A photovoltaic curtain wall intelligent control method based on a digital twinborn algorithm

Abstract

The application discloses a photovoltaic curtain wall intelligent control method based on a digital twinborn algorithm and relates to the technical field of building photovoltaic integration.The application comprises the following steps: S1, multi-source data acquisition and preprocessing;S2, self-evolution multi-level digital twinborn body construction and continuous updating;S3, local state deduction based on a virtual sensor network;S4, multi-agent deep reinforcement learning distributed optimization decision;S5, scene self-adaptive control strategy migration based on meta-learning;S6, interpretable decision output and safety verification;S7, control instruction execution and feedback closed loop.Through a multi-digital twinborn body cooperative operation mechanism and an online incremental learning strategy, the digital twinborn body can continuously adapt to photovoltaic component performance attenuation and environmental changes, effectively overcomes the model drift problem and maintains high-precision prediction capability in long-term operation.

Description

Technical Field

[0001] This invention belongs to the field of building-integrated photovoltaics (BIPV) technology, specifically a smart control method for photovoltaic curtain walls based on digital twin algorithms. Background Technology

[0002] With the acceleration of global urbanization, the proportion of building energy consumption in global total energy consumption continues to rise, making energy conservation and carbon reduction in the construction industry a key issue in addressing climate change. Building Integrated Photovoltaics (BIPV) technology, which directly integrates photovoltaic systems into building curtain walls, not only efficiently utilizes solar energy resources but also reduces buildings' dependence on external energy sources, becoming an important development direction in the field of green building. As a typical form of BIPV, photovoltaic curtain walls integrate photovoltaic power generation with the functions of the building envelope, combining multiple functions such as power generation, shading, heat insulation, and light regulation.

[0003] However, existing intelligent control technologies for photovoltaic curtain walls still have the following prominent problems: 1. In existing technologies, digital twins are mostly built based on initial modeling data, lacking continuous learning and dynamic update mechanisms. They are difficult to adapt to the dynamic changes in photovoltaic modules such as performance degradation, local failures and aging that occur as the service time increases, resulting in a continuous decline in model prediction accuracy over time and control decisions deviating from the actual optimal.

[0004] 2. The operation of photovoltaic curtain walls requires balancing multiple objectives, including power generation efficiency, building energy consumption, indoor lighting comfort, and visual comfort. Significant conflicts often exist between these objectives. Existing control methods typically employ static weights or simple rule-based switching strategies, which struggle to achieve dynamic balance and adaptive adjustment of multiple objectives in complex and ever-changing operating environments. Furthermore, geometric optimization and operational control are often studied separately, limiting the improvement of overall performance.

[0005] 3. Existing technologies typically perform sensing and control on a per-wall or regional basis, neglecting the significant differences in power generation characteristics caused by factors such as localized shading, temperature variations, and individual module performance deviations at different locations within the same curtain wall. While some existing solutions propose using independent sensors to collect shading status for each photovoltaic window, this approach is costly in terms of hardware and presents significant maintenance challenges. Achieving refined sensing and control of localized curtain wall conditions under limited sensor deployment is a pressing technical challenge that needs to be addressed.

[0006] 4. Different buildings have different functions, occupancy patterns, and spatial layouts, resulting in significant differences in curtain wall control requirements. For example, office buildings need to balance lighting comfort and power generation efficiency during working hours, while residential buildings prioritize visual comfort for residents and energy conservation. Existing control methods generally lack the ability to adapt to the specific usage scenarios of buildings and cannot dynamically adjust and optimize target priorities based on actual usage needs.

[0007] 5. Existing prediction models based on end-to-end neural networks are often considered "black boxes," lacking interpretability and making it difficult to verify the rationality and safety of control decisions. In scenarios involving building electrical safety and grid interaction, verification of the safety boundaries of control commands is crucial, but existing technologies do not adequately address this aspect.

