Elsevier

Atmospheric Environment

Volume 98, December 2014, Pages 385-393
Atmospheric Environment

Application of a statistical post-processing technique to a gridded, operational, air quality forecast

https://doi.org/10.1016/j.atmosenv.2014.09.004Get rights and content

Highlights

  • An automated bias correction scheme for air quality forecasting is described.

  • Site specific biases are converted to a gridded field using Kriging.

  • Bias reduced from 7.02 to 0.53 μg m−3 for O3, from −4.00 to −0.13 μg m−3 for PM2.5.

  • Post-processing scheme provides improved model performance out to five days ahead.

Abstract

An automated air quality forecast bias correction scheme based on the short-term persistence of model bias with respect to recent observations is described. The scheme has been implemented in the operational Met Office five day regional air quality forecast for the UK. It has been evaluated against routine hourly pollution observations for a year-long hindcast. The results demonstrate the value of the scheme in improving performance. For the first day of the forecast the post-processing reduces the bias from 7.02 to 0.53 μg m−3 for O3, from −4.70 to −0.63 μg m−3 for NO2, from −4.00 to −0.13 μg m−3 for PM2.5 and from −7.70 to −0.25 μg m−3 for PM10. Other metrics also improve for all species. An analysis of the variation of forecast skill with lead-time is presented and demonstrates that the post-processing increases forecast skill out to five days ahead.

Introduction

Regional air quality forecasts have improved significantly over the last decade or so, due to factors such as (i) the availability of near-real-time boundary fluxes provided by improved global composition models; (ii) increased computing power, allowing improved resolution and greater sophistication in the representation of chemical processes; (iii) improved pollutant emission inventories. For a review of air quality forecast modelling in Europe see Kukkonen et al. (2012).

However despite these advances, the spatially and temporally detailed prediction of atmospheric composition at a given site remains a challenging problem and it is not uncommon for forecasts to contain large errors (see Solazzo et al., 2012). These may arise due to errors in inputs of key model parameters such as actual emissions (as opposed to annual mean values: see Pouliot et al., 2012), initial and boundary conditions for chemical species (Schere et al., 2012) as well as meteorology (Vautard et al., 2012). In such circumstances human forecaster intervention may be required to modify the model predictions, based on recent observations and judgement about how conditions are evolving. Alternatively, automated methods may be employed which offer the possibility of improving forecasts and may minimise or completely remove the need for human intervention (e.g. Rouil et al., 2009). A simple daily persistence forecast (i.e. that today's observed values should be the same as yesterday's) is often used as a reference forecast in meteorological verification (e.g. Jolliffe and Stephenson, 2012). This basic idea can be further developed with varying degrees of sophistication. The most advanced methods of using observations to improve air quality forecasts are those of data assimilation (DA, see Inness et al., 2013). However the high level of additional complexity and computational expense required for constituent DA plus the limited near-real time availability of suitable satellite observations, make the simpler bias correction techniques described in this paper an attractive option.

In this paper we describe a simple scheme for combining model predictions with observations in a post-processing step to generate an improved air quality forecast. In Section 2 we present a short review of various approaches which have been adopted by others and in Section 3 we describe our own scheme. We have evaluated the performance of the scheme over an extended period and the results are presented in Section 4. In Section 5 the results are discussed and possible future developments are described.

Section snippets

Air quality forecast bias correction methodologies

We consider the problem of producing the best gridded field of surface air pollutant concentrations given values from a 3-D numerical model forecast plus a set of recent measurements from a sparse surface observation network. The numerical model may be considered to exhibit both random errors and a systematic bias with respect to the observations. It is frequently observed, for example by Kang et al. (2008) and Savage et al. (2013), that this bias remains approximately constant over the

Air quality forecast model: AQUM

The Met Office air quality forecast is produced using the on-line air quality model AQUM (Air Quality in the Unified Model). This is a limited area configuration of the Met Office Unified Model (MetUM), which has a 12 km horizontal resolution covering a domain containing the UK and nearby western European countries (Fig. 1). There are 38 vertical levels from the surface to 39 km. Lateral boundary conditions for chemistry and aerosols are derived from operational forecasts of the MACC global

