EnerOS is a power-native Agent Operating System for the electric power and energy domain.
It embeds power system domain knowledge, physical constraints, and operational logic as OS kernel primitives,
enabling AI Agents with native understanding, safe decision-making, and autonomous action capabilities in energy scenarios.
General-purpose Agent frameworks face fundamental challenges in the power and energy domain. AI Agent technology is reshaping industries at an unprecedented pace, yet the power and energy domain faces unique challenges — safety constraints are degraded to prompt-level suggestions, and grid topology and electrical coupling relationships are ignored.
Agents don't understand physical quantities like power flow, voltage, and frequency, unable to judge the physical feasibility of decisions.
Safety constraints (N-1, thermal stability, voltage limits) are treated as "prompts" rather than system-level guarantees.
Agents treat the grid as flat data, unable to perceive topology structure and electrical coupling.
Power systems are strongly time-coupled; general frameworks lack first-class support for the temporal dimension.
Transformers, circuit breakers, and inverters each have independent models and protocols, making unified scheduling difficult.
Don't "bolt on" power knowledge to general frameworks — build an OS power-natively from the ground up.
Six core design principles defining the foundational paradigm of a power-native Agent OS.
Power topology, power flow computation, and device models are not bolt-on plugins but native OS abstractions. Agents run on a physical world model of the grid from inception.
Each Agent corresponds to a functional node in the grid (substation, feeder, device), inherently possessing topology awareness and constraint compliance. Inter-Agent communication mirrors information exchange between grid nodes.
Safety constraints (N-1 verification, thermal stability, voltage limits) are enforced by the kernel — no Agent decision can exceed the physically feasible domain. Safety is not a prompt; it's an OS-level hard constraint.
Power systems are strongly time-coupled. EnerOS treats the temporal dimension as a first-class citizen, supporting native operations for real-time data streams, historical lookback, and predictive inference.
Power systems have rigid real-time requirements. EnerOS adopts a dual-execution architecture: the general execution domain handles Agent orchestration and AI inference, while the real-time execution domain guarantees deterministic latency for protection logic and switching operations. The safety domain cannot be blocked by the general domain.
Standardized Agent communication protocols and device access specifications enable plug-and-play for heterogeneous energy devices and multi-vendor systems.
Dual-execution architecture: General Domain + Real-Time Domain. The system is partitioned into two execution domains, balancing complex AI inference computation with hard real-time guarantees for grid control.
Real-time domain tasks have the highest priority; no general domain operation may affect real-time execution determinism.
The real-time domain can directly read decision commands from the general domain, but the general domain cannot directly interfere with real-time domain scheduling.
All general→real-time domain commands must pass through the gateway's constraint verification and priority arbitration.
When the general domain fails, the real-time domain automatically switches to local protection logic, ensuring grid safety doesn't depend on AI.
A five-layer architecture from application to infrastructure, clearly separating responsibilities and abstraction levels.
Six core capabilities that make Agents natively "understand power".
The grid topology graph is EnerOS's core data structure. Agents automatically acquire electrical relationships, upstream/downstream devices, and operational status of their node through topology-aware context — no explicit queries needed.
All Agent decision outputs undergo constraint verification by the power-native kernel — whether power flow converges, voltage exceeds limits, or lines are overloaded. Decisions failing physical constraints are rejected at the kernel level.
Built-in equipment parameter library compliant with Chinese national standards (GB) and IEC standards, covering transformers, lines, switches, inverters, and other core equipment types, with pandapower-compatible format support.
Grid topology-based Agent organization model: Agents within the same substation automatically form collaboration groups; cross-substation Agents communicate structurally via topology paths, avoiding global broadcast chaos.
Real-time data streams, historical lookback, predictive inference — three temporal modes unified at the kernel level. Agents seamlessly switch between "review-perceive-predict" temporal perspectives.
Kernel-level safety guard: N-1 security verification, thermal stability check, voltage limit detection. Safety constraints cannot be bypassed or degraded by Agents — they are the "hard laws" of the operating system.
