🤖 AI

EnergyBridge: Benchmarking Household Energy Management, User Participation, and Grid Flexibility

The paper introduces EnergyBridge, a novel benchmark and agent framework that integrates region-specific energy simulations with an LLM-based user participation simulator to holistically evaluate and optimize household energy management by bridging the gap between grid capacity reporting, resident authorization, and physical demand response execution.

Original authors: Xudong Wu, Zeqing Wu, Jiarui Zhang, Xuhao Fan, Ziang Ding, Yuming Zhuang, Mingqi Yuan, Yilun Du, Hongjie Jia, Yunfei Mu, Jiayu Chen

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xudong Wu, Zeqing Wu, Jiarui Zhang, Xuhao Fan, Ziang Ding, Yuming Zhuang, Mingqi Yuan, Yilun Du, Hongjie Jia, Yunfei Mu, Jiayu Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a giant, bustling city where power plants are the water towers and our homes are the taps. Usually, water flows one way: the tower pushes it out, and we turn on the faucet. But what if the tower gets too full or runs low? In the old days, they'd just build bigger towers. Today, scientists are trying a different trick: asking the taps to close a little bit, just for a few minutes, to help balance the flow. This is called "demand response." It's like a neighborhood agreeing to turn off their sprinklers during a drought so the whole town doesn't run dry.

To make this work, we need "Virtual Power Plants" (VPPs). Think of a VPP not as a physical building with smokestacks, but as a super-smart digital conductor. It gathers thousands of tiny promises from homes—like "I'll delay my washing machine" or "I'll charge my electric car later"—and treats them as one giant battery. But here's the catch: a promise is only as good as the person making it. If a home has the physical ability to turn off their air conditioner but the person inside refuses because they're too hot, the promise is broken. For a long time, computer programs could calculate the physics perfectly but were terrible at guessing what humans would actually agree to do.

This is where a new study called EnergyBridge steps in. The researchers built a digital playground to test a new kind of "smart agent" that acts like a polite, understanding neighbor. Instead of just shouting orders at your appliances, this agent asks, "Hey, if we shift your dishwasher to 8 PM instead of 6 PM, will that be okay with you?" It then simulates the whole process: asking for permission, getting a "yes" or "no," and then actually running the appliances to see if the energy savings happen.

The team tested this agent in two very different digital cities: one modeled after Tianjin, China, and another after Berlin, Germany. They pitted their new agent against older, stricter computer programs and other "smart" helpers. The results were clear: the new agent was much better at getting people to say "yes" to the plan. In fact, in the simulation, it achieved a 40.0 percentage point higher acceptance rate in Tianjin and 42.9 percentage points higher in Berlin compared to the closest competitor.

But the magic didn't stop at just getting a "yes." The agent also learned to predict exactly how much energy it could actually save. When the researchers checked the numbers, the agent's predictions were spot-on. For the events that were accepted, 90.0% of the time, the actual energy saved was within the 20% tolerance range the agent promised. This is a big deal because it means the VPP can trust the numbers. The agent didn't just guess; it used a "memory" of past interactions and a deep understanding of each household's unique habits—like whether they have a baby to put to bed or a car that needs charging by morning.

The researchers also wanted to make sure their digital "neighbor" wasn't just making things up. They invited 584 real humans to play a role-playing game where they acted as different types of homeowners. When these real people were asked if they would accept the agent's plan, they said "yes" at almost the exact same rate as the computer simulation did. The difference between the human answers and the computer's guesses was tiny—only a 5.3-point error on average. This suggests the simulation is a very reliable way to test these ideas before trying them in the real world.

In short, EnergyBridge shows that the future of a flexible, green grid isn't just about better wires or smarter thermostats; it's about better conversations. By building agents that respect human comfort and habits, we can turn our homes into reliable partners for the grid, saving energy without making anyone uncomfortable. The researchers have even shared their code and data, inviting others to build on this bridge between human choices and electrical needs.

Technical Summary: EnergyBridge

Problem Statement

Residential Virtual Power Plants (VPPs) aim to provide grid flexibility by aggregating household demand response (shifting HVAC, EVs, appliances, etc.). However, a critical gap exists in current evaluation frameworks: they typically assess physical controllability (whether a device can be shifted) while omitting human authorization (whether a resident will allow it) and reliable capacity delivery.

Existing benchmarks often treat households as passive assets or rely on aggregated historical data that fails to capture event-specific uncertainties like comfort preferences, domestic routines, and tolerance for automation. Consequently, a technically feasible control action may be revised or rejected by occupants, rendering the promised flexibility undeliverable. There is a lack of a unified framework that connects capacity reporting, household authorization, physical execution, and measured grid delivery within a single closed-loop evaluation.

Methodology: The EnergyBridge Framework

EnergyBridge is introduced as a benchmark and agent framework designed to bridge the gap between VPP requests and verified grid delivery. It operates at the household–VPP interface, modeling the complete "request-to-delivery" process.

