NVIDIA Earth2Studio Enables Custom Ensemble Weather Forecasting for Renewable Energy Planning

NVIDIA has released a detailed tutorial demonstrating how to build custom batched ensemble weather forecasting workflows using Earth2Studio, an open-source Python framework. The tutorial, published on MarkTechPost, walks through creating an ensemble pipeline that integrates atmospheric data from the Global Forecast System (GFS) with NVIDIA’s FourCastNet (FCN) prognostic model, along with a custom wind-power diagnostic for renewable energy applications.

The tutorial addresses a growing need among energy planners, climate researchers and AI developers in India and globally: the ability to generate probabilistic weather forecasts that account for uncertainty, especially for wind energy forecasting. India’s wind energy capacity stood at over 44 GW as of 2025, and accurate ensemble forecasts can help grid operators and project developers optimise turbine operations and power purchase agreements.

How the ensemble forecasting pipeline works

Earth2Studio is NVIDIA’s framework for physics-informed machine learning in Earth sciences. The tutorial uses its low-level iterator, coordinate-mapping, batching and Zarr APIs to construct a custom ensemble execution pipeline. Users start by installing Earth2Studio components while preserving an existing CUDA-enabled PyTorch environment, such as in Google Colab.

The pipeline loads the FCN prognostic model and retrieves atmospheric initial conditions from GFS. It then implements a variable-scaled perturbation system that applies physically appropriate noise amplitudes to different atmospheric variables while retaining an unperturbed control member. This approach produces an ensemble of eight forecasts, each with slightly different initial conditions, allowing users to quantify forecast uncertainty.

Custom wind-power diagnostic model

A notable feature of the tutorial is the creation of a WindPowerCF class, a custom diagnostic model that converts 10-metre wind components into hub-height wind speed and turbine capacity factor. The model uses a power-law shear correction and a standard power curve with cut-in, rated and cut-out wind speeds. This allows researchers to directly estimate potential power output from weather forecasts.

The diagnostic model is compatible with Earth2Studio’s batch processing and coordinate-handling utilities, making it suitable for large-scale simulations across multiple locations. The tutorial demonstrates this by defining a point of interest in New Delhi (28.61°N, 77.21°E) and generating wind capacity factor forecasts.

Verification and visualisation of ensemble forecasts

The tutorial includes a verification step where forecasts are compared against GFS analyses using latitude-weighted RMSE, fair CRPS (continuous ranked probability score), ensemble spread and spread-skill ratios. These metrics help users assess the reliability and skill of the ensemble.

Visualisation techniques include spatial maps of ensemble mean and spread, geopotential-height spaghetti contours (500 hPa), point-based fan charts, wind-capacity-factor forecasts and lead-time skill curves. These outputs are designed to be immediately useful for operational decision-making.

Relevance for India’s energy and climate sector

For Indian readers, this tutorial has direct applications. India’s Ministry of New and Renewable Energy has set a target of 500 GW of non-fossil fuel capacity by 2030, with wind energy playing a significant role. Ensemble weather forecasting can improve the accuracy of wind power predictions, reducing the cost of balancing renewable energy on the grid.

The use of open data from GFS and open-source tools like Earth2Studio lowers the barrier for Indian startups, research institutions and state energy agencies to develop custom forecasting solutions. The framework also supports other AI models beyond FCN, making it adaptable to future developments in weather AI.

Technical requirements and limitations

The tutorial assumes familiarity with Python, PyTorch and basic atmospheric science concepts. It requires a GPU for reasonable performance — the tutorial explicitly warns that running on CPU will be very slow. The code is designed to run in Google Colab with a T4 GPU, which is freely available but has usage limits.

The ensemble size is set to eight members with a batch size of two and eight forecast steps. Users can adjust these parameters based on their compute resources and accuracy requirements.

Datos clave

Aspect	Detail
Framework	NVIDIA Earth2Studio (open-source)
Prognostic model	FourCastNet (FCN)
Initial conditions	GFS (Global Forecast System)
Ensemble size	8 members with control member
Custom diagnostic	WindPowerCF — turbine capacity factor
Verification metrics	RMSE, fair CRPS, spread, spread-skill ratio

Source: MarkTechPost — https://www.marktechpost.com/2026/08/29/building-custom-batched-ensemble-weather-forecasting-with-nvidia-earth2studio/