Operating smart grids by customizing large model agents

Abstract

Operating smart grids is becoming more difficult as renewable energy, electrification, and digital technologies make power systems more complex, data-rich, and fast-changing. This review explains how customizing large model agents for control room tasks could support human operators in monitoring conditions, coordinating across functions, and making timely decisions. It describes a practical pathway for operating smart grids with these agents, including task-focused data preparation, grounded use of grid knowledge, safety checks, staged deployment, and close human supervision. Together, these steps show how large model agents could be introduced into existing grid operations in a careful and useful way.

Power grid operation is becoming more complex with renewable energy, electrification, and digitalization. Hanjiang Dong and colleagues describe how customized large model agents could support human operators in smart-grid control rooms.

Introduction

The electric power grid is undergoing a fundamental transformation in response to the dual imperatives of decarbonization and digitalization¹. As renewable energy penetration rises^2,3, electric vehicles (EV) proliferate^4,5, and distributed energy resources become ubiquitous⁶, modern power systems are evolving into complex, data-rich, and highly dynamic infrastructures. These developments, while enabling a more sustainable energy future⁷, also introduce unprecedented operational challenges⁸. Ensuring reliability, economic efficiency, and flexibility in such an environment stretches the limits of traditional human-centric dispatching paradigms^9,10.

In conventional dispatch control rooms, system operators are tasked with coordinating a wide array of functions, including load balancing, generation scheduling, security assessment, and multi-area interconnection, using a suite of specialized software tools and domain expertise (for details, see Supplementary Note 1). This model has served the power system well for decades. The 2025 Iberian Peninsula blackout underscores how slower-than-needed system responses, stemming from structural fragmentation and reliance on manual operator interventions^11–13, can transform localized faults into regional blackouts under modern grid conditions. Recent research has highlighted the evolving landscape of control room operations, emphasizing the shift from traditional operator-centered workflows to hybrid or autonomous systems. Observational studies from real-world dispatch centers have documented the changing role of grid operators in an increasingly automated environment¹¹, underscoring the opportunities posed by artificial intelligence (AI)-driven control. The concept of AI-based operational assistants is also gaining traction, with early applications explored in multiple settings^14,15. Marot et al.¹⁶ emphasize hypervision interfaces, bidirectional interaction, and decision-support functions tailored to real-time operations. This line of work was further extended in a broader control-center perspective, where operators serve for continuous operation procedures supported by unified interfaces, structured decision-making, and progressively integrated AI-based assistance¹⁷. Notably, the CIGRE C2.42 Technical Brochure outlines foundational AI workflows and reference architectures for power system operations¹⁸, including the roles of digital twins, reinforcement learning, data-driven situational awareness, and human-centered operational support. Several open-source community efforts, such as LF Energy¹⁹, GridFM²⁰, and the Open Power AI Consortium¹³, have emerged to promote transparent, interoperable, and ethics-aligned AI tools for energy systems. These findings provide empirical, observation-based insights into operator workflows, supporting the need for intelligent augmentation tools.

Following the human-centered AI (HCAI)²¹ and explainable AI (XAI) thinking²² (for related work see Supplementary Note 2), recent advances in AI, particularly large language models (LLMs), multi-agent systems, and model context protocol (MCP), offer a transformative opportunity to reimagine grid operations²³. By embedding intelligent agents into control room workflows, it becomes possible to enhance situational awareness, automate routine tasks¹⁴, and support human operators with explainable and adaptive decision-making. More specifically, AI-based decision systems should be designed around trustworthiness, transparency, adaptive human-AI interaction, and continuous learning²⁴, more than pure automation. In particular, the perspective on human-AI interaction in safety-critical network infrastructures highlights the importance of cognitive transparency, function allocation, co-learning, and meaningful human oversight in such settings²⁵. As smart grid contexts evolve continuously, integration of renewables, electrification, prosumers, and cyber-physical interfaces, the supporting software tools must also reconfigure and adapt more frequently. These large model agents (LMAs) offer flexibility in this regard, acting as cognitive interfaces that can dynamically integrate and evolve workflows across operational functions.

To this end, this paper emphasizes an evaluative perspective that LMAs are not designed to replace human operators but to support cognition, trace decisions, and facilitate operator participation through cooperative interaction, aligning AI behavior with human evaluative reasoning. We make the following distinct contributions to the field of AI-augmented power system operation:

• Customizing Operation-Oriented Agents: we propose a systematic pipeline for tailoring large foundation models into domain-specific dispatching agents by integrating task-aware datasets, supervisory control and data acquisition (SCADA) logs, grid operation manuals, and real-time telemetry. The customization strategy includes retrieval-augmented generation (RAG), context engineering, expert-co-designed prompt templates, safety alignment, and progressive deployment, forming an actionable blueprint for practical AI agent integration into real control environments.

• Mixture-of-Agents (MOA) Architecture for Grid Operation: rather than relying on monolithic AI systems, our approach introduces an MOA paradigm where function-specific LMAs are organized into collaborative and hierarchical architectures. Each agent simulates the reasoning and responsibilities of control room desks, enabling interpretability, compositional intelligence, and plug-and-play integration with existing energy management system (EMS)/SCADA systems.

• Progressive Deployment Strategy: We advocate a three-stage strategy: (1) smart-patch assistants targeting high-cognitive-load tasks, (2) horizontally expanding collaborative agents across all control functions, and (3) coordinating via supervisory agents within a federated, lifelong learning framework.

• Operational Alignment with Open Ecosystems: The work bridges AI capabilities and power system safety by aligning instruction-tuned agents with open-source communities, emerging standards such as MCP, and regulatory pathways. This contributes to interoperable, transparent, and trusted AI deployments tailored for the delicate context of national-critical infrastructure.

