A new startup is targeting a gap in AI weather forecasting

Artificial intelligence is making its way into many corners of the energy sector these days, including into the weather forecasting that’s integral for renewables planning. But of course, AI isn’t useful without good data. Today’s weather models have largely been trained on processed or simplified datasets, which fill in gaps to smooth out weather data.

Brightband, a new company out of stealth today is looking to take a new approach, using raw observational data to build an advanced, end-to-end earth system AI platform to improve weather predictions and decision making.

• The top line: Brightband said today it had raised a $10 million Series A led by Prelude Ventures, which will help the company build out its team and tools. The company’s initial focus is on global forecasts up to 15 days in advance, but it hopes to eventually predict weather a month or two out.
• The nuts and bolts: The company will open source its basic forecasting tools, methodology, and data sets, which include data sources like microwave and irradiance data, as well as wind and rain data, for example. Brightband’s business model will depend on selling add-on services, like higher resolution forecasts or longer time periods.
• The current take: There’s been some research indicating that using raw observational data to train AI weather models will result in more accurate predictions, but to date nobody has put that research into practice, explained Brightband co-founder Julian Green, who previously ran moonshot projects for Google X. “We feel like there’s a gap in bringing the best and latest AI technology to this space,” he added.

One of the largest areas of opportunity for Brightband and its add-on services is the energy sector, Green explained.

“They need to operationally get ahead of tomorrow’s renewable energy generation and tomorrow’s heating and cooling demand,” he said. Other sectors already clamoring for better prediction tools include agriculture and transportation, he added. 

For the energy sector in particular, trust is key to widespread use and deployment. To that end, Brightband is also working to publish “benchmarks” against which users of its models can evaluate how well they’re predicting something like a cold snap or a heat wave.

“It's really important that we have this agreed, holistic set of measures and can see whether AI is doing better, or where it’s doing worse,” Green said.

(Editor’s note: Prelude Ventures is also an investor in Latitude Media.)

What’s next

Brightband is currently hiring for geospatial and research engineers to round out its current founding team, an effort funded by its Series A. The company’s founders have been in discussions with potential end users for the last year, Green said, and the feedback has been positive: “When we tell people what we’re up to, they’re like ‘can you hurry up?’”

That’s because of the uncertainty that still exists in forecasting, Green added; there’s so much uncertainty with today’s predictions, he said, that “it’s basically guessing.”

Well-trained AI models, though, can provide a step-change.

“Right now the physics models, because they are so expensive to run and you can only realistically run so many steps, they’re basically not that great after seven to ten days, ” Green said. “AI promises to be an alternative to just looking at the Almanac and saying ‘how much does it usually rain on September the 18th?’”

Because of the problem of training models on processed data sets, they don’t tend to be very reliable past seven or ten days. The slightly longer timeframe is where Brightband hopes it can fill the gaps.

“It looks like AI may be able to go longer and stronger,” Green added.