We met playing ultimate frisbee in Cambridge, UK, and quickly found common interests in statistics and optimization. Thomas had previously worked on wind turbine placement problems, providing experience with power markets, and we discussed them but didn't see an immediate entry point, so Thomas continued his statistics PhD and Rajan worked in ML research and GPU algorithm design at a startup.
A year ago we heard of a need for better wind power forecasts and started to look at the market more closely. We found a gap emerging from the increase in the prevalence of storage, especially lithium-ion, grid-scale batteries. It seemed like an interesting and useful real-world application of machine learning, particularly with the possibility of reducing carbon emissions, so once the business case looked tenable, we decided to go ahead!
Electrical power markets have become increasingly volatile due extreme weather events and increased prevalence of intermittent renewables. In response to this, producers are bringing on more flexible generation assets such as batteries to even out fluctuations in supply, and electrical consumers are aiming to increase their ability to modulate demand to better take advantage of cheap intermittent power. These assets don't fit into the day-ahead markets designed for mostly traditional steam power plants, making it difficult to choose when to use them. Our forecasts help traders better align their use with power availability, who now do so on gut feeling or low-quality coarse-grained forecasts. We hope this will increase the value enough to make transitioning to renewables more financially appealing.
Most standard machine learning approaches struggle in particular with price forecasting due to the limited data, large number of factors, heavy-tails, high noise, and underlying complexity; even given the bids for each producer and consumer, solving for the prices across a power network taking into account transmission, energy balance, and AC power flow constraints relies on an NP-hard mixed-integer programming problem that can take hours to solve. Of course in reality we don't even know the bids ahead of time, and we still haven't won the battle against the heavy tails today!
Our pilot experiences with a major East Coast utility looking to trade power, a major New England utility managing their demand response program, a battery storage operator in Texas, and a wind trader in Texas, have shown us that every participant has differing needs for their particular asset collection, so we dedicate time to each of our customers to make sure that the product is tailored to their needs. Along the way we've developed a generic forecasting system tuned for power markets to speed up customization, but we know we have a long way to go before we support the full range of forecast granularity, location, range, risk metrics etc we've heard interest for. With over 3000 market participants operating in open electricity markets (including Texas, California, New York, New England, and the mid-Atlantic), we’re hoping to hit 7 figure revenue within the next year.
We need huge amounts of storage to facilitate a transition to zero carbon grids long term, so we hope to minimize risk and maximize the reward for building new storage assets.
We’d love to hear your thoughts, questions, and comments!
by Andrey
Azimov