While simulation-based platforms remain rule-centric and fail to exhibit adaptive, interactive, and cognitive capabilities (for simulation engines, see Supplementary Note 3), our proposed LMA framework addresses these limitations by transforming tool-driven workflows into AI-assisted reasoning processes. Looking ahead, the evolution of power system control rooms is poised to shift from manually operated environments toward hybrid intelligent command centers. In this envisioned setting, human operators collaborate with cognitively capable agents, each emulating traditional desk roles.

Challenges of operating smart grids

Traditional control rooms are staffed 24/7 by multiple teams working in rotating 8- or 12-hour shifts¹¹, typically divided into several specialized desks (e.g., Forecaster, Load, Security, Generation, TSO; for specific desk responsibilities see Supplementary Note 4), each managed by one or more operators with distinct responsibilities, as shown in Fig. 1. Operators at each desk make decisions based on outputs from a suite of software tools used for dispatch optimization^26,27, load forecasting^28,29, grid security analysis^30,31, and more. These tools rely on both historical data and real-time system observations. However, individual decisions must be carefully coordinated across desks to achieve unified operational objectives.

Fig. 1. Functional architecture of a power grid control room.

The control room interfaces with large-scale generation facilities, including thermal, hydropower, nuclear, utility-scale solar photovoltaics (PV), onshore/offshore wind farms, and grid-scale energy storage systems, issuing dispatch instructions to maintain supply-demand balance. It ensures the reliable operation of the transmission and distribution networks to manage power flows, prevent overloads, and maintain voltage and frequency stability. It supervises the integration of behind-the-meter or community-based generation units, such as rooftop solar PVs, battery energy storage systems, and microgrids, often through aggregators or automated grid-edge platforms. It monitors consumption trends and may issue demand response signals to various types of users, including electric vehicle (EV) fleets, industrial parks with flexible loads, commercial and residential buildings, hospitals and schools, and other facilities. It coordinates with markets by receiving scheduling and settlement instructions from the market operator and aligns physical dispatch with economic signals. In the control room, multiple specialized desk operators must coordinate effectively to make holistic decisions under time constraints.

Over the years, extensive research has contributed to the development of open-source and commercial software tools for the optimal operation of various grid segments³², including microgrids³³, transmission systems³⁴, and distribution networks^35,36, as well as integrated energy systems such as multi-energy microgrids^37,38. Forecasting tools for electricity demand, EV charging loads^39,40, and renewable generation from wind and solar^41,42 have also advanced. A key emerging challenge lies not in the availability of these tools, but in how to effectively select, integrate, and orchestrate them to meet the increasingly complex operations of the modern power grid. Despite the long-standing effectiveness of traditional control rooms in maintaining grid stability, several inherent limitations in their operational structure are becoming increasingly evident as the power system grows more dynamic.

We acknowledge that contemporary EMS solutions and technical reports (e.g., ⁴³; see Supplementary Notes 2 and 3) have made strides toward integrating forecasting, real-time grid simulation, and congestion management within a cohesive platform. The performance of unifying simulation-based dispatching, forecasting, and optimization functions remains constrained by rigid scenario enumeration, limited real-time adaptability, and weak human-machine interaction capabilities. These systems often rely on static, rule-based workflows and narrowly reactive simulations, lacking cognitive flexibility and contextual reasoning across domains. Moreover, their architectures are typically hard-wired, making adaptation to new grid conditions or operator workflows difficult. The emergent complexity of renewable-dominated, cyber-physical power systems calls for systems that not only integrate functions but also reason across them in an adaptive, explainable, and human-aligned manner. LMAs are proposed in this context to supplement, not supplant, these efforts, offering a paradigm shift toward intelligence-centered orchestration. Rather than hard-coded behavior, LMAs leverage retrieval-based memory, context-aware prompting, and multi-agent collaboration to dynamically coordinate across tools and functions with human-on-the-loop supervision.

Towards large model agents

While modern EMS platforms are increasingly incorporating integrated tools that combine forecasting, simulation, and remedial optimization, their interactions are often predefined and rigid. In contrast, LMAs envision a shift toward flexible, learning-based coordination, where integration is not hard-coded but cognitively adaptive, enabling emergent behaviors under unforeseen conditions. To address the limitations of traditional human-centric dispatching systems (see Supplementary Note 5), the integration of LMAs presents a transformative opportunity. By leveraging advanced AI, multi-agent collaboration, and foundation models enriched with domain-specific knowledge, LMAs offer a paradigm for operating smart grids⁴⁴, together with advantages (see Supplementary Note 6).

To clarify the intended design space of our LMA system, we align with the human control-computer automation space explored in a two-dimensional framework²¹, which locates LMAs within a region of high-computer assistance and sustained human supervision, rather than full autonomy. As illustrated in Fig. 2, this two-dimensional representation maps the trade-off between excessive human dispatch and excessive system automation, highlighting four distinct regimes. Our LMA system is purposefully positioned in the upper-right quadrant, embodying a supervised co-pilot model wherein large model agents operate under sustained human oversight. This design contrasts with both manual legacy systems (e.g., SCADA) and unsupervised black-box automation, which risk either cognitive overload or loss of meaningful human control. By embedding the LMA within this regime, we could maintain expert-driven human agency; enhance situational awareness through structured context; support explainability and validation via transparent multi-turn reasoning; and enable a gradual and adaptive transition from traditional dispatch modes.

Fig. 2. Positioning our large model agent (LMA) system in the human-automation design space.

