Building control method and system based on multi-source heterogeneous data and Pareto decision

Abstract

The invention relates to a building control method and system based on multi-source heterogeneous data and Pareto decision, and belongs to the technical field of intelligent building and automatic control, and the method comprises the steps: obtaining and fusing multi-field state data, such as building state data, biological state data and real-time environment data; adopting a non-standard quantization multi-target reinforcement learning model to calculate a Q vector representing building energy saving, indoor comfort and plant health; screening a Pareto optimal leading edge to generate an instruction set, and arbitrating an optimal action according to a real-time preference weight; and driving control is executed, and the building plug-in is adjusted to the optimal angle orientation. The invention further provides a building control system based on the multi-source heterogeneous data and the Pareto decision. According to the building control method and system based on the multi-source heterogeneous data and the Pareto decision, multi-conflict targets can be optimized in a collaborative mode, the problem of the multi-source heterogeneous data is solved, and the robustness and safety of the system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent buildings and automation control, and particularly relates to a building control method and system based on multi-source heterogeneous data and Pareto decision. BACKGROUND

[0002] With the acceleration of global urbanization, the urban heat island effect is becoming more and more obvious. Its main formation factors include artificial heat sources, building materials, etc. Among them, the artificial heat sources include the heat generated by a large number of traffic, industrial and building activities, thereby causing the urban temperature to rise. The heat island effect caused by building materials is mainly that the materials such as concrete and asphalt used in cities have high heat capacity and can quickly absorb and store heat, causing the urban surface temperature to be higher than the natural ground. In order to alleviate the urban heat island effect and improve the ecological benefits of buildings, building vertical greening (VGS), commonly known as "green wall", can significantly increase the urban green area and reduce the impact of heat island effect through cooling effect, and is widely considered as an effective means to improve the urban heat island effect. However, although the traditional green wall has certain heat insulation and transpiration cooling functions, its physical form is fixed and cannot adapt to the dynamic changes of day and night, seasons and weather conditions, which may block the necessary natural lighting in rainy days or block the beneficial solar radiation heat in winter, so that the comfort level needs to be improved by strengthening the lighting and increasing the heating power consumption of indoor air conditioning equipment, which in turn increases the energy consumption of buildings and is not conducive to energy saving and environmental protection.

[0003] In order to solve the above problems, a building control device integrated with a mechanical driving mechanism has emerged. This kind of device allows the building outer hanging module carrying vegetation to rotate or move around a certain axis, similar to building outer sunshade louvers, aiming to dynamically adjust the light and heat environment of the building by changing the angle. However, most of the current building control devices belong to reactive control, which only relies on single-dimensional real-time sensors such as illumination sensors or air temperature sensors and the control strategy is too simple, and cannot handle "multi-source heterogeneous" data. In addition, when the sensor data drifts due to aging or communication failure, the building control device may even output incorrect instructions, causing mechanical structure damage or falling risk. SUMMARY

[0004] The present application aims to provide a building control method and system based on multi-source heterogeneous data and Pareto decision, to solve the problem of insufficient multi-source heterogeneous data fusion capability of the building control device in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions: A building control method based on multi-source heterogeneous data and Pareto decision, the method comprising: First step: Constructing multi-dimensional heterogeneous state space and collecting data in real time; By deploying distributed sensor networks in the building envelope, greenery devices, and the external environment, real-time multi-field heterogeneous data is collected, and combined with third-party meteorological service data, a multi-dimensional state vector is constructed at all times . The state vector is defined as ; Wherein, represents the building thermal state sub-vector, including the heat flux density through the envelope and the indoor thermal environment parameters; represents the biological physiological state sub-vector, including soil and light parameters reflecting the vitality of vegetation; represents the environmental state sub-vector, including real-time micro-meteorological data and predicted meteorological sequences in the future time window .

[0006] Second step: Decision inference based on non-scalar multi-objective reinforcement learning model; The multi-dimensional state vector is input into the pre-trained non-scalar multi-objective optimization model. The model is constructed based on Markov Decision Process (MDP), and its tuple is defined as .

[0007] In this process, the model does not output a single scalar value, but maps a non-dominated action-value vector set through a multi-objective Q learning algorithm. For each candidate angle action in the action space , the model calculates its corresponding multi-dimensional Q vector , which represents the expected cumulative discounted return in the state after performing action in the building energy saving, comfort, and plant health multiple target dimensions:

[0008] Wherein, is the discount factor, is the vectorized reward function.

[0009] Third step: Generating Pareto optimal instruction set and preference arbitration; Based on the calculated Q vector, the non-dominated sorting algorithm is used to filter out the Pareto optimal frontier to generate a non-dominated instruction set .

[0010] Subsequently, the real-time preference weight vector set by the user or system (corresponding to energy saving, comfort, and plant weight, respectively, and by a scalarization function from optimal actions :

[0011] Fourth step: execute the driving control and safety protection mechanism; convert the selected optimal action into a pulse width modulation (PWM) signal or bus instruction to drive the mechanical actuator of the greening device to adjust to the target angle. At the same time, the real-time monitoring of the environmental wind speed , when it is detected that the preset safety threshold is exceeded, the highest priority safety interrupt is triggered to forcibly override the optimization instruction and reset the device to the minimum wind resistance angle .

[0012] Preferably, the construction of the vectorized reward function specifically includes: defining the reward vector at the moment .

[0013] Energy reward term : based on the instantaneous heat flux (unit: ) collected by the heat flux sensor. To minimize heat transfer through the building envelope (heat resistance in summer, heat preservation in winter), the following penalty function is defined:

[0014] wherein is a normalization coefficient.

[0015] Comfort reward term : based on the indoor predicted mean opinion score . When deviates from the preset comfort interval , a penalty is applied:

[0016] wherein is the comfort penalty coefficient.

[0017] Vegetation health reward term : combined with the photosynthetically active radiation intensity and soil moisture content , a segmented reward function is constructed:

[0018] wherein is the optimal light interval for plants,​ is the light compensation point, is the critical water content, is the wilting coefficient, is the positive weight constant.

[0019] Preferably, the training and updating of the non-quantized multi-objective optimization model employs the following mechanism: The vector iteration update rule based on Bellman equation is adopted. For state and action , the update formula of the vector is:

[0020] wherein, is the learning rate, the operator is used to filter out the representative vector from the set of all possible vectors of the next state based on the current preference distribution or hyper-volume index to solve the propagation problem of multi-objective value. The training process adopts a model-in-loop simulation architecture, jointly using the BCVTB interface to connect the EnergyPlus building energy consumption simulation engine and the reinforcement learning agent, using typical meteorological year (TMY) data for offline pre-training until the average Euclidean distance change rate of the vector set is less than the convergence threshold .

[0021] Preferably, the acquisition of multi-domain heterogeneous state data further comprises a data integrity guaranteeing step: An unsupervised learning algorithm (such as Isolation Forest or Variational Autoencoder VAE) is used to detect anomalies in the sensor array data. When the reconstruction error of a certain sensor data exceeds the threshold, it is determined to be data drift or failure.

[0022] Subsequently, a time series prediction model based on long short-term memory network (LSTM) is used to generate an interpolated value to replace the faulty data, ensuring the integrity of the input state vector .

[0023] Preferably, the greening device is integrated with a photovoltaic power generation module, and the vectorized reward function is expanded to a four-dimensional vector:

[0024] wherein, is the photovoltaic power generation reward item, defined as ,​ Real-time output power of the photovoltaic module. The model will seek a Pareto optimal solution between shading energy saving and maximizing power generation (possibly requiring tilt angle) at the time of decision.

[0025] The application also provides a building control system based on multi-source heterogeneous data and Pareto decision, the system comprising: Multi-field data acquisition and preprocessing module: used for acquiring real-time data through heat flux sensors, photosynthetically active radiation (PAR) sensors, soil moisture sensors and ultrasonic anemometers; and configured with a data cleaning unit for normalizing and abnormal value interpolation of heterogeneous data to generate a standardized multi-dimensional state vector .

[0026] Multi-objective intelligent decision-making module: built-in high-performance edge computing unit, running a pre-trained non-quantized multi-objective optimization model. The module receives as input, calculates the Q vector set corresponding to all actions in the action space through a multi-objective Q learning reasoning engine, and outputs a non-dominated optimal instruction set . The module aims to synergistically optimize the building thermal performance target and the plant vitality target.

[0027] Preference arbitration and safety monitoring module: used for receiving an external input preference weight vector , screening the final execution action from . The module runs a safety monitoring logic in parallel, and when the real-time wind speed or the predicted wind speed exceeds a safety threshold , a high-priority reset interrupt signal is generated to override the intelligent decision-making result.

[0028] Drive control execution module: connects the electric linear drive of the greening device through an industrial bus (such as RS485 or CAN), converts the angle instruction into mechanical displacement, and feeds back the position state of the actuator in real time.

[0029] Compared with the prior art, the application has the following beneficial effects: Multi-objective synergistic optimization and nonlinear trade-off, the application breaks through the limitations of traditional single-target control (such as only considering shading or only considering daylighting). By constructing a vectorized reward function containing , and , and using a multi-objective Q learning algorithm, the system can find a Pareto optimal solution among the three often conflicting targets of building energy saving, indoor comfort and plant health. For example, at noon in summer, the system can automatically trade off the contradiction between "completely closed to insulate" and "properly opened to prevent plant photoinhibition", and maximize the overall benefit.

[0030] The precision control based on biological physiological characteristics is different from the traditional control logic which ignores the plant characteristics, and the photosynthetically active radiation (PAR) interval and soil moisture constraint are explicitly introduced in the reward function. The optimal growth zone and stress zone of the plant are quantified by the formula, so as to ensure that the dynamic adjustment of the greening device will not sacrifice the vitality of the plant, and the key technical problems of high plant maintenance cost and low survival rate in the dynamic greening are solved.

[0031] The data-driven foresight and robustness are realized by fusing real-time microclimate data and future predicted meteorological information. The model predictive control (MPC) is used in the present application, which can cope with the coming bad weather or severe temperature change in advance. Meanwhile, the data integrity guarantee mechanism based on VAE and LSTM is introduced, which can maintain the stability of the decision through interpolation data in the case of sensor failure or data loss, and the high-priority wind speed safety interruption mechanism is combined, so that the survival ability and reliability of the system in the complex outdoor environment are significantly improved.

[0032] The flexible preference adaptability and expansibility are realized by generating a non-dominated instruction set instead of a single action, and the preference weight vector is changed to realize the adjustment of the control strategy. The control strategy (such as seamless switching from the "energy saving priority" mode to the "plant maintenance" mode) can be adjusted in real time without retraining the model. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The building control method flowchart based on multi-source heterogeneous data and Pareto decision of the present application is shown in the figure. Figure 2 The state and vectorized reward function definition diagram of the multi-objective decision optimization (MDP) in the present application is shown in the figure. Figure 3 The principle diagram of the multi-objective Q learning and Pareto front selection in the present application is shown in the figure. Figure 4 The flowchart of the safety protection mode in the present application is shown in the figure. Figure 5 The building control system block diagram based on multi-source heterogeneous data and Pareto decision of the present application is shown in the figure. DETAILED DESCRIPTION

[0034] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only used as examples, and other obvious modifications can be thought by those skilled in the art.

[0035] Please refer to Figure 1 In the first aspect of the present application, a building control method based on multi-source heterogeneous data and Pareto decision is provided, which includes core steps S101 to S104. S101, acquire multi-field heterogeneous state data and generate a multi-dimensional state vector. The system collects three types of data in real time through the deployed sensor network: 1) building state data , such as indoor temperature and humidity, wall heat flux; 2) biological state data , such as soil moisture, soil temperature, electrical conductivity (EC), and key photosynthetically active radiation (PAR) in greenery devices; 3) environmental data , such as real-time wind speed and direction obtained by ultrasonic anemometer. In addition, the system can also obtain predicted weather information in the future (e.g. within 6 hours) through access to third-party weather forecast API. All data are fused into a unified multi-dimensional state vector .

[0036] S102, based on the multi-dimensional state vector, perform multi-objective decision optimization operation. The state vector is input into a pre-trained non-quantitative multi-objective optimization model. The model is based on Markov Decision Process (MDP) and aims to learn a strategy that can optimize the building thermal performance and plant vitality, the two core objectives.

[0037] S103, generate a set of non-dominated instructions. The core of the model is a multi-objective reinforcement learning algorithm (as described in Example 2). Instead of a single optimal action, the output is a set of non-dominated optimal instructions . Each action (i.e. angle instruction) in the set corresponds to a Pareto optimal solution, i.e. without deteriorating any target (e.g. plant health), it is impossible to make another target (e.g. building energy saving) better.

[0038] S104, based on the preference vector, select the optimal action from the non-dominated instruction set and execute it by the building control device. The system selects an optimal action (e.g. angle instruction) from the set according to a real-time specified preference vector (e.g. in the summer heat, the user may set the preference to , corresponding to energy saving, comfort, plant, i.e. highly prefer energy saving). Finally, the control signal is sent to the driving mechanism (e.g. electric push rod) of the building external hanging (e.g. greenery device) to adjust the device to the corresponding angle.

[0039] ​The skilled person can understand that the present application solves the trade-off problem between building energy saving and plant health of the dynamic greening device by an artificial intelligence model, especially multi-objective reinforcement learning. It combines real-time data and predicted data, making the control decision forward-looking and intelligent, overcoming the limitations of traditional static greening and simple threshold control.

[0040] Embodiment 1 Please refer to Figure 2 , in this embodiment, the interaction and execution logic of the non-quantized multi-objective reinforcement learning model in step S102 and the building physical entity are described in detail. The present application converts the environmental physical quantity into a vectorized reward signal with clear physical meaning by deploying a multi-dimensional physical sensor array in the building facade, indoor environment and plant growth medium, so as to directly drive the mechanical actuator to produce accurate angular displacement.

[0041] 1. Electrical signal conversion of building thermal state and mechanical loss protection reward

[0042] This reward is directly coupled to the heat flux sensor array deployed on the wall envelope and window seat. The system converts the micro-millivolt voltage signal of the heat flux sensor into instantaneous heat flux density by real-time acquisition according to the preset electrical-physical sensitivity coefficient. In order to balance the building energy saving and the device self-consumption energy, the present embodiment carries out fine physical modeling on the energy reward item.

[0043] On the physical energy efficiency trade-off mechanism, the system monitors the instantaneous power consumption of the driving motor when performing the rotating action in real time through the current / voltage transformer and obtains the rated power consumption of the motor . Hardware fatigue and energy consumption balance formula:

[0044] Among them and are the weight coefficients of balancing building energy saving and device self-consumption. is a normalization factor, taking the value of the historical maximum heat flux observation value, which is used to map the heat flux data to the interval, preventing the difference in numerical magnitude from affecting the stability of gradient descent. The physical meaning of the formula is to directly link the physical insulation requirements of the building envelope with the electrical loss of the mechanical actuator.

[0045] The industrial edge computing unit uses the formula to identify small heat flux fluctuations. If the predicted energy saving brought by the action is less than the power consumption of the driving motor, the system will suppress unnecessary mechanical instructions, thereby avoiding the overheating risk of the motor due to frequent start-stop and significantly reducing the physical wear of the mechanical transmission bearing.

[0046] 2. Indoor micro-environment physical feedback and human thermal comfort smoothing mechanism

[0047] The system uses temperature and humidity sensors and average radiation thermometers deployed on the indoor work surface to collect real-time thermal environment signals, and combines the preset metabolic rate and clothing thermal resistance parameters of indoor personnel to calculate the real-time thermal comfort index in real time by the edge computing unit . In order to eliminate the step-by-step jitter of the mechanical driving mechanism during angle adjustment, a Gaussian kernel feedback adjustment mechanism is used, and a Gaussian kernel function is used to generate a control gradient reward:

[0048] In the formula, is the real-time thermal comfort index calculated based on indoor temperature , relative humidity RH , average radiation temperature , and set clothing thermal resistance and human metabolic rate; is the target comfort value, usually set to 0 (i.e. thermal neutral state); is the bandwidth parameter of the Gaussian function, used to control the tolerance of comfort deviation, the larger, the more gentle the reward function decreases, indicating a higher tolerance for small deviations; the smaller, the more precise the control required; the value range of this function is When is exactly equal to , the reward is 0 (maximum value); as the deviation increases, the reward value decreases exponentially to -1, thereby imposing significant negative feedback punishment on the harsh thermal environment.

[0049] This function controls the tolerance of comfort deviation through the parameter , providing a continuous and differentiable gradient signal for the driving unit to guide the electric push rod actuator to perform "micron-level" angle compensation fine-tuning, rather than dramatic and discontinuous start-stop actions.

[0050] This formula will effectively avoid the instantaneous stress impact of mechanical linkage rods through physical gradient descent adjustment, significantly extending the service life of precision mechanical transmission systems such as lead screws and push rod bearings, and ensuring that indoor personnel will not experience dramatic fluctuations in body temperature due to sudden changes in sunshade angles.

[0051] 3. Hardware-aware and bio-mechanical synergistic reward for plant physiology

[0052] This is the core step for the system to achieve the synergy of bio-energy balance and mechanical motion control. The system monitors the physical light quantum flux on the plant leaf surface using a photosynthetically active radiation (PAR) sensor and perceives the dielectric constant change of the growth medium through a frequency domain reflectometry (FDR) soil moisture sensor.

[0053] First, define the photosynthesis efficiency factor . Use a variant of the Michaelis-Menten equation to describe the nonlinear relationship between photosynthetically active radiation (PAR) and plant growth rate:

[0054] In the formula, is the real-time photosynthetically active radiation value (PAR) collected by the sensor; is the maximum photosynthetic rate of the plant; is the light compensation point, when the light is lower than this value, the plant's respiration consumption is greater than the photosynthesis output, turning to negative, the coefficient represents the penalty for health decline due to insufficient light; is the light saturation point, when the light exceeds this value, the plant experiences photoinhibition phenomenon, starts to decline, and the coefficient represents the penalty for light burn; is the half-saturation constant. Then, the system introduces a soil moisture constraint factor

[0055] , which directly maps the water absorption pressure state of the vegetation growth medium by reading the level or frequency signal returned by the moisture sensor. The soil moisture constraint factor is:

[0056] In the formula, is the soil volumetric water content, is the critical water threshold, is the slope parameter of the Sigmoid function. This factor is a coefficient, indicating that when the water is sufficient, the plant can fully utilize the light; when the water is scarce ( ), the stomata close, and even if the light is sufficient, the photosynthetic efficiency will drop significantly.

[0057] The final biological limit alignment mechanism is realized through a segmented compound reward function: ​

[0058] When the PAR sensor monitors that the intensity of ultraviolet light or light flux exceeds the plant tolerance threshold (light inhibition zone), the reward function value will have a nonlinear sharp drop. This biological signal will be immediately converted into mechanical driving instructions, forcing the driving mechanism to rotate the device angle, using the physical shadow generated by the building external module itself to cover the vegetation surface, achieving direct physical negative feedback closed-loop control from the vegetation biological survival state to the mechanical actuator motion.

[0059] 4. Physical mapping and signal transformation logic of driving execution layer When the multi-objective optimization model generates a non-dominated instruction set in step S103, the final physical instruction conversion is performed by the preference arbitration module. According to the geometric topological structure of the four-bar linkage or linear driver of the building external greening device, the Pareto optimal angle instruction is converted into the physical displacement stroke of the electric push rod .

[0060] The driving control module sends a pulse width modulation (PWM) signal with a specific duty cycle to the DC motor or servo unit to drive the mechanical arm to produce angular displacement. At the same time, the position feedback signal is obtained through the integrated Hall sensor or potentiometer and connected to the PID controller to ensure the absolute positioning accuracy of the mechanical action, so as to adjust the device to the physical optimal orientation that can synergistically optimize energy saving, comfort and biological health.

[0061] Embodiment 2 Please refer to Figure 3 In the embodiment, the non-scalar multi-objective optimization model described in step S102 is the core of the deep mapping of the virtual reinforcement learning algorithm and the complex building physical environment. Unlike traditional single-target control systems that simplify all performance indicators such as energy consumption, temperature, plant growth, etc. into a single scalar by weighted summation, the agent (Agent) of the system maintains a high-dimensional vectorized Q-value table, aiming to approximate the true Pareto optimal solution set in the physical action space through the non-dominated sorting mechanism.

[0062] The following are the specific implementation details of the combination of the algorithm and the physical entity: S201, Construction of vectorized action-value function and physical state space First, the system needs to establish a data structure that can accommodate multiple physical field feedback. Define the state space and the action space , where the action space directly corresponds to the set of discrete physical angles that the greening device can adjust , The agent establishes a multi-dimensional lookup table or a deep neural network approximator to represent the vectorized action-value function For each physical state and each mechanical action , its value is defined as a 3-dimensional vector (i.e. ):

[0063] where maps the building heat flux to the motor loss, maps the indoor thermal environment index, and maps the physiological survival rate of the vegetation. At initialization, the value of all state-action pairs is set to the zero vector, waiting for the feedback incentive from the physical sensors.

[0064] S202, Exploration and Physical Execution Strategy Based on Dynamic Preference At each step of the training phase , in order to enable the model to understand the trade-off relationship under different weather and user needs, the system adopts the strategy based on dynamic weight sampling.

[0065] In the random exploration phase, an action is randomly selected with a probability , so that the mechanical actuator swings in the full angle range to collect sensor data under extreme working conditions.

[0066] Greedy exploitation and preference mapping: with a probability , the execution strategy is performed. In order to traverse different physical regions of the Pareto frontier, the agent randomly samples a preference weight vector at each step, and satisfies . By calculating the scalar projection value

[0067] The agent can simulate decisions under different scenarios, such as tilting the preference vector towards in hot summer, and tilting the preference vector towards in dry season, ensuring that the model covers the complete Pareto frontier physical surface in training.

[0068] S203, Real-time Interaction with Environment and Acquisition of Multi-field Heterogeneous Rewards When the agent drives the green device to perform an action , the building physical environment transfers from state to , and feeds back a vectorized immediate reward generated by the physical entity. The increment of wall heat conduction perceived by the heat flux sensor and the motor power consumption measured by the current sensor jointly constitute the negative reward item. Based on the calculation of indoor temperature and humidity The bias is transformed into a smoothed gradient penalty signal using a Gaussian kernel function. The sensors and soil moisture sensors provide feedback, reflecting whether the current angle adjustment has caused the plant to be in a state of photoinhibition or water deficit.

[0069] S204. Dynamic Maintenance and Pareto Pruning of Non-Dominated Solution Sets Before updating the Q value, the system needs to determine the next physical state. The potential value of the next state. Due to the conflict of objectives, the optimal value of the next state is no longer a single scalar, but a non-dominated solution set.

[0070] Define vector dominance relations For two utility vectors and If satisfied and Then it is called Dominate The system iterates through the next state. Lower Action Space Construct a candidate set using the Q-vectors corresponding to all actions. And select rigorous Pareto frontiers :

[0071] This set eliminates all dominated inferior solutions and represents the theoretically optimal expected return boundary that different operating modes (e.g., extreme energy saving mode versus plant care mode) can achieve in the future.

[0072] S205, Iteration and Physical Parameter Correction of Vectorized Bellman Equations Based on the obtained Pareto frontier The system's current state-action pair Iterative updates are performed. To address the curse of dimensionality caused by multiple objectives, an update mechanism based on random sampling is adopted.

[0073] from The weights used in the current step are determined by the current weights. Select the best matching target vector The Bellman equation in vector form is corrected. value:

[0074] In the formula, For learning rate, This is the discount factor. This step, through continuous iteration, gradually converges the Q-vector table and accurately describes the complex nonlinear relationship between the actions of greening devices and building energy consumption and environmental quality.

[0075] S206, Physical generation and storage of non-dominated instruction set After about steps of simulation or field training convergence, the model will maintain a non-dominated action instruction set for each physical state . All possible mechanical angle actions are traversed , if the vector generated by the action is not dominated by any other angle action vector, it will be stored in the instruction set .

[0076]

[0077] The set contains a variety of "optimal balance points", such as "extreme shading angle", "optimal photosynthesis angle" and "overall comfort angle". This provides a solid decision-making basis for subsequent steps to make millisecond-level arbitration choices according to the user's real-time preferences (such as energy saving priority or plant beauty priority set through the mobile phone App).

[0078] Through the above detailed steps of S201 to S206, the embodiment realizes an intelligent control logic that can handle high-dimensional target conflicts and does not rely on single scalar feedback, ensuring the multi-target collaborative optimization capability of the dynamic greening device in complex environments.

[0079] Embodiment 3 Please refer to Figure 4 , this embodiment describes in detail how the system builds a "fuse" mechanism independent of the multi-target intelligent decision module through hardware interrupt logic, sensor fusion filtering, and aerodynamic physical modeling. This mode ensures the structural safety of the greening device under extreme weather (such as typhoon, strong convective gust), preventing the risk of mechanical structure fatigue fracture, plastic deformation or falling off from the building exterior wall. The specific implementation process includes the following key steps: S301, Real-time fusion and filtering of multi-source wind speed data

[0080] The system monitors the wind speed component in the environment state in real time . In order to prevent control jitter caused by instantaneous gust or sensor noise, the system does not directly use the original reading of the ultrasonic anemometer , but uses a sliding time window filter to smooth the real-time wind speed. The time window length is defined as (e.g. seconds), and the smoothed real-time wind speed is calculated as follows:

[0081] In the formula,​ The time decay weighting coefficient satisfies This is used to assign higher weight to data at the current moment. Simultaneously, the system obtains future data via API. Forecast gust extremes within a time window (e.g., the next hour) , as a forward-looking early warning indicator.

[0082] S302. Dynamic threshold determination based on aerodynamic model.

[0083] Safety threshold The setting is not a fixed value, but is calculated based on the structural strength limit and aerodynamic characteristics of the greening device. The aerodynamic drag experienced by the device in the wind field... It is directly proportional to the frontal area and the square of the wind speed, and its physical model can be expressed as:

[0084] In the formula, air density, For angle The varying drag coefficient, This represents the total surface area of ​​the device. The system is pre-set with a maximum critical load that the structure can withstand. Therefore, the logic for determining the dynamic wind speed threshold that triggers the safety mode is as follows:

[0085] in, For the theoretical limit wind speed, To predict the safety factor (e.g., take 0.8).

[0086] S303, high-priority interrupt and minimum flow resistance reset.

[0087] once The control system immediately activates a safety interrupt. At this time, the system will forcibly suspend the intelligent decision output based on reinforcement learning in step S102, ignore the current energy saving or plant health goals, and directly send the highest priority reset command to the underlying driver.

[0088] Target security perspective This is defined as the state in which the frontal area of ​​the device is minimized, i.e.:

[0089] Under normal circumstances, Corresponding to (Complete Level) or (Completely closed and flush against the wall), depending on the mechanical hinge design of the device, to ensure that airflow can pass through or glide over the device surface with minimal resistance, preventing plastic deformation or detachment of the mechanical structure.

[0090] S304, recovery mechanism based on hysteresis comparator.

[0091] To avoid the device frequently switching between "safe mode" and "smart mode" near the critical wind speed (i.e. "ping-pong effect"), this embodiment introduces a hysteresis comparator logic with time delay. The system must meet the following two conditions to exit the safe mode: 1. Real-time wind speed decreases to the recovery threshold Below, and For example ; 2. Low wind speed state duration exceeds the holding time (e.g. 5 minutes).

[0092] The Boolean expression of the recovery logic is:

[0093] Only when is true, the system is unlocked, the multi-objective smart decision module is reactivated, and the optimization control of building energy consumption and plant state is restored. Through the above mechanism, the system has high robustness and self-adaptive survival ability when facing typhoons, strong convection and other extreme weather, and also prolongs the service life of the precision mechanical drive system through accurate physical modeling.

[0094] Embodiment 4 This embodiment details the specific implementation process of offline pre-training of non-scalar multi-objective optimization model based on model-in-loop simulation architecture. Since directly training reinforcement learning model on real building has the risk of long cycle, high trial and error cost and possible physical damage to equipment, this invention builds a high-fidelity virtual simulation environment for policy iteration.

[0095] Specifically includes the following steps: S401, construct a coupled physical model of building thermal and dynamic greenery device.

[0096] In the EnergyPlus building energy consumption simulation engine, a virtual model of the target building is established based on the IDF standard. For the dynamic greenery device, it is equivalent to a shading component with variable optical thermal performance.

[0097] Define the equivalent solar heat gain coefficient and equivalent thermal resistance of the greenery device at angle . Considering the nonlinear superposition of plant leaf transpiration and mechanical shading, the following physical parameter calculation model is constructed:

[0098]

[0099] wherein, is the effective transmittance of the device at an angle ; and are the transmittance extremes of the device when fully closed and fully open, respectively; is the leaf overlap factor; is the overall heat transfer coefficient of the envelope; is the base wall thermal resistance; is the air gap thermal resistance as a function of angle; is the equivalent thermal resistance of the plant layer.

[0100] By the above equations, the mechanical action of the dynamic green device is mapped to the change of the thermal parameters recognizable by EnergyPlus, ensuring that the simulation environment can truly feedback the building thermal state .

[0101] S402, build BCVTB collaborative simulation communication interface.

[0102] Using the building control virtual test bench (BCVTB) as middleware, a two-way data communication between the Python deep learning framework (such as PyTorch or TensorFlow) and the EnergyPlus engine is established. The BSD Socket protocol is used for data packet transmission, and the simulation step size is set. The state exchange vector of the collaborative simulation is defined:

[0103] wherein, the control instruction vector sent by the agent to the environment contains the action instruction (i.e. angle ) at the current time; the state vector fed back by the environment to the agent contains the simulated indoor temperature , wall heat flux , and environmental parameters read through the weather file; is the simulation synchronization flag.

[0104] In order to simulate the sensor noise in the real environment, a Gaussian white noise is injected in the feedback state :

[0105] wherein, is the standard deviation matrix corresponding to the sensor, to enhance the robustness of the model in real deployment.

[0106] S403, multi-objective Q-vector iterative training based on typical meteorological year.

[0107] Load the typical meteorological year (TMY3) data file of the target city as the environmental incentive. The agent explores and utilizes based on the multi-objective learning algorithm. At each simulation step , the agent observes the state , selects an action , and obtains a vectorized reward .

[0108] The Q-table (or neural network parameters) is updated using the vectorized Bellman optimal operator. For the multi-objective case, the Q-value update is no longer scalar convergence, but non-dominated sorting optimization of the Q-vector set. The iterative update formula of the Q-vector is as follows:

[0109] In the formula, is the value vector of the state-action pair; is the learning rate; is the discount factor; is the reward vector including energy saving, comfort, and plant health; is the non-dominated sorting operator, used to select the Pareto front set from all possible Q-vectors of the next state .

[0110] To balance exploration and exploitation, the strategy is adopted, and exponentially decays with the number of training rounds :

[0111] S404, model convergence determination and parameter solidification.

[0112] The training termination condition is set as the hyper-volume index of the Q-vector set tends to be stable. Define the difference degree of the Q-vector set between the th iteration and the th iteration :

[0113] When the difference degree of the Q-vector set between the th iteration and the th iteration is less than , the model training is completed.

[0114] After training, the converged Q-vector table (for discrete state space) or the weight parameters of the deep neural network (For continuous state space) The data is exported and packaged into a binary inference file. This file is burned into the edge computing unit of the actual control system, serving as the decision-making core for real-time control, thereby enabling policy migration from virtual simulation to physical entities.

[0115] Example 5 In this embodiment, a photovoltaic (PV) power generation unit (e.g., flexible thin-film solar cell or semi-transparent photovoltaic glass) is integrated on the mechanical structure surface or the back of the blade-bearing module of the greening device. At this point, the system's control logic upgrades from three-dimensional target coordination to four-dimensional target coordination, aiming to find the optimal dynamic balance between building energy conservation, indoor comfort, plant health, and photovoltaic power generation.

[0116] The specific implementation steps and algorithm model construction are as follows: S501. Construct a physical coupling model for photovoltaic production capacity.

[0117] In the multi-objective intelligent decision-making module, the relationship between photovoltaic output power and device adjustment angle is first established. A nonlinear coupling model between them. Real-time theoretical output power of the photovoltaic module. It depends not only on the intensity of solar radiation, but also on the effect of the device angle on the angle of solar incidence and the correction for the temperature of the solar panel. The calculation formula is as follows:

[0118] In the formula, For at any time The device angle is Predicted photovoltaic power generation at that time; The overall conversion efficiency of the photovoltaic system (including inverter efficiency and line loss). The effective light-receiving area of ​​the integrated photovoltaic module; The total radiation intensity on the horizontal surface is derived from meteorological data. Obtain; This is a function of the incident angle correction coefficient, which depends on the angle of the device. Solar altitude angle and solar azimuth This function describes how changes in the device's angle alter the direct radiation component received by the photovoltaic panel. This is the temperature power coefficient of the photovoltaic cell (usually a negative value). The operating temperature of the solar panel. The reference temperature (25°C) is the temperature under standard test conditions.

[0119] Using the above model, the agent can predict and adjust the device to any candidate angle under the current environmental conditions. The exact power generation benefit that can be obtained.

[0120] S502, extend the four-dimensional vectorized reward function.

[0121] In order to incorporate the capacity target into the optimization framework, the original three-dimensional reward function is extended to a four-dimensional vector. The reward vector at time is defined as follows:

[0122] Among them, the new photovoltaic power generation reward term is not a simple linear mapping, but a nonlinear activation function after normalization, to ensure that its numerical magnitude is on the same order of magnitude as the heat flux ) and the comfort index (PMV), avoiding gradient disappearance or dominant problems caused by numerical differences. The calculation formula of

[0123] In the formula, is the rated peak power of the photovoltaic module; is the reference weight coefficient of the capacity target, used to set the relative importance of the target in the Pareto frontier; is the sensitivity adjustment factor; is the hyperbolic tangent activation function, used to smooth the reward value and limit its upper limit.

[0124] S503, augmentation of state space.

[0125] In order to support accurate prediction of photovoltaic capacity, the multi-field data acquisition module needs to augment the photovoltaic related state sub-vector when constructing the state vector . At this time, the total state vector is defined as:

[0126] Among them, specifically contains the following feature parameters: is the open circuit voltage of the photovoltaic array; is the short circuit current of the photovoltaic array; is the photovoltaic panel backboard temperature collected by the patch sensor; is the daily cumulative power generation.

[0127] S504, multi-objective conflict resolution and decision execution.

[0128] In the training and inference phase, the agent faces more complex Pareto trade-offs. The system needs to handle the following typical conflict scenarios: In the summer noon scenario, the building energy saving goal is to tend to close the device ( ), to maximize the shading, block the solar radiation into the room; the photovoltaic power generation goal is to tend to adjust the device to the angle perpendicular to the sunlight (for example, θ ), to obtain the maximum power generation; the plant health goal is to tend to be moderately opened to avoid strong light scorching while ensuring ventilation and heat dissipation.

[0129] At this time, the non-scalar multi-objective Q-learning model calculates the four-dimensional Q vector corresponding to each angle in the action space . Based on the current preference weight vector , the system calculates the comprehensive utility value :

[0130] Among them, is the action (i.e. the target angle) that can make the comprehensive utility value reach the maximum among all possible action options and output. The system will select as the optimal execution action. For example, if the current grid load is high and the peak electricity price is high, the user sets higher, the system may sacrifice part of the shading effect (allowing a small amount of heat to enter), adjust the angle to a position more conducive to power generation, use photovoltaic power generation to offset the cost of increased air conditioning energy consumption, and thus achieve optimal global economy.

[0131] In the overcast light scenario, the photovoltaic power generation capacity is weakened, and the model will automatically reduce the actual influence in decision-making, and instead pay more attention to the plant's capture efficiency of scattered light (PAR) and the natural lighting demand indoors.

[0132] Through the above expansion, the system is not only a dynamic shading and vertical green system, but also a self-adaptive building integrated photovoltaic (BIPV) energy management system, realizing the comprehensive performance optimization of "energy saving-power generation-ecology-comfort" four in one.

[0133] Embodiment 6 Please refer to Figure 5 , in the second aspect of the present application, a building control system based on multi-source heterogeneous data and Pareto decision. The system is a hardware entity implementation of the above control method, adopts a hierarchical distributed architecture design, mainly includes a multi-field data acquisition and preprocessing module, a multi-objective intelligent decision module, a preference selection and arbitration module, and a driving control and execution module. Each module realizes data interaction and collaborative control through high-speed industrial field bus and low-power wide area network, and the specific implementation manner is as follows: 1. Multi-domain data acquisition and preprocessing module, which serves as the perception layer and is responsible for performing step S101. In terms of hardware configuration, the system integrates a set of heterogeneous sensor arrays, including but not limited to heat flux sensors installed on the building facade (for measuring , photosynthetically active radiation (PAR) sensors (spectral response range 400-700 nm), frequency domain reflectometry (FDR) soil moisture sensors, and high-precision ultrasonic anemometers.

[0134] Considering the outdoor environmental noise interference, the module is built-in with a signal conditioning circuit and an FPGA preprocessing unit. To address the dimensional differences of heterogeneous data, the preprocessing unit adopts the Z-score standardization method for real-time normalization processing of the original signal. Let the original sampling value of a sensor at time be , and its normalized output be , the calculation formula is:

[0135] where and are the sliding window mean and standard deviation of the sensor data, is a small constant to prevent the denominator from being zero.

[0136] In addition, for data transmission, the module integrates a LoRaWAN communication module, which uses spread spectrum modulation technology for long-distance transmission. To ensure the integrity of the data packet, the module uses the cyclic redundancy check (CRC) mechanism, with the check polynomial set to . Through the low-power wide-area network technology, the wiring difficulty and signal attenuation problem of the facade sensors of high-rise buildings are effectively solved.

[0137] 2. Multi-objective intelligent decision-making module, which is the core computing hub of the system and is responsible for performing steps S102 and S103. In terms of hardware, an industrial-grade edge computing gateway is used, which is built-in with a high-performance embedded processor (such as NVIDIA Jetson series or ARM architecture SoC with NPU acceleration).

[0138] The pre-trained non-quantized multi-objective Q-learning model is fixed in the memory of this module. The inference engine does not directly output a single action, but through matrix operation and parallel computing of the Q vector matrix corresponding to all candidate actions in the action space under the current state :

[0139] where each row represents an action in energy saving ), comfort ( ), and plant ( ) three dimensions of expected value. The edge computing unit uses the non-dominated sorting algorithm to perform row vector screening on , and outputs the Pareto optimal instruction set . This process is completed in milliseconds at the local edge end, without the need for continuous dependence on cloud computing power, greatly reducing the impact of network delay on control real-time performance.

[0140] 3. Preference selection and arbitration module, which is responsible for performing the decision arbitration function in step S104. It receives the user preference weight vector from the host computer or mobile terminal App, and combines the safety monitoring logic to perform weighted optimization on the instruction set.

[0141] An utility maximization algorithm is run inside the module to calculate the scalar utility value of each action in the Pareto optimal set:

[0142] The system selects the action with the maximum utility as the preliminary instruction.

[0143] At the same time, the module runs an independent safety monitoring thread in parallel. This thread compares the instantaneous wind speed with the safety threshold in real time. The safety arbitration logic generates the final execution instruction using the following Boolean algebraic expression:

[0144] where is the hard-coded mechanical reset angle (such as 0° or 90°, depending on the pneumatic characteristics of the device), and the safety interrupt signal has the highest system priority and can directly override the intelligent decision result.

[0145] 4. Drive control and execution module, which is responsible for performing the physical action execution in step S104. The control signal is transmitted to the distributed motor controller (MCU) through RS485 or CAN bus.

[0146] Since the mechanical structure of the greening device usually adopts a linkage mechanism for driving, there is a nonlinear geometric relationship between the target angle and the stroke of the linear driver. The module integrates a kinematics solving unit inside, which performs inverse kinematics transformation based on the cosine law:

[0147] where , L is the length of the connecting rod, is the initial phase angle, L0 is the length of the driver zero position.

[0148] The calculated target stroke is sent to the PID controller to adjust the PWM (pulse width modulation) duty cycle of the electric push rod :

[0149] wherein, is the position deviation. The actuator is an industrial-grade electric linear driver with a protection level not lower than IP66, with overload protection and Hall position feedback function, to ensure that the green device can be accurately and stably adjusted to the target angle in outdoor harsh working conditions.

[0150] Through the close cooperation of the above four modules, the system realizes closed-loop control from environmental perception, intelligent decision-making, safety arbitration to precise execution, effectively solving the multi-objective collaborative optimization problem of dynamic green devices in complex environments.

[0151] Example 7 In this embodiment, the multi-domain data acquisition and preprocessing module is also integrated with a data integrity guarantee unit to solve the problem of sensor data loss, drift or noise interference in complex outdoor environments, to improve the overall robustness of the control system. The specific implementation process of this unit includes two core stages of data anomaly detection and missing value interpolation, and the specific technical scheme is as follows: First stage: unsupervised anomaly detection based on variational autoencoder (VAE) In order to accurately identify abnormal data in the sensor network (such as the solidification of soil moisture sensor readings due to corrosion, and the attenuation of light sensor readings due to dust covering), this unit constructs a variational autoencoder (VAE) network to perform reconstruction probability analysis on multi-dimensional heterogeneous sensor data.

[0152] 1. Data input and encoding: define the original sensor measurement vector at time t as wherein is the total number of sensors. The encoder (Encoder) network of VAE maps the input data to the distribution parameters of latent variable space , i.e. mean vector and standard deviation vector . The encoding process is represented as a probability distribution , and its neural network calculation formula is:

[0153]

[0154] where, and are the weight matrix and bias term of the encoder respectively.

[0155] 2. Resampling and decoding: With the reparameterization trick, we sample from the latent space where . Subsequently, the decoder network maps the latent variable back to the data space, generating the reconstructed data :

[0156] 3. Anomaly detection criterion: The training objective of the VAE network is to maximize the lower bound of the evidence, whose loss function is defined as the sum of reconstruction error and KL divergence:

[0157] In the detection phase, we calculate the reconstruction probability score of the current time data. If the reconstruction residual of a certain sensor channel exceeds the dynamic threshold based on historical statistical characteristics (e.g. ), it is determined that the sensor data has an anomaly or failure, and is marked as a failure state.

[0158] Second stage: Dynamic interpolation based on spatiotemporal correlation LSTM When data anomalies or missing are detected, the system does not directly discard the samples, but starts a spatiotemporal correlation prediction model based on long short-term memory network (LSTM) for interpolation, to ensure the integrity of the state vector input to the decision module.

[0159] 1. Spatiotemporal feature construction: Construct the input matrix of the interpolation model, which not only contains the time sliding window historical data of the sensor, but also integrates other sensor data with strong spatial correlation (such as using air humidity meter and rainfall data to assist in correcting soil humidity meter) and environmental prediction data.

[0160] 2. LSTM unit operation: Use LSTM network to capture long-distance dependence relationship of time series. For each time step, LSTM unit controls information flow through forget gate , input gate and output gate . The specific state update formula is as follows:

[0161]

[0162]

[0163]

[0164]

[0165]

[0166] wherein, is the cell state, is the hidden layer output, is the Sigmoid activation function, denotes the Hadamard product.

[0167] 3. Imputation value generation and replacement: the final output of the LSTM is mapped through a fully connected layer to generate the estimated value of the failed sensor at the current time . The system utilizes the estimated value . Instead of the original abnormal value .

[0168] Through the above mechanism, for example, when it is detected that a certain soil moisture meter reading is constant for a long time (variance is 0, VAE reconstruction error is high), and the weather data shows that it has just experienced rainfall and the air humidity has increased dramatically, the data integrity guarantee unit will determine that the sensor is faulty, and use the LSTM model to calculate a reasonable soil moisture imputation value according to the rainfall, air humidity and data of the adjacent normal working soil sensors. Finally, the standardized state vector after cleaning and repairing is transmitted to the multi-objective intelligent decision module, ensuring that the control system can still make a Pareto optimal decision under the condition of partial hardware failure, avoiding plant death or energy waste caused by data errors.

[0169] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A building control method based on multi-source heterogeneous data and Pareto decision-making, characterized in that, include: Multi-domain heterogeneous state data is acquired, and cross-validation and anomaly detection are performed on the data using an unsupervised machine learning algorithm based on variational autoencoders to automatically identify and calibrate sensors experiencing data drift or failure. Missing data is then imputed using a time-series prediction model based on long short-term memory networks. Finally, the processed multi-domain heterogeneous state data is fused into a multi-dimensional state vector. The multidimensional state vector At least including: Building status data used to characterize the thermal state of a building. ; Biological state data used to characterize the vitality of internal vegetation on building exteriors. ; Real-time and predicted environmental data used to characterize external environmental conditions ; The multidimensional state vector The input is fed into a pre-trained non-standardized quantized multi-objective optimization model, which is constructed based on a Markov decision process, wherein the tuples of the Markov decision process are... ,in For state space, For the action space, Let be the state transition probability. To vectorize the reward function, Discount factor; The multi-objective optimization model outputs the non-dominated optimal instruction set. Each of the actions And every action This corresponds to a Pareto optimal solution; Based on preference vector From the optimal instruction set Select the optimal action Control signals are sent to one or more drive mechanisms of the building exterior to adjust the building exterior to the optimal angle orientation.

2. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 1, characterized in that, The acquisition of heterogeneous state data across multiple domains includes: The building status data is acquired by heat flux sensors deployed on the building envelope and temperature and humidity sensors deployed indoors. ; The biological state data is acquired by using soil moisture sensors, soil temperature sensors, conductivity sensors, and photosynthetically active radiation sensors deployed in the growth medium of the internal vegetation within the building's exterior. ; Real-time wind speed and direction data are acquired using an ultrasonic anemometer, and data containing at least future weather forecasts is obtained from an API that connects to a third-party weather forecasting application. Hourly forecast weather information together constitutes the environmental data. .

3. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 1, characterized in that, The non-standardized multi-objective optimization model constructs a decision space that interacts with the physical environment in the following way: Configure the state space To map the , and Normalized high-dimensional continuous vectors composed of physical signals are used to characterize the real-time physical conditions of buildings and installations. Configure the action space The set of physically discrete angles that the drive mechanism of the building's external structure can execute. Each element corresponds to a specific mechanical displacement command; Construct the vectorized reward function To quantitatively assess the feedback impact of control actions on the physical environment, in Generate one at a time 3D reward vector : in, This is an energy reward item calculated based on heat flux and electrical energy consumption. This is a comfort bonus calculated based on thermal environment parameters. This is a vegetation health incentive program based on physiological constraints of light and water.

4. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 3, characterized in that, The energy reward items The formula is: in, Indicates the time measured by the heat flux sensor array Instantaneous heat flux density through walls or windows This indicates the suppression of ineffective heat transfer in any direction; The normalization factor is the historical maximum observed heat flux. Indicates the execution of an action The mechanical and electrical energy consumed; Energy consumption corresponding to the rated power of the drive motor; and It is a dimensionless weighting coefficient used to balance the relationship between building energy saving and device self-consumption energy; in, It is a real-time thermal comfort index calculated based on indoor temperature, relative humidity, average radiant temperature, and set clothing thermal resistance and human metabolic rate. The target comfort value; The bandwidth parameter of the Gaussian function is used to control the tolerance for comfort deviations. The larger the value, the more smoothly the reward function decreases, indicating a higher tolerance for small deviations; The smaller the value, the more precise the control required; the range of values ​​for this function is... ; in, To introduce soil moisture constraint factors, It is the photosynthetic efficiency factor.

5. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 1 or 3, characterized in that, The multi-objective optimization model is configured to execute a multi-objective Q-learning algorithm through a processor to establish a mapping decision logic between the physical state of the building and the control actions; The multi-objective Q-learning algorithm constructs a Q-vector representing the value of the control strategy based on the collected heterogeneous state data from multiple domains. : Wherein, the Q vector Characterized in the current physical state of the building Execute mechanical drive action Subsequently, the cumulative discount rewards that can be obtained in the dimensions of building energy efficiency, comfort, and plant health; As a discount factor, Vectorized reward function; Utilize real-time sensor feedback as an immediate reward Based on the non-dominant rule, the above The vector is updated, and the processor's update logic follows the following Bellman equation iterative formula: in, The updated control strategy value assessment value. For learning rate, As a discount factor, In the state Execute action The immediate reward received afterward The next state The set consisting of all the Q vectors mentioned above. It is a vector selection function based on the Pareto front, used to select from the potential physical states at the next time step. Filter out non-dominated control paths; The Pareto pruning algorithm is used in each state. Maintain a non-dominated state A set of vectors Based on this set, an instruction set for driving building attachments is generated to replace the traditional scalar weighted aggregation.

6. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 1, characterized in that, The building control method also includes a safety protection mode configured to respond to extreme weather conditions, which includes: The system monitors the physical sampling values ​​of the wind speed sensor and the predicted wind speed values ​​from external meteorological data in real time, and compares the values ​​with the preset structural safety threshold. When the detected value exceeds the structural safety threshold, a high-priority hardware control signal is generated to physically bypass or block the instruction output channel of the multi-objective optimization model. A forced reset signal is sent directly to the drive mechanism of the building's external structure, driving the mechanical structure of the external structure to rotate to a preset safe angle. This reduces the device's frontal area and air resistance to a physically safe level.

7. The building control method based on multi-source heterogeneous data and Pareto decision-making according to claim 3, characterized in that, The training method for the pre-trained non-standardized multi-objective optimization model is a strategy configuration process for building control devices, executed by a computing platform, and includes: Based on the actual physical and thermal parameters and geometric dimensions of the building's external structure and building envelope, a digital simulation environment that maps the physical entity is constructed. A data communication interface is established between the agent of the non-standardized multi-objective optimization model and the digital simulation environment to simulate the thermodynamic response and mechanical motion state of physical entities under different environmental stimuli. The digital simulation environment is driven by inputting historical meteorological and physical data, and the vectorized reward function is used. Gradient updates and policy optimizations are performed on the decision network of the intelligent agent; When the change in the Q vector set is less than a preset threshold, the training is determined to be converged, the model weight parameters are exported and fixed into the embedded operating unit of the building control device to complete the initial configuration of the building control device.

8. A building control system based on multi-source heterogeneous data and Pareto decision-making, characterized in that, The building control system includes: A multi-domain data acquisition and preprocessing module is configured to acquire and fuse heterogeneous state data from multiple domains. It performs cross-validation and anomaly detection on the heterogeneous state data using an unsupervised machine learning algorithm based on a variational autoencoder to automatically identify and calibrate sensors experiencing data drift or malfunction. It also imputes missing data using a time-series prediction model based on a long short-term memory network, and fuses the processed heterogeneous state data into a multi-dimensional state vector. The multidimensional state vector It should include at least building status data, biological status data, and real-time environmental data; The multi-objective intelligent decision-making module is configured to be based on the multi-dimensional state vector. The system controls a pre-trained, non-standardized multi-objective optimization model to perform multi-objective decision optimization operations, generating a non-dominated optimal instruction set. ; The preference selection and arbitration module is configured to be used based on preference vectors. From the optimal instruction set Select the optimal action ; The drive control and execution module is configured to perform the optimal action. Control signals are sent to one or more drive mechanisms of the building scaffold to adjust the building scaffold to an optimal angle orientation.

9. The building control system based on multi-source heterogeneous data and Pareto decision-making according to claim 8, characterized in that, The multi-domain data acquisition and preprocessing module includes a heat flux sensor, a photosynthetically active radiation sensor, a soil moisture sensor, an ultrasonic anemometer, and an indoor environment sensor; data from each sensor are wirelessly transmitted to the multi-target intelligent decision-making module via a long-distance wide area network protocol or an NB-IoT protocol. The multi-objective intelligent decision-making module includes an industrial-grade edge computing unit or a single-board computer, on which the non-standard quantitative multi-objective optimization model is embedded and runs.

10. The building control system based on multi-source heterogeneous data and Pareto decision-making according to claim 9, characterized in that, The edge computing unit or single-board computer in the multi-objective intelligent decision-making module is configured to execute a multi-objective Q-learning algorithm. The execution logic of the algorithm includes establishing a Q-vector data structure in memory. : Wherein, the Q vector Characterized in the current physical state of the building Execute mechanical drive action Subsequently, the cumulative discount rewards that can be obtained in the dimensions of building energy efficiency, comfort, and plant health; As a discount factor, Vectorized reward function; The edge computing unit is configured to invoke an update and action selection procedure to update the Q-vector based on non-dominated rules, where the Bellman equation iterative formula is: in, The updated control strategy value assessment value. For learning rate, As a discount factor, In the state Execute action The immediate reward received afterward The next state The set consisting of all the Q vectors mentioned above. It is a vector selection function based on the Pareto front, used to select from the potential physical states at the next time step. Filter out non-dominated control paths; The edge computing unit is also configured to run a Pareto pruning algorithm program in each state. Maintain a set of non-dominated Q-vectors. This replaces the traditional scalar-weighted aggregation.

Images