Capturing uncertainties of household decision making with machine learning in an agent-based model

Aug 4, 2026

The document presents an approach to model uncertainties in household-level decisions within national energy system simulations. It combines the agent-based model AMIRIS with stochastic optimization and a diffusion model to represent flexibility technologies, including PV and storage, heat pumps, and electromobility. Micro-models generate household operating time series under varying inputs, which train machine learning to predict aggregated load, while separate diffusion captures expansion dynamics up to 2045. Coupling enables system-level integration of uncertain individual behavior that would be computationally infeasible otherwise, and results summarize sector-specific findings and best ML performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Motivation

Methodology

Micro-model aggregation and ML training

Diffusion and stochastic optimization coupling

Results

Capturing uncertainties of household decision making with machine learning in an agent-based model Ulrich FREY(1), Evelyn SPERBER(1), A. Achraf EL GHAZI(1), Fabia MIORELLI(1), Christoph SCHIMECZEK(1), Stephanie STUMPF(2), Anil KAYA(2), Steffen REBENNACK(2) (1) Deutsches Zentrum für Luft-und Raumfahrt, (2) Karlsruher Institut für Technologie Motivation Precise modelling of uncertainties in energy system analysis is central for future energy systems. However, it is difficult to bring individual decisions to the system level. We combine the agent-based model AMIRIS (Schimeczek et al. 2023) with a stochastic optimisation and a diffusion model for three flexibility providing technologies, namely PV and storage, heat pumps (Sperber et al. 2020) and electromobility (Wulff et al. 2021) . These three micro-models are aggregated with machine learning (ML) to fit the system level. Methodology Each micro-model generates a time series for the households' operating decisions by varying the input, such as different weather years and optimising individual costs. This is the basis for training a machine learning model to predict the aggregated load from thousands of household decisions (see Figure 1) . Figure 1: Aggregation of individual household decisions by machine learning to predict the total load A separate diffusion model calculates the respective expansion dynamics of the three sectors. The supply side is modelled using a stochastic optimisation model (Rebennack 2014) . It enables the portfolios of the operators of power plant parks in Germany to be modelled up to the year 2045. The diffusion model, the stochastic optimisation, and AMIRIS are coupled to map the uncertainties of the expansion dynamics, the construction of power plants and the actor-based market behaviour. This makes it possible to integrate these uncertain individual decisions into a larger model at system level. This would otherwise be completely unfeasible for computational reasons. Results We will report on the individual results of all three models – PVS, heat pumps and electromobility. For example, the heat pump model shows a cost-optimised shift in electricity consumption without compromising thermal comfort (Sperber et al. 2020) . The investment decisions of the coupled runs and the best ML model for the aggregated load are also presented. Overall, the model coupling shows that individual decisions and their uncertainties can be modelled very well in national energy system analysis simulations. References Rebennack, Steffen (2014): Generation expansion planning under uncertainty with emissions quotas. In: Electric Power Systems Research 114, S. 78–85. DOI: 10. 1016/j.epsr.2014.04.010. Schimeczek, Christoph; Nienhaus, Kristina; Frey, Ulrich; Sperber, Evelyn; Sarfarazi, Seyedfarzad; Nitsch, Felix et al. (2023): AMIRIS: Agent-based Market model for the Investigation of Renewable and Integrated energy Systems. In: JOSS 8 (84), S. 5041. DOI: 10. 21105/joss.05041. Sperber, Evelyn; Frey, Ulrich; Bertsch, Valentin (2020): Reduced-order models for assessing demand response with heat pumps – Insights from the German energy system. In: Energy and Buildings 223, S. 110144. DOI: 10. 1016/j.enbuild.2020.110144. Wulff, Niklas; Miorelli, Fabia; Gils, Hans Christian; Jochem, Patrick (2021): Vehicle Energy Consumption in Python (VencoPy): Presenting and Demonstrating an Open-Source Tool to Calculate Electric Vehicle Charging Flexibility. In: Energies 14 (14), S. 4349. DOI: 10.3390/en14144349 .

FAQs

Why is capturing uncertainty in household decisions important for energy system analysis?

Precise uncertainty modelling is central for future energy system studies, but translating individual decisions into system-level effects is difficult. The approach integrates household uncertainty into larger simulations efficiently.

How is machine learning used in the modelling workflow?

Micro-models generate household operating time series by varying inputs and optimizing individual costs. Machine learning is then trained to predict aggregated load from thousands of household decisions.

What roles do the diffusion model and stochastic optimisation play?

A diffusion model calculates expansion dynamics for the three technology sectors, while stochastic optimisation models the supply side and operator portfolios up to 2045. Together with AMIRIS, they map uncertainties in expansion, construction, and actor-based market behaviour.

Aug 4, 2026

Upload by Emma Wilson