Scaling wind power forecasting for energy trading

September 22, 2020

Energy trading in liberalised markets is particularly interesting from the perspective of wind energy producers because of the non-dispatchable nature of wind. This means that wind energy producers need to forecast how much they will produce in the future in order to place their bids. The fundamental issue at hand is that over or under predicting the wind power generation will lead to balancing costs that result in penalties for the energy producer. This means that better forecasts can directly translate to economical savings.

Another important factor is that these balancing costs generally increase with wind penetration as shown in the image below. EWEA expects wind penetration to grow in the EU from the current 15% to 24% by 2030, so it’s expected for these balancing costs to further increase in the future.

Increase in balancing costs with wind penetration. Image source

Overall, there is a strong incentive for energy producers that are already in the liberalised market to keep on the lookout for the best forecasters, but also for producers entering the liberalised market to look for the best forecasters available. And this is only going to get more important as balancing costs and associated penalties increase due to wind penetration growth. 

Real-world case example

There is no doubt that better forecasts can result in saving for wind energy producers. So at Jungle, we are committed to developing the best forecasts available using state-of-the-art ML algorithms as discussed on a previous blog post. There, we mentioned a project with a large utility where we developed a model for day-ahead wind power forecasting that we will now further explore. 

The project focused on a portfolio of 0.5GW in Romania and the objective was to optimise towards the Normalised Mean Absolute Error (NMAE), which corresponds to the mean absolute error normalised by the installed capacity of the wind farm. The model developed by Jungle was able to improve the forecasting performance by 15% when compared to the utility’s average forecasters. 

Example of Jungle’s forecast against the Utility’s forecast

One of the motivations for the project is that the imbalance costs in Romania are significantly higher than in most European countries. This means that an improvement in forecasting accuracy can directly translate into economical savings. To make a retrospective calculation of the economical impact we can use the average imbalance cost in Romania of 9€/MWh and the European average capacity factor of 26% in 2019.  From these figures results that:

Reducing the NMAE by 15% for a 0.5GW portfolio leads to 154k€/year savings.

It is important to note that we considered a symmetric penalisation scheme, i.e. a surplus forecasting error is penalised the same way as an under-prediction error. This simplifies the calculations but isn’t necessarily what happens in the real world. That’s why at Jungle we develop full probabilistic forecasts as can be seen in the previous figure. In this project not only was the accuracy of our deterministic forecasts better, but also the calibration of our probabilistic forecasts, as can be seen in the figure below. 

Jungle’s calibration curve against the Utility’s calibration curve

The calibration curves in the image show for each of the predicted percentiles what is the actual data that was within that percentile. In this case, Jungle’s calibration curve outperforms the Utility’s because it is closer to the ideal calibration curve. This means that even in more complex scenarios with asymmetric penalisation our customers can count on our well-calibrated probabilistic forecasts to mitigate the risk involved in energy trading. We will explore this topic further on a future blog post, so stay tuned!

Wind fleet impact 

The real-world case shown above is an example of the many that happen at wind fleet scale. In fact, the energy trading balancing costs can easily scale to the millions of euros. The plot below shows the average yearly revenue loss (in million EUR) due to imbalance costs.

Yearly imbalance costs according to forecast solution NMAE and fleet size

We can see above that a 3 GW portfolio incurs an average balancing cost of 5 million EUR/year even when a good forecasting solution with an NMAE of 8% is available. For larger wind fleets these penalties quickly surpass the 10 million EUR/year. As expected, it is even of more importance to reduce the NMAE for large wind fleets since the financial impact quickly grows.

Jungle can help our customers capture back some of these costs by providing more accurate forecasts. In the plot below we show the cost reduction that may be achieved. For this example, we considered a base NMAE of 12% and across the y-axis we can see how different percentual improvements would translate to savings. 

Yearly imbalance cost reduction according to forecast improvement and fleet size

For example, an improvement of 15% which means going down from 12% to 9% NMAE, for a 3GW wind portfolio would translate to savings of 1 million EUR/year. Of course that for bigger portfolios and bigger improvements, which are possible depending on the current performance and on the specificities of the wind sites, quickly translate to much higher savings.

Main advantages for our users

  • Very low entry barriers
  • Powerful state-of-the-art AI models specifically tailored for your applications
  • Scalable solution to any wind turbine model and wind farm location
  • Trustworthy probabilistic forecasts

Get in touch

If you enjoyed our approach to power forecasting and energy trading, feel free to reach out! Let us help you reduce uncertainty from your operations! Send us a message to hello@jungle.ai, and we will be in touch shortly after! It’s that simple.

Keep an intelligent pulse on your assets.
start small, build big, together creating value from day one
Start Now