Can energy system modeling benefit from artificial neural networks? Application of two-stage metamodels to reduce computation of security of supply assessments

Computers & Industrial Engineering, vol. 142, pp. 106334

Abstract

Assessments of security of electricity supply are of high necessity for decision-makers in politics and industry. Based on scientifically sound and reliable advices, better decisions can be made concerning potential interventions in the energy system. With growing shares of intermittent renewable energy generation, probabilistic methods with high temporal resolution are increasingly important to simulate key indicators of security of electricity supply. However, these simulations are computationally complex and thus very time demanding. For this reason, we propose a two-stage metamodeling approach to reduce computational effort while maintaining high levels of accuracy. In the first step, we represent probability distribution curves of available power plant capacities using sigmoid functions. In the second step, we approximate the relevant regression coefficients. We apply both, linear regression methods and more advanced approaches based on artificial neural networks in this second stage. Our results indicate that computational time can be reduced from ~10 h to ~2 min. Further, the accuracy of the linear regression approach is relatively high and does not comprise the interpretation of relevant key indicators for the assessment of security of electricity supply. However, our results also show that no further accuracy gains can be achieved using artificial neural networks, indicating strong linear relationships. Overall, our approach reduces modeling complexity and therefore allows the investigation of a higher number of different scenarios, thus allowing for deepened insights about the future energy system. Based on our results, we can state that rather simple, but tailor-made approximation methods can outperform more sophisticated approaches if the latter are not suitable for the specific use-case. For future research, the application of artificial neural networks to approximate non-linear relationships within the assessment of security of electricity supply is highly recommended.

Authors 5

1. Lars Nolting corresponding Aachen E.ON Energy Research Center Institute for Future Energy Consumer

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, E.ON Energy Research Center, Institute for Future Energy Consumer, Needs and Behavior (FCN), Mathieustrasse 10, 52074 Aachen, Germany

2. Thomas F. Spiegel corresponding

   Hochschule Düsseldorf University of Applied Sciences

   Affiliation as printed

   University of Applied Sciences Duesseldorf, Centre of Innovative Energy Systems, Muensterstr. 156, 40476 Duesseldorf, Germany

3. Marius Reich corresponding

   Hochschule Düsseldorf University of Applied Sciences

   Affiliation as printed

   University of Applied Sciences Duesseldorf, Centre of Innovative Energy Systems, Muensterstr. 156, 40476 Duesseldorf, Germany

4. Mario Adam corresponding

   Hochschule Düsseldorf University of Applied Sciences

   Affiliation as printed

   University of Applied Sciences Duesseldorf, Centre of Innovative Energy Systems, Muensterstr. 156, 40476 Duesseldorf, Germany

5. Aaron Praktiknjo corresponding Aachen E.ON Energy Research Center Institute for Future Energy Consumer

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, E.ON Energy Research Center, Institute for Future Energy Consumer, Needs and Behavior (FCN), Mathieustrasse 10, 52074 Aachen, Germany