Improving and understanding probabilistic precipitation forecasts using machine learning

Uncertainty in numerical weather prediction (NWP) arises due to the initial state not being fully known and physical processes not being perfectly represented within the models. Precipitation is challenging to predict because it is non-linear with complex drivers from the atmosphere and so varies quickly even on a local scale. This means even advanced NWP models struggle to predict precipitation with the correct intensity at the right time or location. This study aims to explore whether machine learning (ML) can rediagnose precipitation rates based on vertical profiles of temperature, humidity and wind, thus replicating the precipitation calculated by cloud and precipitation parametrization schemes that are used in NWP models to represent the unresolved microphysical processes. A small but high-quality dataset comprised of days with widespread precipitation has been curated for developing an initial model, with in depth exploratory data analysis carried out to understand any trends in the model input data and assess the need for feature engineering. Vertical profiles of atmospheric variables (temperature, humidity, wind) taken from 6-hour forecasts of the Met Office Unified Model global ensemble (MOGREPS-G) provide input features for the ML model, and the target variable (or truth) is instantaneous precipitation intensity measured by the UK radar network at a 1km resolution. The two data sources are aligned onto the same grid by calculating the fractions of the MOGREPS-G ~20km cell containing radar precipitation in five precipitation intensity bands, with bounds informed by domain experts.

Each MOGREPS-G ensemble member is used to generate a ML prediction of the fractional precipitation coverage that exceeds each intensity threshold, then an ensemble average of these fractions is calculated for each intensity threshold. These values can be considered as ML generated ensemble probabilities. They can then be compared with the true fractional coverage from radar, as well as precipitation probabilities from MOGREPS-G to identify similarities and differences in their behaviour. Explainable AI techniques are applied to better understand the decisions made by the ML model when creating predictions.  The aim is to understand the potential of using ML for improving precipitation forecasts, either through complementing NWP outputs with ML outputs, or by using ML as a tool for improving the understanding of the drivers of errors in NWP precipitation forecasts. Initial results look promising and a number of avenues for further development have been identified following consultation with domain experts.