Results

Operationally, AQUM is run at 23:30Z with the SPPO usually run around 01:30Z. Observations are therefore available up to and including 01Z for the first full forecast day. Here we show the results from a year-long hindcast (1st July 2012–30th June 2013) which simulates these operational timings. The system was run for one month prior to the first date. This allowed the 30 day mean residual to be calculated and then used for correcting forecasts out to a five day leadtime, as described in

Conclusions

This paper has described a new bias correction technique which has been used for improving a five day gridded air quality model forecast. This technique, referred to as ‘SPPO’, is based on the short-term persistence of model bias for pollutant predictions, which is estimated from a comparison of recent measured and modelled concentrations. A gridded field of residual values is derived using Kriging of the residuals calculated at observation sites. The bias correction is extended to five days by

Acknowledgements

Near-real time observations are kindly provided by Ricardo-AEA, on behalf of Defra.

References (25)

  • L. Delle Monache et al.

    A Kalman-filter bias correction method applied to deterministic, ensemble averaged and probabilistic forecasts of surface ozone

    Tellus B

    (2008)
  • B. Denby et al.

    Interpolation and Assimilation Methods for European Scale Air Quality Assessment and Mapping. Part I: Review and Recommendations

    (2005)
  • Cited by (27)

    • Evaluation of air quality forecasting system FORAIR-IT over Europe and Italy at high resolution for year 2017

      2022, Atmospheric Pollution Research
      Citation Excerpt :

      Knowing in advance the probability of high concentration episodes allows for taking actions at different levels (starting from policy-administration level to private sector and citizen level), reducing the risks for human health. In the last decades, real-time air quality forecasting (RT-AQF) systems were developed to predict short-term pollutant concentrations with valid forecasts varying between 1 and 5 days, and they are widely used nowadays (Ha et al., 2020; Miskell et al., 2019; Neal et al., 2014; Peng et al., 2018). In general, forecasting differs from nowcasting because the first aims to predict the future state of a system, while the latter has the objective to assess its current state (Varotsos et al., 2020, 2021).

    • CMAQ-CNN: A new-generation of post-processing techniques for chemical transport models using deep neural networks

      2022, Atmospheric Environment
      Citation Excerpt :

      One results from the linear structure of these models; that is, they are reliable for only the post-processing of regional forecasts; they cannot be generalized to an entire domain (Neal et al., 2014; Kang et al., 2008) (i.e., the continental United States, as studied here). In addition, to generate stable corrections, these models require no change in modeling processes (Neal et al., 2014). In the past few years, however, CTM modeling processes have significantly improved, owing to the enhancement of our understanding of secondary processes in the atmosphere (Appel et al., 2017) (thanks to several comprehensive measurement campaigns) and advancements in the modeling of the feedback between weather and chemical processes (Wong et al., 2012) (i.e., two-way modeling).

    • Selection of bias correction models for improving the daily PM<inf>10</inf> forecasts of WRF-EURAD in Porto, Portugal

      2017, Atmospheric Pollution Research
      Citation Excerpt :

      Simulated SO2 concentrations were compared to the measurements at 7 monitoring stations run by the Baoshan Environmental Protection Bureau (www.hbj.baoshan.gov.cn) and it was found that the normalized mean bias is −48.34%. More references regarding the biases of deterministic air quality systems can be found in (Monteiro et al., 2013; Neal et al., 2014; Silibello et al., 2015). To improve the forecasts at critical locations of high public concern (e.g. densely populated areas, hospitals, and schools, etc.) provided with air quality monitoring station, application of bias correction techniques has become increasingly popular to perform local correction of the deterministic model outputs with empirical formulae (Borrego et al., 2011; Kang et al., 2008; Konovalov et al., 2009; McKeen et al., 2005; Monteiro et al., 2013; Neal et al., 2014; Ridder et al., 2012; Silibello et al., 2015; Struzewska et al., 2016; Wilczak et al., 2006).

    • Mortality and emergency hospitalizations associated with atmospheric particulate matter episodes across the UK in spring 2014

      2016, Environment International
      Citation Excerpt :

      A detailed description and evaluation of the model is presented in Savage et al. (2013). Although the AQUM model currently has no data assimilation for aerosol or chemical species, a bias correction is applied to forecasts of pollutant concentration, based on a comparison of model predictions and recent observations from the AURN (Neal et al., 2014). Hourly AURN observations are also used to evaluate the forecast model predictions in near-real-time.

    View all citing articles on Scopus
    View full text