[0008] The information disclosed above in this background section is only for enhancing the understanding of the background technology of this invention, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a smart control method for photovoltaic curtain walls based on digital twin algorithms. By constructing a multi-level digital twin with self-evolution capabilities, introducing a distributed optimization decision framework based on multi-agent deep reinforcement learning, designing a scene-adaptive control strategy based on meta-learning, and establishing an interpretable control mechanism based on physical information encoding, this invention achieves high-precision, adaptive, refined, and safe and reliable smart control of the photovoltaic curtain wall system.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a smart control method for photovoltaic curtain walls based on digital twin algorithms, comprising the following steps: S1: Collect multi-source sensing data of the photovoltaic curtain wall system, including curtain wall-level environmental parameters, string-level electrical parameters, component-level local state parameters, and building operation data; S2: Construct a self-evolving multi-level digital twin, which includes a photovoltaic array twin model, a thermal-optical coupling twin model, and a component health status twin model, and achieves continuous self-evolution and updating through a multi-digital twin collaborative operation mechanism and an online incremental learning strategy; S3: Based on the deployed real sensor data and the digital twin, construct a virtual sensor network to deduce the local state parameters of all components of the photovoltaic curtain wall; S4: Employ a multi-agent deep reinforcement learning framework, using the output of the digital twin as the state input, to generate a multi-objective collaborative optimization control strategy. The multi-agent includes multiple control region agents and a global coordination agent. S5: Utilize a meta-learning-based policy transfer mechanism to quickly adapt pre-trained meta-policies to the target building usage scenario and generate personalized control policies. S6: Perform interpretability tracing and multi-dimensional security boundary verification on the generated control strategy, and issue control commands that have passed security verification to physical actuators; S7: After the actuator completes its action, it collects the system response data and feeds it back to the digital twin to form a closed-loop control.

[0011] Preferably, in S2, the photovoltaic array twin model adopts a hybrid prediction model with physical information encoding, including a data-driven external parameter prediction sub-model and a physical-driven power calculation sub-model; the external parameter prediction sub-model performs time-series prediction of solar irradiance and component temperature based on a long short-term memory network, and the power calculation sub-model calculates the output power of the photovoltaic module based on a single diode equivalent circuit model.

[0012] Preferably, in S2, the twin model of the component health status is constructed based on Bayesian inference and Kalman filtering. By performing time-series analysis on the residual between the actual output power and the theoretical output power of the photovoltaic module, the equivalent series resistance increment, efficiency decay factor and local shading state parameters of each component are estimated online.

[0013] Preferably, in S2, the self-evolutionary update mechanism specifically involves: simultaneously maintaining multiple twin model instances with different initialization parameters, and fusing the prediction results of each model through a weighted integration method; when the deviation between the actual observed data and the model prediction value exceeds a preset threshold, an online incremental learning strategy is adopted to adjust only a portion of the network weights to update the model.

[0014] Preferably, in step S3, the method for constructing the virtual sensor network includes: training a deep neural network model using limited real sensor data that has been deployed; pre-training the model using synthetic data generated in the simulation environment through a transfer learning mechanism, and then fine-tuning the model using real data; and adding a physical consistency constraint term based on Kirchhoff's current law and the law of conservation of energy to the model loss function.

[0015] Preferably, in step S4, the multi-agent deep reinforcement learning framework models the photovoltaic curtain wall control problem as a distributed partially observable Markov decision process. The observation states of the control area agents include the output of the photovoltaic array twin model, the output of the thermal-optical coupling twin model, weather forecast data, and building usage scenario parameters. The actions of the control area agents include photovoltaic module tilt adjustment, azimuth angle fine-tuning, and bypass switch status control. The actions of the global coordination agent include multi-objective optimization weight vector allocation for each control area.

[0016] Preferably, the multi-agent deep reinforcement learning framework employs a multi-agent proximal policy optimization algorithm and is trained in a "centralized training-distributed execution" manner: during the training phase, each agent can access global state information to learn policies, and during the execution phase, each agent in the control area makes independent decisions based solely on local observation information; during the training process, the simulated running data generated by the digital twin is used as the training environment.