This two-dimensional framework, adapted from human-centered artificial intelligence (HCAI) principles²¹, illustrates the balance between human control and system automation in power grid operations. Our supervised LMA system occupies the upper-right quadrant, promoting meaningful human-AI cooperation. In contrast, traditional supervisory control and data acquisition (SCADA) represents low automation with high human workload, as fully autonomous agents risk opacity and exclusion of human oversight. The diagram outlines a potential pathway from legacy systems towards transparent, co-adaptive grid intelligence. We noted that LMAs do not lead to operator deskilling but serve as scaffolds to enhance learning and retain decision traceability.

Customizing operation-oriented agents

Overview of agents in grid operations

As the electric grid becomes increasingly complex, with the integration of renewable energy sources, demand-side participation, and distributed energy resources^45,46, there is a growing need for intelligent systems that can assist human operators in managing this complexity. The term LMA reflects a novel paradigm that transcends traditional AI assistants, offering not merely task augmentation but role emulation, agent specialization, and hierarchical cooperation, tightly aligned with human operators’ workflows (for the role of agents in grid dispatching see Supplementary Note 7).

As shown in Fig. 3, we advocate for a progressive and modular approach to deploying intelligent agents in grid operations, with human-AI collaboration (for the design principle see Supplementary Note 7).

• Identify the highest-cognitive-load functions in the dispatch workflow, those that are prone to error, require extensive human interpretation, or demand constant attention.

• Deploy targeted LMAs to support the operators in those specific functions, effectively acting as smart patches that increase resilience and reduce operator burden⁴⁷.

• Iteratively expand LMA coverage across other desks and functions, building a network of intelligent agents that can collaborate with each other and with human operators.

• Towards full automation, gradually shift from human-in-the-loop to human-on-the-loop paradigms⁴⁸, where operators supervise increasingly autonomous agent workflows.

This strategy envisions a future where machines can eventually do all the dispatching work, not through a monolithic AI replacing humans, but through a composite system of intelligent agents⁴⁹, each mastering a part of the workflow and working together in a reliable and human-aligned way. In line with the growing concern of AI-induced deskilling, LMAs are designed not to automate reasoning away from humans, but to scaffold learning and support operator development through explainable action paths, memory-based dialogs, and continuous collaborative planning.

Fig. 3. A progressive strategy for deploying large model agents (LMAs) in grid operations.

In Stage 1, specialized LMAs are introduced as intelligent assistants to support specific high-burden tasks (e.g., forecasting, contingency analysis), acting as smart patches that reduce operator workload and improve resilience. In Stage 2, LMAs are gradually extended across different operational desks (e.g., Load, Generation, Security), forming a collaborative network of agents that communicate, coordinate, and align decisions across domains. In Stage 3, functional agents are organized into a modular system resembling traditional control room roles, enabling end-to-end autonomous dispatching and real-time adaptation under the supervision of human operators.

Data preparation and task definition

Figure. 4 shows the procedures for customizing operation-oriented agents, forming the following framework of data preparation, prompt design, context design, RAG for grid knowledge⁵⁰, task instruction tuning, evaluation, safety alignment, and deployment into control rooms. The customization of LMAs for power grid dispatching begins with a rigorous understanding of the operational tasks and the preparation of high-quality, context-rich data. Unlike general-purpose natural language tasks, power system operations involve domain-specific language, structured and unstructured data, real-time dynamics, and safety-critical decision-making. Therefore, the success of LMAs in this domain hinges on the careful definition of dispatching tasks and the systematic preparation of corresponding datasets.

Fig. 4. A comprehensive framework for customizing operation-oriented large model agents (LMAs) in power grid operations.

In task-oriented data preparation, grid operations are translated into model-accessible tasks using structured datasets derived from manuals, logs, supervisory control and data acquisition (SCADA) systems, and historical dispatch records. In prompt design, well-crafted, expert-informed prompts, including few-shot examples, multi-turn dialogs, and chain-of-thought reasoning, bridge the gap between general-purpose models and specific dispatch scenarios. In context engineering, operational context such as grid topology, real-time states, and dispatch policies is injected via long-context memory, external tool calls, and agent memory mechanisms to ensure situational awareness. In retrieval-augmented generation (RAG) for grid knowledge, external domain-specific documents like emergency playbooks, grid codes, and past incidents are retrieved and fused into the model input for grounded decision-making. In task instruction tuning, models are fine-tuned using supervised learning, preference optimization, or reinforcement learning with feedback to learn dispatch logic, constraints, and expert preferences. In evaluation and safety alignment, model outputs are evaluated on accuracy, interpretability, and stability, with guardrails, verification tools, and human-on-the-loop oversight ensuring accountability. For deployment into control rooms, agents are embedded into real-world operations via application programming interfaces (APIs), user interface (UI) assistants, and function-calling systems, enabling safe and collaborative interaction with operators.

Grid dispatching encompasses a wide range of operational tasks (e.g., load forecasting, resource coordination, contingency analysis and emergency handling, multi-area Interconnection, and maintenance and outage scheduling; for definition see Supplementary Note 8), each requiring distinct reasoning capabilities and contextual awareness. To enable LMAs to effectively assist in these tasks, each must be translated into language-based, model-accessible formats, such as prompt-response pairs or multi-turn dialogs. The quality and diversity of data are critical for grounding LMAs in the operational realities of power systems. These heterogeneous sources must be cleaned, annotated, and transformed into usable datasets. Based on these data datasets, task-oriented datasets, together with semantic representations (for statement see Supplementary Note 9), can be constructed to enhance LMAs from instruction tuning, retrieval, and in-context learning. Data sources include operational manuals and grid codes; maintenance logs and operator notes; historical dispatch records; SCADA/EMS logs; weather data; sensor and equipment status data; event logs and emergency drills; and simulation engines⁵¹.

