type of model is calculated using training and a testing dataset from the period 2005-2011. The results of each type of model show that the linear regression with interactions following the BIC criterion is the best with a R2 of 0.42 and 0.40 for the training and testing period, respectively. At the three sites, the prediction is in better agreement for winter than for summer, corresponding to higher means of PM10. Consequently, a future increase in the seasonal means mainly would contribute during the peak hours of working days. Discrepancies between observations and model results occur in those hours when PM10 levels are very different from the consecutive hours. This is often associated with unexpected traffic congestions that the disaggregating method of the vehicle intensity, from annual to hourly values, is not able to reproduce. When means are calculated on the basis of longer time intervals, the error between observations and calculations diminish. On a daily basis, the model fit improves by 20%, and on a seasonal basis the improvement reaches 40%. For summer the correlation coefficients (R) are 0.90, 0.95 and 0.80 at inland, centre and coastal sites, respectively and for winter the R values are 0.85, 0.86 and 0.95. The quality of the predictions of the models developed regarding episodic concentration levels is undermined by their inability to correctly spot some distinct very high concentrations events. This verifies the general assumption explained by the unexpected traffic congestions at some locations. However, following Grivas and Chaloulakou (2006) and Kukkonen et al. (2003), there are approaches aiming at increasing the frequency of extreme values for the training period in order to be repeated more times for the improvement of the prediction. Acknowledgements
9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Financial support for this work is gratefully acknowledged to the K-EGOKITZEN project and TECNALIA Research & Innovation and – Iñaki Goenaga (FCT-IG) Technology Centres Foundation. Also, thanks to Drs: Alexander Mahura and Ole Christensen (DMI), Jon Saenz and Gabriel Ibarra-Berastegi (University of the Basque Country), Marivi Albizu (Basque Government), Juan-Angel Acero (Tecnalia) and Olivier Mestre (MeteoFrance) for the constructive discussions about the development of the methodology and the selection of the dataset. References Abdul-Wahab S, Bakheit C, Al-Alawi S. (2005) Principal component and multiple regression analysis in modelling of ground-level ozone and factors affecting its concentrations. Env. Modelling and Software 20:1263-1271. Acero Juan-Angel (2012) Urban Climate Modelling: development of urban climate evaluation methods for urban planning purposes. Dissertation, University of Kassel Allende Landa J (1982) Áreas metropolitanas y contaminación atmosférica. El caso del Gran Bilbao. Ciudad y Territorio. Revista de Ciencia Urbana Nº 52 Agirre-Basurko E, Ibarra-Berastegi G, Madariaga I (2006) Regression and multilayer perceptron-based models to forecast hourly O3 and NO2 levels in the Bilbao area. Environ. Model. & Soft, 21: 430-446 Biscay Country Council and Public Works Department “Basque Traffic Evolution Document 2010” (2010) ZURE S.A. 1348-2011. Basque Meteorological Agency (EUSKALMET). http://www.euskalmet.euskadi.net/s075853x/es/meteorologia/home.apl?e=5 Basque Statistical institute (EUSTAT) (2011) Beckers JM, Rixen M (2003) EOF calculations and data filling from incomplete oceanographic datasets. Atmospheric and Oceanic Technology 20: 1839–1856. Christensen JH, Christensen OB, Lopez P, van Meijgaard E, Botzet M (1996) The HIRHAM4 regional atmospheric climate model. Danish Meteorological Institute (DMI) Technical Report 96-4. Delfino RJ, Zeiger RS, Seltzer JM, Street DH, McLaren C (2002) Association of ashma symptoms with Peak particulate air pollution and effect modification by anti-inflamatory medication use. Environmental Health Perspectives 110:A607-A617. Diputación Foral de Bizkaia (DBF) (2010) Evolución del tráfico en las carreteras de Bizkaia. Biscay Autonomous Council & Department of Public Works Escudero M, Querol X, Alastuey A, Pérez N, Ferreira F, Cuevas E, Rodriguez S, Alonso S (2007) A methodology for the quantification of the net African dust load in air quality monitoring Networks. Atmos. Env. 41: 5516-5524 Gangoiti G, Alonso L, Navazo M, Albizuri A, Perez-Landa G, Matabuena M, Valdenebro V, Maruri M, García J, Millán M (2002) Regional transport of pollutants over the Bay of Biscay: analysis o fan ozone episode under a blocking anticyclone in west-central Europe. Atmos. Env. 36: 1349-1361 Gangoiti G, Albizuri A, Alonso L, Navazo M, Matabuena M, Valdenebro V, Garcia J, Millán M (2006) Sub-continental transport mechanisms and pathways Turing two ozone episodes in Northern Spain. Atmos. Chem. Phys 6:1469-1484 Gong D, Ho C, Chen D, Qian Y, Choi Y, Kim J (2007) Weekly cycle of aerosol-meteorology interaction over China. Jour. Of Geophys. Research Doi: 10.1029/2007JD008888
10
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