[0017] Preferably, in step S5, the meta-learning-based policy transfer mechanism includes: constructing a scene encoder to encode building function type, spatial function layout, typical personnel usage patterns, and climate region characteristics into low-dimensional scene embedding vectors; employing a model-independent meta-learning algorithm to perform meta-training on simulation task distributions covering multiple building types and climate regions to obtain meta-policy parameters; and when deployed to the target building, performing a small number of gradient updates on the meta-policy network using the scene embedding vectors as context conditions to quickly generate personalized control policies adapted to the current scene.

[0018] Preferably, in step S6, the multi-dimensional safety boundary verification includes: voltage / current safety boundary verification, hot spot risk assessment, building electrical safety verification, and mechanical safety verification; if any safety verification fails, the current control command is rejected and a safety rollback strategy is triggered.

[0019] A smart control system for photovoltaic curtain walls based on a digital twin algorithm, used to implement the above method, includes: A multi-level sensing network module is used to collect multi-source sensing data and building operation data from the photovoltaic curtain wall system; The digital twin engine module is used to build and run the self-evolving multi-level digital twins, including photovoltaic array twin models, thermal-optical coupling twin models, and component health status twin models, as well as to realize the collaborative operation and online incremental learning of multiple digital twins; The virtual sensor network module is used to extrapolate the local state parameters of all components of the curtain wall based on limited real sensor data. A multi-agent reinforcement learning decision-making module is used to generate multi-objective collaborative optimization control strategies; The meta-learning strategy adaptation module is used to quickly adapt pre-trained meta-policies to the target building usage scenario; The interpretability and security verification module is used to perform physical source interpretation of control decisions and multi-dimensional security boundary verification. The actuator control module is used to send control commands that have passed safety verification to the photovoltaic bracket angle adjustment motor and bypass switch controller, and to receive execution feedback.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention enables digital twins to continuously adapt to photovoltaic module performance degradation and environmental changes through a multi-digital twin collaborative operation mechanism and an online incremental learning strategy, effectively overcoming model drift problems and maintaining high-precision prediction capabilities during long-term operation.

[0021] 2. This invention uses a multi-agent deep reinforcement learning framework to achieve dynamic adaptive weight allocation for four objectives: power generation efficiency, building thermal comfort, indoor lighting comfort, and visual comfort, thus solving the problem of poor adaptability of traditional static weight methods in changing operating environments.

[0022] 3. This invention utilizes virtual sensor network technology, requiring only a small number of real sensors to achieve high-precision simulation of the local state of all photovoltaic modules in the entire curtain wall system, significantly reducing hardware deployment costs and maintenance difficulty, while also significantly improving the ability for refined control.

[0023] 4. Through the meta-learning strategy transfer mechanism, the system can quickly adapt to building scenarios of different functional types and climate regions in a short period of time without having to retrain from scratch, which greatly shortens the deployment cycle and improves the scenario adaptability of the control strategy.

[0024] 5. This invention makes the control decision-making process transparent and traceable through a decision tracing mechanism based on physical information encoding and a multi-dimensional safety boundary verification mechanism, effectively avoiding electrical safety, hot spot risks and mechanical safety risks, and ensuring the long-term safe and reliable operation of the system.

