PROGNOS: A Renewed Statistical Postprocessing Infrastructure and Opportunity for AI Applications to Weather and Air Quality Forecasting for the Meteorological Service of Canada (MSC)

PROGNOS, a recent MSC initiative, is intended to replace the current weather and air quality operational post-processing system (UMOS) with a more versatile, modular and innovative system. Advantages of the proposed design include reduced system maintenance cost, facilitated adaptation to frequent numerical model updates, improved ability to apply new post-processing strategies and to better serve research and development projects. PROGNOS has extensible statistical modeling and machine learning capabilities. It currently issues real-time experimental forecasts of surface PM2.5, O₃ and NO₂ pollutant concentrations as well as surface air temperature and dew point temperature for point locations across Canada; such preliminary PROGNOS forecasts will be showcased. Post-processing will eventually be applied to additional meteorological fields and numerical models. The system is developed following an iterative and incremental approach. In its current development phase, batch updates of multiple linear regression models occur weekly using parallel processing in a cluster computing environment. Less flexible, but more computationally efficient, online updating methods are also considered as alternative post-processing options for future development. Several statistical modeling and machine learning approaches have been explored, including Kalman filter and random forest prototypes for air quality forecasts. The evaluation of such techniques is facilitated by the systems modular design. Logging, parameterisation, diagnostic and visualization features are also being developed and tested for improved system traceability, user-friendliness, transparency and portability. A standalone database was developed and is being populated with real time observation data for up-to-date and seamless time series extraction for model training and post-processing. Medium to long term objectives include integrating a complementary data quality control module, seasonal and other model transition schemes as well as gridded post-processing.