To assess what constitutes a minimal viable dataset for agent customization, we define three key dimensions of adequacy: (1) modal diversity (covering telemetry, context, and procedural logs), (2) task relevance (data linked to high-burden dispatch tasks), and (3) semantic density (information-dense records like annotated contingency plans). Rather than requiring full hindsight telemetry, early-stage LMAs can be bootstrapped from curated operational manuals, past incident reports, and synthetic data generated via simulation engines. Given that high-quality operations data remains limited in many utility contexts (for minimal data quantity, see Supplementary Note 9), all agent customization workflows include explicit data volume tracking to support controlled scaling. For each task class (e.g., contingency handling, re-dispatch), it is suggested to log quantities of prompt pairs, distribution of source types (e.g., EMS logs, maintenance notes), and estimated coverage across grid regions.

In addition to task-oriented datasets derived from manuals, logs, and simulations, we highlight the importance of adversarially generated and curriculum-structured data⁵². As detailed in Supplementary Note 9, these samples are not limited to manually enumerated disturbances, but may arise from iterative simulator interaction, parameter randomization, or agent-driven search for failure-inducing conditions. Such data are especially valuable for uncovering blind spots in LMA reasoning and for reducing the dependence of training coverage on human assumptions alone. In early-stage deployments⁵³, these curriculum-generated samples can serve as a bridge between static rule-based test cases and more adaptive, open-ended stress testing.

Prompt design for scenarios

LMAs are effective when guided by well-designed prompts that simulate the complexity of real-world grid dispatching tasks. In the context of power system operations, prompt engineering is not a mere input formatting exercise; it is a core strategy for aligning general-purpose language models with the specific reasoning, coordination, and safety requirements of electric grid scenarios. This section explores advanced prompt design techniques tailored for grid operations, including few-shot learning setups, multi-agent collaboration prompts¹⁵, and the incorporation of chain-of-thought reasoning to enhance interpretability and dispatching quality.

Few-shot prompting introduces the model to task structure and expected outputs by providing a small number of carefully selected examples directly in the prompt for a typical few-shot template see Supplementary Prompt Examples 1 and 2. This is particularly effective for complex tasks such as load re-dispatching in response to generator outages, reactive power compensation dispatching, and contingency response generation. These examples teach the model the task format, domain vocabulary, and types of decisions expected, without requiring fine-tuning.

Grid operations are inherently multi-agent: different desks (e.g., Load, Generation, Security) work together to achieve a coordinated outcome. Prompting can simulate this environment by framing dialogs between virtual agents⁵⁴, each representing a control room role (see Supplementary Prompt Example 3). This format encourages the model to simulate reasoning across perspectives, negotiate constraints, and form consensus, just like human operators.

Chain-of-thought prompting enhances transparency and accuracy by encouraging the model to think step-by-step before producing a final plan (see Supplementary Prompt Example 4). This is particularly useful in situations requiring multi-stage reasoning, such as evaluating multiple contingencies, coordinating time-dependent dispatch schedules, and balancing competing priorities (e.g., cost vs. reliability). This structured reasoning allows operators to understand how the model reached its suggestion, increasing trust and ease of validation.

Crafting high-quality prompts often requires close collaboration with domain experts who understand the exact language, constraints, and thought processes used in real dispatch scenarios. (Supplementary Prompt Example 5) serves as a template co-designed with grid operation experts to guide the model in generating a load re-dispatch plan following prediction deviation. By using such templates across different grid conditions (e.g., weather events, equipment outages, load anomalies)⁵⁵, LMAs can be trained and evaluated under realistic and diverse operational contexts.

To minimize additional workload on human operators, LMAs are designed to function primarily as assistive copilots that extract prompts automatically from real-time SCADA data and contextual logs (see Supplementary Note 10), and agents emulate existing dispatch desk responsibilities in a virtualized manner, embedding themselves as supportive extensions to current workflows. Furthermore, all LMA prompts are auto-extracted from telemetry data, and operator feedback is minimal and unobtrusive, requiring no free-text input and no additional workload.

Context engineering: memory, tools, and environment

While prompt design determines how a model is asked, context engineering defines what the model knows at the time of decision-making. In power system dispatching, context is king: decisions must account for grid topology, real-time operating conditions, historical trends, safety constraints, and regulatory rules, all of which form a rich, dynamic background for every dispatch action. To enable an LMA to function reliably in this environment, we must engineer a context pipeline that can (1) represent the right information at the right time, (2) maintain state awareness across multiple interactions, and (3) interface with specialized tools for dispatching and verification (for tool use see Supplementary Note 11), considering human factors (for design strategies see Supplementary Note 2).

Drawing from supervision failure in highly automated systems⁵⁶, we recognize that passive oversight by human operators is unsustainable for safety-critical infrastructure. Accordingly, our LMA architecture embeds active cognitive involvement through human-AI dialog structures more than control gates. By facilitating explainable workflows, guided feedback mechanisms, and role-appropriate interaction modalities, human operators remain cognitively engaged, situationally aware, and empowered to shape (not just monitor) the system’s operation.

To simulate or assist actual dispatch workflows, the model must be provided with structured and unstructured context, involving grid topology (e.g., substations, transmission lines, generation units, constraints, and regional interconnections), real-time status: (load levels, generation availability, outage notifications, weather impacts), and dispatch rules and policies (operating manuals, reserve requirements, ramp rate constraints, environmental limits). These inputs can be embedded in a prompt (see Supplementary Prompt Example 6) or passed as external documents in retrieval-augmented systems.

