Weather forecasts cannot always be guaranteed. Understanding the range of possible outcomes is therefore imperative for making decisions such as issuing public warnings or managing renewable energy. Traditional forecasting relies on numerical weather forecasting, which performs physics-based simulations of the atmosphere.Recently, machine learning weather prediction has improved the accuracy of single predictions. However, there is still little that can indicate uncertainty or risk. A new model called GenCast is trying to change that.According to a study published in Nature, GenCast, a model trained on decades of reanalysis data, is able to generate 15-day probabilistic global forecasts that are not only aware of uncertainties, but also provide an ensemble of plausible weather changes. It operates at high resolution and can predict a large number of variables in a short period of time. This amazing achievement has the potential to make predictions more accurate and faster, and therefore more valuable from a planning and safety perspective, not just in specific regions but everywhere in the world.
The future of weather forecasting: How machine learning is changing forecasting
GenCast is a machine learning system trained on decades of global reanalysis data. Employ a generative approach to present many different future weather scenarios instead of just one. The model operates at a resolution of 0.25 degrees latitude and longitude and advances in 12-hour increments for up to 15 days. To represent uncertainty, GenCast creates an ensemble of probabilistic forecasts, each representing an expected weather path. The model derives spatial and temporal correlations from historical data, along with the power spectra of atmospheric variables. Because GenCast is trained on reanalysis, it benefits from the best historical reconstructions while making predictions much faster than traditional ensemble simulations.
Satellite data and AI: a powerful combination for global predictions
GenCast has three main advantages.
- Generate clear, discrete weather trajectories that preserve natural variability in both space and time.
- Its marginal predictive distribution is very accurate. This means that the predicted probabilities correspond to the observed frequencies, making risk estimates more reliable.
- Because GenCast reflects spatial and temporal dependencies, it can better predict regional phenomena such as wind power generation or the path of a tropical cyclone.
Moreover, this model is not very slow. You can generate a complete 15-day global ensemble within minutes, allowing for faster updates and better accessibility. Therefore, faster predictions can be used to support the real-time decision-making processes of emergency managers, power grid operators, and other users who require timely probabilistic information.
Performance comparison: GenCast vs. ENS
GenCast significantly outperformed ECMWF's top operational ensemble, ENS, on a large number of targets in the experiment. Comparisons revealed that GenCast performed better on nearly every variable and lead time tested, and was able to make better predictions, especially for extreme events. Furthermore, the accuracy of tropical cyclone track predictions and local wind power predictions will also be enhanced, with accurate identification of the spatial pattern of seams being paramount in these regions. Such improvements indicate that ML-based probabilistic prediction systems, when trained on high-quality reanalysis data, have the potential to perform at least as well as, or better than, state-of-the-art physically-based ensembles. GenCast's faster generation times allow it to deliver updates more frequently. This is a huge advantage in rapidly changing weather conditions.
Use and impact
GenCast's speed and skill yield many practical scenarios. Power grid operators use probabilistic wind and solar forecasts to balance energy supply and demand and enable lower power curtailments. Emergency managers can more quickly estimate local risks such as flooding, heat waves, and storms, allowing them to order evacuations and proactively deploy personnel. It is becoming possible to price weather risks more efficiently and plan for losses in the insurance and financial sectors. Scientists will be given the ability to perform extensive scenario studies to understand the impacts of climate change and the required strength of infrastructure.GenCast shows that modern generative machine learning can provide fast and reliable probabilistic weather forecasts that rival the best physically-based ensembles. GenCast captures uncertainty by building an ensemble of realistic weather trajectories to enhance predictions of extreme events and support practical applications from energy to emergency response. Increasing resolution and reducing computational costs remain challenges, but model extraction and operational fine-tuning provide a clear path forward. As these techniques mature, they may change how forecasts are created and used, making probabilistic weather information more accurate, timely, and accessible to decision makers around the world.