González-Aparicio I, Nuterman R, Korsholm U, Mahura A, Acero JA, Hidalgo J, Baklanov A (2010) Land-Use Database Processing Approach for Meso-Scale Urban NWP Model Initialization. Sci Report 10-02. Digital ISBN: 978-87-7478-593-4 Henry R, Hidy G (1979) Multivariate Analysis of Particulate Sulfate and other Air Quality variables by Principal Components – Part I. Annual data from Los Angeles and New York. Atmos. Env. 13: 1581-1596 Hörmann S, Pfeiler B, Stadlober E (2005) Analysis and Prediction of Particulate Matter PM10 for the Winter Season in Graz. Austrian Jour. Of Stats 34:307-326 Ibarra-Berastegi G, Elias A, Barona A, Saenz J, Ezcurra A, Argandoña J (2008) From diagnosis to prognosis for forecasting air pollution using neural Networks: Air pollution monitoring in Bilbao. Env. Modelling and Software 23:622-637 Ibarra-Berastegi G, Elias A, Agirre E, Uria J (2001) Long-term changes in ozone and traffic in Bilbao. Atmos. Environ. 35:5581-5592 Inza-Aguirre A, Albizu M, Querol X, Ortega L, Gil J (2004) Estudio del material particulado en áreas de fondo urbano con influencia de emisiones industriales (País Vasco). IX Congreso de Ingeniería Ambiental - Bilbao. Proma. Inza-Aguirre A (2010) Estudio de series temporales y composición química del material particulado atmosférico en distintas áreas del País Vasco. Dissertation University of the Basque Country Kaprara A, Karatzas K, Moussiopoulos N, Viras L (2003) Investigating week-end air quality observations with the air of Fourier analysis in Athens, Greece. Int. J. Environment and Pollution 19:171-176 Labein (2006) Elaboración de planes de acción municipales de calidad del aire en la CAPV según el RD 1073/2002. Labein (2009) Actualización de la elaboración de planes de acción municipales de calidad del aire en la CAPV según el RD 1073/2002. Labein (1995-2003) Inventario de Emisiones de Compuestos Orgánicos Volátiles de la Comunidad Autónoma del País Vasco. GOBIERNO VASCO (Departamento de Ordenación del Territorio, Vivienda y Medio Ambiente). Labein (1996-2010) Inventario de Gases de Efecto Invernader en la CAPV. Environmental Management Public Society, IHOBE. Labein (2004-2008) Inventario de emisiones de Partículas, Gases Acidificantes y Precursores del Ozono de la CAPV. Environmental Management Public Society, IHOBE. Lanki T, Hoek G, Timonen KL, Peters A, Tiitanen P, Vanninen E, Pekkanen J (2008) Hourly variation in the fine particle exposure is associated with transiently increase risk of ST segment depression. Occupational Environmental Medicine 65:782-786 Millán M, Otamendi E, Alonso L, Ureta I (1987) Experimental Characterization of Atmospheric Diffusion in Complex Terrain with Land-Sea Interactions. JACPA 37: 807-811 Millán M, Alonso L, Legarreta J, Albizu M, Ureta I, Egusquiaguirre C (1984) A Fumigation Episode in an Industrialized Estuary: Bilbao, November 1981. Atmos. Env. 18: 563-572 Moreno T, Lavín J, Querol X, Alastuey A, Viana M, Gibbons W (2009) Control son hourly variations in urban background air pollutant concentrations. Atm. Env 43:4178-4186 Nakicenovic N, Swart R (2000) Intergovenmental Panel of Climate Change Emission Scenarios. Cambridge University Press Pires J, Martins F, Sousa S, Alvim-Ferraz M, Pereira M (2008) Prediction of the Daily Mean PM10 Concentrations Using Linear Models. American Jour of Environ. Sci. 5:445-453
11
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Querol X, Alastuey A, Rodriguez S, Viana M, Artiñano B, Salvador P, Mantilla E, García S, Fernandez R, Rosa J, Sanchez A, Menéndez M, Gil J (2004) Levels of particulate matter in rural, urban and industrial sites in Spain. Sci. of the Total Env. 334-335:359-376 Andrew Robinson (2011) IcebreakeR, Department of Mathematics and Statistics. University of Melbourne Rodriguez S, Querol X, Alastuey A, Viana M, Alarcón M, Mantilla E, Ruiz C (2004) Comparative PM10-PM2.5 source contribution study at rural, urban and industrial sites Turing PM episodes in Eastern Spain. Sci. of Total Env 328:95-113 Tecnalia, Limia-Martin (2011) Inventario de Gases de Efecto Invernadero en la CAPV. Environmental Management Public Society, IHOBE. Tecnalia, Limia-Martin (2011) Inventario de emisiones de Partículas, Gases Acidificantes y Precursores del Ozono de la CAPV. Environmental Management Public Society, IHOBE. Viana M, Querol X, Alastuey A, Gil J, Menéndez M (2006) Identification of PM sources by Principal Component Analysis (PCA) coupled with wind direction data. Chemosphere 65: 2411-2418. Viana M, Querol X, Alastuey A, Gangoiti G, Menéndez M (2003) PM levels in the Basque Country (Northern Spain): analysis of a 5-year data record and interpretation of seasonal variations. Atmos. Env. 37:2879-2891 Voukantis D, Karatzas K, Kukkonen J, Räsänen T, Karppinen A, Kolehmainen M (2011) Intercomparison of air quality data using principal component analysis, and forecasting of PM10 and PM2.5 concentrations using artificial neural networks, in Thessaloniki and Helsinki. Sci. of the Total Env. 409:1266-1276 Wilmott CJ, Ackleson SG, Davis RE, Feddema JJ, Klink KM, Legates DR, O‟Donnell J, Rowe MC (1985) Statistics for the evaluation and comparison of models. Jo. of Geog. Res. 90:89959005 Yusof N, Ghazali N, Ramli N, Yahaya A, Sansuddin N, Madhoun W (2008) Correlation of PM10 concentration and Weather Parameters in Conjunction with Haze Event in Seberang Perai, Penanag ICCBT 20:211-220 Zorraquino J (1971) Datos y comentarios sobre la Contaminación del Aire en Erandio (Bilbao) 1969. DYNA 1 Zorraquino J (1970) Algunos datos y comentarios sobre la contaminación del aire en Erandio (Bilbao) en 1968. DYNA 1
