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

Regional carbon-aware operation requires nodal carbon intensity (NCI) signals that remain valid under heterogeneous operational inputs, topology changes, and network congestion; however, existing workflows for carbon-flow tracing and marginal-emission analysis are still difficult to operationalize because data integration, model configuration, and result auditing rely heavily on manual intervention. This paper proposes a large language model (LLM)-agent-in-the-loop framework in which the language model is restricted to orchestration, while dispatch optimization, physical verification, and carbon attribution are executed by deterministic modules. The framework combines a direct-current optimal power flow (DC-OPF)-based modeling layer, a four-layer verifier with Karush-Kuhn-Tucker (KKT)-inspired diagnostics, and a unified engine that computes average carbon intensity (ACI) and marginal carbon intensity (MCI) anchored to the same verified base operating point and network model. In the PJM 5-bus benchmark prompt-bank validation, the framework achieves task-pass rates of 1.00 across structured, semi-structured, and tested anomalous prompt-bank inputs, with the main parser advantage appearing on semi-structured operational notes, while the verifier reduces unsafe acceptance from 0.429 under a V2 verifier without Layer-4 certificate diagnostics to 0.048 and raises resolved recall from 0.571 to 0.952 on injected-error artifacts. Across PJM 5-bus and IEEE 14-bus benchmark attribution cases, under congestion the bounded MCI diagnostic standard deviation reaches 283.9 kg/MWh and the bounded node-to-node MCI diagnostic spread reaches 980.0 kg/MWh, while high-renewable scenarios lower mean ACI by 24% on PJM 5-bus and 30% on IEEE 14-bus. These results demonstrate that LLM-based orchestration can improve the auditability and benchmark-level operational usability of nodal carbon accounting without replacing physics-based computation.