GridAI — Battery Fleet Coordination Protocol

GridAI solves the herding problem in distributed energy resource (DER) coordination.

The Problem

When thousands of home batteries follow the same price signal, they all discharge simultaneously and create a new synchronised demand spike instead of smoothing the existing one. This is the herding problem — and it pushes voltage above legal limits (AS IEC 60038:2022 band: 0.94–1.10 pu) at the edge of the distribution network.

The Solution

A priority-based coordination protocol where the Coordinator allocates each battery's dispatch slot using global fleet state to desynchronise the fleet, producing a flatter aggregate demand curve while respecting voltage limits and owner preferences.

Critical design point: desynchronisation comes from fleet heterogeneity (varied private thresholds and SOC), not from negotiation alone. The protocol channels heterogeneity; on a homogeneous fleet it only weakly desynchronises.

Architecture

Agents propose. Physics decides. GridAI separates agent coordination from physical acceptance: agents propose dispatch schedules, but voltage/SOC/kW validation decides whether a schedule is accepted, repaired, or reported as residual risk.

flowchart TB
    subgraph Layer1["Layer 1 — Python Simulation Core"]
        direction LR
        F[Feeder<br/>LV radial, 60 homes] --> S[Simulator<br/>288×5min steps]
        P[Profiles<br/>Load, PV, battery] --> S
        ST[Strategies<br/>Naive / Gossip] --> S
        S --> O[Outputs<br/>scenario_*.json]
    end

    subgraph Layer2["Layer 2 — Band Agent Chain"]
        direction TB
        A1[Forecaster<br/>Analyses risk window] -- handoff:risk_window --> A2
        A2[Coordinator<br/>Runs gossip dispatch] -- handoff:plan+trajectory --> A3
        A3[Compliance<br/>Audits voltage breaches] -- handoff:escalation/approval --> A4
        A4[Operator<br/>Governance decision] -- broadcast:decision --> ALL
        BAND[Band SDK<br/>@mention routing<br/>Native audit trail] -.-> A1 & A2 & A3 & A4
    end

    subgraph Layer3["Layer 3 — HTML/Canvas Viz"]
        D[Dashboard<br/>Side-by-side playback<br/>Voltage sparklines<br/>Compliance cards]
    end

    O --> A1
    O --> D

Loading

Layers

Layer	Status	Description
Layer 1	✅ Done	Python simulation: LV feeder, battery agents, naive & gossip protocols, AEMO 2012 load data
Layer 2	✅ Done	Four agents (Forecaster, Coordinator, Compliance, Operator) collaborating over real Band SDK as the transport bus. Verified live against app.band.ai with full parity to mock.
Layer 3	✅ Done	Self-contained HTML/Canvas dashboard — side-by-side naive vs gossip playback, voltage sparklines, Band audit-trail panels, compliance cards. Deployed on GitHub Pages.

Key Results

Metric	Naive (price-following)	Gossip-style coordination
Battery-herding overvoltage events	471	0
Overvoltage steps (evening peak)	14 steps	0 steps
Max simultaneous discharge (synchrony)	1.000 (60/60 homes)	0.167 (10/60 homes)
Convergence rounds	—	1–2
Peak aggregate demand change (heterogeneous)	baseline	−0.9%

All numbers verified against AEMO 2012 Victorian summer data and backed by 89 automated tests.

Quick Start

cd gridai
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Run all scenarios and write outputs/
python3 sim/runner.py

# Run 4-agent Band chain (mock, offline)
python3 agents/run_agents.py

# Run tests
pytest tests/ -v

# Run final validation reproducibility report
python sim/runner.py          # ensures all outputs exist
python scripts/run_final_validation.py

Run against real Band

cp .env.example .env
# Fill in your Band API keys from app.band.ai
export USE_REAL_BAND=true
python3 agents/run_agents.py

Output Files

All simulation outputs written to outputs/:

• scenario_naive_homogeneous.json — herding baseline, identical thresholds
• scenario_naive_heterogeneous.json — herding baseline, varied thresholds
• scenario_gossip_homogeneous.json — protocol on homogeneous fleet
• scenario_gossip_heterogeneous.json — protocol on heterogeneous fleet
• scenario_*_aemo.json — same scenarios driven by AEMO 2012 Victorian load
• summary.json / summary_aemo.json — headline metrics
• band_audit_*.json / compliance_decision_*.json — agent chain records

Agent-chain outputs written to outputs/:

• band_audit_naive_aemo.json — full 8-entry Band audit log
• band_audit_gossip_aemo.json — same for gossip scenario
• compliance_decision_*.json — escalation/approval records
• band_parity_report.md — mock vs real Band parity verification

Project Structure

gridai/
├── sim/                  # Layer 1 — Simulation core
│   ├── feeder.py         #   LV radial feeder (N=60, prefix-sum voltage)
│   ├── profiles.py       #   Load, PV, battery profiles, make_homes()
│   ├── strategies.py     #   Naive & gossip protocols
│   ├── simulator.py      #   24h simulation engine
│   ├── aemo.py           #   AEMO 2012 Victorian load data
│   └── runner.py         #   Scenario runner
├── agents/               # Layer 2 — Band agent chain
│   ├── band_interface.py #   Abstract Band interface
│   ├── mock_band.py      #   In-process mock implementation
│   ├── real_band.py      #   Real Band SDK transport bus
│   ├── forecaster.py     #   Risk-window analysis agent
│   ├── coordinator.py    #   Gossip dispatch coordinator
│   ├── compliance.py     #   Voltage compliance auditor
│   ├── grid_operator.py  #   Human governance operator
│   └── run_agents.py     #   Chain runner
├── viz/                  # Layer 3 — HTML/Canvas visualisation
│   ├── gridai_demo.html  #   Self-contained dashboard (deployed)
│   ├── build_demo.py     #   Generates dashboard from JSON
│   └── screenshots/      #   Verified render captures
├── tests/                # 89 tests (pytest)
│   ├── test_simulation.py
│   ├── test_aemo.py
│   ├── test_agents.py
│   └── test_coherence_verification.py
├── outputs/              # Generated scenario + agent data
├── submission/           # lablab.ai submission package
└── docs/                 # GitHub Pages deployment (mirror of viz/)

Limitations

GridAI is a hackathon prototype, not a production DERMS. The current public results are simulation-based and do not use live feeder telemetry. The feeder model is simplified and does not yet include full three-phase unbalanced power-flow validation. The coordination architecture is priority-based/hybrid through a Coordinator and should not be described as fully decentralised peer-to-peer control. The current public baseline reports residual undervoltage in the relevant scenario (435 battery-herding undervolt events in the gossip-heterogeneous AEMO case); this is disclosed as a limitation and future voltage-support target. Future work includes feeder-specific validation, improved voltage-support optimisation, lower fragmentation/synchrony, and deployment-grade safety testing.

Submissions & Links

• Live demo: https://dexflex66.github.io/gridai/
• Pitch video: final/gridai_submission_video_TRUEFINAL.mp4 (90s, 1920×1080)
• Slide deck: submission/gridai_lablab_band_of_agents_2026/assets/gridai_pitch_deck.pdf
• Submission form: SUBMISSION.md
• Narration script: viz/NARRATION.md

Hackathon

lablab.ai Band of Agents Hackathon · Track: Regulated and High-Stakes Workflows. Submission deadline: June 19, 2026 15:00 UTC.