12
Figure Captions: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Figure 1: Topography of the Bilbao metropolitan area and the air quality stations selected for the study. The a) station is the coastal station (Getxo); the centred station is Erandio (b) and c) is the inland station (Basauri). The rural station to detect the regional PM10 is d) Mundaka. Figure 2: Wind roses of the Basque Country Region. Extracted from the Spanish Meteorological Agency (AEMET) Figure 3: Box plots of the RMSE of the training (A.) and testing (T) period for the three models (LIN, INT, and INT-BIC) at the urban station. Figure 4: Linearity and Independency between the input explanatory quantitative dataset. Figure 5: Residuals plots to determine the suitability of the model performance: a) Residuals vs. fitted; b) scale-location; c) Normal Q-Q plot and d) residuals vs. leverage. Figure 6: Scatter plots at the urban site. The points falling inside the first quadrant correspond to the underestimation of the model and the fourth to the overestimation. The second and third quadrants mean that the model is able to reproduce high and low values of PM10 concentrations. Figure 7: Predicted and observed PM10 mean distribution during the diurnal cycle for both seasons winter and summer on working days (a), Saturdays (b) and Sundays (c). The figures show the diurnal cycle at the centre station. Figure 8: Comparison of the predicted and observed PM10 levels for summer (left side) and winter time (right side) at the centre station. The a) figures are based on hourly frequency during the two high polluted episodes in summer 2009 (13th-18th July) and in winter 2010 (20th-29th January). The b) figures correspond with daily means and the c) figures with the seasonal means during the six-year period.
13
Authors' Response to Reviewers' Comments Click here to download Authors' Response to Reviewers' Comments: ReviewersComments.doc
Ref.: Ms. No. ESPR-D-12-01028 An hourly PM10 diagnosis model for the Bilbao metropolitan area using a linear regression methodology Environmental Science and Pollution Research Answers To Reviewers Reviewer #2: The methodology is completely out of date. Nowadays, the neural networks and statiscal tools are merely complementary tools to the full comprehensive 4D Air Quality Models which provides a full description of the atmospheric dynamics using the Navier-Stokes partial differential equations. The main limitation of these techniques - even for climate purposes as it seems to be the final purpose of the authors - is the difficulties to explain the causes of the pollution. So that, for health studies is critical to have a detailed and accurate description of the formation of the pollution (PM in this case). So authors are encouraged to compare the results presented in this contribution with the simulations produced by comprehensive 4D air quality models (Eulerian approach) to understand the pollution process and apply the exposure techniques for health effects studies. In addition to this, three stations are not representative at all of the complexity of the pollution particle fields in urban environment. The urban environment is particularly complex due to the influence of many local factors and many more urban and non-urban stations should be taken into account unless a different approach is used (i.e., 4D air quality models). We know that there are a large panel of modelling tools and, before starting this study, we have extensively disccused and evaluated the most suitable methodology. We know well the 4D modelling techniques, personnaly I work with the Enviro-HIRLAM modelling system (Baklanov A, Korsholm U, Mahura A, Petersen C, Gross A (2008) ENVIRO-HIRLAM on-line coupled modelling of urban meteorology and air pollution, Adv. Sci. Res., 2,41-46). Another option that was also considered was to use the Meso-NH model from Meteo-France. Those models are a 4D online air quality and meteorological meso scale models which provides the full description of the atmospheric dynamics and composition at regional and urban scales;. For exemple we have decided to use this kind of tool to study the urban boundary layer of the Bilbao metropolitan (González-Aparicio I, Hidalgo J, Baklanov A, Korsholm U, Nuterman R, Mahura A and Santa-Coloma O, (2012) Urban Boundary Layer Analysis in the Complex Coastal Terrain of Bilbao using Enviro-HIRLAM Theoretical and Applied Climatology, DOI: 10.1007/s00704-012-0808-6). In the present study the objective was to obtain a tool that in a second step will allow us to evaluate the impact of future projections of climate on PM10 concentrations. The dynamic modelisation approach is, from our point of view, not realistic in terms of computational resources. Currently when we want to model long periods at the local scale we perform off-line simulations using the climate scenarios to force the surface module but this approach is not possible when using a chemical module. Statistical modelling allows us to obtain a diagnostic model that not require expensive computational resources and that can be evaluated at daily frequency that is enough for our future analysis based on the changes at seasonal frequency.
