GraphCast Data is Essential

The tech giant’s graph neural network is systematically benchmarked and up to 15% more accurate than traditional methods.

Weather forecasts are the foundational layer for any power market participant. They are the basis for any load, solar and wind forecast, which themselves are a key component in price formation. On any horizon, they drive billion-dollar investment and operational decisions every day. Most weather forecasts today are derived from numerical weather predictions (NWP), meaning they are based on our best understanding of how physical models drive future weather.

In 2020, Google DeepMind published a paper that introduced a widely available benchmarking and evaluation criteria, WeatherBench, for weather forecasting. It leverages the ERA5 reanalysis dataset, which is the ‘best guess’ for actual weather across the globe for the past 40 years. At the time, the ECMWF operational model generated the best forecasts based on the suite of metrics compared to leading machine learning approaches. That benchmark laid the groundwork for comparing traditional and ML-based forecasts — and it revealed where machine learning had room to grow.

GraphCast’s inevitable rise

Two years later, Google DeepMind released a new model, GraphCast, that provides forecasts for over 100 weather variables at 0.25 degrees resolution (every ~28KM) for the next 10 days at 6H granularity.  Its name is inherited from the type of algorithm used to generate forecasts: a graph neural network (GNN), which differs from past algorithms in how it structures input data. Instead of modelling historical observations as a data table or as a time series, a GNN uses a graph architecture where weather variables from a grid are intelligently connected to each other. The resultant graph used for training defines a sparse, meaningful structure that captures atmospheric interactions.

GraphCast is a fundamentally different approach from previous attempts to model weather: it includes building a graph of different coordinates and training a model to predict how they interact.

The forecasts from GraphCast have been systematically benchmarked against ECMWF’s leading operational model, and outperform on 90% of metrics across all lead times. The reduction in error (RMSE) for key variables defined in WeatherBench is up to 15%.  The reason for the improvement in the past five years is thanks to improved architectures for physical system forecasting (GNNs), technological advances in processing, and the hard work of many at Google for developing the model.  

Its applications in downstream forecasting are broad: it outputs many variables at many locations, which may not be helpful for analysts looking at charts but is a perfect use case for downstream graph neural networks simulating the power grid.

What’s next for GraphCast

The current version of GraphCast will be the least accurate version of itself. As any who used early versions of LLMs in late 2022 will know, these models tend to improve over time and given the relatively small parameter size of GraphCast (37 million parameters, which is roughly the size of a small LLM), there remain huge opportunities for more accurate forecasting.

GraphCast outperforms a leading NWP from ECMWF, HRES, on all lead times by varying amounts.

Typical weather models introduce probabilistic intervals by perturbing initial conditions to simulate uncertainty. The ‘butterfly effect’ makes using different inputs today provide a large distribution of weather scenarios tomorrow (the ensemble method). The current GraphCast model does not support this sort of scenario planning and only outputs deterministic forecasts. One exciting next step is to introduce an evolution of GraphCast that outputs probabilities, which themselves will be key inputs into downstream power market forecasts.

It's worth noting that ECMWF has released its own AIFS (AI forecasting model) with some success, which is also based on a graph-neural network backbone.

Open for everyone

Despite what may be commercially available, GraphCast forecasts are available for free with some limitations on granularity and use. They can be requested through the Google Weather API and include forecasts for every 6H for the next 10 days. Interpolation is necessary for most power market use cases (trading hourly).

Historical vintages of GraphCast data are freely distributed under a Creative Commons BY 4.0 license, permitting reuse and redistribution with attribution. Real-time forecasts (forward looking) are more restricted and are only available for free if they are either a) for internal use only (most trading desks) or b) are used in a derivative product as a value-added service and not redistributed. Google’s terms discourage straightforward reselling of the data.

Conclusion

Traditional weather forecasting methods are an important part of power markets analytics and that will not change soon. However, new methods with innovative algorithms offer an alternative perspective for traders alongside typical numerical weather prediction models. The new ML-based methods should continue to improve with more data and compute. Probabilistic intervals with an already broad coverage of weather variables will support downstream machine-learning in ways that traditional methods cannot. And progress will be measured rigorously along the way by a wide array of users thanks to the open terms of use for one of the biggest advances in weather forecasting in the past few decades.

Enertel AI uses GraphCast data to provide short-term energy and ancillary price forecasts for utilities, independent power producers (IPPs), and asset developers to inform trading strategies. You can view our data catalog, view a clickthrough demo of our product for operators, or request a sample of backcast data for your assets.

Sources

GraphCast

WeatherBench

GraphCast Terms of Use