A modular system of 12 Rust Crates, each with clear responsibilities, loosely coupled and highly cohesive. Built on Rust for memory safety and high performance.
Unified types, errors, config
Grid topology graph modeling & analysis
Newton-Raphson power flow solver
Safety constraint verification & enforcement
Equipment parameter model library
Time-series data storage & query
Event-driven communication bus
Safety gateway & command control
Device communication & protocol adaptation
CLI / HTTP API service
Python bridge (cnpower/pandapower)
Topology-powerflow unified pipeline (planned)
Intelligent application scenarios covering the full lifecycle of power and energy.
Day-ahead / intra-day / real-time dispatch based on load forecasting and renewable generation, achieving optimal balance between economics and safety.
Equipment condition monitoring, fault diagnosis, and maintenance decisions — shifting from reactive repair to predictive maintenance.
Network expansion and equipment selection under load growth forecasting, supporting new power system construction.
Spot market bidding strategies and settlement analysis, adapting to the wave of power market reform.
Fault location, isolation, and non-fault area power restoration — millisecond-level response improving supply reliability.
Energy optimization and demand response for commercial and industrial users, supporting dual carbon goals.
EnerOS leads simultaneously across three dimensions: power physics modeling, AI nativeness, and safety constraints.
Built on Rust, start in a few commands. Prerequisites: Rust 1.70+ and Cargo.
14 phases of continuous evolution, from kernel foundation to deterministic decision loop.
Topology engine, power flow kernel, equipment model library
Agent lifecycle management, memory system, tool engine
Topology-aware injection, constraint verification guard, event bus
Multi-agent collaboration protocol, topology-structured communication
SCADA / IEC 61850 / IEC 104 / MQTT protocol adapters
Dispatch Agent, Operation Agent, Self-Healing Agent, domain collaboration protocol
SCADA data pipeline, DC-OPF, state estimation, web dashboard
E2E integration tests, ApiClient real HTTP, SQLite persistence
Deadlock fix, SelfHealingAgent interlock verification, clippy zero warnings
IEEE 14-bus accuracy verification, LlmReasoningEngine, degradation fallback
rig-core 0.38 integration, 4 power system tools materialized, unified reasoning engine
Priority command queue, real-time executor, watchdog timeout protection
Structured action output, feasibility projection, 3-stage decision pipeline, feedback re-reasoning
Eliminate "phantom loops", FeedbackLoop integration, 5 end-to-end closed-loop integration tests
From energy internet to embodied agents, EnerOS defines the intelligent foundation for next-generation power systems.
High penetration of renewables and power electronics brings fundamental grid transformation. Source-grid-load-storage coordinated control, virtual power plant aggregation, and microgrid autonomy will become the norm, urgently requiring agents that natively understand power physics to manage increasingly complex grids.
Deep fusion of large language models and domain reasoning engines, evolving from "knowledge retrieval" to "autonomous decision-making". LLMs provide intent understanding and natural language interaction, while EnerOS kernel provides physically feasible domain guarantees — together realizing trustworthy energy agents.
Multi-energy coupling of electricity, gas, heat, and transport — energy systems evolving toward internet architecture. EnerOS's graph-native architecture naturally fits multi-energy network topology modeling, providing a unified foundation for cross-energy-type intelligent scheduling.
Dual carbon goals drive deep energy restructuring; national unified power market construction accelerates. EnerOS's Trading Agent and Planning Agent will become the most active intelligent participants in the new power market.
From "software Agents" to "embodied Agents" — inspection robots, smart switchgear, and autonomous maintenance equipment will become embodied agents in the grid. EnerOS's Agent-as-Grid-Node paradigm natively supports intelligent integration of physical entities.
Just as Linux is to the internet and Android is to mobile, EnerOS aims to be the OS foundation of the energy intelligence era — a universal platform enabling thousands of energy agents to collaborate safely, efficiently, and natively.
Five core principles at the engineering implementation level.
Build an OS power-natively from the ground up. Converge Energy, Orchestrate Intelligence.