Conference Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 50 - Industry Applications for Environmental Statistics
Type: Topic Contributed
Date/Time: Sunday, August 7, 2022 : 4:00 PM to 5:50 PM
Sponsor: Section on Statistics and the Environment
Abstract #322330
Title: Using Physics-Informed Machine Learning to Generate Global Medium-Range Probabilistic Weather Forecasts
Author(s): Steven Joel Brey and Ashley Payne* and Stelios flampouris
Companies: Tomorrow.io and Tomorrow.io and Tomorrow.io
Keywords: ensemble; weather; crps; forecasting; time series; climate
Abstract:

A weather forecast is useful when it aids decision-making. Due to limited weather observations, and an incomplete understanding of atmospheric processes, perfect weather forecasts do not exist. The utility of weather forecasts increases when the uncertainty of the forecasting system is well characterized. Traditionally, this uncertainty is quantified by running many expensive numerical weather prediction (NWP) simulations (i.e. ensembles). However, these simulations are overconfident and biased. Machine learning post-processing approaches can learn systematic and flow-regime specific biases and other shortcomings of NWP-based forecasts. We post-process NWP forecasts with a physics-informed multi-task neural network that produces global probabilistic weather forecasts. This computationally efficient method blends multiple NWP models and auxiliary features with spatiotemporally matched observations to learn a shared covariance structure. This framework generates probabilistic realizations of multiple downscaled surface-corrected outputs at any given location and forecast lead time. Out-of-sample validation shows our model minimizes error, bias, and increases statistical reliability.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2022 program