Air Quality Forecasting

Improving 48-h predictions of fine particulate matter (PM2.5) over the US

In an effort funded by NASA, NCAR and its partners have developed a new capability to produce 48-hour detailed forecasts of ground-level ozone and fine particulate matter. The new forecasting capability combines satellite and in-situ observations with state-of-the-art air-quality modeling. It is generating more detailed, probabilistic air-quality forecasts compared to the current forecasts, which provide just a single-value prediction and do not specify the uncertainty associated with the prediction. Just as a weather forecast, for example, might warn of an 80% chance of rain in the afternoon, new air quality forecasts may warn of an 80% chance of high ozone levels during certain times of the day, while the current forecasts only tell whether ozone will be high or low. Such detailed forecasts can significantly enhance decision-making in air-quality management. The system has been set up over the US but can be easily applied to any part of the world.

The first objective of the ongoing project is to improve the initialization of the National Oceanic and Atmospheric Administration (NOAA) / National Centers for Environmental Prediction (NCEP) operational air-quality system, which is based on the Community Multiscale Air Quality (CMAQ) model, through chemical data assimilation of satellite retrieval products with the Community Gridpoint Statistical Interpolation (GSI) system. We use GSI to assimilate retrievals of aerosol optical depth from the NASA Aqua/Terra Moderate Resolution Imaging Spectroradiometer (MODIS) satellite instruments to improve predictions of particulate matter of aerodynamic matter of less than 2.5 µm (PM2.5) over the US. The assimilation of MODIS AOD in CMAQ improves the model’s ability to simulate day to day variability, i.e., the correlation coefficient by ~67% and reduces the mean bias by ~38% (Figure 1).

Figure 1: Top left panel shows EPA PM2.5 monitoring sites used for evaluation of CMAQ simulated PM2.5 mass concentrations. The comparisons of the observed and CMAQ simulated diurnal and daily variability of PM2.5 averaged over all the sites during 15 July to 14 August 2014 for three CMAQ experiments are shown in the top right and bottom panels, respectively. Standard deviation in the average observed values range from 4.8 to 11.9 µg/m3, and those in CMAQ average value range from 2.7 to 7.5 µg/m3. BKG represents the CMAQ experiment without assimilation and ASSIM represents CMAQ experiment with AOD assimilation.
Figure 1: Top left panel shows EPA PM2.5 monitoring sites used for evaluation of CMAQ-simulated PM2.5 mass concentrations. The comparisons of the observed and CMAQ-simulated diurnal and daily variability of PM2.5 averaged over all the sites during 15 July to 14 August 2014 for three CMAQ experiments are shown in the top right and bottom panels, respectively. Standard deviation in the average observed values range from 4.8 to 11.9 µg/m3, and those in CMAQ’s average value range from 2.7 to 7.5 µg/m3. BKG represents the CMAQ experiment without assimilation and ASSIM represents CMAQ experiment with AOD assimilation.

The second objective is to improve CMAQ’s deterministic predictions and to reliably quantify their uncertainty with analog-based post-processing methods applied to the deterministic predictions. The AnEn technique (Delle Monache et al. 2013) has been extensively tested for the probabilistic prediction of both meteorological variables and renewable energy (Alessandrini et al. 2015). The AnEn is built from a historical set of deterministic predictions and observations of the quantity to be predicted. For each forecast lead time and location, the ensemble prediction of a given variable is constituted by a set of measurements from the past (i.e., 1-hour averages of Oconcentrations). These measurements are those concurrent to past deterministic predictions for the same lead time and location, chosen based on their similarity to the current forecast. The forecast variables used to identify the past forecast similar to the current one are called analog predictors. In this application we use as predictors wind speed, wind direction, 2-m temperature, cloud cover, and PM2.5 and Oconcentrations forecasts over the continental US generated with CMAQ. The AnEn has been successful tested for the deterministic and probabilistic prediction of both PM2.5 and O3. The Analog-predictor weights are obtained independently at each station by an optimization algorithm which minimizes the continuous ranked probability score (CRPS) over May 2015.

In Figure 2 the AnEn Oconcentration forecasts are plotted through quantile ranges together with the observations and the CMAQ predictions. In this example, the AnEn mean can correct the CMAQ forecast, especially during the night when the O3 concentration decreases.

Figure 2: Root Mean Squared Error (RMSE) and BIAS as a function of forecast lead time for the AnEn mean (red) and CMAQ (black) forecasts of O3 concentrations over all the available stations for the period of June 2015-September 2015. The vertical bars indicate the 5-95% boot strap confidence intervals.
Figure 2: Root mean squared error (RMSE) and bias as a function of forecast lead time for the AnEn mean (red) and CMAQ (black) forecasts of Oconcentrations over all the available stations for the period of June 2015-September 2015. The vertical bars indicate the 5-95% boot strap confidence intervals.

In Figure 2, we compared the CMAQ and the AnEn mean forecasts. The AnEn mean consistently improves the CMAQ forecast by significantly decreasing both the RMSE and the bias.  We have also looked at several attributes of probabilistic predictions, and similarly to other applications, the analog ensemble is statistically consistent, and it is able to reliably quantify the uncertainty of the prediction.

The third objective is the extrapolation of deterministic and probabilistic point-based predictions to a two-dimensional grid over the US with a Barnes-type iterative objective analysis scheme. Figure 3 shows an example of this technique, which is currently considered for operational implementation as part of the NOAA operational air quality prediction system.

Figure 3: CMAQ Ozone gridded forecast data on the left and corresponding observed data on the right for August 29, 2017, 22 UTC.
Figure 3: CMAQ Ozone gridded forecast data on the left and corresponding observed data on the right for August 29, 2017, 22 UTC.

