5 Forecasting PM10

AEAT-3844

Introduction
An important new task following the introduction of the new air pollution bandings in November 1997 is the forecasting of episodes of elevated concentrations of airborne particles, specifically PM₁₀. The forecasting of PM₁₀ episodes is a challenge because of the range of sources that contribute to ambient particle concentrations. The relative contributions from these sources must be understood before a reliable forecasting strategy can be implemented. The sources of PM₁₀ in the UK have been reviewed by the Quality of Urban Air Review Group (QUARG, 1996) and source attribution methods involving the combination of data from several DETR monitoring networks has recently been developed at NETCEN (Stedman, 1998). Ambient PM₁₀ can be considered to be made up from three distinct types of particles:

• Primary combustion particles - these are particles derived from vehicle exhaust or stationary combustion sources. These particles are generally black due to the presence of carbon.
  
• Secondary particles - consisting of sulphates and nitrates from the atmospheric oxidation of sulphur dioxide and oxides of nitrogen. These particles and their precursors can be transported over distances of 100s of km.
  
• Coarse particles - in contrast to the primary combustion or secondary particles, these resuspended dust, soil, mechanically derived and sea salt particles are generally larger than 2.5 mm in diameter.
  

We have used an empirical regression method to disentangle the different origins of measured daily PM₁₀ particle concentrations. An equation of the following form can be obtained:

[measured PM₁₀] = A.[measured black smoke] + B.[measured sulphate] + C

with these three terms representing, primary combustion, secondary and coarse particle concentrations, all in mgm^-2. This method was originally applied (Stedman, 1998) on a network mean basis (mean of all sites on a daily basis) and regression coefficients have subsequently been derived for the cities listed in Table 5.1.

Table 5.1. Regression coefficients for PM₁₀ data

	Smoke coefficient, A	SO4 coefficient, B	Intercept, C	r2
London Bloomsbury	0.64	2.26	10.96	0.78
Birmingham Centre	0.59	2.41	8.30	0.71
Bristol Centre	1.03	2.35	10.83	0.70
Manchester Piccadilly	0.60	2.46	9.77	0.74
Newcastle Centre	0.66	3.13	7.73	0.84
Belfast Centre	0.71	2.30	9.21	0.79
Cardiff Centre	0.86	1.71	13.07	0.73
Leeds Centre	1.00	2.58	4.56	0.84
Network Mean	1.00	3.00	5.00	0.84

The forecasting of episodes of elevated PM₁₀ concentrations therefore requires the use of appropriate models to forecast these different contributions on a daily basis. The regression method described by Stedman (1998) significantly simplifies the derivation and testing of the required models because the individual models can be compared with measured black smoke and sulphate data. The models that have been developed for forecasting primary combustion and secondary PM₁₀ are discussed below. This is followed by an investigation of the reliability of the combination of these two models during a 10 week period during the winter of 1997/98.

Fixed daily mean concentrations have been considered here in order to simplify the analysis because black smoke and sulphate concentrations have been measured on a daily basis. It is recognised that the daily maximum of running 24-hour averages can be higher than fixed daily means. The methods described here represent the initial work that has been undertaken to provide a method for forecasting PM₁₀. There is a continuing research programme to refine these techniques.

Primary Combustion PM₁₀
Primary combustion derived PM₁₀ originates, in most UK cities, from the same emission sources as oxides of nitrogen (NO_x), including vehicle exhausts. High concentrations of both of these species tend to be associated with poor dispersion of low level emissions caused by light winds and low mixing heights. The box model that is used to forecast hourly NO_x concentrations can therefore be adapted to provide forecasts of primary combustion PM₁₀. The following multi-stage analysis was carried out using daily monitored and forecast pollutant concentrations for the one year period from April 1995 to March 1996. This period was chosen because it includes the major secondary particle episodes in early 1996 and several primary particle episodes in late 1995.

• The regressions listed in Table 5.1 provide a coefficient for the relationship between measured daily mean black smoke concentrations and primary combustion PM₁₀ for individual cities.
  
• An empirical relationship between measured daily mean NO_x concentrations and black smoke was also derived for each city.
  
• An empirical relationship between the daily mean NO_x concentration predicted by the urban box model and measured black smoke was then derived and checked for consistency with the relationship with measured NO_x. Daily mean NO_x was calculated as the average of the 24 individual hourly forecast concentrations.
  
• The end result of this analysis is a composite coefficient which can be used to calculate estimates of daily mean primary combustion PM₁₀ from predicted NO_x.
  The coefficient for the relationship between black smoke concentrations and primary PM₁₀ is typically between 0.6 and 1.0 depending on the blackness of the particles in each city, which will depend on the fuel usage. The values of the empirical coefficient between measured NO_x and primary PM₁₀ is however reasonably consistent for GB cities, with values ranging from 0.08 to 0.13. The coefficients for the relationship between forecast NO_x and primary PM₁₀ is, however, more variable, depending on the absolute values derived from the box model.
  