1. The Agent Workflow

The EnergyBridge agent functions through a sequential, closed-loop process:

Pre-Event Capacity Reporting: The agent retrieves historical events (using region, household type, weather, and load context) to generate a data-driven estimate of reliable capacity before an event occurs.
Plan Synthesis & Personalization: The agent combines household memory (profiles, calendars, past feedback) with control tools (MPC, MILP, or rule-based schedulers) to generate a feasible device-level plan. It then generates a household-facing explanation.
Authorization Gate: The plan is presented to a User Participation Simulator. Household members (modeled as personas) evaluate the plan based on comfort, cost, and task constraints, resulting in an Accept, Revise, or Reject decision.
- Accept: The plan proceeds to execution.
- Revise: Feedback is stored, and the agent regenerates the proposal.
- Reject: The system reverts to an ordinary fallback policy.
Physical Execution: Authorized plans are executed in a high-fidelity EnergyPlus simulation environment (configured for Tianjin and Berlin), modeling HVAC, EV charging, water heating, and task appliances.
Verification: The realized delivery is compared against the pre-event capacity report to calculate accuracy and reliability metrics.

2. Simulation Environment

Physical Layer: Uses EnergyPlus 25.1 with region-specific building models (Tianjin following GB 50176-2016; Berlin following DIN 4108-4:2020-11). It simulates 10-minute timesteps for HVAC, EVs, water heaters, and task appliances.
User Layer: A User Participation Simulator based on Large Language Models (LLMs). It utilizes six-dimensional role cards (schedule, comfort, task, price, control, grid) to represent diverse household personas. The simulator is calibrated against 584 human role-play judgments to ensure alignment with real-world decision-making patterns.

3. Evaluation Metrics

The benchmark evaluates three distinct levels of flexibility:

Physical Feasibility ( $C_{phys}$ ): Theoretical capacity under technical constraints.
Authorized Capacity ( $C_{auth}$ ): Capacity after accounting for household consent ( $\alpha_{i,e}$ ).
Delivered Capacity ( $C_{del}$ ): Actual metered reduction after execution, accounting for the realization gap ( $\eta_{i,e}$ ).

Key metrics include simulated gate-acceptance rates, post-event user scores (1–5 scale), event-window energy consumption, and Reliable Capacity Coverage (the fraction of requests that are both authorized and accurately reported).

Key Contributions

End-to-End Benchmark for Human-Authorized Agents: EnergyBridge formulates residential flexibility as a closed-loop task requiring agents to reason over preferences, plan under constraints, secure consent, execute physically, and verify outcomes. It moves beyond isolated evaluations of language responses or control performance.
Human-Grounded Evaluation: The framework includes a calibrated User Participation Simulator validated against 584 human responses, achieving a 5.3-point mean absolute acceptance error in method ordering. It releases human data, code, and physical models for reproducible research.
Cross-Layer Flexibility Agent: The proposed agent integrates household grounding, constraint-aware planning, consent-gated execution, and retrieval-augmented capacity reporting, enabling the reliable conversion of potential flexibility into grid resources.

Experimental Results

The study evaluated EnergyBridge against five baselines: the HEMA agent, MPC Dynamic, Rule+MILP, PPO (Deep RL), and a standard fallback policy across two regions (Tianjin, Berlin) and five household personas.

Authorization & User Satisfaction: EnergyBridge achieved the highest simulated acceptance rate (77.1% vs. 37.1% for HEMA in Tianjin) and the highest user scores (4.21 vs. 4.00 for HEMA).
Grid Performance: It achieved the lowest event-window energy (35.2% lower than HEMA in Tianjin) while maintaining 100% task completion.
Capacity Reporting: EnergyBridge demonstrated superior Reliable Capacity Coverage (64.3%) compared to HEMA (25.7%) and MPC (17.1%), indicating it is more likely to both get consent and deliver the promised capacity.
Human-LLM Alignment: The LLM-based simulator successfully reproduced the aggregate human acceptance ordering (EnergyBridge > HEMA > MPC) with a 5.3-point mean absolute error, validating its use as a proxy for human decision-making in this context.
Ablation Studies: Removing household grounding components (feedback memory, calendar context, or personalized explanations) significantly reduced acceptance rates, confirming that personalization is critical for authorization, not just for generating explanations.

Significance and Claims

The paper claims that existing research often evaluates flexibility stages (reporting, authorization, delivery) in isolation, which fails to determine if household flexibility constitutes a reliable grid resource. EnergyBridge demonstrates that connecting these stages in a closed-loop process is essential.

The authors assert that:

Authorization is the bottleneck: Physical capability alone is insufficient; the ability to secure and sustain household consent is the primary determinant of reliable flexibility.
Context matters: Accurate flexibility reporting requires household and event context (preferences, routines) rather than simple aggregation.
Simulation is viable: A human-calibrated LLM simulator can effectively approximate aggregate human responses for benchmarking, provided it is validated against matched human data.

The work concludes that evaluating household authorization, physical execution, and capacity-reporting accuracy as a single process is necessary to advance the deployment of residential VPPs. The authors release all benchmark environments, human evaluation data, and baseline implementations to support reproducible, human-centered grid-flexibility research.