[0025] 6. Through the synergistic effect of multi-level intelligent control strategies, this invention is expected to improve the overall power generation efficiency of the photovoltaic curtain wall system by 15% to 25% and reduce the overall energy consumption of the building by 10% to 20% while ensuring the comfort of the indoor environment, thus significantly improving the technical and economic performance of the photovoltaic curtain wall system. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0027] Figure 1 This is a flowchart of the control method of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1:

[0029] A smart control method for photovoltaic curtain walls based on digital twin algorithms includes the following steps: S1. Multi-source data acquisition and preprocessing Deploying a multi-level sensing network in a photovoltaic curtain wall system includes: Curtain wall level sensors: solar irradiance sensors, ambient temperature sensors, wind speed sensors, and indoor illuminance sensors installed on the facades of the curtain wall facing various directions; String-level sensors: current sensors, voltage sensors, and temperature sensors installed at the output terminals of each photovoltaic string; Component-level sensors: Miniature irradiance sensors and component back temperature sensors are installed at key nodes of the curtain wall in a sparse deployment manner.

[0030] Simultaneously, building operation data is collected, including information on the location and number of people indoors, energy consumption data of the air conditioning system, energy consumption data of the lighting system, historical weather data, and weather forecast data.

[0031] Data cleaning, timestamp alignment, outlier removal, and normalization preprocessing are performed on the collected multi-source data to construct a multi-source time-series dataset.

[0032] S2. Construction and continuous updating of self-evolving multi-level digital twins Construct a four-layer digital twin architecture comprising a physical layer, a connectivity layer, a virtual layer, and a decision layer.

[0033] Within the virtual layer, the following three mutually coupled digital twin sub-models are constructed: (1) Photovoltaic array twin model: A hybrid prediction model based on physical information encoding, which decomposes the end-to-end prediction network into a data-driven external parameter prediction sub-model and a physical-driven power calculation sub-model. The external parameter prediction sub-model uses a Long Short-Term Memory (LSTM) network to predict external parameters such as solar irradiance and module temperature within future time windows; the power calculation sub-model uses a single-diode equivalent circuit model to calculate the theoretical output power of each photovoltaic module based on the predicted external parameters and module physical parameters. The output expression of this model is: Among them, P PV For the output power of photovoltaic modules, I ph I0 is the photogenerated current, I0 is the reverse saturation current, q is the electron charge, V is the output voltage, I is the output current, and R0 is the output current. s For series resistance, R sh is the parallel resistance, n is the diode ideality factor, k is the Boltzmann constant, and T is the absolute temperature of the component.

[0034] The specific expression for the single-diode equivalent circuit model in the photovoltaic array twin model is as follows: Where I is the output current of the photovoltaic module (A); V is the output voltage of the photovoltaic module (V); I ph I0 is the photocurrent (A), which is proportional to the solar irradiance; I0 is the diode reverse saturation current (A); q is the electron charge; R0 s R is the equivalent series resistance (Ω); sh Ω is the equivalent parallel resistance; n is the diode ideality factor, which typically ranges from 1 to 2; k is the Boltzmann constant; and T is the absolute temperature of the photovoltaic module (K).

[0035] (2) Thermal-optical coupled twin model: Convolutional Long Short-Term Memory (ConvLSTM) network is used to construct the model. The three-dimensional geometric information of the curtain wall, real-time irradiance distribution, outdoor meteorological parameters and indoor air conditioning operation status are used as inputs. The output is the temperature field distribution of each area of ​​the curtain wall and the indoor natural lighting illuminance distribution. This model can predict the impact of photovoltaic curtain walls on building heat load and indoor light environment under different control states.

[0036] (3) Twin model of component health status: Based on Bayesian inference and Kalman filtering, the model is constructed by performing time series analysis on the residual between the actual output power and the theoretical output power of the photovoltaic module, estimating the equivalent series resistance increment, efficiency decay factor and local shading state parameters of each module online, so as to realize the real-time assessment of the health status of the module and the prediction of degradation trend.

[0037] The component health state twin model uses Kalman filtering to estimate the degradation state online, and its state-space model is as follows: Equations of state: Observation equation: Among them, F k Let z be the state transition matrix, describing the linearized dynamics of the degradation process; k Let H be the observation vector at time k, and take the residual sequence of the actual output power and the theoretical output power; k The observation matrix maps the state vector to the observation space; w k-1 For process noise; v k To mitigate noise, a Kalman filter-based prediction-update recursive process is used to estimate the health status parameters of each photovoltaic module in real time.

[0038] A multi-digital twin collaborative operation mechanism is introduced, simultaneously maintaining multiple twin model instances with different initialization parameters. The prediction results of each model are fused through a weighted ensemble method. An online incremental learning strategy is adopted. When the deviation between the actual observed data and the model's predicted values ​​exceeds a preset threshold, the model is triggered to make incremental updates, adjusting only some network weights without retraining all parameters. This achieves lightweight and continuous evolution of the digital twin, effectively overcoming the model drift problem.

[0039] S3. Local state inference based on virtual sensor networks To address the issue of sparse sensor deployment at the component level in photovoltaic curtain walls, a virtual sensor network based on transfer learning and physical constraints is constructed.

[0040] First, using real data collected by the deployed finite component-level and string-level sensors, a deep neural network model is trained, which takes string-level electrical parameters (current, voltage, power) and curtain wall-level environmental parameters (total irradiance, ambient temperature) as inputs and local state parameters of each component (local irradiance, component temperature, degree of shading) as outputs.

[0041] Secondly, a transfer learning mechanism is introduced, which uses a large amount of synthetic data generated in the simulation environment to pre-train the model, and then uses a small amount of real collected data in the actual deployment scenario to fine-tune the model, which greatly reduces the dependence on the number of actual sensors deployed.

[0042] Furthermore, a physical constraint term based on Kirchhoff's current law is added to the model loss function to minimize the consistency error of the current prediction values ​​of each component within the same string, as well as a thermal balance constraint term based on the law of energy conservation, to ensure that the derivation results conform to physical laws.

[0043] The total loss function of the virtual sensor network consists of two parts: data fitting loss and physical consistency constraint loss. L MSE The mean squared error loss between the predicted and actual values; L KCL As a constraint term of Kirchhoff's current law, it is required that the sum of the predicted current values ​​of all components within the same string be consistent with the measured current of the string; L Energy As an energy conservation constraint, it requires a balance between input light energy and output electrical energy and heat dissipation; , , which is the regularization weight coefficient for the physical constraint term.

[0044] Ultimately, a virtual sensor network covering all photovoltaic modules of the curtain wall is formed, enabling high-precision local state global simulation driven by a small number of real sensors.

[0045] S4, Multi-agent Deep Reinforcement Learning Distributed Optimization Decision The photovoltaic curtain wall control problem is modeled as a distributed partially observable Markov decision process, specifically including: Multi-agent partitioning: The curtain wall is divided into multiple control zones based on orientation and height, with each control zone corresponding to a control agent; a global coordination agent is also set up. Each control zone agent is responsible for the angle adjustment and bypass switch control of the photovoltaic modules within its zone, while the global coordination agent is responsible for multi-objective weight allocation and the formulation of inter-zone collaborative strategies. Angle adjustment is achieved through adjustable photovoltaic brackets, and bypass switch control is used to isolate locally shaded modules.

[0046] State space: For the i-th agent in the control region, its observation state o i This includes: the predicted power generation and health status parameters of each component output by the photovoltaic array twin model of the region; the regional temperature field and daylighting distribution output by the thermal-photovoltaic coupled twin model; the hourly irradiance and temperature prediction for the next 24 hours output by the weather forecast data; and the current time period type and occupancy status output by the building usage scenario parameters.

[0047] Action Space: The actions of the control region agent include: photovoltaic module tilt adjustment, photovoltaic module azimuth fine-tuning, and bypass switch status. The actions of the global coordination agent include: the multi-objective optimization weight vector W of each control region. t =[w E , w T , w L , w V ] T These correspond to the weights for power generation efficiency, thermal comfort, lighting comfort, and visual comfort, respectively.