Dispatch decisions often require reasoning over multi-area data (e.g., balancing across North, Central, and East grids) and multi-hour timelines (e.g., hourly dispatch plans for the next 24 hours). To address this, we can employ (1) extended context windows via long-context LLMs (e.g., models with 128k+ tokens); (2) dynamic context selection, where only relevant regional/state data is loaded per interaction; and (3) hierarchical summarization or vector memory to compress past decisions and trends. This ensures that LMAs can reason over large-scale, multi-source input without losing coherence or exceeding token limits.

Electric grid operation is inherently iterative and collaborative. An operator (or agent) rarely acts in isolation, especially during contingency situations or forecasting deviations⁵⁷. By invoking tools mid-prompt (e.g., via function-calling or API calls), LMAs can iteratively refine their reasoning (see Supplementary Prompt Example 7). Memory allows the agent to simulate continuity, learning, and situational awareness, making it a more trustworthy assistant. Thus, agent memory is crucial:

• Conversation History: Recall decisions made in earlier rounds.

• Outcome Tracking: Whether previous plans succeeded or failed.

• Adjustment Rationale: Why certain changes were made.

Consider a scenario where an LMA assists a human operator over multiple rounds of dispatching (see Supplementary Prompt Example 8). The agent is provided with the full dispatch context (from data preparation) and is asked to propose, revise, and explain dispatch plans across time. It demonstrates how context, memory, tools, and human feedback can be orchestrated into a multi-turn, high-stakes decision workflow, where the LMA evolves from a passive responder into a proactive dispatching assistant.

RAG for grid knowledge

LLMs, despite their impressive generalization capabilities, are not designed to memorize and reliably retrieve fine-grained, dynamic, or domain-specific knowledge, especially in safety-critical domains like power system operations. To address this limitation, RAG provides a powerful framework that combines the linguistic fluency of LLMs with the factual accuracy of external knowledge bases. In the context of power grid dispatching, RAG enables an LMA to consult structured and unstructured documents¹⁴, such as operational manuals, contingency plans, regulatory codes, and incident logs, before forming a response²³. This enhances both precision and trustworthiness in high-stakes decision-making, such as grid dispatching. It improves factual accuracy by grounding model outputs in authoritative documents, reducing hallucinations. RAG also supports expert alignment by enabling the integration of human-generated knowledge without the need for model retraining⁵⁸. Its scalable architecture allows new documents to be added to the retrieval index seamlessly. Furthermore, RAG promotes transparency and explainability by presenting the retrieved sources alongside its recommendations.

Given a domain-specific vector knowledge base (see Supplementary Note 9), a typical RAG pipeline for grid dispatch operates in four main stages:

• Query formulation: the user or agent issues a natural-language query (e.g., “What are the dispatch alternatives if G3 goes offline during peak hours?”)⁵⁹.

• Embedding and retrieval: the query is embedded into the same vector space as the knowledge base; a similarity search retrieves the top-K relevant documents or passages.

• Contextual fusion: retrieved documents are formatted into a context window; The original query is concatenated with this context, forming a single prompt.

• Answer generation: a language model (e.g., GPT-4, Claude, or a custom LLM) generates an informed, context-aware response using both the query and retrieved knowledge.

This workflow allows the agent to stay grounded in verified knowledge while still leveraging its reasoning and synthesis abilities. Then, RAG empowers the LMA to (1) reference standard procedures and historical precedent; (2) adapt recommendations to real-time constraints, and (3) generate grounded, explainable, and auditable plans, thus answering a realistic dispatch question (see Supplementary Prompt Example 9).

Fine-tuning and instruction tuning for operations tasks

While prompt engineering and RAG can improve a large model agent’s performance in grid operations, these methods alone may not be sufficient for deep alignment with the language, logic, and safety constraints of power system operations. To fully adapt a base model to the dispatch domain, fine-tuning, especially instruction tuning, is an essential step. Fine-tuning allows us to specialize a language model by exposing it to curated examples from real-world dispatch tasks⁶⁰, enabling it to internalize domain-specific terminology, operational logic, and decision-making patterns. Fine-tuning is not a one-time process. As grid conditions evolve and new operational policies emerge, continual fine-tuning or online adaptation may become necessary⁶¹. This also opens up the potential for fine-tuning with synthetic data generated by expert agents or simulators, enabling safe scaling of model capabilities without relying solely on historical dispatch records.

Fine-tuning in this context serves several key objectives:

• Sensitivity to Professional Language: recognize and generate domain-specific terminology, abbreviations, and structured formats (e.g., “AGC”, “N-1 contingency”, “ramp rate 20 MW/min”).

• Operational Logic Alignment: learn the causal and temporal reasoning involved in dispatching, such as how load forecasts, generator availability, and reserve constraints interact.

• Compliance with Safety and Reliability Constraints: internalize rules from manuals and grid codes to avoid unsafe or non-permissible recommendations.

• Improved Multi-Turn Consistency: enhance the model’s ability to participate in sustained, goal-driven dialogs across multiple rounds of dispatching.

Based on the constructed dataset (e.g., instruction-style prompts, historical dispatch dialogs, action-outcome pairs, annotated edge cases; see Supplementary Note 9), there are several fine-tuning methods available, each with unique strengths and trade-offs. As listed in Supplementary Table 1, we compare three of the most commonly used approaches and assess their suitability for power system dispatching. The supervised fine-tuning (SFT) is used for initial alignment with domain language and operational structure⁶². The direct preference optimization (DPO) is suitable when multiple dispatch options exist, and human preference (e.g., economic vs. secure dispatch) must be reflected. The reinforcement learning from human feedback (RLHF) dominates in high-risk, long-horizon dispatching tasks (e.g., black-start recovery, region-wide load restoration)⁶³, where reward signals can be designed around safety, stability, and compliance.

