Using Artificial Intelligence to forecast flexibility potentials
The core of the project: a machine learning model that helps to predict flexibility potentials. With this goal in mind, sensor data was used to first of all determine the current values for the available flexibility of the individual plants across a defined testing period. This allowed training data for the prediction of flexibility to be collected over a period of almost a year. In the case of the P2H plants, an average flexibility of 40kW was determined for the 29 plants which were connected up to the demonstration unit. For the three CHP plants that were connected up to the demonstration unit, evidence of the availability of flexibility was provided in up to 9 percent of the times tested during operation.
The availability of flexibility is, however, dependent on the weather to a large extent because the intention is to balance out weather-dependent power generation and consumption through the systems provided – their utilisation will thus also be influenced by current and expected weather conditions. As well as the operating data collected, we therefore also recorded the corresponding weather data so that we could train a so-called support vector regression model to allow us to predict the flexibility of the individual systems. The method serves as a way of describing and quantifying statistical dependencies. The result: on the basis of weather forecasts for the coming three days, the trained model is able to make predictions with regard to flexibility potentials.
Integrating prognoses and measurement data
A crucial factor if the operation of the P2H and CHP plants is to be beneficial to the grid is the integration of the flexibility potentials into an overall system into which current electricity and gas network measurement data and plant data are also fed. In the DESIGNETZ project, this was provided by the so-called system cockpit, which serves as a central element for the exchange of data and the initiation of requests for flexibility. In addition to setting up the demonstration units, the project team from umlaut was also responsible for the realisation of the data connections to the system cockpit, as well as the implementation of requests for flexibility (see Figure 2).
Such aspects of data connection, data preparation and data management regularly come up as hurdles in data analytics and AI projects, aside from the actual development of the model, and have to be resolved by the project team. As well as having a good understanding of data and modelling, it is therefore crucial that those working on such projects also possess interdisciplinary knowledge and skills in the area of energy and systems technology if project implementation is to be successful.