An LLM-agent-based framework for calculating nodal carbon intensity in regional power systems

Area of Science:

• Computational Power Systems
• Artificial Intelligence in Energy
• Environmental Engineering

Background:

• Regional carbon-aware operation requires reliable nodal carbon intensity (NCI) signals.
• Existing carbon-flow tracing and marginal-emission analysis methods are difficult to operationalize due to manual intervention in data integration, model configuration, and auditing.
• Challenges include heterogeneous operational inputs, topology changes, and network congestion impacting NCI signal validity.

Purpose of the Study:

• To propose a novel large language model (LLM)-agent-in-the-loop framework for operationalizing nodal carbon accounting.
• To enhance the auditability and benchmark-level operational usability of carbon accounting in power grids.
• To integrate LLM-based orchestration with deterministic modules for dispatch optimization, physical verification, and carbon attribution.

Main Methods:

• A framework combining a direct-current optimal power flow (DC-OPF)-based modeling layer.
• A four-layer verifier with Karush-Kuhn-Tucker (KKT)-inspired diagnostics for physical verification.
• A unified engine to compute average carbon intensity (ACI) and marginal carbon intensity (MCI) anchored to a verified operating point and network model.

Main Results:

• Achieved 1.00 task-pass rates on PJM 5-bus benchmark validation across structured, semi-structured, and anomalous inputs.
• The verifier reduced unsafe acceptance from 0.429 to 0.048 and improved resolved recall from 0.571 to 0.952 on injected errors.
• Under congestion, bounded MCI diagnostic standard deviation reached 283.9 kg/MWh; high-renewable scenarios lowered mean ACI by 24% (PJM 5-bus) and 30% (IEEE 14-bus).

Conclusions:

• LLM-based orchestration significantly improves the auditability and operational usability of nodal carbon accounting.
• The proposed framework effectively integrates AI orchestration with physics-based computation for reliable carbon analysis.
• Demonstrated the framework's robustness and accuracy in handling complex power system operational scenarios and carbon attribution.