用物理仿真 + AI Agent，把热力站做成 AI 原生平台。
Physics-grounded AI control for district heating substations.
Modelica 提供可信的物理世界，FMI/FMU 打通 Python 控制层，MPC / RL Agent 在仿真里学策略——像 BOPTEST 之于建筑，HeatStationGym 之于热力站。
Modelica gives a trustworthy physical world, FMI/FMU bridges Python control, MPC / RL Agents learn policies in simulation — BOPTEST for buildings, HeatStationGym for substations.
把物理世界变成 AI 能用的世界
Turn the physical world into something AI can use
像 Tesla 的视觉系统把摄像头画面变成点云、可行驶区域、3D 障碍物——同一时刻三种视角：左边是现实里的换热站，中间是 AI 看到的物理抽象（带方程的世界模型），右边是 AI 做出的预测与控制决策。这就是世界模型干的事。
Like Tesla's vision stack turns camera frames into point clouds, drivable surfaces, 3D obstacles — three lenses on the same instant: left = the physical station, middle = the AI's abstraction with equations, right = the AI's prediction and control. That's what a world model is.
热力站的三个真实痛点
Three real pains in substation operations
调控靠经验
Tuning lives in the head of a senior operator
每个站的策略依赖老师傅；节能空间大但难复用，新站从零开始。
Each station depends on tribal knowledge; savings exist but don't scale to new stations.
数字化止于看板
Dashboards, not decisions
SCADA 收了大量数据，但没人用得起来；AI 落不进控制环。
SCADA collects everything, but data rarely closes the control loop.
建模门槛高
Physical modeling is expensive
Modelica/CFD 是专业工具，每个站建一遍要几周到几个月。
Modelica/CFD is expert tooling; modeling a station can take weeks to months.
五层架构
Five-layer architecture
v0 模型：一个最简换热站
v0 model: a minimal heat substation
一次侧热水经控制阀进入板式换热器，把热量传给二次侧；二次侧循环泵把热水送到末端用户。控制变量是一次侧阀门开度 u_valve，扰动是室外温度 T_oa 与用户负荷 Q_load。
Primary hot water enters a plate HX through a control valve and transfers heat to the secondary loop; a circulation pump delivers heat to building loads. Control = primary valve opening u_valve; disturbances = outdoor temp T_oa and load Q_load.
在 web/public/station-photo.jpg 放入真实换热站照片
Drop a real station photo at web/public/station-photo.jpg
（建议 4:3，至少 1200px 宽）
(4:3, at least 1200px wide)
现实里有阀门、泵、管网、换热器、保温层；模型里只保留控制和能量平衡决定的部分。AI 不需要看见钢铁，但需要一个可信的能量账本。
Reality has valves, pumps, piping, exchangers, insulation. The model keeps only what control and energy balance need. AI doesn't need to see the steel — it needs a trustworthy energy ledger.
查看 Modelica 源码核心方程View core Modelica equations
m_dot_pri = u_valve * m_dot_pri_max; C_min = min(m_dot_pri * cp, m_dot_sec * cp); Q_hx = eps_hx * C_min * (T_pri_in - T_sec_ret); T_sec_ret = T_sec_sup - Q_load / (m_dot_sec * cp); C_sec * der(T_sec_sup) = Q_hx - Q_load;
v1 路线：用 AixLib 组件级重建
v1 roadmap: rebuild with AixLib at component level
v0 是集总参数模型，够 demo 但不够工程。v1 把每个组件——阀门、泵、传感器、换热器——换成 AixLib 的工业级实现，保留同样的 FMI 接口，控制器代码零改动。
v0 is a lumped model — enough for demo, not engineering. v1 swaps every component (valve, pump, sensors, HX) for AixLib's industrial implementations, keeping the same FMI surface; controller code is unchanged.
v0 Demo：MPC vs PI，24 小时
v0 Demo: MPC vs PI, 24 hours
把上面的模型导出为 FMU，从 Python 控制；MPC 与带 feedforward 的工业 PI 对比。
Export the model above as an FMU and control it from Python; MPC vs industrial PI with feedforward.
| KPIKPI | PI | MPC |
|---|---|---|
| 跟踪 RMSTracking RMS | 10.30 °C | 0.008 °C |
| 满足率 (±1°C)Setpoint hit rate (±1°C) | 71.3 % | 100 % |
| 未满足负荷Unmet load | +0.034 MWh | −0.003 MWh |
| 阀门总变化Valve total variation | 8.8 | 1.1 |
PI 在头 7 小时出现冷启动极限环（典型 PI 现场调参痛点），MPC 开机即稳。这不是把 PI 调烂——是 MPC 真实的工程价值：不需要逐站手调。
PI shows a cold-start limit cycle in the first 7 hours (a real on-site tuning pain). MPC stays steady from start — that's MPC's actual value: no per-station tuning required.
v1.5：从单站走向源-网-站
v1.5: from single station to source-network-station
真正的供热价值在网级：冷水波传播、N-1 工况、安全约束触发的策略协同。下面是一个源 + 一次网 + 两站的耦合模型——上面所有的反事实、故障、Agent 工具，都可以无缝挂上来。
The real value in district heating lives at the network level: cold-water plug propagation, N-1 contingencies, safety-driven coordinated control. Below: a coupled source + primary network + 2-station model. All the what-if / fault / agent tooling above generalizes to this.
