As part of the efforts to reduce greenhouse gas emissions, thus mitigating climate change, large and stable sources like nuclear, coal, and steam-gas power plants are increasingly being replaced by numerous renewable sources on a smaller scale. Photovoltaics and windmills may be able to produce with zero emissions, however, they can’t be forced to produce energy when it’s needed to meet demands. And yet, the share of renewable sources in gross electricity consumption rose from around 16% in 2004 to 41% by 2022. In the Czech Republic, for example, this share last year was only 15.5%. However, according to the state-approved climate-energy plan, it should double by 2030. And that means maintaining the energy network’s stability will become more complicated once again.
Loads of available tools
Quality forecasts can make this task easier for TSOs. They also offer cost savings created by the imbalances between production and consumption. With the share increase in renewable resources, the expenses for the so-called “performance balance” rise significantly, and the customers are the ones then paying the difference in their invoices. At the same time, Europe already suffers from expensive energy, so every efficiency improvement counts.
Today, we’re capable of making forecasts with the help of more and more advanced information technologies. More powerful computers expand our possibilities, along with the introduction of AI and machine learning. Many methods exist today aside from classic regression models, such as random decision forests and neural networks. At Unicorn, we work together with our Unicorn University to pursue their applicability for forecasting imbalance nettings (IN) in the energy network. The result proved that interesting improvements can be made by introducing AI. During the monitored time period, the forecasting ability of our model was 14% better than that of older solutions. The forecasted period was the three following quarter hours, which is the period based on the new fifteen-minute trading interval. As expected, the results were the best during the first fifteen minutes, while the error rate grew at later times. Of course, precision wasn’t the only parameter; we were also interested in the robustness of the models. In this parameter, the robustness of the neural networks proved problematic, as they failed more often at later times.
Frequent training is necessary
Models based on machine learning take priority because they know how to develop continuously. Of course, it’s necessary to allow them to train in such a way that they can constantly take into account the fact that dependencies between input parameters can change over time. Consider the following as an illustration: the models used by real estate agencies to forecast the developments in stock prices are trained on a day-to-day basis. To a certain extent, we drew inspiration from the world of finance when examining the methods that could help improve IN forecasts.
We also can’t leave out the fundamental significance of the input data’s quality. Our testing clearly showed that a larger amount of data would help refine the results, especially with a finer geographical breakdown and a smaller time resolution. For instance, information about the weather that would impact energy consumption is fundamental as well as the renewable resources’ performance. Compared to the regional level, more detailed information on sunlight, temperature, and wind would enable one to better consider the local differences and improve the production forecasts from regionally dispersed renewable sources. In this same way, a longer time series of the actual quarter-hourly IN or more precise data on the actual production of photovoltaic power plants would help more than hourly information for the entire zone. Luckily, refining the database shouldn’t be an unsolvable task.
The energy sector is going through a fundamental transformation and the entire system is becoming more and more complicated. The most modern computing methods combined with AI aren’t a miracle solution to all the resulting problems. When it comes to forecasting INs in the energy network, however, they can bring about significant improvements. And yet, their application in this area is still in the beginning stages. That’s why at Unicorn, we believe in this path and know how to guide anyone interested from the energy sector through the case study phase based on academic research, onward to data prep and integration after creating the model, all the way to training and maintaining it during operation. Our prerequisites include many years of experience in software development and our thorough knowledge of the energy sector that we’re co-creating together in Europe.
This is a featured article.