[0048] Reward Function: Design a composite reward function that integrates multiple objectives. Among them, P PV,t P represents the actual power generation. max For the rated maximum power, T in,t Indoor temperature, T set To set the temperature, T tol For temperature tolerance range, UDI t DGP is the effective natural daylight index. t C is the glare probability index. act,t For the cost of action of the implementing agency, This represents the penalty coefficient for the action cost.

[0049] The Multi-Agent Proximity Policy Optimization (Multi-Agent PPO) algorithm is used for training. The training objective function of the Multi-Agent PPO algorithm is: Where θ represents the policy network parameters; r t (θ) represents the probability ratio between the old and new strategies; Let be the generalized advantage estimate (GAE) value at time t; The trimming hyperparameter is typically set to 0.1 to 0.2. This represents the empirical expectation for the time step.

[0050] For the agent in the i-th control region, its advantage function is estimated using the output of a centralized critic network: in, Discount factor; For GAE smoothing parameters; It is a centralized value network, with global observation information as the input; Global state information accessible during the training phase.

[0051] The control area agent and the global coordination agent adopt a "centralized training-distributed execution" framework: during the training phase, each agent can access global state information provided by the digital twin for policy learning; during the execution phase, each control area agent makes independent decisions based solely on local observation information. During training, a large amount of simulated operational data generated by the digital twin is used as the training environment, significantly reducing the cost of trial and error in the real environment.

[0052] S5. Meta-learning-based scenario adaptive control policy transfer To address the personalized adaptation needs of different building usage scenarios, a rapid policy transfer mechanism based on meta-learning is introduced: First, a scene encoder is constructed: the building usage scene features, including building function type, spatial function layout, typical personnel usage patterns, and climate region features, are encoded into a low-dimensional scene embedding vector z. s .

[0053] Secondly, the meta-policy is trained: the Model-Independent Meta-Learning (MAML) algorithm is used to perform meta-training on simulation task distributions covering multiple building types and climate regions, learning a set of meta-policy parameters θ that can quickly adapt to new scenarios. meta .

[0054] The inner and outer optimization processes of the Model-Independent Meta-Learning (MAML) algorithm are as follows: Inner layer adaptation update: Outer meta update: Where θ represents the initial parameters of the meta-policy network; α represents the inner layer learning rate (task-specific adaptation step size); and β represents the outer layer meta-learning rate.

[0055] Secondly, scene-adaptive fine-tuning: When the system is deployed on a new specific building, the scene embedding vector z of that building is obtained through the scene encoder. s Using this as a contextual condition, the meta-policy network is fed in for a small number of gradient updates, which can quickly generate personalized control strategies that adapt to the current building usage scenario.

[0056] S6. Explainable Decision Output and Security Verification To address the "black box" problem of end-to-end models, the following interpretability mechanism is established: (1) Decision traceability based on physical information coding: Since the photovoltaic array twin model adopts a physical-data hybrid architecture, its internal state variables have clear physical meanings. Each time a control decision is output, the system synchronously outputs the traceability information of the decision basis, including the health status assessment value of the photovoltaic modules in the current control area, the predicted irradiance distribution, the shading status diagnosis result and the building load prediction value, so that the decision process is traceable and verifiable.

[0057] (2) Safety boundary verification mechanism: Before the control command is issued to the physical actuator, the control command is pre-execution simulated and verified using a digital twin. Specifically, this includes: Voltage / current safety boundary verification: Ensure that the operating voltage and current of each photovoltaic string do not exceed the equipment's safety rating after the control command is executed; Hot spot risk assessment: For components with the risk of local shading, verify whether the bypass switch control command effectively avoids the hot spot effect and ensures that the hot spot temperature is within a safe range; Building electrical safety verification: Ensure that the impact of curtain wall power generation fluctuations on the building electrical system is within an acceptable range, and avoid large power surges caused by control commands; Mechanical safety verification: For angle adjustment actions, verify whether the angle change within adjacent adjustment cycles exceeds the maximum permissible rate of change of the mechanical structure.

