Human-Supervised Artificial Intelligence Researchers in Electric Power Systems: Knowledge, Physics, Simulation, and Governance for Large Language Models
Organizations: China Southern Power Grid Research Institute Co., Ltd. Guangzhou 510663, China
Abstract
Conclusions in power-system research require a clear evidential basis. Researchers need to understand the data sources, modelling assumptions, and solution procedures to assess the conditions under which a conclusion holds. Large language models (LLMs) can help researchers extract information from documents and translate research requests into code and tool calls, but their outputs still require physical validation. Focusing on human-supervised artificial intelligence (AI) researchers, this review examines power-system LLM applications and related scientific workflows, together with physics-grounded power AI studies, to discuss the conditions for using LLMs in scientific research. In the proposed knowledge-physics-simulation-governance framework, LLMs translate information and coordinate tasks, simulators and validation tools test models, and experts retain authority over critical assumptions, safety boundaries, and final conclusions. Research processes should preserve materials for review so that claims can be traced to their sources, computations can be reproduced, and failed attempts and approvals can be examined. On this basis, a staged development roadmap and requirements for research reporting are proposed. Evaluation should focus on research artifacts and their validation to assess whether LLMs can improve the efficiency of power-system research within clearly defined permissions.
Figures & tables
| Literature stream | Main research content | Role in this framework |
|---|---|---|
| Power-system LLM studies | Review LLM and generative-AI applications in energy and grids [ 16 , 17 , 18 , 29 , 30 , 31 , 32 , 33 , 34 , 35 ] ; study modelling, simulation, optimization, digital twins, regulatory compliance, markets, and design [ 22 , 23 , 24 , 25 , 26 , 27 , 36 , 37 , 38 , 39 , 40 , 41 ] ; propose power-agent architectures and controller-design concepts [ 28 , 42 ] | Provide direct application evidence to examine which research tasks language models can support and how those tasks connect |
| AI-for-science workflows | Discuss research agents and scientific discovery [ 43 , 44 , 45 , 46 , 47 , 48 ] ; explore reasoning, hypothesis generation, and preliminary experiments through retrieval, knowledge representation, and tool execution [ 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 ] | Provide workflow design references for organizing multistep tasks, handling execution feedback, and arranging human review |
| Physics-grounded power AI | Study state estimation, optimal power flow, and power-model validation through physics-informed learning and topology modelling [ 1 , 2 , 3 , 4 , 5 , 6 ] | Establish validation requirements for equations, network connectivity, constraints, and numerical solution procedures |
| Artifact category | Power-system examples | Required validation | Consequences if absent |
|---|---|---|---|
| Evidence artifacts | Retain original standard clauses or event reports; link data sources and literature claims to their citation context | Are sources recoverable and versions and access conditions clear? Does their scope support the claim? | Unsupported statements enter model design |
| Modelling artifacts | Specify network connectivity and demand assumptions, units and normalization bases, study conditions, and applicable rules | Are variable types and units consistent? Have boundary conditions, applicability, and uncertainty been confirmed? | Undocumented assumptions change the actual research question |
| Execution records | Preserve actual cases, scripts, solver settings, and outputs; archive residuals, convergence logs, and failed runs | Are data and dependency versions available? Are solver tolerances, residuals, and constraint violations consistent with execution status? | Computations are hard to rerun or debug and results cannot be tested adequately |
| Governance records | Record data authorization, operational permissions, tool calls, and human interventions; explain rejected assumptions and claim approvals | Are approvers identified? Are credentials and sensitive data controlled, and have prompt-injection checks been performed? | Responsibility within the workflow is unclear |
| Stage | Main capability | Trustworthy research artifacts | Governance measures |
|---|---|---|---|
| LLM assistant | Explain concepts from literature and draft code or text for human use | Notes with sources and uncertainty statements, and draft code | Human checking before use |
| Traceable artifact builder | Extract source-linked evidence and represent assumptions as typed data structures | Source-supported claims and variable definitions with types and units | Citation, schema, and unit checks |
| Simulation-connected agent | Run generated cases and propose repairs from tool errors | Executable cases with corresponding solver, error, and validation records | Sandboxed execution and physical validators |
| Closed-loop research agent | Repeat simulations according to an experimental plan, compare results, and record negative cases | Experimental plans and complete run histories, including failures and comparison reports | Human approval of critical experimental decisions and research claims |
| Human-supervised AI researcher | Coordinate research tasks under expert authorization and use evidence and execution results in reporting | Research packages with sources, decision processes, and approval records | Expert decision authority, accountability, and institutional review |