State-based load profile generation for modeling energetic flexibility

Förderer and Schmeck Energy Informatics (2019), 2(Suppl 1): 18 https://doi.org/10.1186/s42162-019-0077-z Energy Informatics RESEARCH Open Access State-based load profile generation for modeling energetic flexibility Kevin Förderer1* and Hartmut Schmeck1,2 From The 8th DACH+ Conference on Energy Informatics, Salzburg, Austria. 26-27 September, 2019 *Correspondence: 1 FZI Research Center for Information Technology, Haid-und-Neu-Str. 10–14, 76131 Karlsruhe, Germany Full list of author information is available at the end of the article Abstract Communicating the energetic flexibility of distributed energy resources (DERs) is a key requirement for enabling explicit and targeted requests to steer their behavior. The approach presented in this paper allows the generation of load profiles that are likely to be feasible, which means the load profiles can be reproduced by the respective DERs. It also allows to conduct a targeted search for specific load profiles. Aside from load profiles for individual DERs, load profiles for aggregates of multiple DERs can be generated. We evaluate the approach by training and testing artificial neural networks (ANNs) for three configurations of DERs. Even for aggregates of multiple DERs, ratios of feasible load profiles to the total number of generated load profiles of over 99% can be achieved. The trained ANNs act as surrogate models for the represented DERs. Using these models, a demand side manager is able to determine beneficial load profiles. The resulting load profiles can then be used as target schedules which the respective DERs must follow. Keywords: Smart grid, Flexibility, Distributed energy resources, Demand side management, Machine learning Introduction With the growing use of variable renewable energy, like wind and solar power, influencing electricity demand becomes increasingly relevant for balancing electricity supply and demand. Distributed energy resources (DERs), such as battery storage systems (BESSs) and combined heat and power plants (CHP plants), are sources of flexibility that may be used by a demand side manager (DSMgr) to steer electricity demand. Following the notion of Bremer et al. (2010), Mauser et al. (2017) and Sawall et al. (2018) the energetic flexibility of a DER can be understood as the set of load profiles the DER is technically able to attain while performing its duties, i.e., satisfying all constraints. Each load profile in this set is called feasible. While this understanding of energetic flexibility is by no means restricted to electricity, we focus on electric load profiles in this paper. However, the addition of further commodities is rather simple. In order to exploit the flexibility of DERs a DSMgr needs to know how their operation may be influenced. The primary goal of the approach presented in this paper is to enable a DSMgr to plan and steer the behavior of diverse DERs that are managed by a © 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. Förderer and Schmeck Energy Informatics (2019), 2(Suppl 1): 18 local energy management system (EMS). This is achieved by providing a method to generate load profiles with a high likelihood of being feasible. These load profiles can then act as target schedules and be communicated to the respective EMSs. The algorithm for generating load profiles uses special models that allow a targeted search for feasible load profiles. Hence, the DSMgr is able to shape the load profiles according to their needs. The employed models in combination act as a model for the energetic flexibility of the represented DERs. In this paper, we use machine learning approaches, and more precisely artificial neural networks (ANNs), to create the required models. The major benefit of our approach is the generic representation of flexibility, allowing to use a single interface for all kinds of flexible devices and even aggregates of multiple DERs. Also, by using machine learning models, the models for future applications may potentially be learned from data directly captured from real DERs, eliminating the need for handcrafting and formulating physical models. All scripts and models, including the simulation models for generating the training data and the neural models, as well as the results presented in this paper have been published on GitHub (see “Availability of data and materials”). The paper is structured as follows: An overview of and a comparison with related approaches and applications of similar models in the context of electric load profiles is given in “Related work”. The “State-based load profile generation” section introduces the approach for generating feasible load profiles investigated in this paper. The simulation models used for generating the training data and the chosen parameters are presented in “Simulation models and parameters”. “Neural models” presents the ANNs used in the evaluation of the approach. “Evaluation setup” provides further details of the implementation of the load profile generation process that has been used for evaluating the approach. Results of the evaluation are presented in “Results”, which is followed by a “Conclusion”. Related work Load profiles can be generated in various ways and for different reasons. In Hoogsteen et al. (2016) a bottom-up approach is used to create a household load profile generator for evaluating demand side management approaches. Among other things, the generator supports different household configurations, occupancy profiles and several classes of flexible devices. Markov chains are another option for generating load profiles. For example, a Markov chain with 24 states is used in McLoughlin et al. (2010) to generate domestic load profiles for households in Ireland. Hidden Markov models are used in Akkaya et al. (2016) to generate randomized control sequences for lighting appliances. In all cases the respective goal is to produce load profiles or control sequences that are similar to real ones, e.g., in terms of statistical properties. More examples for Markov models can be found in Tao et al. (2017), including a Markov chain model that can be used to evaluate the capacity and estimate the availability of a BESS that stores photovoltaic generation (Song et al. 2013). In contrast, the goal of the approach presented in this paper is to allow some external party, namely the DSMgr, to explore and select load profiles that are likely to be feasible. This is achieved by explicitly estimating the state of the represented DERs at any considered point in time. While the same is possible using finite-state machines (Costanzo et al. 2012) and other (...truncated)