Modeling and Communicating Flexibility in Smart Grids Using Artificial Neural Networks as Surrogate Models

Increasing shares of renewable energies and the transition towards electric vehicles pose major challenges to the energy system. In order to tackle these in an economically sensible way, the flexibility of distributed energy resources (DERs), such as battery energy storage systems, combined heat and power plants, and heat pumps, needs to be exploited. Modeling and communicating this flexibility is a fundamental step when trying to achieve control over DERs. The literature proposes and makes use of many different approaches, not only for the exploitation itself, but also in terms of models.

In the first step, this thesis presents an extensive literature review and a general framework for classifying exploitation approaches and the communicated models. Often, the employed models only apply to specific types of DERs, or the models are so abstract that they neglect constraints and only roughly outline the true flexibility. Surrogate models, which are learned from data, can pose as generic DER models and may potentially be trained in a fully automated process.

In this thesis, the idea of encoding the flexibility of DERs into ANNs is systematically investigated. ... mehrBased on the presented framework, a set of ANN-based surrogate modeling approaches is derived and outlined, of which some are only applicable for specific use cases. In order to establish a baseline for the approximation quality, one of the most versatile identified approaches is evaluated in order to assess how well a set of reference models is approximated. If this versatile model is able to capture the flexibility well, a more specific model can be expected to do so even better.

The results show that simple DERs are very closely approximated, and for more complex DERs and combinations of multiple DERs, a high approximation quality can be achieved by introducing buffers. Additionally, the investigated approach has been tested in scheduling tasks for multiple different DERs, showing that it is indeed possible to use ANN-based surrogates for the flexibility of DERs to derive load schedules. Finally, the computational complexity of utilizing the different approaches for controlling DERs is compared.