For energy suppliers and large consumers
Know tomorrow's demand before the market closes at noon, including how sure the forecast is. Free, open-source forecasting software, set up for you by the sktime developers.
No obligation. We look at your numbers together.
How much power should we buy for tomorrow?
Sound familiar?
In Germany the day-ahead market for tomorrow closes at noon today. Every kilowatt hour you got wrong is then bought or sold at short notice, or settled at the imbalance price, which is very volatile. It shows up in five places: (Narajewski, Energies, 2022)
Day-ahead buying
Last week's pattern misses the weather, holidays, and changes in production. sktime forecasts demand for every hour of tomorrow, so you buy what you will actually use.
Imbalance costs
Every deviation from your schedule is settled afterwards at the imbalance price, which can swing sharply from one quarter hour to the next. sktime forecasts per quarter hour and shows the likely range, so you see where the risk sits.
Peak-load charges
Large consumers usually pay part of their grid fees based on their highest load of the year. A forecast that flags a peak a day early gives your team time to shift or switch off load.
Weather and holidays
Demand reacts to temperature, public holidays, and the days around them. sktime learns these patterns from your history and adds weather forecasts, so unusual days come as less of a surprise.
Changing load
Last year's data does not show these changes yet. sktime compares many methods on your most recent data, so the forecast keeps up with how your load really looks today.
Who it is for
Energy suppliers and balancing group managers
A better forecast for the whole portfolio means less balancing energy and fewer trades at short notice. You see the likely range for every hour of tomorrow before the market closes.
Large electricity consumers
A forecast per site shows tomorrow's demand and its peak before the deadline. Your energy team buys closer to what the site will use and sees expensive peaks coming.
What the numbers say
Studies and practice reports show what machine-learning forecasts achieve for energy demand. sktime offers many machine-learning forecasting methods in one open-source toolkit, so you can use them without building everything from scratch.
40%
In a test run, Fraunhofer IOSB reports cutting the average day-ahead error of the quarter-hourly balancing-group forecast for the regional grid operator Albwerk by about 40% compared with its previous provider. Albwerk says the balancing energy volume and cost fell as well.
ZfK, 202428%
A neural-network forecast for the next day saved a large Italian industrial company about 28% of its balancing costs compared with the method it used before.
Salvatori et al., 20192.4%
On German national power demand, the best transfer-learning neural network in the study was off by 2.4% on average. A simple forecast that repeats last week was off by 4.3%. A single site is harder to forecast than a whole country.
Tzortzis et al., 2023These figures come from studies and trial reports on machine-learning forecasting, not from sktime projects. Results depend on your data and on how good your current forecast already is. We work out what is realistic for your business in a first conversation.
sktime is open source: no license fees per user or site.
Your numbers
A back-of-the-envelope estimate. Enter your annual power consumption and what a miss costs you per megawatt hour, and see a rough range in euros. The rest are our assumptions, based on public studies.
–/ year
– of your power bill at €89/MWh
Not included: lower peak-load charges, less risk, and the cost of setup and support. After a first conversation we give you a fixed price for a trial project.
Three steps. Consumption and price gap come from you. The percentages are fixed assumptions, based on the sources below.
1. Energy a simple forecast gets wrong
annual consumption × 4% error of a simple forecast
4% is the average day-ahead error of a simple forecast that repeats last week, for Germany as a whole. A single site or a small portfolio is harder to forecast, which would raise this figure. If you already have a good forecast, your current error is lower and so is the saving. At 500 GWh, 4% is 20 GWh a year bought or sold at short notice.
2. What it costs you
energy a simple forecast gets wrong × your extra cost per MWh
The extra cost is what you pay on top of the day-ahead price when you fix a miss at short notice. It depends on your market and contract, so you set it yourself. We start at €15/MWh as an illustration. For scale, the German day-ahead price averaged €89/MWh in 2025.
3. What better forecasts save
cost of errors × 25 to 40% less
At the grid operator Albwerk, forecast errors fell by about 40% against the previous provider. Costs do not fall one-to-one with errors. A manufacturer in Italy saved about 28% of its balancing costs. We use 25% (cautious, our own assumption) to 40% (optimistic) as the range.
Annual savings = step 3. The expected value is the midpoint of the cautious and optimistic case.
Simplifications: errors in both directions are treated alike, every megawatt hour of error costs the same, and the cost of errors falls by the same percentage we use in step 3.
This is an illustration, and we do not promise these results. The reduction rates are based on reports on other companies, not on sktime projects. The baseline is a simple forecast, so the saving is smaller if your current forecast is already good. The extra cost of a miss varies a lot between markets and contracts. In the call we will explore options for your own use case.
Before you ask
How we start
We look at your data and at where forecast errors cost you the most today. From that, we set the scope and a fixed price for the pilot.
Over two to six months, we build a first forecasting system on your own load data and compare it with your current forecast. You get benchmarks, all the code, and a report, and decide from there how to continue.
The forecasts go into your daily buying and scheduling. Under a service agreement, sktime Enterprise Solutions maintains the system and stays your contact when questions come up.
Your first step
sktime is free and open source. In a free 60-minute call, the people behind sktime look at how you buy power and plan load today, and what better forecasts could be worth for your business.
sktime is downloaded over a million times a month and built by more than 500 contributors worldwide. Its core team now sets it up for you.
No obligation, no sales pressure.
What happens in the call
How you buy power and plan load today, and where forecast errors cost you the most.
What you already have, and what a first forecast would need.
A first, rough estimate of what it could be worth, and what a pilot would look like.