RainForests: Calibrating Precipitation Forecasts with Machine Learning

Accurate rainfall forecasting is vital for public safety, yet precipitation remains one of the most challenging variables to predict. While the development of ensemble Numerical Weather Prediction (NWP) models enables the kind of probabilistic forecasting that is so much more descriptive of rainfall characteristics than simple deterministic forecasts, accurate and efficient calibration of the ensembles presents its own challenges. Rainfall’s skewed distribution and discontinuous nature in both time and space, the importance and impact of extreme events, and its sub-grid scale variability within most operational NWP model resolutions limit the effectiveness of most parametric and non-parametric approaches. We introduce RainForests, an efficient machine-learning approach to calibrating ensemble NWP model rainfall forecasts using gradient-boosted decision trees (GBDTs). RainForests produces probabilistic forecasts from global-scale ensembles, while still representing the important sub-grid variability characteristic to rainfall by using select feature variables that influence variability. Building upon the ECPoint approach recently adopted at the European Centre for Medium-Range Weather Forecasting (ECMWF), RainForests maintains ECPoint’s computational efficiency and adaptability, while addressing some limitations of the modelling approach. Notably, the addition of GBDTs automate tree construction and training, and lead-time specific models increase forecasting performance by adjusting the calibration for varying uncertainty over time.  We demonstrate that RainForests outperforms the widely-used Ensemble Model Output Statistics (EMOS) approach using logistic regression. Rainforests is now implemented in the Integrated Model Post-Processing and Verification System (IMPROVER), an open-source probabilistic post-processing system created by the UK Met Office and now developed and maintained in partnership with the Bureau of Meteorology. With IMPROVER soon to replace a legacy post-processing system for production forecasting at the Bureau, we show that RainForests makes significant accuracy gains over the Bureau’s previous rainfall calibration approach, with the 3-hour precipitation forecast achieving one additional lead day of skill and much greater consistency with the daily accumulation forecast, showing its suitability to real-world forecasting conditions.