16B.1 NOAA Operational Air Quality Forecasting Guidance - Using Artificial Intelligence to Achieve Faster and Adaptive AQ Forecasting

Thursday, 1 February 2024: 4:30 PM
338 (The Baltimore Convention Center)
Jennifer Sleeman, APL, Laurel, MD; and I. Stajner, C. Ribaudo, D. Chung, C. Ashcraft, R. Chen, C. Tang, C. Kofroth, Q. H. Dang, M. Halem, K. Wang, J. Huang, H. C. Huang, R. Montuoro, C. A. Keller, J. M. Wilczak, I. V. Djalalova, J. McQueen, B. Baker, V. Krasnopolsky, W. Putman, and M. Hughes

The National Oceanic and Atmospheric Administration (NOAA) is building Artificial Intelligence (AI) models to overcome computational limitations of the current operational air quality (AQ) forecasting methods used for AQ guidance. As NOAA moves towards finer resolution and in order to better represent extreme air quality events, current methods could benefit from AI emulators that are able to both reduce the computation load and adapt to changing operational conditions in real time. We describe the progress made on a number of NOAA AI efforts in order to evolve how AQ forecasting guidance is produced. We describe a deep learning U-NET model used for bias correction, which has resulted in RMSE scores that improve upon the forecast and that are competitive with the current Kalman-filtering method used. We also describe a novel 3-D U-Net model for emulating the transport of chemical and aerosol species for the Community Multiscale Air Quality (CMAQ) modeling system. This model learns a translation of input to output representing the functionality of the transport, and is able to do so across all 64 levels spanning the model domain in vertical direction. We describe the challenges of processing the chemical and aerosol tracer concentration data and the effects of including additional features such as meteorological variables, in order to get the best performance from the model. We also describe a comparison between this model and a version of the model that is built to be a deep learning continual learner. It is envisioned that if deployed in operations, the continual learner model will evolve over time to accommodate extreme air quality events that may occur. We describe how this work has been further developed to support emulating the CMAQ numerical model entirely. We share speed-up and accuracy results when the deep learning full model emulator is incorporated into the CMAQ process flow. Called every 6 hours by the CMAQ model, the emulator will perform hourly forecasts for 72 hour segments. To perform this emulation we evaluated the use of two separate models as emulators, a conditional Generative Adversarial Network (cGAN) and NVIDIA’s FourcastNet. Both models were adapted and tailored to work with the NOAA’s AQ data. We show preliminary results for forecasting both ozone and PM2.5 using these two models. Finally, we give a brief overview of our newest work related to subseasonal to seasonal forecasting of wildfire emissions using deep probabilistic models.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner