Intelligent building envelope dynamic temperature adjusting system based on reinforcement learning

Abstract

The invention relates to the technical field of dynamic optimization of building thermal resistance, and discloses an intelligent building envelope structure dynamic temperature adjustment system based on reinforcement learning. The system comprises a data acquisition layer, a multi-agent decision-making layer, an envelope structure execution layer and an energy consumption optimization layer. The data acquisition layer is used for detecting meteorological data of a current area, state data of an enclosure structure and temperature and humidity of an indoor environment. According to the system, through innovative material-algorithm collaborative design, dynamic optimization and accurate regulation and control of thermal performance of a building are achieved, dynamic stepless regulation of thermal resistance of an enclosure structure body is achieved for the first time, the core pain point that static design is difficult to adapt to dynamic climate change is solved, the system has self-learning ability and physical constraint guarantee, and the system has good application prospects. The method has wide application prospects in the fields of super high-rise buildings, data centers, medical clean spaces and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dynamic optimization of building thermal resistance, in particular to an intelligent building envelope dynamic temperature regulation system based on reinforcement learning. BACKGROUND

[0002] The severe situation of global energy consumption, according to the International Energy Agency (IEA), global building operation energy consumption accounts for 36% of total terminal energy consumption (2022 data), of which heating, ventilation and air conditioning (HVAC) accounts for 50-60%. China's urbanization rate has broken through 65% (2023), the stock of glass curtain wall buildings exceeds 1 billion square meters, and the annual increase in air conditioning load is 50 million kW, equivalent to the annual power generation of 5 Three Gorges power plants. In 2023, record-breaking high temperatures occurred in many parts of the world (such as Beijing 43℃, Spain 47℃), and traditional building thermal inertia design is difficult to cope with continuous heat waves, and real-time adaptive systems are needed for regulation, and the peak-valley difference of electricity price is expanding, and buildings need to have demand response capability to actively reduce load during peak electricity price period. Renewable energy volatility (such as photovoltaic power generation intermittency) requires building envelopes to have energy buffer function, and edge computing hardware computing power needs to be improved, making real-time physical field simulation + AI decision-making possible.

[0003] Limitations of traditional energy-saving technologies, existing building envelopes (such as double-layer Low-E glass, rock wool insulation layer) are designed based on the most unfavorable working conditions, and static thermal design cannot adapt to dynamic climate conditions and indoor load changes, resulting in high air conditioning energy consumption, which cannot dynamically adapt to changes in day and night temperature difference and seasonal changes; traditional PCM application relies on natural heat storage and release, and cannot actively control the rhythm, such as insufficient heat storage during the day leading to air conditioning overload at night, and existing technologies (CN119684975A) provide a formula and production process for improving the thermal conductivity of phase change materials and refrigerant, which does not solve the problem of dynamic control, and the singleness of intelligent control systems: CN119468418A discloses an air conditioning equipment group energy-saving control method and system based on big data driving, relating to the field of air conditioning equipment group control, but ignores the optimization of the thermal performance of the envelope itself, resulting in that the energy-saving potential has not been fully tapped. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides an intelligent building envelope dynamic temperature regulation system based on reinforcement learning, which has the advantages of realizing dynamic optimization of building thermal resistance, and solves the above technical problems.

[0005] To achieve the above purpose, the present application provides the following technical scheme: an intelligent building envelope dynamic temperature regulation system based on reinforcement learning, comprising a data acquisition layer, a multi-agent decision layer, an envelope execution layer and an energy consumption optimization layer. The data acquisition layer is used to detect current regional meteorological data, building envelope state data and temperature and humidity of indoor environment; The multi-agent decision layer includes an edge computing unit and a multi-agent reinforcement learning engine, the edge computing unit is used to process all data collected in the data acquisition layer and perform preprocessing, the multi-agent reinforcement learning engine is based on the data obtained after preprocessing of the edge computing unit, and outputs commands to the building execution layer through the MADDPG algorithm carried; The building execution layer adjusts the temperature of the building based on the command output of the intelligent agent decision layer; The energy consumption optimization layer includes a digital twin model and an incremental learning module, the digital twin model uses EnergyPlus for real-time simulation, and the incremental learning module is used to feedback to the multi-agent decision layer.

[0006] The MADDPG algorithm in the multi-agent reinforcement learning engine coordinates the output of the agent through the Actor network and the Critic network, wherein the local observation state of the first i agent is specifically expressed as follows: Where, t represents the discrete time period of policy interaction and decision in the reinforcement learning system, is the core index of reinforcement learning training in the MADDPG algorithm, and is used for: state transition , action , k Indicates the continuous iteration round or optimization step in physical simulation, parameter calibration, prediction optimization, the optimization step in EnergyPlus parameter calibration, the evolution of all k is "offline" "model internal iteration", and does not participate in actual execution of policy reasoning, i Refers to the first i building facade area, And are the indoor temperature and humidity corresponding to the first i building facade area, , are the outdoor temperature and humidity corresponding to the first i building facade area, is the solar radiation intensity received by the first i building facade area, is the real-time electricity price corresponding to the moment, t is the heat load demand of the first building facade area, i is t ​The temporal features corresponding to each moment are obtained by concatenating the local observation states of 1 to N agents, as shown in the following expression: Where: s represents the concatenated set of all agents, s1 is the state vector (15-dimensional) of the first agent, and so on, and the action expression for each agent is as follows: in: This represents the set of concatenated action vectors. ~ These represent the action vectors corresponding to agents 1 through N, respectively. For the first i The driving voltage of the VO2 coating controlled by an intelligent agent. For the first i The opening degree of the PCM ventilation valve is controlled by an intelligent agent.

[0007] As a preferred technical solution of the present invention, the specific means of controlling the execution layer of the enclosure structure include variable thermal conductive coating drive and PCM ventilation valve regulation; The variable thermal conductivity coating drive is used to control the voltage across the VO2 film to adjust the thermal conductivity of the VO2 film. The PCM ventilation valve is controlled by controlling the stepper motor to drive the louvers, thereby controlling the rate of heat storage and release.

[0008] As a preferred embodiment of the present invention, the digital twin model is simulated in real time using EnergyPlus, and the building thermodynamic model is updated every 15 minutes. The specific expression is as follows: in: This includes parameters related to the thermal conductivity of the wall, the U-value of the window, the thickness of the insulation layer, the heat capacity of the material, the specific heat capacity of each component, the thermal diffusivity of the wall, the internal surface thermal resistance, the external surface heat transfer coefficient, the heat capacity and enthalpy of the phase change material (PCM) layer, the thermal conductivity of the VO2 thin film, and other parameters related to the thermal performance of the material. argmin represents the parameter corresponding to the minimum value. for k The corresponding measured indoor temperature, heat flux on the wall or curtain wall surface, total energy consumption of the area, and electrical input power (power consumption of VO2 drive circuit and PCM valve motor) at any given time. This is the simulation output of the EnergyPlus model under the current input conditions, executed every 15 minutes. The simulation output generated by EnergyPlus for the digital twin model is as follows: in: fork the measured indoor temperature corresponding to the time, the wall or curtain wall surface heat flux, the total energy consumption of the region and the electrical input power (VO2 driving circuit and PCM valve motor power consumption).

[0009] wherein: θ is the thermal parameter, is the input condition, is the forward simulation function of the EnergyPlus model, and the digital twin model feeds back the output to the Critic network at time k the future Δt load and temperature for the next wherein: is the predicted future temperature, thermal load, is the input condition corresponding to time k, and the output of the digital twin model is fed back to the Critic network.

[0010] Compared with the prior art, the present application provides an intelligent building envelope dynamic temperature regulation system based on reinforcement learning, which has the following beneficial effects: The present application realizes dynamic optimization and precise regulation and control of building thermal performance through innovative material-algorithm collaborative design, and for the first time realizes dynamic stepless regulation of the envelope thermal resistance, solves the core pain point that static design is difficult to adapt to dynamic changes in climate, has self-learning ability and physical constraint protection, and has broad application prospects in super high-rise buildings, data centers and medical clean spaces. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a structural schematic diagram of the present application. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0013] Please refer to Figure 1 , an intelligent building envelope dynamic temperature regulation system based on reinforcement learning, comprising a data acquisition layer, a multi-agent decision-making layer, an envelope execution layer and an energy consumption optimization layer. The data acquisition layer is respectively meteorological data acquisition, building envelope state monitoring and indoor environment perception. The meteorological data acquisition can combine local weather forecast and outdoor roof micro weather station to collect temperature, radiation and wind speed and the like around the building; the building envelope state monitoring is longitudinally laid along the outer wall by distributed optical fiber temperature measurement (DTS), and there is one measuring point per meter; the indoor environment perception is deployed in a 10m*10m grid by a wireless temperature and humidity sensor network (Zigbee protocol); The multi-agent decision layer includes an edge computing unit and a multi-agent reinforcement learning engine. The edge computing unit is used for processing all the data collected in the data acquisition layer and performing preprocessing. The multi-agent reinforcement learning engine is based on the data obtained after preprocessing by the edge computing unit, and outputs commands to the building envelope execution layer through the MADDPG algorithm. Compared with the prior art, the preferred technical scheme of the present application has more advantages. The multi-agent framework can simultaneously optimize the heat distribution and cooperation relationship between different facades of the building, breaking through the limitations of single-fan isolated optimization. The state input dimension and multi-modal perception (15 dimensions) are used, and the richer environmental and economic inputs enable the Actor to make fine temperature adjustment decisions in higher dimensions. Compared with the fixed K1 and K2 weight values in the prior art, the weight parameters are introduced α, β, γ The dynamic adjustment of the operation target is realized. The Critic is enhanced by using digital twinning and online simulation, and EnergyPlus real-time calibration and short-term prediction, joint state-action and simulation prediction and least square fitting of thermal parameters every 15 minutes Compared with the prior art, the experience formula, no measured state transition and no dynamic model updating are more reasonable. In terms of algorithm, the double Critic + target network soft update in MADDPG is adopted , which improves the robustness and convergence performance in a multi-dimensional complex environment. The soft update combined with simulation enhancement ensures smooth evolution of the strategy. The above improvements comprehensively surpass the AC network control mode in the comparative patent, which is based on a single Agent, a single temperature input and a fixed strategy, fully embodying the creative breakthrough of the present application in algorithm architecture, system integration and application scenarios. The multi-agent decision layer is divided into an edge computing unit and a multi-agent reinforcement learning engine. The edge computing unit receives data from the data acquisition layer and performs data preprocessing. NVIDIA Jetson AGX Orin with 32GB memory and supporting TensorRT acceleration inference can be used. The multi-agent reinforcement learning engine uses the MADDPG algorithm (Multi-Agent Deep Deterministic Policy Gradient), which is an algorithm for multi-agent reinforcement learning (MARL). The joint state and action of all Agents are input, and the Q value is output for evaluation. Finally, the corresponding decision is made, and the command is output to the execution layer.

[0014] The specific description of MADDPG in the intelligent building envelope dynamic temperature adjustment system introduces the multi-agent deep deterministic policy gradient (MADDPG) framework. Under the centralized training and decentralized execution (CTDE) paradigm, the algorithm coordinates the continuous control problem of multiple agents by each agent having an independent actor network and a centralized critic network. The local observation state of each agent is: i where t represents the discrete time period of policy interaction and decision-making in the reinforcement learning system, and is the core index of reinforcement learning training in the MADDPG algorithm, which is used for state transition , action , k represents the continuous iteration round or optimization step in physical simulation, parameter calibration, and prediction optimization, the optimization step in EnergyPlus parameter calibration, and the evolution of all k is "offline" "model internal iteration" and does not participate in actual execution of policy reasoning, i refers to the i th building facade area, and are the indoor temperature and humidity of the i th building facade area, , are the outdoor temperature and humidity of the i th building facade area, is the solar radiation intensity received by the i th building facade area, is the real-time electricity price corresponding to the time t , is the heat load demand of the i th building facade area, is the time feature corresponding to the time t , where s represents the set of all agent spatch states, ~ represent the state vectors of 1~N agents, respectively, and the others are similar, and the action expression corresponding to each agent is as follows: where: ~ represent the action vectors corresponding to 1~N agents, ​VO2 coating driving voltage controlled by the i th agent, VO2 coating driving voltage controlled by the i th agent, PCM damper opening controlled by the th agent; Critic network updates: i The Critic network takes joint state and joint action as input to estimate the state-action value function of the th agent. The Critic parameters are updated by minimizing the mean squared Bellman residual loss function: Target value is defined as where is the discount factor, , are the target network parameters of Critic and Actor respectively, D is the experience replay buffer; Actor network updates: The Actor network generates actions for the i th agent given local observations. Its parameters are optimized by policy gradient ascent: This gradient, which utilizes global information provided by the Critic network, guides the Actor to adjust the output actions to improve the overall return.

[0015] Target network soft updates: The trained MADDPG model is exported in TensorFlow Lite format. To stabilize training, the target network parameters and adopt a soft update strategy, fixing the first L-1 layers of the Actor and Critic, and only updating the Lth layer parameters ϕ online: Reward function design: The immediate reward of the i th agent is composed of energy saving, comfort, and cost indicators: where is the energy consumption under agent control, is the comfort temperature set point, for the rate of temperature fluctuation, α , β , γ is a weight coefficient; The building execution layer adjusts the temperature of the building based on the command output of the agent decision layer; The building execution layer is divided into a variable thermal conductive coating drive and a PCM ventilation valve regulation. The variable thermal conductive coating drive can adjust the thermal conductivity by controlling the voltage. Specifically, when a voltage is applied across the VO2 thin film, Joule heat is generated, causing the temperature of the thin film to rise. When the temperature reaches the phase transition temperature, the VO2 thin film changes from an insulating state to a metallic state. By controlling the size and duration of the voltage, the temperature of the VO2 thin film can be accurately controlled. The PCM ventilation valve regulation controls the rate of heat storage and release by controlling the step motor to drive the louvers. The thermal resistance is dynamically distributed, and the air conditioning energy consumption is monitored; The energy consumption optimization layer is divided into a digital twin model and an incremental learning module. The digital twin model can be simulated in real time using EnergyPlus, and the building thermodynamic model is updated every 15 minutes to accurately predict current and future thermal loads and energy saving potential. The digital twin is defined by EnergyPlus as follows: The simulation output generated by EnergyPlus for the digital twin is as follows: wherein: θ is the thermal parameter (wall thermal conductivity, window U value, etc.), is the input condition (weather, load, control action), the calibration module solves the building thermal model parameters θ by constructing an optimization problem with the objective of minimizing the error between the simulation output and the measured observation value, so that the digital twin model remains consistent with the real building, which is mathematically expressed as: wherein: including wall thermal conductivity, window U value, insulation layer thickness, material heat capacity, specific heat capacity of each component, wall thermal diffusivity, internal surface thermal resistance, external surface heat transfer coefficient, phase change material (PCM) layer heat capacity and phase change enthalpy, VO2 thin film thermal conductivity and other parameters related to material thermal performance, is k the measured indoor temperature, heat flux, etc. at the moment, the measured values include temperature and humidity, heat flux, energy consumption, etc. physical quantities collected by field sensors; the simulation output is calculated by the EnergyPlus model; The energy consumption optimization layer includes a digital twin model and an incremental learning module. The digital twin model is simulated in real time using EnergyPlus, and the incremental learning module is used to feedback to the multi-agent decision layer; Digital twin at time k To the future Δt The prediction of load and temperature for the next 15 minutes (typical) is: This prediction is incorporated into the target value calculation of the MADDPG Critic, which estimates the Q value of the next state; The purpose of real-time simulation of the digital twin model is to correct the key parameters in the model, such as heat transfer coefficients and equipment efficiency, in time by combining the physical simulation of EnergyPlus (simulating the structure, materials, HVAC system, etc. of the building) with real-time sensor data (temperature, humidity, radiation, wind speed, etc.) to keep the virtual model consistent with the real building behavior. Based on the current state and weather forecast, the digital twin can predict the thermal load and temperature changes for the next 15-60 minutes, providing short-term policy evaluation and "virtual commissioning" capabilities for the RL agent, reducing the cost of trial and error in the field. Strategy simulation verification: Before the control is officially issued at the execution layer, the system can first simulate multiple action combinations (such as different VO2 coating voltages and PCM valve openings) in parallel in the digital twin to estimate their energy consumption and temperature response, to help the Critic network in MADDPG more accurately evaluate the Q value; The incremental learning module uses TensorFlowLite for fine-tuning, and updates the MARL model weights with new data every week or month to adapt to environmental changes and equipment aging; The fine-tuning process and purpose of the incremental learning module, the trained MARL model is exported in TensorFlowLite format and deployed on the edge computing unit to support low-latency policy inference and updates. The system summarizes new state-action-reward triplets generated in actual operation every week or several days, uses the incremental training capability of TensorFlowLite (freezes part of the network layers and only opens the last one or two layers for fine-tuning) to perform several training epochs on the model to adapt to new environments or load changes, replaces the Actor and Critic target network parameters in the edge unit with the fine-tuned new weights, and ensures that subsequent inference remains high-performance in the latest environment; The incremental learning module uses TensorFlowLite for fine-tuning, and updates the MARL model weights with new data every week or month to adapt to environmental changes and equipment aging; Initialization: build Actor, Critic and corresponding target network for each Agent; initialize experience replay pool. Centralized training loop: all Agents interact with the environment and collect to the replay pool; periodically sample from the pool to update the Critic and the Actor; perform soft updates. Decentralized execution: after training converges, each Agent only calls the local Actor, and in actual operation, it is based on independent action generation, online control with low communication demand. This MADDPG scheme combines a central Critic with a local Actor to achieve coordinated control of VO2 coating and PCM valve in building envelope, balancing energy saving and comfort, and is suitable for large-scale multi-Agent scenarios; Overall process State acquisition: obtain real state values (including temperature, humidity, radiation, electricity price, etc.); Digital twin calibration: ; Bionic prediction: ; Policy inference: ; Action execution: issue VO2 coating driving voltage and PCM ventilation valve opening ; Feedback sampling: ; Experience storage: ; Small batch update: Critic and Actor are updated according to the formula; Incremental fine-tuning (weekly / daily): ; Target network software update, closed combination of digital twin and MADDPG, through continuous simulation calibration and online fine-tuning, to ensure that the temperature regulation strategy remains optimal in the real building environment at all times; Material hardware arrangement (1) Intelligent envelope hardware, material selection of phase change material (PCM) layer: octadecane (C 18 H 38 , melting point 28℃), phase change enthalpy ≥180 kJ / kg, suitable for human comfort temperature range, modifier is 5% expanded graphite, thermal conductivity coefficient from 0.2 to 5 W / m·K, accelerating phase change response; Packaging process: microencapsulation uses interfacial polymerization to prepare urea-formaldehyde resin shell (thickness 2-5μm); Construction method: mix with silicate thermal insulation mortar at a volume ratio of 1:4, mechanically spray to a thickness of 20mm; (2) Variable thermal conductivity coating, VO2 film preparation: magnetron sputtering method, thickness 300±50nm, phase transition temperature 68℃ (can be reduced to 40℃ by doping tungsten), thermal conductivity coefficient switching ratio 3:1, visible light transmittance >40% (satisfies lighting requirements); (3) Drive circuit, low power design, 12V DC power supply, maximum current 0.1A (single flat meter power consumption ≤1.2W); Multi-agent reinforcement learning (MARL) algorithm, agent division strategy: divide the building facade into an agent every 10 floors (such as a 200m high building divided into 20 agents), and each agent only obtains the sensor data of the region (reduces the state space dimension); Algorithm implementation process, create Actor and Critic networks (including target network) for each agent during initialization.

[0016] Training loop, select the corresponding operation for each agent according to the current strategy.

[0017] Execute the operation, and the environment returns the rewards and next state of all agents.

[0018] Store the global transition experience (joint state, action, reward, etc.) in the replay pool.

[0019] Sample a batch of data from the replay pool and update the Critic and Actor. Critic update: minimize TD error; Actor update: maximize the Q value evaluated by Critic through gradient ascent.

[0020] Make the corresponding decision and output the command to the execution layer; Layer execution command, the execution layer is the physical response terminal of the system, responsible for converting the AI control instructions generated by the decision layer into material property changes and mechanical actions, including the following two modules: variable thermal conductivity coating (VO2 film) drive, VO2 undergoes an insulator-metal phase transition at 68°C, and the thermal conductivity coefficient jumps from 0.5 W / m·K (insulating state) to 3.0 W / m·K (metallic state). By applying voltage, the phase transition can be triggered artificially (without natural heating); Control is divided into low thermal conductivity mode (0.5 W / m·K), which reflects solar radiation and reduces heat transfer (summer day), and high thermal conductivity mode (3.0 W / m·K), which accelerates heat dissipation (winter night or data center emergency heat dissipation); Execution steps Receive instructions: the edge computing node sends voltage instructions through PLC; Voltage conversion: the drive circuit converts digital signals to analog voltage (0-10VDC, accuracy ±0.1V); Thin film response: voltage is applied to the VO2 thin film electrode to trigger carrier migration and change the crystal lattice structure (response time ≤100ms); State feedback: real-time monitoring of thin film resistance value (normal range: insulating state 10 6Ω→metallic state 10²Ω), ensuring that the phase transition is complete; At the same time, the PCM ventilation valve is adjusted according to the opening degree command and converted into the number of pulses of the stepper motor (32 pulses per 1% opening degree). Motor drive, ULN drive board controls motor phase (four-phase eight-beat mode), speed 15 RPM. Position calibration, limit switch (Omron SS-5GL) resets zero point to avoid cumulative error. State feedback, Hall sensor (A) detects blade angle and real-time feedback of actual opening degree (accuracy ±1%). VO2 coating electrode wiring, silver paste printed electrode (line width 0.5 mm, spacing 10 mm) is used to avoid shading; the electrode end is connected to a waterproof junction box (IP67 level). PCM ventilation valve deployment: 4 ventilation valves per square meter of curtain wall (spacing 500 mm); air duct design: installed at an angle of 30° to drain dust by gravity. Control signal transmission, communication protocol: VO2 drive: Modbus RTU over RS-485 (address code allocation, baud rate); ventilation valve control: CAN bus (message ID distinguishes regions). Fault tolerance mechanism: heartbeat packet detection (every 5 seconds), 3 times of timeout triggers safety mode (VO2=1.5 W / m·K, ventilation valve=50%). Maintenance and fault handling: VO2 coating is wiped with ethanol every quarter to maintain light transmittance, and electrode resistance is detected every year (normal value 10²-10 6 Ω); PCM ventilation valve removes dust from blades every month (compressed air blowing); lubricate gear set every half year (food-grade silicone grease). The execution layer converts AI decisions into precise responses in the physical world through the physical property regulation of VO2 film and the mechanical action of PCM ventilation valve. Its technical parameters (such as thermal conductivity switching ratio 6:1, ventilation valve response time 8 seconds) and collaborative control strategy solve the pain points of traditional building envelope rigidity and slow response, providing reliable technical support for dynamic energy saving. Extended application and future upgrade, photovoltaic integration, superimpose transparent perovskite solar cell (efficiency >15%) on VO2 coating to achieve energy self-sufficiency; digital twin fusion, pre-train MARL algorithm through BIM model to shorten on-site debugging cycle to 24 hours.

[0021] Energy saving index: air conditioning energy consumption is reduced by 40% (compared with traditional curtain wall).

[0022] Comfort improvement: indoor temperature fluctuation range is reduced from ±3℃ to ±0.5℃.

[0023] Economic: 3-5 years payback period (peak-valley electricity price difference > 0.8 yuan / kWh).

[0024] This technology will reshape the paradigm of building energy saving, driving the industry from "equipment energy saving" to "structural intelligence", with significant social, environmental and commercial value.

[0025] Significant reduction in building energy consumption, air conditioning energy saving rate more than 40%, through the dynamic reflection of VO2 coating and the cooperative control of PCM heat storage and release, reducing solar radiation heat gain and cold loss.

[0026] Improve indoor thermal comfort, temperature fluctuation range reduced by 80%, temperature fluctuation reduced to ±0.7℃, meet ISO AA level comfort standard (fluctuation ≤1.0℃). Thermal inhomogeneity is improved, multi-agent regional control reduces the temperature difference between building facades from 5℃ to 1℃, eliminating local overheating problems such as "west sun".

[0027] Economic advantage is outstanding, investment payback period is short, operation and maintenance cost is low, there is no moving wear parts (VO2 is a solid film), annual maintenance cost is less than 5 yuan / ㎡, which is lower than the maintenance cost of 15-20 yuan / ㎡ per year of traditional air conditioning system.

[0028] Environmental and social benefits: significant carbon emission reduction, helping to achieve the "double carbon" goal, through demand response (such as night cold storage), reducing the pressure of peak-valley difference of power grid, improving the consumption capacity of renewable energy, and alleviating the contradiction between power supply and demand.

[0029] Resource sustainability is good, PCM microcapsule recyclability is more than 90% (traditional insulation material recyclability is less than 30%), VO2 coating service life is more than 15 years, reducing the frequency of building renovation.

[0030] Technological innovation: dynamic thermal resistance adjustment, realizing real-time stepless adjustment of building envelope thermal conductivity (0.5-3.0 W / m·K), breaking through the limitation of static thermal performance of traditional materials.

[0031] Good reliability and adaptability, still maintaining reflectivity > 85% at 45℃ high temperature (traditional curtain wall reflectivity < 60%), automatically switching to safety mode (ventilation valve half open, VO2 medium heat conduction state) in typhoon and rainstorm weather, and long-term stability, VO2 coating performance decay < 5% after 100,000 cycles, PCM microcapsule cycle life > 10,000 times (national standard requirement).

[0032] Shanghai Global Financial Center Intelligent Curtain Wall Renovation Project Project background: Shanghai Global Financial Center (492 meters high, 101 floors), glass curtain wall area of 120,000 square meters, annual average air conditioning energy consumption of 120 million kWh, accounting for 35% of operating cost.

[0033] Pain point analysis: The surface temperature of the curtain wall reaches 65℃ in summer, and the peak cooling load is 8MW; the existing Low-E glass + sunshade curtain solution only saves 15% energy and affects lighting; the peak-valley electricity price difference reaches 1.0 yuan / kWh (peak period 10:00-15:00).

[0034] Patent technology implementation, hardware modification: Use a magnetron sputtering device, VO2 coating area 80,000 square meters, thickness 320 nm, driving voltage 0-10VDC; PCM microcapsule layer uses construction unmanned aerial vehicle or manual spraying thickness 20mm, phase change enthalpy 185kJ / kg, ventilation valve density: 4 / ㎡.

[0035] Sensor deployment: Optical fiber temperature measurement points (spacing 10m), weather station roof + every 30-layer overhanging platform, optical fiber pipe laying (bending radius >10cm), LoRa base station coverage (transmission distance 500m).

[0036] AI system deployment, edge computing node, deployment location every 20-layer equipment room (total 5 units).

[0037] Algorithm configuration: MADDPG parameters (single intelligent agent).

[0038] state_dim=15 action_dim=2 hidden_dim=256 batch_size=1024 Training data: historical data, one-year hourly energy consumption records (17,520 groups), simulation data, EnergyPlus generated extreme weather scenarios (typhoon, heat wave).

[0039] Key event analysis, 10:00-12:00 (peak electricity price), VO2 switches to 0.8V (thermal conductivity 1.0W / m·K), reflectivity increases to 82%, ventilation valve opening 70%, uses natural convection to dissipate heat, reduces mechanical refrigeration demand; air conditioning power from 2.8MW to 1.6MW (energy saving 43%).

[0040] 14:00 (extreme high temperature): AI detects that the glass temperature exceeds the threshold (60℃), triggers emergency mode: VO2 forced 0.5V (minimum thermal conductivity), reflectivity 88%; ventilation valve fully open (90%), start PCM layer active water cooling cycle (alternative patent technology).

[0041] Energy consumption comparison, monthly average refrigeration energy consumption before modification 9.8 million kWh, after modification 5.88 million kWh, reduced by 40%; indoor temperature compliance rate, before modification 78%, after modification 95%, increased by 17%.

[0042] Economic comparison, the cost of transformation is 10 million yuan (267 yuan / ㎡), 392 million kWh x 1.0 yuan = 392 million yuan / year, 8.2 years to recover, reduce 392 million kWh x 0.785 kg CO2 / kWh = 3074 tons of carbon emissions.

[0043] This example verifies the feasibility of the patented technology in super high-rise buildings, and the technical value of 40% energy saving is far more than traditional energy saving reconstruction (usually <20%); the commercial value of 8-year payback period meets the investment expectations of commercial real estate.

[0044] The social value of a single building reduces carbon tons per year, equivalent to planting 160,000 trees.

[0045] This case can provide a replicable technical paradigm for similar buildings (such as Guangzhou East Tower and Shenzhen Ping'an Center).

[0046] This scheme realizes the revolutionary upgrade of building envelope from "passive insulation" to "active temperature regulation" through the deep integration of dynamic material property control and distributed AI decision-making. Compared with existing technologies, it has significant advantages in energy consumption, comfort, and economy, and promotes the industry from "equipment energy saving" to "structure intelligence", with significant social, environmental and commercial value.

[0047] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A reinforcement learning based intelligent building envelope dynamic temperature regulating system, characterized in that: The system comprises a data collection layer, a multi-agent decision-making layer, a building envelope execution layer and an energy consumption optimization layer. The data collection layer is used to detect current regional meteorological data, building envelope state data and indoor temperature and humidity. The multi-agent decision-making layer comprises an edge computing unit and a multi-agent reinforcement learning engine. The edge computing unit is used to process all data collected in the data collection layer and perform preprocessing. The multi-agent reinforcement learning engine is based on the data obtained after preprocessing by the edge computing unit and outputs commands to the building envelope execution layer through the MADDPG algorithm.

2. The intelligent building envelope dynamic temperature regulating system based on reinforcement learning according to claim 1, characterized in that: The MADDPG algorithm in the multi-agent reinforcement learning engine coordinates the output actions of the agents through an Actor network and a Critic network, wherein the local observation state of the i th agent is specifically expressed as follows: i ​ Where t represents the discrete time period of policy interaction and decision-making in the reinforcement learning system, and is the core index for reinforcement learning training in the MADDPG algorithm, used for: state transition ,action , k This indicates the number of consecutive iterations, and the optimization step size in EnergyPlus parameter calibration. i Refers to the first i Each building facade area and The first i Indoor temperature and humidity corresponding to each building facade area , For the first i The outdoor temperature and humidity corresponding to each building facade area. For the first i The intensity of solar radiation received by each building facade area for t The real-time electricity price corresponding to that moment. For the first i Heat load requirements of each building facade area for t The temporal features corresponding to each moment are obtained by concatenating the local observation states of 1 to N agents, as shown in the following expression: Wherein: s represents the set of all agents after splicing, Respectively represent 1~N agent state vector, and so on, and each agent corresponding action expression is as follows:​ in: This represents the set of concatenated action vectors. ~ These represent the action vectors corresponding to agents 1 through N, respectively. For the first i The driving voltage of the VO2 coating controlled by an intelligent agent. For the first i The opening degree of the PCM ventilation valve is controlled by an intelligent agent.

3. The intelligent building envelope dynamic temperature regulating system based on reinforcement learning of claim 1, wherein: The building envelope execution layer adjusts the temperature of the building envelope based on the command output of the agent decision-making layer. The energy consumption optimization layer comprises a digital twin model and an incremental learning module. The digital twin model is simulated in real time by EnergyPlus.

4. The intelligent building envelope dynamic temperature regulating system based on reinforcement learning of claim 2, wherein: The incremental learning module is used to feedback to the multi-agent decision-making layer. wherein: including parameters related to the thermal performance of the materials, argmin denotes the minimum value corresponding parameter, is k the measured indoor temperature, the wall or curtain wall surface heat flux, the total energy consumption of the region and the electrical input power corresponding to the moment, is the simulation output of the EnergyPlus model under the current input condition, which is executed every 15 minutes, and the simulation output generated by EnergyPlus is wherein: represents k measured indoor temperature at the time, represents wall or curtain wall surface heat flux, represents total energy consumption of the area, represents electrical input power; wherein: The specific means of the building envelope execution layer for control include variable heat conduction coating driving and PCM ventilation valve regulation. is a thermal parameter, is an input condition, is a forward simulation function of the EnergyPlus model, the digital twin model at time k the prediction of load and temperature for the next The variable heat conduction coating driving is used to control the voltage across the VO2 thin film to adjust the heat conduction coefficient of the VO2 thin film. The PCM ventilation valve regulation controls the rate of heat storage and release by controlling the step motor to drive the louvers. The digital twin model is simulated in real time by EnergyPlus. The building thermodynamic model is updated every 15 minutes, and the specific expression is as follows: θ Δt minutes is: wherein: is the predicted future temperature, thermal load, is the input condition corresponding to the kth moment, and the output of the digital twin model is fed back to the Critic network.

Images