GenCast: Diffusion-based ensemble forecasting for medium-range weather (2312.15796v2)

Abstract: Weather forecasts are fundamentally uncertain, so predicting the range of probable weather scenarios is crucial for important decisions, from warning the public about hazardous weather, to planning renewable energy use. Here, we introduce GenCast, a probabilistic weather model with greater skill and speed than the top operational medium-range weather forecast in the world, the European Centre for Medium-Range Forecasts (ECMWF)'s ensemble forecast, ENS. Unlike traditional approaches, which are based on numerical weather prediction (NWP), GenCast is a machine learning weather prediction (MLWP) method, trained on decades of reanalysis data. GenCast generates an ensemble of stochastic 15-day global forecasts, at 12-hour steps and 0.25 degree latitude-longitude resolution, for over 80 surface and atmospheric variables, in 8 minutes. It has greater skill than ENS on 97.4% of 1320 targets we evaluated, and better predicts extreme weather, tropical cyclones, and wind power production. This work helps open the next chapter in operational weather forecasting, where critical weather-dependent decisions are made with greater accuracy and efficiency.

Overview of GenCast Model

GenCast is a machine learning-based generative model designed specifically for ensemble weather forecasting. It aims to produce a set of possible future weather scenarios, known as ensemble forecasts, that are not only accurate but also coherent over time and space. GenCast outperforms its counterparts by generating high-resolution global weather predictions with exceptional skill and reliability, even for rare and extreme events.

Methodology and Architecture

GenCast employs a diffusion model, a type of neural network that generates forecasts by refining an initial state comprised mostly of noise. The model is fed historical weather data and trained to predict future states. Its architecture is similar to previous models like GraphCast, but notably, it transitions from a spatial transformer network to a sparse transformer, enhancing its ability to preserve high-frequency content.

Results and Performance

Through an extensive evaluation using multiple metrics, GenCast demonstrates superior forecasting skill compared to both the European Centre for Medium-range Weather Forecasts (ECMWF)’s top operational ensemble forecast, ENS, and a leading ML model GraphCast-Perturbed. Notably, GenCast maintains high performance across all examined weather variables and up to 15 days into the future. It even excels at forecasting extreme weather events, a critical capability for many applications.

Physical Consistency and Reliability

One of the strengths of GenCast is its ability to generate physically consistent forecasts. It preserves the high-frequency spatial structure and produces reliable ensemble forecasts with spread and rank histograms comparable to or better than operational models. This ability is especially important for ensuring that forecasts remain plausible and useful for real-world decision-making.

Implications and Future Work

With its rapid computation on a single Cloud TPU device, GenCast opens new opportunities for producing large ensemble weather forecasts efficiently. However, it still requires high-quality initial conditions from traditional weather prediction systems. Moving forward, GenCast signifies a substantial advancement in the applicability and utility of ML for weather forecasting, potentially revolutionizing how weather predictions are made and used.