AI Model Gives Energy Traders Hourly Forecasts 7 Months Out

A tech firm is expanding the possibility of electric-grid forecasting by offering hourly projections of US power demand seven months into the future. That will provide energy traders a new view compared to 15-day weather projections and broad seasonal outlooks. 

To do so, Amperon Holdings Inc. is leveraging artificial intelligence and machine learning to create a range of hourly demand for this extended period that’ll be updated upon the daily release of global weather models created by Europe’s biggest forecasting center, said Sean Kelly, a former power trader who co-founded Amperon in 2018. The firm’s customers will start receiving these forecasts on Wednesday in addition to the typical 15-day outlook Amperon offers. 

Predicting electricity usage has become harder in recent years as more extreme storms strain grids. The rise of heat pumps, rooftop solar, batteries and electric vehicles reshapes demand, and the AI boom is poised to cause usage to rocket higher. Demand from data centers will more than double by 2035, increasing its share of total US electricity usage to 8.6% from 3.5% today, according to BloombergNEF. 

Utilities and power retailers are eager for greater visibility into weather conditions and changing consumer behavior to hedge against price shocks. Energy speculators are also seeking a financial edge, which advanced forecasting may help them eke out. 

“Hourly granularity is what people are really craving,” said Kelly, noting that includes municipalities and all types of utilities and energy traders. “It serves as an insurance policy. They want to keep the lights on for their clients.” 

Among power companies using Houston-based Amperon’s forecasts are PG&E Corp., Orsted AS, AES Corp., and Eversource Energy. The company is backed by investors including Energize Capital, HSBC Holdings Plc, National Grid Plc and Tokyo Gas. 

Ultimately, the accuracy of the new projections depends on the underlying weather forecasts, Kelly said. Amperon’s models take the weather forecast and use machine learning to come up with hourly demand curves, in part with the help of historical data. The models also factor in solar and wind conditions. Amperon’s models also adjust to how homes and businesses change their behavior in response to grid stress. 

The weather outlook is based on predictions from existing models developed by the European Centre for Medium-Range Weather Forecasting, or ECMWF. The group’s models provide forecast conditions every six hours, out to seven months. Traders frequently use ECMWF’s forecasts, though their accuracy declines the further out it extends, as it does with all forecasts.  

Using machine learning, data scientists at Amperon have been able to boost the granularity of the ECMWF models and provide hourly temperature forecasts over a seven-month period, an approach that the company says its customers have asked for.

There are early signs that Amperon’s new approach can provide accurate forecasts. Backtests show that one of its models successfully predicted a June demand spike on the 13-state grid operated by PJM Interconnection 32 days in advance, as well as the impacts of a January 2024 blast of Arctic air that hit Texas 38 days ahead of its arrival. Models were also able to forecast that this summer was poised to be a dud for Texas power traders betting on price spikes. 

Hedge funds and other speculative traders are looking for such insights so they can place bets on the power market even at a split-second faster than other traders, said Kelly.

To contact the authors of this story:

Naureen S Malik in New York at nmalik28@bloomberg.net

Lauren Rosenthal in New York at lrosenthal21@bloomberg.net

Generated by readers, the comments included herein do not reflect the views and opinions of Rigzone. All comments are subject to editorial review. Off-topic, inappropriate or insulting comments will be removed.