Register now After registration you will be able to apply for this opportunity online.
Large Language Model (LLM) Agent-Based Modeling (ABM) of Residential Energy Behavior under Flexible Load Management Scenarios
Household energy consumption patterns, accounting for ~25% of European electricity demand, play a pivotal role in demand flexibility to support the grids under increasing intermittent renewable generations. The specific patterns of household appliance usage and time preferences can be the complex consequence of asset and facility conditions, household economic status, resident occupational and recreational lifestyles, and local social-organizational context. We have also been working towards integrating household energy models with socio-economic survey data to emulate these complex and heterogeneous patterns. Agent-based modeling (ABM) via large language models (LLMs) is a promising approach to reflect individual household properties and simulate complex human-like reasoning, behavioral adaption, and interactions in this process via LLM. In this project, we will leverage existing LLM-agent frameworks to simulate Swiss households’ energy behavior using our collected demographic and time-use survey data, and gain understanding of populational behavioural shifts & individual reactions to different demand response policy and extreme weather scenarios.
**Background**
As global energy systems transition toward decarbonization, flexible load management becomes a critical mechanism for balancing intermittent renewable generation and ensuring grid stability. Household energy consumption patterns, accounting for ~25% of European electricity demand, play a pivotal role in this process. The specific patterns of household appliance usage and time preferences can be the complex consequence of asset and facility conditions, household economic status, resident occupational and recreational lifestyles, and local social-organizational context. As part of the UrbanTwin (https://urbantwin.ch/) project, we have also been working towards integrating household energy models with socio-economic models that emulate these complex and heterogeneous patterns.
Agent-based modeling (ABM) is considered an effective approach to reflect individual household properties and their interactions. However, conventional ABM often relies on simplified assumptions and rigid parametrization that fail to capture the nuanced decision-making processes of households or realistic reactions under extreme scenarios. Recent advances in large language models (LLMs) offer transformative potential. Unlike rigid conventional ABM, LLM agents can simulate human-like reasoning, behavioral adaption, and interactions. Existing studies already demonstrated LLM agents' capability to replicate complex social dynamics and multi-stakeholder interactions. Recent advances in the community of computational sociology and progress of opensource LLMs have enabled cost-effective deployment of thousands of agents for city-scale simulations while preserving individual heterogeneity and high fidelity to human behavior. Based on the household demographic information and time-use survey collected from the UrbanTwin studies, we plan to leverage LLM agents to simulate representative urban scenarios related to household demand flexibility and extract policy insights.
**Project Objectives**
This project aims to:
1. Leveraging existing scalable LLM-agent frameworks to simulate Swiss households' energy behavior using our collected demographic and time-use survey data.
2. Investigate behavioral shifts in the population and individual reactions under scenarios related to household demand flexibility, such as including dynamic tariffs Home Energy Management System (HEMS) adoption campaigns, and extreme weather-induced supply constraints, etc.
3. Derive policy insights by linking simulated agent decisions to real-world initiatives and interventions.
**Key Responsibilities**
- Configure agent profiles according to survey data and implement simulation using existing LLM-agent simulation platforms
- Design scenario-specific prompting strategies (such as virtual surveys) to emulate policy/event impacts and record agent reactions
(if planned as a Master thesis:)
- Validate simulations against populational energy use patterns from the collected survey data
- Identify emergent behaviors and conduct corresponding statistical analysis
**Candidate Requirements**
- Familiarity with Python; Experience with LLM APIs, Dockers, and machine-learning experiments can be a plus
- Experience with (or passion for) energy transition/social statistics & computation/ABM is preferred
- Strong autonomy in exploring simulation setups and troubleshooting computational challenges
**Benefits & Support**
- (Co-)authoring high-impact publications in energy/computer science/computational sociology journals/conferences
- Hands-on experience with cutting-edge LLM applications
- Access to dedicated clusters for large-scale simulations
**Background**
As global energy systems transition toward decarbonization, flexible load management becomes a critical mechanism for balancing intermittent renewable generation and ensuring grid stability. Household energy consumption patterns, accounting for ~25% of European electricity demand, play a pivotal role in this process. The specific patterns of household appliance usage and time preferences can be the complex consequence of asset and facility conditions, household economic status, resident occupational and recreational lifestyles, and local social-organizational context. As part of the UrbanTwin (https://urbantwin.ch/) project, we have also been working towards integrating household energy models with socio-economic models that emulate these complex and heterogeneous patterns.
Agent-based modeling (ABM) is considered an effective approach to reflect individual household properties and their interactions. However, conventional ABM often relies on simplified assumptions and rigid parametrization that fail to capture the nuanced decision-making processes of households or realistic reactions under extreme scenarios. Recent advances in large language models (LLMs) offer transformative potential. Unlike rigid conventional ABM, LLM agents can simulate human-like reasoning, behavioral adaption, and interactions. Existing studies already demonstrated LLM agents' capability to replicate complex social dynamics and multi-stakeholder interactions. Recent advances in the community of computational sociology and progress of opensource LLMs have enabled cost-effective deployment of thousands of agents for city-scale simulations while preserving individual heterogeneity and high fidelity to human behavior. Based on the household demographic information and time-use survey collected from the UrbanTwin studies, we plan to leverage LLM agents to simulate representative urban scenarios related to household demand flexibility and extract policy insights.
**Project Objectives**
This project aims to: 1. Leveraging existing scalable LLM-agent frameworks to simulate Swiss households' energy behavior using our collected demographic and time-use survey data. 2. Investigate behavioral shifts in the population and individual reactions under scenarios related to household demand flexibility, such as including dynamic tariffs Home Energy Management System (HEMS) adoption campaigns, and extreme weather-induced supply constraints, etc. 3. Derive policy insights by linking simulated agent decisions to real-world initiatives and interventions.
**Key Responsibilities**
- Configure agent profiles according to survey data and implement simulation using existing LLM-agent simulation platforms - Design scenario-specific prompting strategies (such as virtual surveys) to emulate policy/event impacts and record agent reactions (if planned as a Master thesis:) - Validate simulations against populational energy use patterns from the collected survey data - Identify emergent behaviors and conduct corresponding statistical analysis
**Candidate Requirements**
- Familiarity with Python; Experience with LLM APIs, Dockers, and machine-learning experiments can be a plus - Experience with (or passion for) energy transition/social statistics & computation/ABM is preferred - Strong autonomy in exploring simulation setups and troubleshooting computational challenges
**Benefits & Support**
- (Co-)authoring high-impact publications in energy/computer science/computational sociology journals/conferences - Hands-on experience with cutting-edge LLM applications - Access to dedicated clusters for large-scale simulations
Not specified
If you are interested, please submit an CV to us via E-Mail: vasantha.ramani@epfl.ch, yufei.zhang@epfl.ch
If you are interested, please submit an CV to us via E-Mail: vasantha.ramani@epfl.ch, yufei.zhang@epfl.ch
