Toward Calibrated Ensembles of Neural Weather Model Forecasts

Abstract

Neural Weather Models (NWM) are novel data-driven weather forecasting tools based on neural networks, that have recently achieved comparable deterministic forecast skill to current operational approaches using significantly less real-time computational resources. They require short inference times, which can potentially improve the characterization of forecast uncertainty by designing very large ensembles, which is of paramount importance for for example, extreme events, and critical for various socio-economic sectors. Here we propose a novel ensemble design for NWMs spanning two main sources of uncertainty: model uncertainty and initial condition uncertainty. For the model uncertainty, we propose an effective strategy for creating a diverse ensemble of NWMs that captures uncertainty in key model parameters. For the initial condition uncertainty, we explore the “breeding of growing modes” for the first time on NWMs, a technique traditionally used for operational numerical weather predictions to estimate the initial condition uncertainty. The combination of these two types of uncertainty produces an ensemble of NWM-based forecasts that is shown to improve upon benchmark probabilistic NWM and is competitive with the 50-member ensemble of the European Centre for Medium-Range Weather Forecasts based on the Integrated Forecasting System (IFS), in terms of both error and calibration. Results are particularly promising over land for three key variables: total column water vapor, surface wind and surface air temperature. The proposed strategy is scalable, enabling the generation of very large ensembles ($>$100) with potential applications for extreme events.