The proposed effort is led by NCAR, in collaboration with NOAA, CU Boulder, and the University of Maryland. Currently, NOAA/NCEP is running operationally in real-time the deterministic analog-based correction for the prediction of ground level ozone and surface PM2.5.

Air quality forecasting system for Delhi

NCAR and the Indian Institute for Tropical Meteorology (IITM), an autonomous institution of the Ministry of Earth Sciences of India, have jointly developed an air-quality forecasting system to enhance the decision-making activity in the area of air quality. The forecasting system synergistically integrates MODIS AOD retrievals and in situ measurements of fine (PM2.5) and coarse (PM10) particulate matter from 48 stations in the National Capital Region (NCR) of India with the WRF-Chem modeling system to improve forecasts of surface PM2.5 in Delhi. The system was tested for the October-November of 2016 and 2017 (Figure 4). The assimilation of MODIS AOD significantly improved the performance of WRF-Chem model in simulating PM2.5 mass concentrations in Delhi.

This air quality forecasting system has been providing operational air quality forecasts to the public and the decision-makers since October 2018. A website has been developed by the IITM to disseminate the forecasts, their near-real time verification, and warning bulletin and messages. The website can be accessed here: https://ews.tropmet.res.in/

Figure 4: Verification of averaged 72 h WRF-Chem forecasted PM2.5 with (ASM) and without (BKG) assimilation of MODIS AOD against PM2.5 observations performed by the Central Pollution Control Board (CPCB) of India. The shades areas represent standard deviation in the average values.
Figure 4:Verification of averaged 72-h WRF-Chem forecasted PM2.5 with (ASM) and without (BKG) assimilation of MODIS AOD against PM2.5 observations performed by the Central Pollution Control Board (CPCB) of India. The shaded areas represent standard deviation in the average values.

A Quasi-operational air quality forecasting system for the CONUS

The Atmospheric Chemistry Observations and Modeling (ACOM) laboratory and the Research Application Laboratory (RAL) of NCAR have jointly developed a quasi-operational air quality forecasting system for the conterminous United States (CONUS). The system is designed to support field campaigns and research groups across the US such as NASA’s Tropospheric Ozone Lidar Network (TOLNET), to offer additional information to air-quality forecasters across the nation, to extend NCAR’s current global atmospheric chemistry prediction capability, and to provide a long-term model output for use in research projects and health studies. These forecast products are available via the following website: https://www.acom.ucar.edu/firex-aq/forecast.shtml

Figure 5: An example of the hourly evaluation of WRF-Chem predicted surface ozone against the EPA AirNOW measurements.
Figure 5: An example of the hourly evaluation of WRF-Chem predicted surface ozone against EPA AirNOW measurements.

A near real-time evaluation system has also been set-up for the CONUS air quality forecasts. The evaluation system downloads surface ozone and PM2.5 observations from the EPA AirNOW monitoring network and evaluates WRF-Chem forecasts every hour. An example of hourly evaluation of WRF-Chem forecast is shown in Figure 5. Customized plots meeting the requirements of the TOLNET team are also created and posted on the website. In addition to hourly evaluation, monthly evaluation of first and second days of forecasts is also performed and provided on the website. We also provide our forecasts in Google Earth files to support the flight planning during the field campaigns.

Figure 6: Top 5 contributors (as percentages) to CO and BC levels over Delhi during four seasons namely winter monsoon (wmn: November to March), transition period spring (tps: April-May), summer monsoon (smn: June-September), and transition period autumn (tpa: October) in 2014. The dark green label represent contribution from rest of India.
Figure 6: Top 5 contributors (as percentages) to CO and BC levels over Delhi during the winter monsoon (wmn: November to March), transition period in spring (tps: April-May), summer monsoon (smn: June-September), and transition period in autumn (tpa: October) in 2014. The dark green label represent contribution from rest of India.

Quantifying Inter-state transport of air pollutants in India

Quantifying transboundary transport of air pollution is an important component of state-level air quality management because air quality in a state depends on emissions within that state and upwind states. In an effort funded by the World Resource Institute (WRI), NCAR has employed a computationally efficient tagged-tracer approach in the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) to quantify interstate transport of air pollution in India during 2014. Carbon monoxide (CO) and black carbon (BC) are used as the tracers because they represent air pollutants with lifetimes ranging from about a week to more than a month. WRF-Chem simulations are evaluated against ground-based and space-borne retrievals of aerosols and trace gases. The results suggest that a potentially important role for interstate governance in managing India’s air quality. States in the Indo-Gangetic Plain (IGP) exhibit a strong linear relationship between the direct emissions and near-surface CO and BC mass concentrations, indicating that local emissions dominate in determining surface CO and BC in these states, but interstate transport contributes 43-48% and 38-46% of CO and BC, respectively, in the IGP. The linear relationship starts decreasing as we move to other parts of India, indicating that interstate transport starts to dominate as we move away from the IGP. The contribution of interstate transport to surface CO and BC mass concentrations is estimated to be 20-82% in different states and seasons. Figure 6 demonstrates how interstate transport affects CO and BC during different seasons in Delhi. These results could be used to curb local emissions and create joint mitigation plans among states to improve the air quality in India.

2019 Accomplishments:

  • Transition of the Delhi air quality forecasting system to the Indian Institute for Tropical Meteorology
  • Addition of surface PM2.5 and PM10 assimilation in the Delhi air quality forecasting system
  • Development of CONUS air-quality forecasting system in collaboration with ACOM
  • Development of a tagged tracer technique for quantification of transboundary transport of air pollutants in India

2020 goals: 

  • Develop an analog-based post-processing system for Delhi air quality forecasting system
  • Develop assimilation capability for the CONUS air quality forecasting system. 
  • Develop a chemical reanalysis system for the CONUS