Evaluation and safety alignment

LMAs in power grid operations, where safety, reliability, and accountability are paramount, require a robust framework for evaluation and alignment. Unlike general-purpose chatbots, grid dispatch agents must be held to rigorous standards to ensure they do not generate unsafe, inaccurate, or misleading outputs. To this end, this section outlines the key dimensions of model evaluation, the design of safety mechanisms, and the critical role of human-on-the-loop alignment in operational deployment.

To ensure that LMAs are developed and evaluated in a scalable and transparent manner, we explicitly monitor the quantity, diversity, and granularity of human-related operational data used in task instruction tuning, context construction, and safety alignment. Each evaluation cycle logs the source modalities (e.g., SCADA logs, forecast records, post-event reports), example counts per task, and data coverage across operational zones. Table 1 lists our multi-dimensional evaluation criteria (for statement see Supplementary Note 12), within which dataset adequacy is explicitly tracked and reported.

Table 1.

Evaluation dimensions

Dimension	Description
Accuracy	The degree to which model outputs align with correct dispatch actions or expert-validated plans. This includes correct application of constraints, valid unit selection, and feasible load/generation balancing.
Latency	Time taken to produce a response, especially for real-time or emergency scenarios. Acceptable latency may range from seconds (for advisory roles) to sub-seconds (for automation support).
Robustness	Stability under input perturbations, response to unforeseen contingencies, distributional shifts (e.g., untrained regions, weather anomalies).
Confidence	Indicators of uncertainty or low confidence are used to trigger human oversight. This includes token-level entropy, retrieval sparsity, and disagreement between reasoning paths. Also includes model self-awareness about the reliability of its own plan or recommendation (e.g., signaling low-confidence suggestions), and structured escalation to human review when uncertainty or risk exceeds a threshold.
Trustworthiness	Whether the model not only explains why a decision was made but also adheres to principles of trustworthy AI. This includes interpretability through chain-of-thought reasoning and document citation (via RAG), output consistency and validation, alignment with operational constraints, human-on-the-loop oversight, and prevention of unsafe or unjustified actions.
Consistency	Whether the model produces consistent, non-fluctuating responses under similar prompts or slightly varied conditions. Also includes behavioral stability across successive training or fine-tuning stages, ensuring that updates do not cause regressions or unpredictable shifts. This is essential for decision reproducibility and safe model iteration.
Data accounting	The data volume, source diversity, and domain coverage used for agent customization are tracked at each deployment stage. For each task class (e.g., re-dispatch, forecasting), metrics such as prompt pair counts, source modality ratio, and regional representation rate are monitored to ensure transparency and informed scaling.
Socio-technical fit	The degree to which the LMA supports real control-room practice without degrading human situational awareness, coordination quality, workload balance, trust calibration, or shift continuity. This includes effects on team communication, interruption handling, handover quality, and operator learning in realistic dispatch environments.

This table outlines key evaluation metrics for large model agents operating in smart grid environments, including accuracy, latency, robustness, confidence, trustworthiness, consistency, data accounting, and socio-technical fit. These dimensions reflect not only model-centric performance but also collaborative human-AI interaction quality and reliability under real-world uncertainty and operational stress.

Based on the above metrics, LMA-powered dispatch systems can be assessed via a three-stage framework (see Supplementary Note 12) encompassing human-system interactions and operational environments. LMAs can first be tested in vitro under curated datasets and simulator-based validation, then evaluated in situ through controlled but realistic human-AI operational workflows, and finally assessed in vivo through longitudinal analysis after real-world deployment. Importantly, high-fidelity control room replicas are not an additional stage beyond this framework; rather, they constitute an advanced experimental implementation for rigorous in situ evaluation. We emphasize that such a three-stage evaluation scheme remains only a starting point for the socio-technical assessment of LMAs in real dispatch settings. In safety-critical infrastructures, evaluation must also examine team coordination, workload redistribution, trust calibration, interruption management, handover quality, shift-to-shift continuity, and the ways in which agent behavior reshapes operator cognition and organizational routines over time.

To prevent LMAs from generating unsafe or unauthorized dispatch recommendations, safety alignment strategies must be implemented (for constraint-aware decoding, tool-based verification, permission control, error pattern detection, and fallback to human oversight, see Supplementary Note 12). However, a structured handover protocol alone is not sufficient to guarantee safe takeover in time-critical grid operations. If operators remain passive for extended periods, an abrupt transfer of responsibility may occur before they have reconstructed an adequate understanding of system context, priorities, and pending risks (for details see Supplementary Note 2). LMAs should be designed to keep operators cognitively on the loop even during prolonged periods of automated assistance, by continuously exposing them to concise state summaries, rationale traces, uncertainty indicators, and projected system trajectories rather than only issuing alerts at the moment of escalation. From a systems-theoretic safety perspective, this requirement can also be interpreted as maintaining an adequate control loop between the human supervisor and the AI-supported dispatch process. In the spirit of systems-theoretic process analysis (STPA)⁶⁴, unsafe handover may arise not merely from the absence of a protocol, but from flawed control structures, delayed or incomplete feedback, and mismatches between the operator’s mental model and the actual system state (for STPA see Supplementary Note 12).