Reviewer #3: General comments:
The paper aims to compare the performance of three linear models in prediction of PM10 concentrations. However, the authors only presented the results of one model. Although authors present the figures and equation of one model; throuthought the manuscript, it is described the performance of the three models. However, the statistics of the three stations have been included in the Table 3. More consideration is given to the model for the urban station because the data extracted from the Regional Climate models (RCM) has a grid-cell resolution of 25x25 km and only one grid cover the area of Bilbao. In this study it is explained (in the page 5) that at the coastal station, the observed PM10 concentrations are 10% (in summer) and 18% (in winter) lower than at the urban station. The observed PM10 levels at the inland sub-urban site are 35% (in summer) and 20% (in winter) higher than in the city centre. The spatial variability can be evaluated using only the reference values obtained with model at the urban station and weighted as a function of the other model. Citations and references are not in agreement with journal guidelines. Both citations and references have been change according to journal guidelines
I think that the authors used multiple linear regression and not simple linear regression (definition on page 5). Yes, it was a mistyped. It is multiple linear regression How is Bayesian Information Criteria calculated? Here the BIC is a function of the number of observations n, the sum of square errors (SSE), the pure error variance fitting the full model (σ2), and the number of independent variables k ≤ p + 1 where k includes the intercept
Specific comments: Page 2: Change 1999/30/CE to 1999/30/EC (the same comment for 2008/50/CE). The regression models referred by the authors should be written as: Multiple Linear Regression, Principal Component Regression and Independent Component Regression. It has been changed in the text according reviewer‟s comments. Page 3: In line 44 - What is the text in brackets? There was no text in brackets, it was mistyped. Page 5: In Figure 2, the authors should translate the used words to English. The Figure 2 has been improved according to the reviewer‟s comments.
What was the reason to divide the wind direction in these quarters?
The wind direction was divided into the quarters, each quarter selected include the main wind patters of the area: i.e a wind direction from 45º and 359º are completely different in frequency and strength and must be treated separately: (1) from 291º to 20º includes: NE and E winds which are the less frequent and very calm (2) from 21º to 111º includes: SE and S winds which are more frequent and strong (3) from 112º to 201ºincludes SW and W winds which are frequent but calm (4) from 201º to 290º includes NW and N winds whic are very frequent and strong. Was the wind direction used as an explanatory variable in the models? The wind direction was used as explanatory variable being a “lagged variable”; i.e. including the four quarters selected as categorical variable. It had no sense to add the wind direction (in degrees) as explanatory variable because very far values between each other identify winds of the same direction e.g 0 º, 360º and 4º.
"This process has been carried out 7 times (each time considers one different year for the testing period)..." - Does this procedure have sense (predict the PM10 concentrations with models developed with data obtained after the test period)? It was misexplained in the text and It has been re-written . The process that have been carried out 7 times is the entire process for the training and testing period: for example, the 1 st iteration consists on training period (2005-2010) and testing period (2011); the 2nd iteration consists on training period (2006-2011) and the testing period 2005; the 3rd iteration consists on training period (2005, 2007-2011) and the testing period 2006, etc. Before: “Each type of model is built selecting two different periods: the training (it consists on 6 years) and the testing dataset (it consists on 1 year). This process has been carried out 7 times (each time considers one different year for the testing period) and finally it is selected the best model of the iterations.” Now: “Each type of model is built selecting two different periods: the training (it consists on 6 years) and the testing dataset (it consists on 1 year). This process has been carried out 7 times (each time it is considered one different year for the testing period and the rest of years for the training period) and finally it is selected the best model of the iterations” What is the meaning of "adjusted coefficient of determination"? The method of calculation of the R2 and the RMSE has been included in the manuscript, and the meaning of the adjusted coefficient of determination. “The adjusted coefficient of determination (R2) is used to describe how well a regression line fits a set of data it is the square of the correlation coefficient (R). Where n is the number of observations, Xi is the observations and Yi is the calculated value at each time step (i); and and
are the mean of the observations and calculated sample, respectively”
In the last sentences, Figure 3 was referred having the box-plots of the RMSE values for training and test periods. However, the models were only compared with coefficients of determination (the RMSE values should be discussed). Regarding the R2 values, which data was used to calculate them? Following the reviewer‟s suggestion, the RMSE values has been discussed in the text: Before: “A good linear model has small root mean square error (RMSE) and a high adjusted coefficient of determination (R2) close to 1. These coefficients evaluate the degree of agreement between modelled vs. observed PM10 concentrations. The best model gives the highest R 2, with the lowest RMSE both in the training and testing periods. The Figure 3 presents the box-plots of the RMSE of each type of model at the urban station. The INT-BIC (R2 = 0.42) fits better than the INT (R2 = 0.36) and than the LIN- based model (R2 = 0.29)” Now: “A good linear model has small root mean square error (RMSE) and a high adjusted coefficient of determination (R2) close to 1. These coefficients evaluate the degree of agreement between modelled vs. observed PM10 concentrations. The best model gives the highest R 2, with the lowest RMSE both in the training and testing periods. The Figure 3 presents the box-plots of the RMSE of each type of model at the urban station. It can be observed that comparing the testing and training period of each type of model, the LIN and INT-BIC based-models has the smallest difference between them. Thus, the INT-based model is ruled out. In addition, the INTBIC based-model has lower RMSE both in the testing and training period than the LIN-based model. Moreover, the INT-BIC (R2 = 0.42) fits better than the INT (R2 = 0.36) and than the LINbased model (R2 = 0.29)”. The data used to calculate the R2 values were the entire dataset of the observations and the calculated PM10 values Page 6:
Does Figure 4 present all data (7 years)? Are the wind speed data represented by integer numbers? The previous Figure 4 presented a random sample (5000 values per variable) of the total dataset to show the non-linearity between independent variables. However, it has been changed the Figure 4 including all the data set (30620 values per variable). The wind speed are float numbers, they are not integer numbers.
