MetDesk revolutionises energy trading with rapid AI forecasts powered by NVIDIA technology

Artificial intelligence (AI), with its remarkable ability to analyse vast datasets, is rapidly changing the landscape of weather forecasting. MetDesk is announcing a collaboration with NVIDIA to expedite the delivery of weather forecasts using AI.

Harnessing NVIDIA Earth-2, an AI platform for climate and weather simulations, MetDesk is the first commercial company to market with generating and productising deterministic and ensemble AI forecasts. Whilst deterministic forecasts are useful in short range forecasting, ensemble models are more skilful from approximately day 7 onwards. This is why MetDesk is utilising NVIDIA Earth-2 platform tools to create a machine learning (ML) ensemble model, with a proprietary set of initial conditions for the 51 member ensemble. Technical details on how we did that can be found here.

Accelerating weather predictions

MetDesk’s implementation of the NVIDIA FourCastNet model (MD-FCN) model not only achieves a performance competitive with the global Numerical Weather Prediction (NWP) models such as the GFS & ECMWF, but the huge win is the reduction in computational processing time from 6 hours (ECMWF) to just 1 hour. Furthermore, MetDesk are taking steps to integrate NVIDIA NIM, a set of easy-to-use inference microservices designed for secure, reliable deployment of high-performance AI model inferencing, to further reduce the computational time to just 2 minutes. An industry where accuracy and speed of weather forecasts are of great importance is energy trading.

Competitive edge for energy traders

The weather forecast has a huge influence on the price of electricity and gas. When temperatures are predicted to be colder than normal in winter, heating demand will increase and therefore so do the prices of power and gas. In times like this, when the ability to get forecast data to traders as quickly as possible matters, a few seconds can make all the difference. This leap in forecast speed currently gives traders an additional source of information five hours ahead of the next model run and soon will be available before other commercial providers’ deterministic model runs are even complete. This results in a huge competitive advantage, enabling traders to anticipate market movements and adjust their position accordingly, seizing opportunities before competitors do.

“Fast, accurate meteorological information can have critical impacts on global markets,” said Bhoomi Gadhia, Senior Product Marketing Manager for Earth-2, NVIDIA. “MetDesk’s platform, powered by NVIDIA Earth-2, delivers weather forecasts with high accuracy and magnitudes of improvement in speed, enabling energy traders to react to imminent weather conditions faster than previously thought possible, to manage risks effectively, and to optimise their trading strategies.”

Recent performance

Julian Telford, Head of Energy Forecasting, MetDesk, explains how the models are used in-house: “The MD-FCN model is considered each day by MetDesk’s team of energy meteorologists and is valued to use in addition to traditional NWP models such as the ECMWF and GFS models. The MD-FCN Ensemble provides us with a range of possibilities and therefore risks, which are then discussed with clients around the globe daily.” He goes on to explain some recent wins from the model: “It accurately predicted the timing of the forecast breakdown in heat across northern Europe in Week 26 (week commencing 24th June 2024) whilst also correctly signalling a cooler interlude into the Northeast US later in the week.”

Sub-seasonal forecasting

MetDesk has taken it one step further. It’s not just the 15-day forecast that influences the market; traders monitor weather weeks and even months in advance. This is why MetDesk has taken a position to also generate 50-day forecasts (MD-FCN50) leveraging NVIDIA technology. The huge speed advantages of ML forecast systems compared to traditional NWP mean that it’s possible to receive the MD-FCN50 model output nearly 12 hours earlier than the ECMWF’s 46-day forecast system – delivering data within the European daytime trading period rather than later the same evening after markets have closed.

Future prospects

There have been significant advancements in ML models in a relatively short amount of time, and what’s clear is we are only scratching the surface of AI capabilities in weather forecasting. Dan Harding, Head of Research and Development, MetDesk notes, “Over the past year, there has been a meteoric rise in the use of ML for weather forecasting. Using our in-house GPU hardware, we’re able to create large ML ensemble forecasts in a fraction of the time of traditional NWP models. As the field continues to evolve, our aim is to stay at the forefront and integrate these advancements into MetDesk products for our clients.”

The future of AI in weather forecasting is bright, and MetDesk is leading the charge, continuously evolving to provide cutting-edge tools for energy traders.