Operational Machine Learning Post-Processing of Short-Range Temperature, Humidity, Wind Speed and Gust Forecasts

Abstract

Statistical methods can be used to create bias correction models that learn from past forecast errors and reduce systematic errors in real-time forecasts. This study presents a machine learning (ML) approach using extreme gradient-boosted (XGBoost) trees to address biases in a numerical weather prediction (NWP) nowcast model for key meteorological parameters: 2-m temperature, 2-m relative humidity, 10-m wind speed, and 10-m wind gust. These ML models have been integrated into the Finnish Meteorological Institute's (FMI) operational nowcasting framework, Smartmet nowcast. Results show that, even with a relatively modest set of meteorological predictors, the ML bias correction method significantly improves forecast accuracy, reducing the root mean square error (RMSE) by 24%–29% compared to the direct NWP model output. The implementation of this new bias correction method not only improves the quality of FMI's short-range forecasts, but also extends the availability of bias-corrected data for longer forecast lead times, offering substantial improvements over the previously implemented bias correction method. The codebase for this machine learning bias correction is available at (https://github.com/fmidev/snwc_bc).