In Table 2, the commas should be replaced by points (this comment can also be applied to Table 3). The numbers should be presented without "E". The Table has been modified In Equation 1, the "Cte" should be defined. It has been defined: where „Cte‟ is the constant value in the equation
I would like that the authors discuss the deviation of the line y=x presented by the residuals (for high values) in Q-Q plot. Can we conclude that the residual errors are normally distributed? Following reviewer‟s suggestion, the Q-Q plot has been discussed and added in the text in page 7, section 3: Results and discussion in the “Validation of the model performance”: Before: “The Figure 5 in the lower left is the standard Q-Q plot, which is a graphical method for comparing two probability distributions by plotting their quantiles against each other. Both distributions are similar, the points in the Q-Q plot are approximately lie on the line y=x. It suggests that the residual errors are normally distributed.” Now:
“The Figure 5 in the lower left is the standard Q-Q plot, which is a graphical method to compare two probability distributions plotting their quantiles against each other. It can be observed that data from the distribution follow the normal curve (y=x) fairly closely until the second theoretical quantile that is light tile, that means most of the observations piles near the upper boundary, being skewed to the right. These points are the high PM10 observed levels which are not able to reproduce by the model. It counts for less than the 15% of the total amount of observations, thus, it can be said that it is normally distributed.” Page 7:
How was the model reproducibility tested? Why was the correlation values used to evaluate the model performance (and not the RMSE and R2)?
The model reproducibility was tested looking the descriptive statistics (mean, standard deviation and R) of the table 3. As well, the physical mechanisms of the area are well known by the authors because of the descripcion of the area in the literature review (e.g. Millán et al, 1984; Millán et al. 1984; Acero et al. 2012 and González-Aparicio et al., 2012). The correlation values were changed into R2 values to evaluate the model performance to have coherence with the selection of the type of model. Before: “On a daily basis, the model fit improves by 20% while on a seasonal basis the improvement reaches 40%. For summer the correlation coefficients (R) are 0.90, 0.95 and 0.80 at the inland, centre and coastal site, respectively. For winter the R values are 0.85, 0.86 and 0.90. The mean seasonal errors are reduced to 0.1 µg m-3 and 0.2 µg m-3 for summer and winter, respectively.” Now: “On a daily basis, the model fit improves by 20% while on a seasonal basis the improvement reaches 40%. For summer the R2 are 0.81, 0.90 and 0.64 at the inland, centre and coastal site, respectively. For winter the R values are 0.72, 0.74 and 0.81. The mean seasonal errors are reduced to 0.1 µg m-3 and 0.2 µg m-3 for summer and winter, respectively.”
In Figure 6, why did the authors refer the quadrants? The points below the line y=x correspond to the underestimation of the model, while the points above this line correspond to the overestimation.
Authors referred to the quadrants in order to show how the model behaves with highest or lowest values. For example, the first quadrant (upper left) means that the model understimates highest observed values and the fourth quadrant (lower right) means that the model overestimates the lowest observed values.
Regarding the Figure 7, the authors did not explain the daily profile of PM10 concentrations. Why were high concentration values observed at night (Figure 7b and 7c)?
The discussion of the daily profile of the PM10 concentrations are explained in the “2.2 Historical 2005 to 2011 PM10 records analysis” since it belongs to the PM10 observed variability. The aim of Figure 7 and section “diurnal cycle, daily and seasonal predictions” was to demonstrate that the models fit with the PM10 variability during the day, week and seasons. The concentrations at night of Figure 7b and Figure 7c are similar to the Figure 7a, ranging between 25 and 30 µg m-3 between 20 and 23 UTC. In the three cases the PM10 levels increased from 16 UTC. This increase is related to the traffic intensities based on the Basque Traffic Evolution document analysis and also to the properties of the atmospheric boundary layer, which decreases its height during the late evening hours and favours the stagnation of the
pollutants, as it was explained in section “2.2 Historical 2005 to 2011 PM10 records analysis”. The legend interval of the figure has been modified for better understanding. Reviewer #4: Introduction Clarify motivation to use Linear regression models in spite multi-layer Artificial Neural Networks, non-linear statistical models (generalized additive models, Random forests, multiple regression models,) or deterministic models. Add a review of few papers showing comparison between different models to strength the choice of Linear regression models. Following the reviewers suggestions this section has been modified: Before: “The objective of the study is to obtain a diagnosis PM10 data-driven model for the Bilbao metropolitan area that depends on climatic and urban parameters. This model will allow, in a near-future, to study PM10 levels based on future climatic evolution and urban growth scenarios for Bilbao to evaluate the climatic and air quality shifts under a context of climate change. There is a large series of statistical methods available in the literature to obtain this kind of model. Neural Networks for gaseous pollutants (Voukantsis et al. 2011; IbarraBerastegi et al. 2008; Agirre et al. 2006); Models for gaseous and aerosols: Multiple Linear Regression (Henry and Hidy 1979); Principal Component Regression (Abdul-Wahab et al. 2003); Independent Component Regression (Pires et al. 2008) and finally Cluster-wise methods (such us Poggi and Portier 2011). Linear regression models have been used extensively in air quality studies: Yusof et al. 2008 obtained hourly PM10 predicting equations with a correlation coefficient (R) of 0.60 for Seberang Perai (Penang). Voukantis et al. (2011) obtained R of 0.650.80 at different stations in Thessaloniki, (Greece) of an hourly based frequency model. At a daily basis, Pires et al. (2008) obtained in the application of several PM10 forecasting methodologies R of 0.60-0.79 in Oporto (Portugal) and Hörmann et al. (2005) obtained an R 0.81 of the regression equation for Graz (Austria).” Now: “The objective of the study is to obtain a diagnosis PM10 data-driven model for the Bilbao metropolitan area that depends on climatic and urban parameters. This model is intended to be used in a near-future to study PM10 levels based on future climatic evolution and urban growth scenarios for Bilbao to evaluate the climatic and air quality shifts under a context of climate change. There is a large series of methods available in the literature (e.g. Neural Networks for gaseous pollutants (Voukantsis et al. 2011; Ibarra-Berastegi et al. 2008; Agirre et al. 2006); Models for gaseous and aerosols: Multiple Linear Regression (Henry and Hidy 1979); Principal Component Regression (Abdul-Wahab et al. 2003); Independent Component Regression (Pires et al. 2008) and finally Cluster-wise methods (such us Poggi and Portier 2011)). Here three main considerations drove the choice of the method: a method which not requires expensive computational resources is needed to be used in long-term simulations; not built with black boxes to modify the independent variables for the future projections generation (as is the case for Neuronal Networks behaviour.) and proved robustness in the atmospheric sciences. In this sense Linear Regression Models, used extensively in air quality studies, is the statistical method chosen to obtain the PM10 diagnostic model for Bilbao.”