More specifically, safety nets include hallucination detection classifiers, retrieval-aware confidence scoring, output validation via dispatch simulators, and mandatory operator pre-approval for low-certainty outputs. We adopt conservative default fallbacks: if the agent is unsure, outputs are explicitly flagged and halted unless escalated by the human-on-ohe-loop and robustness testing (see Supplementary Note 12). Beyond one-shot robustness testing, we further emphasize that adversarial evaluation should be organized as a progressive curriculum rather than a fixed benchmark suite. We recommend an auto-curriculum-inspired evaluation pipeline where the difficulty, diversity, and coupling of scenarios increase as the LMA system improves. Such curricula may include cascading contingencies, delayed or conflicting telemetry, multi-zone coordination failures, operator-agent disagreement, and simulator-generated edge cases targeted at the model’s current weaknesses.

Deployment and integration into control rooms

The true value of LMAs in power system dispatching emerges only when they are seamlessly embedded into real-world operational environments. This requires thoughtful system integration, multi-agent collaboration design¹⁵, human-aligned workflows, and strict safety gating mechanisms. In this section, we explore how LMAs can be deployed in control rooms through APIs, user interface (UI) assistants, and workflow orchestration (see Supplementary Note 13), while emphasizing human-on-the-loop control, function calling, and emerging standards, particularly the MCP as follows. We thus can build an AI assistant that is flexible, safe, and human-aligned, supporting operators in making faster, more informed, and auditable decisions.

LMAs can be deployed not just as a single assistant, but as a system of coordinated agents (e.g., Forecaster Agent, Load Desk Agent, Generation Desk Agent, Security Desk Agent, TSO Agent; for definition see Supplementary Note 13). These agents can interact through structured dialog or via a central orchestration model, simulating real-world control room collaboration. This supports scalability and modularity, allowing each agent to specialize and evolve independently. Operational deployment demands continuous learning and adaptation. This flexible workflow creates a closed-loop system that combines automated reasoning with human judgment^47,65, ensuring reliability and accountability. Each interaction between operator and model completes via feedback (confidence scoring, plan validation loop, and replay and audit; see Supplementary Note 13). To address operator workload constraints, the LMA system decouples feedback from real-time operation by adopting flexible timing and granularity mechanisms. While basic validation (e.g., one-click approval, revision flags) is embedded inline during dispatch workflows and designed to take no more than a few seconds, deeper feedback, such as rationale elaboration or scenario annotation, can be deferred to off-shift or post-event review sessions.

Emergency or safety-sensitive scenarios activate structured handover mechanisms that prioritize operational transparency and rapid human control. Upon triggering, the LMA compiles a compact handover log (e.g., contextual summary, rationale trail, ranked list, visual replay) for a posteriori feedback and reporting (for trigger conditions see Supplementary Note 13). The purpose of this handover package is not merely informational completeness, but rapid restoration of operator situational awareness under emergency conditions. To this end, we should privilege state abstraction, temporal ordering, risk salience, and action relevance over raw data volume. Safe intervention depends on preserving comprehension and projection, not only perception of isolated signals^64,66.

To bridge the gap between natural language recommendations and system-level execution, LMAs can generate structured scripts (e.g., JSON; see Supplementary Note 13) that conform to the EMS’s control interface. To support multi-tool, multi-data-source integration, we adopt the MCP, an open standard introduced by Anthropic in early 2024. MCP provides (1) a structured, declarative interface for passing context (e.g., telemetry, forecasts, topology) to LMAs; (2) a standardized API layer for invoking external tools (e.g., simulators, diagnostics, prediction engines); and (3) a universal schema for tool registration, metadata, and permission levels Anthropic envisions MCP as the connector for AI, a universal interface for connecting any model to any tool or dataset, in a secure, interpretable, and extensible way. In our deployment architecture, the LMA connects to grid state data via MCP telemetry schema, dispatch simulation tools via MCP function schema, and operator feedback and scoring via MCP interaction schema. MCP can further enhance interoperability across vendors, extensibility across use cases, and long-term maintainability of the AI agent infrastructure through multi-channel integration and security considerations (see Supplementary Note 13).

A key consideration in LMA deployment is affordability and staged implementation. To reduce the initial barrier, we suggest a plug-and-play integration strategy that enables gradual rollout alongside existing SCADA/EMS systems. More specifically, we frame the economic case in terms of cost categories and value pathways (see Supplementary Note 13 for details). We therefore view a positive return on investment (ROI) within the first several years of deployment as a plausible outcome, particularly when LMAs are first targeted at high-cognitive-load and coordination-intensive tasks (see Supplementary Note 13). The qualitative ROI dimensions and value pathways are summarized in Supplementary Table 2.

Prospects of mixture-of-agents

Building upon the progressive deployment strategy illustrated in Fig. 3, the MOA paradigm refers to a collaborative system of specialized AI agents, each emulating the roles of traditional control room operators. Rather than relying on a singular, monolithic model, this approach distributes intelligence across multiple desk-specific agents, enabling modularity, scalability, and resilient coordination in complex grid environments. This approach inherits the hierarchical logic of traditional protect-control-optimize (PCO) architectures⁶⁷, wherein operational safety and stability form the foundation upon which adaptive optimization is layered. By reinterpreting this structure with modular AI agents, we retain the multi-layered control hierarchy while enabling agentic reasoning, collaborative planning, and continual learning as capabilities not present in static PCO implementations.