  Figure 5.1 shows examples of the range of coefficients obtained relationships between measured primary PM₁₀ concentrations and both measured and forecast NO_x. The coefficients are reasonably consistent in London for measured and forecast NO_x but in Cardiff the modelled NO_x concentrations systematically underestimate the measured NO_x concentrations, giving the two graphs very different slopes. The Cardiff error is taken into account when the model is used in forecasting.
  
  
  Secondary PM₁₀
  A trajectory model has be developed to provide estimates of secondary PM₁₀ concentrations. This model makes use of the same 96-hour back trajectories and temperature forecasts as the ozone forecasting model (Stedman and Williams, 1992). The simplified chemical scheme has been adapted from Derwent et al (1988), and includes a fixed oxidation rate of 1% per hour for the conversion of SO₂ to particulate sulphate.
  
  The model was run using trajectories from 1995 and 1996 and the results compared with measurement network data in order to provide an indication of the reliability of the modelled estimates. Daily measurements of particulate sulphate concentration are available from rural monitoring sites within the acid deposition monitoring networks (RGAR 1997).
  
  To validate the model we investigated the accuracy of the modelled sulphate. The modelled sulphate values were compared with measured particulate sulphate values at seven rural sites: Eskdalemuir, High Muffles, Lough Navar, Strathvaich Dam, Yarner Wood, Sibton and Lullington Heath. Particulate sulphate is not measured at Sibton or Lullington Heath so we have compared the modelled results at these sites with measured concentrations from the nearby sites of Stoke Ferry and Barcombe Mills respectively. At these seven sites we compared data for summer 1995 and 1996, and also looked at distinct episodes for both years. Correlation values (see Tables 5.2 and 5.3) are reasonable and the pattern is generally similar. The model tends to overestimate the peak episode concentrations, despite modelled background being zero at all sites, and generally the modelled values are often within a few mg/m³ of the measured data. The modelled means for the summer periods (Table 5.4) are below the measured means due to the zero background. Generally on an episode day the spatial pattern of the modelled results is very similar to that of the measurements. At Lullington Heath/Barcombe Mills the model does not perform so well: the modelled values are much too low and the correlation negligible, except in two of the episodes. It is not clear why this should happen here, but since sulphate concentrations are considered to be due to long range transport it is likely that there are unknown local factors. It is also possible that there are inaccuracies in the emission inventory used (Lubkert and de Tilly, 1989) and this will be addressed in future work. The good agreement between measured and modelled data found at the Sibton/Stoke Ferry shows that lack of co-location is unlikely to be the cause of the difference.
  
  The secondary particle contribution to PM₁₀ measured at the DETR automatic monitoring network sites is likely to be dominated by ammonium sulphate and sodium nitrate. We have investigated using two alternative methods of deriving an estimate of secondary PM₁₀ from the trajectory model results. The method for estimating secondary PM₁₀ from sulphate measurements presented in QUARG (1996) based on an analysis of sulphate and nitrate measurements for the UK is equivalent to multiplying sulphate by a factor of 2.5 to take into account the presence of both nitrate and counterions. Similar factors were also found by Stedman (1998) when comparing measured sulphate and PM₁₀ concentrations at DETR monitoring sites. We have also derived a second estimate of secondary PM₁₀ by adding up the modelled estimates of the concentrations of the following species: sulphate, nitrate and ammonium nitrate.
  
  In order to compare these modelled estimates of secondary PM₁₀ with PM₁₀ measurements we have followed the approach suggested by Stedman (1998) and derived an estimate of secondary PM₁₀ by subtracting a primary concentration derived from black smoke measurements and a constant coarse particle concentration from the measured PM₁₀, using the equation and coefficients listed in Section 5.1.
  
  Table 5.2: Correlation coefficient (r²) of measured and modelled sulphate values for timeseries
  

  Site	Summer 1995	Summer 1996
  Eskdalemuir (ES)	0.34	0.50
  High Muffles (HM)	0.58	0.46
  Lullington Heath/Barcombe Mills (LH/BM)	0.13	0.13
  Lough Navar (LN)	0.54	0.50
  Sibton/Stoke Ferry (SB/SF)	0.48	0.58
  Strath Vaich Dam (SV)	0.43	0.44
  Yarner Wood (YW)	0.50	0.41

  
  
  Table 5.3: Correlation coefficient (r²) of measured and modelled sulphate values for episodes
  

  Site	April/May 95	March 96	April 96	July 96	August 96
  ES	0.69	0.22	0.12	0.45	0.30
  HM	0.71	0.62	0.24	0.69	0.40
  LH/BM	0.43	0.65	0.12	0.38	-0.49
  LN	0.36	0.66	0.26	0.59	0.50
  SB/SF	0.40	0.72	0.70	0.29	0.70
  SV	0.81	0.47	0.52	0.85	0.37
  YW	0.42	0.26	0.57	0.44	0.79