Results and discussions Since successful predictions of episodic high PM10 concentration is of particular importance for authorities and ordinary people, it's better to extend the evaluation of models performance to prediction of high level PM10 concentration events, using the performance measures suggested by Wilks (1995), e.g. POD (the probability of detection), FAR (the false alarm rate), CSI (the critical success index), as Grivas and Chaloulakou (2006) did. (Grivas G and Chaloulakou A, 2006: Artificial neural network models for prediction of PM10 hourly concentrations, in the Greater Area of Athens, Greece, Atmospheric Environment) Following the reviewers suggestions an additional sub-section in the results and discussion section to assess the prediction of high- concentrations events have been included. “Prediction of high-concentration events: This evaluation of the models’ performance is of particular interest since successful and on-time predictions of episodic PM10 concentration can be of use in assuming activity for the protection of public health. It has been already reported that health effects are more linked with short term exposure, rather than with the 24-h averaged exposure (Gold et al. 2004). In addition, some Member States of the European Union have national legislation that requires 1-h time resolution (CAFE, 2004). Here, following Grivas and Chaloulakou (2006); Chaloulakou et al. (1999) and Wilks (1995) the performance measures used are the probability of detection (POD – percentage of successful forecast over total high concentration events), the false alarm rate (FAR- fraction of false forecasts over total forecasts, the critical success index) and the CSI-ratio of correct forecasts over total potential risk events) and the accuracy (Ac – percent of forecast that correctly predicted the event or no-event) The measures are applied to the entire period analysed for both seasons. Since the 1-h limit value are not legislated, the characterization of an hourly PM10 concentration as high, in the present study will be made based on the criteria selected by Grivas and Chaloulakou (2006) for each station with similar environmental characteristics: the limit is made in the case of being 2 standard deviation levels above the average measured concentration levels in each site. As a result, using the table 3, at the urban site the PM10 limit is 50 µg m-3; 65 µg m-3 at the inland and 43 µg m-3 at the coastal site. The Table 4 presents the measures with the criterion followed to assess the model. For example, the model developed for the urban site correctly forecasts the 65% of the events or non-events but the 65% of the forecasts for high events are not materialized. However, almost the 25% of the episodes are correctly predicted. Extrapolating to the rest of the sites the behaviour of each measure is in the same order. The quality of the predictions of the models developed regarding episodic concentration levels is undermined by their inability to correctly spot some distinct very high concentrations events. This verifies the general assumption explained by the unexpected traffic congestions at some location. Following Grivas and Chaloulakou (2006) and Kukkonen et al. (2003), there are approaches aiming at increasing the frequency of extreme values for the training period in order to be repeated more times.”
Table 4: Measures for the models‟ evaluation performance for high episodic events, the verification of PM10 forecasts utilise a standard contingency table Measures Ac POD FAR CSI Ac = (a+d)/n *100
Urban
Inland 87 25 65 17
Coast 84 24 65 18
80 22 60 17
FAR = b/(b+d)*100 CSI = d/(b+c+d)*100 POD = d/ (c+d)*100
Figures
Figure 2 - bad quality, the legend is written not in English; Figures 5 and 6 - increase the size of axis title and numbers; figures are fuzzy, increase quality; Figures 7 and 8 - it is difficult to recognize both lines, make one dash line.