源温变化要走过 τ_A / τ_B 才能到达对应站
Source temp changes propagate to A/B with delays τ_A / τ_B
任一站开度变化通过一次网影响另一站可用流量与压力
Any station's valve change shifts flow / pressure available to the other
p_A / p_B 低于 0.3 MPa 触发协同回退（A 优先保供，B 退让）
If p_A or p_B drops below 0.3 MPa, coordinated back-off (A prioritized, B yields)
网络级 sandbox（源 + 两站 + 一次网）
Network sandbox (source + two stations + primary network)
把模型当 sandbox：反事实、故障、预报
Use the model as a sandbox: what-if, fault, forecast
物理仿真和世界模型的区别在这里：它不只是跑出结果，还能回答 '如果……会怎样' '哪个设备坏了' '未来 24 小时怎么走'。下面是三个互动模式，全部在浏览器里实时算。
This is where physical simulation becomes a world model: it answers 'what-if', 'which fault matches these symptoms', 'how will the next 24h play out'. All three modes run live in your browser.
让 AI Agent 用这个模型
Let an AI agent use the model
同样的物理模型，挂上 Claude Sonnet 4.6，给它 3 个工具（反事实仿真、故障注入、未来预报）。问一个自然语言问题，Agent 自己决定调哪个工具、怎么解读结果。这就是 '世界模型' 真正的差异。
Same physics model, behind Claude Sonnet 4.6 with 3 tools (counterfactual sim, fault injection, forecast). Ask in natural language; the agent picks tools and interprets results. This is what makes it a world model, not just a simulator.
问问这个世界模型
Ask the world model
gpt-5.1 · 5 toolsv2 · Modeling Agent：从点表自动生成模型
v2 · Modeling Agent: SCADA point list → Modelica model
一个站房从图纸/点表到可仿真的 Modelica 模型，传统做法需要 1-2 周工程师手工建模。这里换一个思路：LLM 不直接写 Modelica（容易幻觉），而是把点表抽取成一个有类型约束的拓扑（YAML），由确定性 JS 做 DOF / 量纲 / 接线预检查，再由模板生成 .mo 文件。下面试试看——粘贴一个点表，几秒钟出模型。
Traditionally, going from a point list / P&ID to a runnable Modelica model takes engineers 1–2 weeks. New approach: the LLM never writes Modelica directly (hallucinates), it fills a typed topology (YAML). Deterministic JS runs DOF / unit / connectivity checks, then a template renders the .mo file. Paste a point list below — model in seconds.
为什么这个 v0 已经是真护城河（不是 demo 装饰）
Why this v0 is already a real moat (not just a demo)
① LLM 输出受 Zod schema 约束 → tag 必须出现在原始点表里，不能幻觉新名字；② JS 做的预检查（DOF / 量纲 / 连接）覆盖了 LLM 最容易出错的物理一致性环节；③ 同一 schema 接入 AixLib 或更精细模板就能产生组件级模型；④ 校准是下一步——接 SCADA 历史数据做参数辨识，把 eps_hx、C_sec 这些默认值变成站特异值。① LLM output is Zod-constrained → every tag must exist verbatim in the original list, no invented names; ② JS pre-checks (DOF / units / connectivity) cover the physics-consistency failure modes LLMs are worst at; ③ Same schema can drive AixLib component-level templates; ④ Calibration is next — fit eps_hx and C_sec from SCADA history to make defaults station-specific.路线图
Roadmap
- v0HeatStationGym 雏形 · FMU + MPC + 对比 KPIHeatStationGym v0 · FMU + MPC + comparison KPIs已完成done
- v1接入真实 SCADA · 参数辨识 · 残差对比Real SCADA · parameter identification · residual analysis
- v2Modeling Agent · 点表→YAML→Modelica 自动生成Modeling Agent · point-list → YAML → Modelica auto-gen
- v3RL baseline · 多回路 · 一次网拓扑RL baseline · multi-loop · primary network topology
- v4多站示范部署 · 商业闭环Multi-station pilot deployment · commercial close
联系我们
Contact
微信