RainForests: A Machine Learning Approach to Calibrating NWP Precipitation Forecasts

• DOI: 10.1175/waf-d-23-0211.1
• Publication Date: 2024-09-12T12:37:52Z
• Authors: Belinda Trotta; Benjamin Owen; Jiaping Liu; Gary Weymouth; Thomas Gale; Timothy Hume; Anja Schubert; James Canvin; Daniel Mentiplay; Jennifer Whelan; Robert Johnson
• Keywords: Quantitative precipitation forecast

Abstract Probabilistic forecasts derived from ensemble prediction systems (EPS) have become the standard basis for many products and services produced by modern operational forecasting centres. However statistical post-processing is generally required to ensure desired properties expected probability-based outputs. Precipitation, a core component of any forecast, particularly challenging calibrate due its discontinuous nature extreme skew in rainfall amounts. A skillful system must maintain accuracy low-to-moderate precipitation amounts, but preserve resolvability high-to-extreme which, though rare, are important forecast accurately interest public safety. Existing machine-learning approaches calibration address this problem, each has drawbacks design, training approaches, and/or performance. We describe RainForests, approach calibrating using gradient-boosted decision trees. The model based on ecPoint recently developed at ECMWF Hewson Pillosu (2021), uses models place semi-subjective trees ecPoint, along with some other improvements structure. evaluate RainForests Australian domain against simple benchmarks, show that it outperforms both overall skill high conditions, while being computationally efficient enough be used an system.