Operational Data-Driven Monthly Forecasting at Microsoft

Recent research on data-driven global weather forecasting has produced deep-learning models such as FourCastNet, Pangu-Weather, and GraphCast, which offer comparable or even better performance compared to state-of-the-art numerical weather predictions from the European Centre for Medium-range Weather Forecasts (ECMWF). These models now show enough promise to be investigated for operational forecasting; indeed, ECMWF has begun producing operational forecasts with Pangu-Weather and FourCastNet for research purposes. We present our operational deep-learning-based forecast system which powers 30-day forecasts for Microsoft Start Weather. Our approach focuses on building a stable sub-seasonal forecasting model that produces realistic forecasts with minimal bias and long-term drift, as compared to the operational ECMWF long-range forecasts. In particular, we use a large multi-model ensemble to sample distributions of future weather patterns, and introduce a coupling mechanism to add ocean forcing. Our ensemble prediction system achieves week 3-6 skill scores within 10% of, and in some cases exceeding, those of ECMWF’s model for several surface variables of interest.