Assessment and calibration of extreme precipitation probabilities in S2S forecast models

Extreme rainfall events have large societal impacts, and advance warning of their occurrence is valuable. However, numerical weather prediction (NWP) models struggle to represent rainfall extremes and are known to have seasonally and regionally varying biases. The presence of model biases has led to the development of a variety of post-processing methods to calibrate probabilistic and deterministic forecasts. These methods generally apply corrections that depend on target period and lead time but otherwise are the same from one forecast to another. On the other hand, model deficiencies tend to be state dependent and replect poorly represented processes. Consequently,

process-based calibration methods that depend on the background forecast state (e.g., temperature or moisture related quantities) and whose corrections vary from one forecast to another represent a new source of potential skill improvements.

Here we propose to assess the extent to which the observed dependence of extreme rainfall on large-scale environments such as near-surface temperature, dew-point temperature and convective available potential energy (CAPE) is represented in submonthly (week-2 to week-4) forecasts. Conditional distributions and quantile regression plots will be evaluated in the SubX and S2S forecast datasets for a range of lead-times (week-2 to week-4), varying space and time averages, and different magnitudes of extremes. State-dependent errors (e.g., one that depend on temperature, dew point, or CAPE) will be corrected using novel process-based calibration methods that are designed to produce reliable forecast probabilities. These new methods will be constructed in the context of regression-based methods by including temperature, dew point, or CAPE as additional predictors. Performance will be assessed using rigorous skill comparison methods.

Our assessment and process-based calibration will focus on warm-season precipitation, but we will also apply the same approaches to snowfall, where the relationship to large scale environment is less established, and where there has been little to no development of calibrated probabilistic snowfall forecasts at lead times beyond a few days.

The main outcomes of the proposed work will include: (1) a complete assessment of how rainfall extremes and snowfall scale with temperature and other quantities in extended-range-class NWP models; (2) a systematic process-oriented identification of conditional (e.g., stratified by temperature) and unconditional biases in the model representation of rainfall extremes that inform model development and novel process-based calibration techniques; (3) a rigorous head-to-head statistical comparison of the many forecasts calibration methods currently available, including those developed here, for a range of lead times and time/space averages.

The proposed project is closely aligned with the overall competition goal to ”advance predictive capability and understanding of precipitation on the subseasonal to seasonal scale.” The project will provide objective methods ”to identify and address sources of model bias across NOAA’s modeling suite”, and will include a novel ”systematic process-oriented evaluation of the biases”. The new process-based calibration methods will have the form of ”statistical regression, error reduction, and bias correction schema” and will explicitly contain ”assessments of uncertainty for various phenomena particularly precipitation.”