Towards domain-specific surrogate models for smart grid co-simulation

Balduin et al. Energy Informatics (2019), 2(Suppl 1): 27 https://doi.org/10.1186/s42162-019-0082-2 R ES EA R CH Energy Informatics Open Access Towards domain-specific surrogate models for smart grid co-simulation Stephan Balduin1* , Martin Tröschel1 and Sebastian Lehnhoff1,2 From The 8th DACH+ Conference on Energy Informatics Salzburg, Austria. 26–27 September 2019 *Correspondence: 1 OFFIS – Institute for Information Technology, Escherweg 2, 26121 Oldenburg, Germany Full list of author information is available at the end of the article Abstract Surrogate models are used to reduce the computational effort required to simulate complex systems. The power grid can be considered as such a complex system with a large number of interdependent inputs. With artificial neural networks and deep learning, it is possible to build high-dimensional approximation models. However, a large data set is also required for the training process. This paper presents an approach to sample input data and create a deep learning surrogate model for a low voltage grid. Challenges are discussed and the model is evaluated under different conditions. The results show that the model performs well from a machine learning point of view, but has domain-specific weaknesses. Keywords: Surrogate model, Deep learning, Co-simulation, Smart grid Introduction As a safety-critical infrastructure, the power grid is not suitable for testing new algorithms or other technologies. At the same time, there is no test system that is comparable to the power grid in terms of functionality and behavior (Nieße et al. 2014). Simulation and cosimulation are therefore key tools for the development and testing of new technologies and methods to transform the power grid into a smart grid. In general, a distinction is made between static and dynamic power grid simulation. In the static simulation, a steady state analysis is performed. In a dynamic simulation, however, transient effects can be observed. An important factor for simulation systems is the time required for the simulation. Frequent questions include: Does the simulation end in reasonable time? Is a step computable in real time (or even faster)? In small systems (e.g. a small distribution grid), runtime may not be a real issue. In larger simulation systems (e.g. the entire German transmission grid), however, time becomes a critical factor. A possible solution to the problem are surrogate models, i.e. data-driven approximations of the system to be simulated. The use of a surrogate model represents a trade-off between accuracy and runtime of the system. Many components can be involved in power grid simulation systems, such as photovoltaic (PV) panels, combined heat and power (CHP) plants, consumers such as households or commercial facilities. In steady state simulation, some components can be replaced by time series because their behavior is not © The Author(s). 2019 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. von Tüllenburg et al. Energy Informatics (2019), 2(Suppl 1): 27 controllable. However, others can reduce or increase consumption or generation to ensure a stable state of the grid. In order to simulate a typical low voltage (LV) grid, household models, PV or CHP models at household level, and models of smaller commercial facilities may be needed (Papathanassiou et al. 2005). The simulation of such a LV grid can take a while, even if some components are replaced by time series. The total runtime depends on the time resolution and the number of simulation steps. For instance, the simulation system used in this paper took about 15 min to simulate a year with time steps of 15 min. The next step for a large-scale simulation is to connect several of those LV grids to a medium voltage (MV) grid. Then PV or wind farms and larger commercial or industrial facilities are added (Buchholz et al. 2004). The resulting simulation system may not be simulated in reasonable time. Replacing individual components with surrogate models is a less promising approach. The components themselves are not that slow and the time-consuming part is the need to simulate a large number of components (Koch et al.). This paper proposes to replace whole LV grids with a single surrogate model built with a deep neural network. This includes the components connected to the grid. As described in Balduin (2018), the idea is to integrate domain knowledge and characteristics of the power grid into the surrogate modeling process. The grid has a large number of inputs with strong interactions, such as load and generation, which may depend on consumer behavior or weather conditions. Information about the simulation setup and interdependencies of the components should be used to reduce the dimensionality of the problem. The individual components are to be abstracted by a correlation model built from this information. The contribution of this paper is to provide a benchmark model and evaluation environment. This model is built without domain knowledge. The suitability of the model is investigated in various simulation experiments, which represent the evaluation environment. The rest of this paper is structured as follows: “Related work” section provides a brief introduction to surrogate models, deep learning, power grid simulation, and power flow calculations. In addition, related work in the field of surrogate models for the power grid will be presented. “Methodology” section presents the simulation setup and describes the construction of the deep learning model. Furthermore, the experimental setup and hypothesis are defined. In “Building the model” section, sampling and model building is documented. “Case study” section documents the conduct of the case study and “Conclusion and outlook” section concludes this paper and presents future work. Related work Surrogate modeling is well documented in the literature, e.g. in Myers et al. (2016); Kleijnen (2015), or Siebertz et al. (2017). An approximation function y = f (x1 , x2 , . . . , xk ) +  is called a surrogate model. The xi are the inputs, y marks an output, and  is the error between f and the true but unknown response function. A sample is an arbitrary, but fixed assignment of the inputs and the corresponding output of the real system. f is built with a sufficient number of samples, called the training data set. One way to create a surrogate model is machine learning, e.g. nearest neighbors, support vector machines, or artificial neural networks (ANN). But there are also other methods and each has advantages and disadvantages. In the German research project D-Flex (Koch et al.) a large simul (...truncated)