03.12.2019 | Axpo is using data science to support the marketing of wind power across Europe
Axpo has developed a machine-learning application that provides traders with a risk-optimised trading recommendation based on the latest forecasts. Axpo is thus marketing the wind turbines’ expected energy production with minor deviations from the actual energy volume produced. This results in higher returns overall.
The opportunities offered by advanced analytics are also playing an increasingly important role in energy trading, and this is precisely why Axpo has launched what is known as a ‘Big Data project’ to actively exploit this development’s potential. In Spain alone, Axpo is currently marketing a wind power portfolio of 3250 MW. Machine-learning models are being used to support the most profitable sales possible on the market.
A wind farm produces different amounts of electricity. How much exactly depends on the weather, the turbines’ condition, the location and other criteria. Traders must therefore make assumptions about how much electricity will potentially be generated and then post their production forecast on the so called spot exchange one day in advance, on the day-ahead market. At the same time, they report the expected production output to the respective grid operator. If the wind farm produces more or less energy in reality than predicted, Axpo is penalized in the form of balancing power costs. Axpo therefore has a keen interest in reducing these forecasting errors and avoiding the consequential costs. But how does it go about doing this? It does so primarily through trading activities on what is known as the ‘intraday market’, where production changes can be managed until shortly before delivery.
This is why an application that calculates more accurate wind power forecasts and also optimises the trading strategy was developed as part of a Big Data project. The project is being managed by Gaudenz Koeppel, Head Models & Optimisation in the Trading & Sales division. The specifically developed algorithm recalculates the forecasts on an hourly basis using latest weather and market data, prepares all the available information and as such supports traders decision taking. Because the application uses machine learning and automatically retrains every night, the resulting trading recommendations evolve continuously and gradually adapt to market conditions.
The application has been used at Axpo Iberia since April 2019. The results are promising to the extent that the model has already been adapted for Italy and the Scandinavian countries, where it is currently in test operation. The models run on a cloud-based platform, that – among other benefits – is also a suitable enabler for the cross-divisional cooperation.
Professional information processing is a key component of successful energy trading. The developments in advanced analytics – the prediction of future events and behaviours using machine learning – are creating new opportunities that have already been rated as significant for Axpo’s trading business. As the advantages and initial results are convincing, an analytics strategy has been adopted in the meantime, aiming at integrating advanced analytics in trading in a coordinated manner. The implementation of this strategy is already ongoing. The project is also part of Axpo Group’s digitalisation portfolio, so the other business areas can benefit from the gained insights.
1. To what extent does the use of advanced analytics differ from today’s activities?
To a large extent, advanced analytics is just the consequent development of existing approaches. A major difference, however, is that building machine-learning models is a very time and resource-consuming task, whereas the effective daily operation requires only a few hours per week. With the Big Data project, we are providing resources, best practice and expertise to the business, such that these new methods can be developed efficiently and but also ready to be scaled to interested business units. In other words, we want to pool the respective efforts and make them available to the entire company.
2. What makes our model so special?
The model is programmed in a way that we can configure it for all the countries in which we operate wind farms. This means that we basically have the same model throughout the entire company, but configured for the respective local market rules. If we improve the model, it will benefit all the concerned countries immediately.
3. Will there still be a need for traders in future if a model is calculating the best recommendations?
Yes, absolutely. We always need people who understand the market and are capable of defining suitable model requirements. Pattern recognition is one of the integral elements of machine learning. In energy trading, the amount of available data – and thus reliable patterns – is far lower than e.g. for a self-driving car that has to process several images per second. In addition, market behaviour is continuously changing due to human actions. The model therefore has to be regularly calibrated and fine-tuned to market changes by the traders. The interaction between model and trader is thus increasing. As a result, new requirements emerge, which is why we are currently developing targeted training courses to enable our colleagues to efficiently and profitably use the models in day-to-day operations. At the same time, we also hope that this will lead to a valuable dialogue between traders and data scientists with regard to developing new models.
Advanced analytics basically refers to the use of technologically advanced methods from the field of data science, i.e. structured and partially autonomous analysis of data and unstructured information in combination with machine learning or neural networks. The use of advanced analytics allows companies both to identify the patterns of past events but also to calculate forecasts or decision recommendations. With the help of machine learning, for example, possible scenarios can be identified, including the ideal reaction in each scenario, and the moment a decision needs to be taken, only the matching scenario has to be identified to know the optimal decision. While this certainly saves time, it also creates awareness of the spectrum of possible changes.
Machine learning is an important field in computer science and a component of artificial intelligence. Algorithms based on machine learning learn certain patterns from existing data sets, then use them to classify and/or describe potential future behaviours. Machine-learning programs have to relearn at regular intervals to ensure that patterns are always up to date and representative.