  
  Table 5.4: mean values of measured and modelled sulphate for timeseries (mgSO₄ m^-3)
  

  	Summer 1995		Summer 1996
  Site	Measured	Modelled	Measured	Modelled
  ES	2.5	2.0	2.2	2.7
  HM	3.4	2.8	3.2	3.6
  LH/BM	4.5	0.8	4.4	1.4
  LN	2.5	1.3	1.8	1.3
  SB/SF	3.6	5.0	3.9	4.0
  SV	1.8	0.8	1.6	1.5
  YW	4.9	3.2	3.0	2.3

  
  We compared the model results derived using both B.[modelled sulphate] and [total modelled aerosol] with secondary PM₁₀ at 3 sites: London Bloomsbury, Birmingham Centre and Newcastle Centre. Model results are available for London and Birmingham but for Newcastle we used the model results from High Muffles, as the closest trajectory arrival point. At all 3 sites the two sets of model results were very nearly equivalent most of the time, although in many cases the total aerosol model gives slightly lower peak concentrations and in all cases the correlation with measured data was higher for the total aerosol model than for the B.sulphate model. In London the modelled values matched the "measured" secondary PM₁₀ well a lot of the time and correlation is reasonable for both (see Table 5.5). Newcastle also shows a good relationship between "measured" and modelled values, with a higher correlation than for London in 1996. Birmingham shows the lowest correlation between modelled and measured values of secondary PM₁₀, although the values are still indicative of the good predictive ability of the model. As with Newcastle and London the correlation is higher in 1996 than in 1995. At all sites most episodes are predicted, some to within a few mg/m³ eg on the 2/5/95, 1/8/95 and 20/8/96 in London (see Figure 5.2).
  
  The early part of 1996 was exceptional in terms of the magnitude and duration of the secondary particle contribution to PM₁₀ episode concentrations in the UK. In particular, March 1996 saw the largest secondary PM₁₀ episode since monitoring began in 1992 (Stedman, 1997) and this episode is excellently modelled at all three urban sites, as well as at the rural particulate sulphate sites, as discussed above. This clearly demonstrates the success of the model.
  
  These preliminary investigations lead us to believe that the model can be extremely useful in predicting urban secondary PM₁₀ episodes. The model has been automated and has been used in the forecasting of PM₁₀ since December 1997.
  
  Table 5.5: Correlation coefficient (r²) of measured and modelled secondary PM₁₀ values for timeseries
  

  	Summer 1995		Summer 1996
  Site	B x sulphate	Total aerosol	B x sulphate	Total aerosol
  London	0.44	0.48	0.49	0.56
  Birmingham	0.43	0.46	0.44	0.52
  Newcastle	0.44	0.51	0.54	0.56

  
  
  Combining Secondary And Primary PM₁₀
  The previous sections have described the two modelling methods that have been developed to forecast the primary and secondary components of daily PM₁₀ concentrations. Both of these types of models have been used since December 1997 to aid the NETCEN air quality experts in the preparation of the daily forecast for PM₁₀. Some example results for selected cities for the 10 week period from December 1997 to mid February 1998 are shown in Table 5.6.
  
  Table 5.6: Measured and predicted mean PM₁₀ concentrations (mgm^-3) and correlation coefficients for daily means, 1 December 1997 to 13 February 1998 (provisional data)
  

  monitoring site	measured mean	predicted mean	r2
  London Bloomsbury	23	28	0.17
  London Camden Roadside	27		0.19
  London Marylebone Road	33		0.17
  Birmingham Centre	21	21	0.36
  Birmingham East	17		0.34
  Wolverhampton Centre	20		0.41
  Bristol Centre	23	24	0.12
  Manchester Piccadilly	23	24	0.27
  Salford Eccles	23		0.45
  Bury Roadside	28		0.36
  Newcastle Centre	18	17	0.25
  Belfast Centre	24	22	0.37
  Derry	28		0.47
  Cardiff Centre	24	22	0.14
  Leeds Centre	-	29
  Bradford Centre	26		0.45

  
  The agreement between the measured and predicted mean concentrations is generally reasonably good, with the concentrations at roadside and kerbside sites tending to be under predicted as would be expected. The correlation between the measured and predicted daily means varies from reasonable to poor. Timeseries plots for selected sites are shown in Figure 5.3. Some periods of high concentrations due to a combination of primary and secondary particles, such as the end of January, were correctly predicted at many sites. High concentrations due to primary particles were also reasonably well predicted on several occasions, for example in Birmingham and Bristol in early December and in Belfast and Derry on several days. The absolute magnitude of the data for the largest episode in Derry are in question, but the general shape of the data is still correctly modelled.
  
  Overall the modelling methods that have been developed to assist the air quality experts in forecasting PM₁₀ concentrations are proving valuable and further work will be undertaken to refine these methods, including evaluation of a new inventory and new chemical scheme for the secondary PM₁₀ model.
  

  INDEX

  ---

  To other DETR publications