The figures has been improved according to the reviewer‟s suggestions
Table
Tables: Table 1: Explanatory variables in the regression model Variable
(LABEL)
Description
unit
Temperature
T
Hourly mean temperature
ºC
Radiation
RN
Hourly short wave incoming radiation
W m-2
Specific Humidity
Q
Hourly mean specific humidity
Kg of water vapour/kg of air
Wind Speed
FF
m s-1
Lagged Wind Direction
DD
Hourly mean wind speed Hourly mean wind direction (lagged variables)
Log (NO2)
logNO2
Logarithm of Hourly mean NO2
µg m-3
Mean Vehicle intensity
TRAF
Vehicles hour-1
Precipitation
RRYN
Disaggregating from Daily Mean Intensity Hourly mean precipitation /No precipitation
Hour
HH
From 00 to 23 UTC
Categorical
Season (winter or summer)
MM
Winter (12,1,2) Summer (5,6,7)
Categorical
Working Day, Saturday and Sunday
WD
Working day / Weekend
Categorical
Categorical
Categorical
Table 2: Coefficients of the model at the city centre R*10siduals:
Min
1Q -51.359
Co*10ffici*10nts:
*10stimat*10
M*10dian -10.955
Std. *10rror
3Q -2.6
t valu*10
Max 7.372
Pr (>|t|)
252.692 Signif
-176
2.41
-7.294
3.09*10-13 ***
153
0.628
24.317
2.00*10-16 ***
MM2
-5.85
2.24
-2.608
0.009116 **
MM6
-16.6
3.17
-5.233
1.68*10-07 ***
MM7
-20.5
3.56
-5.762
8.40*10-09 ***
MM8
-22.6
3.37
-6.708
2.01*10-11 ***
MM12
1.93
2.20
0.88
HH1
2.18
0.786
2.771
0.005588 **
HH2
0.596
0.787
0.758
0.448347
HH3
-1.19
0.787
-1.51
HH4
-5.59
0.788
-7.087
1.40*10-12 ***
HH5
-7.24
0.791
-9.145
2.00*10-16 ***
HH6
-6.77
0.795
-8.521
2.00*10-16 ***
HH7
-6.12
0.797
-7.678
1.67*10-14 ***
HH8
-6.66
0.802
-8.311
2.00*10-16 ***
HH9
-7.57
0.813
-9.311
2.00*10-16 ***
HH10
-9.02
0.831
-10.851
2.00*10-16 ***
HH11
-7.33
0.849
-8.638
2.00*10-16 ***
HH12
-8.78
0.858
-10.236
2.00*10-16 ***
HH13
-8.85
0.859
-10.305
2.00*10-16 ***
HH14
-9.56
0.848
-11.27
2.00*10-16 ***
HH15
-9.53
0.833
-11.444
2.00*10-16 ***
HH16
-1.03
0.821
-12.602
2.00*10-16 ***
HH17
-10.8
0.814
-13.31
2.00*10-16 ***
HH18
-11.4
0.809
-14.11
2.00*10-16 ***
HH19
-10.5
0.805
-12.987
2.00*10-16 ***
HH20
-10.3
0.800
-12.897
2.00*10-16 ***
HH21
-11.2
0794
-14.088
2.00*10-16 ***
HH22
-8.58
0.789
-10.874
2.00*10-16 ***
HH23
-3.65
0.787
-4.645
3.42*10-06 ***
17.7
2.64
6.73
1.73*10-11 ***
-4.14
2.35
-1.76
Ct*10 logNO2
RRYN1 DD2
0.379012
0.131052
0.078355 .
DD3
-28.2
1.77
-15.925
2.00*10-16 ***
DD4
-12.8
2.02
-6.334
2.43*10-10 ***
T
4.09
2.00*10-01
20.401
2.00*10-16 ***
Q
3
+02
-11.402
2.00*10-16 ***
-03
-6.421
1.38*10-10 ***
5.82
4.64*10-1
12.544
2.00*10-16 ***
3.08
7.12*10
-1
4.32
1.56*10-5 ***
-1
7.388
1.54*10-13 ***
RN logNO2:MM2 logNO2:MM6
-4.37*10 -1.51*10
-2
3.83*10
2.35*10
logNO2:MM7
5.85
7.92*10
logNO2:MM8
1.18
7.48*10-01
1.576
0.114971
2.73
4.55*10
-1
5.995
2.06*10-9 ***
6.80*10
-1
-13.309
2.00*10-16 ***
5.18*10
-1
1.952
3.94*10
-1
30.935
2.00*10-16 ***
4.25*10
-1
16.455
2.00*10-16 ***
-1.00
4.93*10
-2
-20.279
2.00*10-16 ***
-0.559
1.18*10-1
-4.739
2.16*10-6 ***
-1.27
8.73*10
-2
-14.514
2.00*10-16 ***
1.05*10
-1
-9.556
2.00*10-16 ***
logNO2:MM12 logNO2:RRYN1 logNO2:DD2 logNO2:DD3 logNO2:DD4 logNO2:T DD2:T DD3:T DD4:T
-9.05 1.01 12.2 6.99
-1.01
0.050907 .
MM2:DD2
1.70
1.52
1.116
MM6:DD2
13.1
1.91
6.893
5.59*10-12 ***
MM7:DD2
12.5
2.19
5.695
1.25*10-8 ***
MM8:DD2
11.8
2.13
5.557
2.78*10-8 ***
MM12:DD2
1.28
1.52
0.847
0.397206
MM2:DD3
-2.61
1.03
-2.524
0.01161 *
MM6:DD3
17
1.40
12.188
2.00*10-16 ***
MM7:DD3
14.4
1.57
9.176
2.00*10-16 ***
MM8:DD3
13.8
1.49
9.24
2.00*10-16 ***
MM12:DD3
-1.51
1.01
-1.502
MM2:DD4
-2.70
1.26
-2.143
MM6:DD4
10.7
1.53
6.965
3.37*10-12 ***
MM7:DD4
12.3
1.69
7.242
4.53*10-13 ***
MM8:DD4
10.5
1.67
6.296
3.11*10-10 ***
MM12:DD4
0.264449
0.133118 0.032119 *
1.10
1.23
0.898
MM2:Q
-1.80*103
3.17*102
-5.689
1.29*10-8 ***
MM6:Q
-5.15*102
3.02*102
-1.706
0.087998 .
