The entire value chain in the energy industry is changing rapidly. Production- and business practices are being turned upside-down, as we witness several simultaneous shifts and changes. From the upsurge in popularity of renewable energy sources and the rise of the prosumer, to advances in technology and economical challenges. With such substantial changes come challenges, but also opportunities.
Over the last decade, Europe has seen a tremendous increase in the popularity and use of renewable energy sources. Although this has to be seen as a positive development, it remains a fact that some renewable energy sources, for instance wind and solar are challenging for producers. Being at the mercy of nature means that production may be hard to plan. Additionally the lack of good storage options presents another obstacle.
The consequences of these issues is that nominations become more difficult to place accurately, leading to high penalties for producers. The scale of the problem varies from country to country, one factor being how often producers receive updated weather data. In Norway, producers only get updated weather data every six hours. A lot can change in six hours and every year Norwegian producers pay vast amounts of money in penalties for placing inaccurate production nominations.
In close cooperation with Norwegian power producer TrønderEnergi, Powel has sought to solve the problem with wind through advanced data analytics. A six-week project that utilised advanced machine learning techniques in order to optimise wind power production has been hailed a success and has led to TrønderEnergi deciding on implementing this solution across all their wind farms.
The trouble with wind power
TrønderEnergi produce both hydro- and wind power and they currently operate four wind farms. For this particular project, the large Bessaker wind park was chosen due to its size and location. Situated on a steep hill in a mountainous area in the Trøndelag region of Norway, the surrounding terrain makes it their most complex wind farm when it comes to forecasting. The farm consists of 25 turbines with an annual production output of 175 GWh, making it the third largest wind farm in Norway. Additionally, a smaller wind farm, Valsneset, consisting of five turbins, was also part of the project.
TrønderEnergi wanted to investigate if machine learning could be applied to predict intraday production forecasts better than traditional methods. The aim was to reduce the imbalance costs typically associated with less accurate nominations. A first of its kind for both them and Powel, this was a high-risk project with an uncertain outcome. The goal was to make intermittent renewable energy production both more predictable and more profitable through machine learning and analytics.
“Planning wind production in Norway is difficult partly because the weather prognosis we work from is only updated every six hours. As a result, intraday and day ahead rarely have accurate enough predictions, resulting in loss of revenue. Although many companies, including us, have been working with wind power production for some time and are managing increasingly more accuracy in predictions, there is still a long way to go. This is essentially where the idea for this project came from,” says Magne Røen from TrønderEnergi.
Combining domain knowledge with machine learning expertise and predictive analytics, machine learning techniques were applied to see how we could optimise production further.
Data requirements for effective machine learning
When planning wind power production you need knowledge of when, how powerful and in what direction the wind is blowing. Looking at vast amounts of data, both weather-, production- and forecast, and using new algorithms, the search for previously undetected patterns in the wind began.
“The basic idea behind this project was that if we were able to find a pattern locally, we could create a mini weather prognosis centre linked directly to the wind parks. Hopefully, by doing localised analysis we would end up with much more accurate forecasts,” says Stein Petter Agersborg, Business Manager, Smart Energy in Powel.
The hypothesis was that by linking three sets of data sources; data for tomorrow's or the following hour’s weather forecasts for the wind farms, data from surrounding weather observation station, and data from previous weather reports and subsequent actual production, we could achieve a significant improvement in the discrepancy between reported and actual production.
Utilising machine learning, a data model was built and vast amounts of data was scrutinised to find new patterns. The data model compared weather data for two different years, starting with the most important variables and comparing measured data to predicted data, for example when it came to the speed of the wind.
Enabling utilities to change
The result of the project, at the end of six weeks, was very uplifting and showed great potential for the use of machine learning techniques in such cases.
“By applying niche predictive analytics techniques, we managed to reduce uncertainty in wind power production by more than 45%, resulting in the imbalance costs being more than halved. We achieved what we set out to do with this project, which was proving that utilising machine learning allows for complex predictions to be made with high accuracy,” says Agersborg.
Following the success of the project, Powel has taken it a step further. Combining this philosophy with advanced algorithms and the ability to process big data, we have developed a new software solution, which TrønderEnergi will use for all their wind power production.
Powel believe we are the first company to be able to offer such a solution to customers worldwide. It is further proof that Norwegian technological environments are both innovative and groundbreaking and has massive potential for contributing to green growth with accompanying green numbers.