Every day, Norwegian hydropower producers solve almost the same scheduling puzzle from scratch, with no memory of yesterday's answer. An international research project led by SINTEF is looking at how AI could help the system learn from experience.
The water in Norway's reservoirs is one of the country's most valuable resources, and getting the most out of it depends on a daily decision: which generating units to run, and when. This is deceptively difficult because of the number of factors involved. Producers have to weigh the value of releasing water now against saving it for later, respect the physical limits of each watercourse, and honour environmental constraints, such as minimal flow rates for specific rivers. They have to do all this while positioning themselves in the energy and capacity markets. For decades, SINTEF's SHOP model has been the tool of choice for this problem, and is used by most Nordic hydropower producers.
Here is the catch: the tools don’t learn. Traditional optimisation methods solve each day's problem independently, accumulating no experience, even when it changes only slightly from one day to the next. The InterOpt project is built on the insight that this repetition is precisely what machine learning is good at.
Accuracy vs speed
At the heart of daily scheduling sits the unit commitment problem, deciding the on/off status of each generator across the hours ahead. Solving it accurately requires mixed-integer
programming (MIP), where binary on/off variables guarantee realistic, feasible schedules. This works but it is slow to compute. Relaxing those variables makes the problem fast, but the results can drift into the unrealistic: production below a unit's minimum, or schedules that break environmental rules.
Today's compromise is a fixed rule of thumb that suppresses the worst cases without really fixing them. InterOpt proposes instead to run the fast, relaxed model first, then use machine learning to predict what the accurate commitment decisions would have been. The speed of one approach, the realism of the other.
The potential gains are significant. Numerical studies across large-scale power systems have shown machine learning delivering speed-ups ranging from 2× to 260× with no observed loss in solution quality.
A consortium spanning three continents
A big part of solving a problem like this is assembling the right people to test it from every angle. InterOpt’s second workshop in Trondheim yesterday brought together teams approaching machine-learning-assisted scheduling from various angles. Alongside SINTEF and NTNU, The project draws in research groups from Spain (Universidad Politécnica de Madrid), Canada (Université du Québec à Chicoutimi) and Brazil (CEPEL), the last two bringing experience from two of the world's largest hydropower systems. The Norwegian hydropower producers ANEO, Hydro Energi, and Å Energi all have a seat at the table in InterOpt – keeping the work anchored to the realities of daily operation – with ANEO represented at yesterday's workshop.
InterOpt is a Knowledge-building Project for Industry, partly funded by the Research Council of Norway and running from 2024 to 2028. For an industry where a single percentage point of improvement translates into substantial returns, getting hydropower's planning tools to finally learn from experience could prove very valuable indeed.
