Get in touch

How AI forecasts handle volatile energy prices

September 16th, 2021

With energy prices across the Nordics and Europe becoming ever more volatile, we take a look under the hood of Montel AI's Spot price forecasts. With the help of data scientist Tobias Foslid, we analyse where the model performed well, where it performed badly, and how it learnt from its mistakes.

Pattern Recognition, Artificial Neural Networks, Evolutionary Algorithms, Monte Carlo Methods, Machine Learning... 

Over the past few years AI and machine learning have crept into the energy sector, with their presence especially noticeable in market analysis and forecasting. Proponents say that AI offers a unique ability to generate predictions that learn from errors and increase in accuracy with each repetition. Detractors and some traditional analysts say that machine learning is unsophisticated and susceptible to errors.  

This is not a debate that we’re able (or intending) to resolve in this blog; indeed, we can expect it to continue for many years to come. What is interesting however, is to assess how AI predictions perform, especially under volatile circumstances.  

With that in mind, we spoke to our colleagues at Montel AI to see how their model has performed over a particularly volatile period (summer 2021) for both European and Nordic power markets. They came back with three examples which show not just how the model performs well, but also where initial predictions were off and how it learns from its mistakes. 

Example 1 – Low Volatility Day, Nordic System Price 


The first example is from July of 2021, and a day when there was limited volatility. ”This day (07/07) we saw wetter weather forecasts, but a hydrological deficit of 5.7TWh,” explains Tobias Foslid, a Montel AI data scientist. The expectation was for fairly stable prices, which is what the Montel AI prediction accurately forecast, with a relative error of only 5%.” The Montel AI prediction was very close to the actual price, despite fundamental outlooks being typically below the actual price on that day,” he adds. This shows how AI models can handle conflicting factors when predicting prices.  

Tobias also reminds us that it’s sometimes difficult to shed light on what’s happening inside an AI model, given the way that they function.  “We’re not domain specialists,” he says. “For us it’s all about the numbers.” It’s an important consideration to bear in mind: AI focuses on a learning process, rather than tailoring inputs or datasets.

Example 2 – High Volatility Day, NO2 


The next example is proof of that statement. On 31/08/21, a combination of factors (dry weather, hydrological deficit, scheduled maintenance for nuclear power plants, bullish wind forecasts, high prices in neighbouring continental markets) meant that extreme and ‘peaky’ spot prices were experienced in the Nordic region, and especially in the NO2 zone. The prices were (at that time) the highest recorded in eight and half years.  

“We see that the AI model predicts increasingly high prices, but it fails to settle at the correct level, or account for the strong daily peaks,” says Tobias. There’s one potential explanation, he thinks. This summer was the first year in which the Nordlink (NO2-DE) cable has been operational, which meant that continental price peaks had an especially strong impact on the NO2 zone. Flows for the cable are still limited in most situations because of bottleneck issues in the German grid. “The errors in the AI prediction might be partly due to the lack of training data that the model had for such situations,” he says. "But what makes AI unique is that it will automatically learn from these mistakes, that's part of the process."

(In general, the model performed well throughout August of that year, with an average error of €3.8 (5%) as prices averaged €72 in NO2.)

Example 3 – High Volatility Day, Nordic System Price 


For the final example, we move into early September 2021, and to another high volatility day, with conflicting factors of wetter weather and lower than usual hydrological balance driving uncertainty. As was the case throughout late summer, the market was taking cues from continental Europe and especially Germany. “This shows that the model is capable of learning from previous mistakes,” says Tobias, “It was able to perform with an average error of €4, or an hourly MAPE of 5.5% on a high volatility day.” 

Learning by doing 

These three examples show that AI Models can be effective predictive tools, even during periods of high volatility. Crucial to this is their ability to learn. Tobias is positive about the potential for AI to improve the energy market. “I think we’re just beginning to scratch the surface,” he says. “The potential for AI models to be used across energy markets – from production forecasts, inflow predictions, price estimates – is huge.” And the value they’ll bring will only increase as they are used more widely and we learn how to apply this versatile and flexible technology.  

The potential for AI models to be used across energy markets is huge. AI could be the solution for anything that would benefit from continuous improvement.

Alongside their spot forecasts (four of which are published in advance of each gate closure), Montel AI currently offers solar, wind and run of river production forecasts; as well as reservoir inflow,  consumption and load predictions. And they’re eager to do more, even outside of the ‘traditional’ energy market, says Tobias. “If you need a forecast for anything that would benefit from continuous improvement, then AI could be the solution.”