4-9 September 2022, Bonn, Germany
EMS Annual Meeting Abstracts
Vol. 19, EMS2022-367, 2022, updated on 10 Jan 2024
EMS Annual Meeting 2022
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

RainForests: A novel Machine Learning approach to calibrating rainfall forecasts

Benjamin Owen, Belinda Trotta, Jaiping Liu, Thomas Gale, Anja Schubert, and Gary Weymouth
Benjamin Owen et al.
  • Bureau of Meteorology, Australia

Improvements to skill (and other characteristics) of quantitative probabilistic rainfall forecasting for weather and hydrological purposes are a high priority at the Australian Bureau of Meteorology (the Bureau), highlighted by a series of major floods in eastern Australia in early 2022.  Post-processed ensemble numerical weather prediction (NWP) rainfall guidance is key to increasing forecast automation in routine conditions and providing better guidance for high impact rainfall events. 

The Bureau's existing NWP post-processing system has markedly increased weather rainfall forecast skill in recent years. However, to enable further forecast improvements and greater integration of rainfall processing for weather and hydrological purposes, a more general approach with fewer calibration steps was required. To these ends, we have developed 'RainForests': a multi-ensemble rainfall processing system, utilising gradient boosted decision tree (GBDT) ensembles for forecast calibration.

RainForests is inspired by the ECPoint method of Hewson and Pillosu (2021). RainForests, like ECPoint, is a non-parametric and generally non-local method which uses decision trees to create situation-dependent error distributions for each input (analogous to an extension of Bayesian Joint Probabilities) that can be used to calibrate grid-scale rainfall guidance to point-scale.

Key features of RainForests:

  • Uses a series of GBDT ensembles to construct the error distribution in place of a single manually trained decision tree. This produces robust outputs which are near-continuous relative to inputs, allows for rapid retraining on new data or addition of feature variables, and utilises open source GBDT software.
  • Uses additive error (delta = obs - forecast) in place of the forecast error ratio. This allows for calibration when forecast rainfall is zero. Error distributions vary with forecast rainfall amount (and other predictors).
  • Models are trained and verified using both rain gauge and gauge-calibrated radar data, each of which have their uncertainties, strengths and weaknesses. Basic Bureau and RainForests QC is applied to the data to reduce gross errors.
  • Error distributions are pooled for each NWP model (e.g. ECMWF ensemble), from around 100 individual input ensemble and deterministic ('ensemble of one') members in total. Resulting calibrated probability distributions for each model are then blended in probability space.

Additionally, RainForests uses and contributes to capabilities in the Integrated Model post-PROcessing and VERification (IMPROVER) system being developed in collaboration with the UK Met Office.

Initial RainForests outputs have comparable skill to the Bureau's existing post-processing system.  Improvements are planned and will be reported on.  It is also planned to provide calibrated rainfall ensemble members, derived from RainForests outputs, to downstream applications.  This aims to support production of ensemble multiple-variable indices, hydrological applications and aggregation of rainfall forecasts in space and time.

How to cite: Owen, B., Trotta, B., Liu, J., Gale, T., Schubert, A., and Weymouth, G.: RainForests: A novel Machine Learning approach to calibrating rainfall forecasts, EMS Annual Meeting 2022, Bonn, Germany, 5–9 Sep 2022, EMS2022-367, https://doi.org/10.5194/ems2022-367, 2022.

Supporters & sponsors