As shown in Fig. 5, this session explores an integrated roadmap illustrating how power system operations can progressively evolve toward human-supervised autonomous control through a layered MOA architecture. This paradigm envisions a transformation across three stages: (1) hybrid operations, where LMAs support real-time analysis under human supervision; (2) functional specialization and collaboration, where agents emulate specific control desks and cooperate via structured dialog; and (3) hierarchical coordination, where mixtures of agents operate across spatial or organizational scales. Alongside the evolution, four foundational dimensions enable safe, scalable, and sustainable deployment: continual and federated learning, which empowers agents to adapt over time while preserving data privacy; trustworthy AI, which ensures robustness, explainability, and human-on-the-loop governance; plug-and-play interoperation, which allows seamless integration with SCADA/EMS systems; and open ecosystems, which foster transparency, regulatory alignment, and broad collaboration. These interacting forces define a feasible path toward self-driving grid operations, where agents serve not as replacements but as intelligent teammates—interpretable, extensible, and anchored in operational reality.

Fig. 5. A layered roadmap to self-driving grid operations.

This diagram illustrates a progressive and modular architecture for building self-driving grid operations through hybrid intelligence. The left vertical stream represents the gradual evolution from operator-led hybrid interaction to agent specialization and fully autonomous multi-agent orchestration. The right column highlights four foundational pillars enabling this transition: lifelong learning, trustworthy design, interoperable deployment, and open co-evolution. Together, these components form an integrated ecosystem for safe, efficient, and human-supervised automation of power grid operations.

For MOA systems, the importance of the training regime becomes even more pronounced. When multiple specialized agents interact, coordinate, disagree, or compete for shared operational priorities, the effective training distribution is no longer determined by isolated task prompts alone, but by the joint dynamics of the agent population and the environment. Future LMA-based dispatch systems may benefit from an auto-curriculum mechanism at the multi-agent level: as one agent or subgroup becomes more capable, other agents and the simulator can generate increasingly challenging coordination demands, forcing the system to develop more diverse and robust operational behaviors. In this respect, the role of a simulator extends beyond physical validation to that of a curriculum engine that continuously reshapes the difficulty landscape for collaborative grid operation.

Outlook

By automating routine tasks, reducing coordination delays, shortening operator training cycles, and minimizing the risk of systemic outages, LMAs offer a pathway to improve the economic efficiency of grid operations. While the development and deployment of LMAs involves initial investment (e.g., hardware upgrades, agent customization, integration with legacy EMS/SCADA systems, and safety assurance), the long-term benefits are expected to outweigh the costs. Governments and utilities may also benefit from economies of scale and shared infrastructure under federated learning and open-interface standards, which further reduce per-unit customization cost. In addition, we believe deeper AI integration in power systems will depend not only on data pretraining or tool calling, but on direct engagement with physics simulators as active learning environments. Through such simulator-enhanced training, future LMAs could evolve from passive advisors into physically aligned co-pilots informed by experiential knowledge.

Looking ahead, several key opportunities and challenges must be addressed to realize the full potential of LMA-enabled smart grids. First, the development of trustworthy agents, capable of robust performance under uncertainty, transparent reasoning, and strict adherence to safety constraints, is essential for building confidence among operators and regulators. Second, based on the modularization and specialization of agents, Hierarchical coordination architectures, supported by open interface standards, particularly the MCP, will be critical for aligning local decisions with system-wide optimization goals. Third, continual and federated learning will allow agents to evolve with the grid, adapting to new technologies such as vehicle-to-grid integration, flexible demand resources, and emerging energy storage systems. Fourth, socio-technical evaluation should advance toward high-fidelity in situ testing through control room environment replicas, including representative user interfaces, alarm streams, communication channels, shift handover procedures, and realistic time pressure. Such platforms support controlled human-on-the-loop experiments, benchmark socio-technical performance, and provide an evidence base for safe regulatory adoption. Fifth, the design of training regimes for LMA-based grid agents will combine expert knowledge with auto-curriculum-inspired scenario generation, where simulators, surrogate adversaries, and interacting agents continually construct harder and more diverse operational samples. Finally, the transition to self-driving power systems will not occur as a single leap, but as a staged evolution, from advisory AI assistants, to semi-autonomous agent collaboration, and ultimately toward fully autonomous dispatch in structured or low-risk scenarios. By fostering an open ecosystem of interoperable agents, aligned with regulatory standards and grounded in real-world knowledge, we move closer to realizing the vision of a resilient, efficient, and autonomous power grid that can meet the demands of the energy transition with confidence and clarity.

Author contributions

H.D. contributed conceptualization, methodology, investigation, writing—original draft, visualization, writing—review & editing; H.Y. contributed conceptualization, methodology, writing—review & editing, visualization; Y.M. contributed methodology, investigation, writing—review & editing; J.Z. contributed conceptualization, writing—review & editing, supervision, project administration, funding acquisition; K.C. contributed conceptualization, writing—review & editing, project administration; H. H. contributed conceptualization, methodology, investigation, writing—review & editing; C.S. contributed conceptualization, investigation, writing—review & editing; S.L. contributed writing—review & editing, visualization; Z.W. contributed writing—review & editing; and C.Y.C. contributed methodology, writing—review & editing, supervision.

Peer review

Peer review information

Communications Engineering thanks Antoine Marot and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: [Wenjie Wang, Rosamund Daw].

Funding

H.Y. discloses support for the research of this work from the Basic Research Operation Fund of Dalian University of Technology [grant number DUT25RC(3)110]. J.Z. discloses support for the research of this work from the Guangdong Basic and Applied Basic Research Foundation [grant number 2022B1515250006] and the National Foreign Experts Program [grant number H20250116]. All other authors declare no relevant funding.

Data availability

No datasets were generated or analyzed during the current study.

Code availability

No custom code or software was used in this study.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44172-026-00709-1.