MM7:Q
-1.30*10
3
3.39*102
-3.84
MM8:Q
-1.83*102
3.25*102
-0.562
MM12:Q
-2.17*103
2.78*102
-7.79
6.92*10-15 ***
logNO2:Q
8.22*102
9.45*101
8.69
2.00*10-16 ***
RRYN1:T
5.06*10
-1
-1
4.823
1.42*10-06 ***
logNO2:RN
3.83*10-3
8.06*10-4
4.752
2.02*10-06 ***
88.4
11.1
7.961
1.78*10-15 ***
T:Q
1.05*10
Significant codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 18.61 on 26902 degrees of freedom Multiple R2: 0.3931, Adjusted R2: 0.3914 F-statistic: 238.6 on 73 and 26902 DF, p-value: < 2.2e-16
0.369102
0.000123 *** 0.574387
Table 3: Descriptive statistics for the overall performance of the model for the period from 2005 to 2011 at coastal, urban (in bold) and inland station 2005-2011
Winter Obs.
Summer
Mod.
Obs.
Mod.
0.8; 0.8; 0.7
R
0.7; 0.7;0.6
Mean
29.5;30.8; 32.3
30.2; 31.5; 33.6
23.5; 26.0; 28.0
24.0; 26.1; 27.0
Stad.Dev
19.8;20.3;21.0
10.6;11.7;12.4
10.8;11.3;12.2
6.5;7.1;8.0
RRYN=0 Obs.
RRYN=1
Mod.
Obs.
Mod.
0.8;0.8;0.7
R
0.7;0.7;0.6
Mean
28.0;31.5;34.3
29.5;30.4;36.2
19.5;22.2;23.5
22.3;25.7;26.8
Stad.Dev
15.2;17.8;21.5
8.6;10.4;12.5
10.5;12.0;17.0
7.5;8.7;9.2
Working Day Obs.
Saturday
Mod.
Obs. 0.8;0.8;0.7
R
Sunday
Mod.
Obs.
Mod.
0.7;0.7;0.6
0.7;0.7;0.6
Mean
29.5;30.2;33.5
30.0;30.6;33.2
22.5;25.8;26.0
23.0;26.2;26.5
18.0;22.0;22.5
19.2;22.8;23.2
Stad.Dev
15.3;17.6;18.5
9.9;10.1;10.5
12.5;14.3;15.0
8.0;9.2;10.0
13.0;11.4;13.2
6.5;7.8;8.0
DD1 Obs.
Mod.
Stad.Dev
Obs.
DD3
Mod.
0.6;0.6;0.5
R Mean
DD2 Obs.
DD4 Mod.
0.6;0.6;0.5
Obs.
Mod.
0.8;0.8;0.7
0.8;0.8;0.7
22.8;23.7;24.5
20.5;21.5;22.6
23.5;24.8;25.6
24.0;25.0;26.0
29.5;30.6;32.1
29.4;30.6;32.0
23.2;25.1;26.5 25.6;28.5;29.0
7.8;8.1;8.5
3.5;4.2;5.5
9.5;10.6;11.2
6.5;7.1;7.2
15.5;18.1;19.5
9.6;10.2;10.5
17.5;17.9;18.5 11.5;12.6;13.0
Table 4: Measures for the models’ evaluation performance for high episodic events: accuracy (Ac); probability of detection (POD). false alarm rate (FAR) and the critical success index (CSI). The verification of PM10 forecasts use a standard contingency table included at the end of the table. Measures Ac POD FAR CSI Ac = (a+d)/n *100 POD = d/ (c+d)*100 FAR = b/(b+d)*100 CSI = d/(b+c+d)*100
Urban (%)
Inland (%) 87 25 65 17
Coast (%) 84 24 65 18
80 22 60 17
Figure Click here to download high resolution image
Figure
Calm 1-2-3
4
>5 Strength
5 10 15 20
Frequency % N
NW
NE
E
W
SW
SE S
T.INT+BIC
Type of model
A.INT+BIC
T.LIN
A.LIN
T.INT
A.INT
16
18
20
22
RMSE (microg m-3) 24
Figure
Figure
2
4
0.005
0.020
800
950
0
1000
2000 250
0
4
0 100
PM10
5 10
0
2
logNO2
0.020
0
FF
1000
0.005
Q
950
0
400
RN
40
800
PP
1000
0
2000
20
T
0
TRAF 0 100
250
0
5 10
0
400
1000
0
20
40
Figure Click here to download high resolution image
Figure
300 I II Observations
250 200 150 100 50 0
0
50
III IV 100 150 200 Predictions
250
300
Mean (microg m-3)
Figure
45
45
45
40
40
40
35
35
35
30
30
30
25
25
25
20
20
20
15
15
15
10
10
10
0
3
6
9 12 15 18 21 HH
0
3
6
9 12 15 18 21 HH Observations Predictions
0
3
6
9 12 15 18 21 HH
Hourly PM10 levels Episode I (20-29 January 2010)
Hourly PM10 levels Episode II (13-18 July 2009)
100
100
0 1 21 41 61
Hours
Hours Daily means in winter 2010
Daily means in summer 2009
Mean (microg m-3)
Mean (microg m-3)
1 16 1 18 1 20 1 22 1
0 81 91 10 1 11 1 12 1 13 1 14 1
20
1 11 21 31 41 51 61 71
20
1
40
14
40
60
1
60
80
12
80
10
Mean (microg m-3)
120
Mean (microg m-3)
120
81
Figure
Days Winter means
Days
Mean (microg m-3)
Mean (microg m-3)
Summer means
Years
Observations Predictions
Years