[0058] If any security check fails, the current control command is rejected, and a security rollback strategy is triggered—either maintaining the current state or executing a preset security control mode.

[0059] S7. Control command execution and feedback closed loop Control commands that have passed safety verification are sent to physical actuators, including photovoltaic bracket angle adjustment motors and bypass switch controllers. After the actuators complete their actions, the sensor network collects system response data and feeds it back to the digital twin for status updates, forming a fully closed-loop intelligent control process of "perception-deduction-decision-verification-execution-feedback".

[0060] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent control of photovoltaic curtain walls based on digital twin algorithms, characterized in that: Includes the following steps: S1: Collect multi-source sensing data of the photovoltaic curtain wall system, including curtain wall-level environmental parameters, string-level electrical parameters, component-level local state parameters, and building operation data; S2: Construct a self-evolving multi-level digital twin, which includes a photovoltaic array twin model, a thermal-optical coupling twin model, and a component health status twin model, and achieves continuous self-evolution and updating through a multi-digital twin collaborative operation mechanism and an online incremental learning strategy; S3: Based on the deployed real sensor data and the digital twin, construct a virtual sensor network to deduce the local state parameters of all components of the photovoltaic curtain wall; S4: Employ a multi-agent deep reinforcement learning framework, using the output of the digital twin as the state input, to generate a multi-objective collaborative optimization control strategy. The multi-agent includes multiple control region agents and a global coordination agent. S5: Utilize a meta-learning-based policy transfer mechanism to quickly adapt pre-trained meta-policies to the target building usage scenario and generate personalized control policies. S6: Perform interpretability tracing and multi-dimensional security boundary verification on the generated control strategy, and issue control commands that have passed security verification to physical actuators; S7: After the actuator completes its action, it collects the system response data and feeds it back to the digital twin to form a closed-loop control.

2. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S2, the photovoltaic array twin model adopts a hybrid prediction model with physical information encoding, including a data-driven external parameter prediction sub-model and a physical-driven power calculation sub-model. The external parameter prediction sub-model performs time-series prediction of solar irradiance and module temperature based on a long short-term memory network, and the power calculation sub-model calculates the output power of the photovoltaic module based on a single diode equivalent circuit model.

3. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S2, the twin model of the component health status is constructed based on Bayesian inference and Kalman filtering. By performing time-series analysis on the residual between the actual output power and the theoretical output power of the photovoltaic module, the equivalent series resistance increment, efficiency decay factor and local shading state parameters of each module are estimated online.

4. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S2, the self-evolutionary update mechanism is as follows: simultaneously maintaining multiple twin model instances with different initialization parameters, and fusing the prediction results of each model through a weighted integration method; when the deviation between the actual observed data and the model prediction value exceeds a preset threshold, an online incremental learning strategy is adopted to adjust only a portion of the network weights to update the model.

5. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S3, the method for constructing a virtual sensor network includes: training a deep neural network model using limited real sensor data that has been deployed; pre-training the model using synthetic data generated in the simulation environment through a transfer learning mechanism, and then fine-tuning the model using real data; and adding a physical consistency constraint term based on Kirchhoff's current law and the law of conservation of energy to the model loss function.

6. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S4, the multi-agent deep reinforcement learning framework models the photovoltaic curtain wall control problem as a distributed partially observable Markov decision process. The observation states of the control area agents include the output of the photovoltaic array twin model, the output of the thermal-optical coupling twin model, weather forecast data, and building usage scenario parameters. The actions of the control area agents include photovoltaic module tilt adjustment, azimuth angle fine-tuning, and bypass switch status control. The actions of the global coordination agent include multi-objective optimization weight vector allocation for each control area.

7. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 6, characterized in that: The multi-agent deep reinforcement learning framework adopts a multi-agent proximal policy optimization algorithm and is trained in a "centralized training-distributed execution" manner: during the training phase, each agent can access global state information to learn policies, and during the execution phase, each agent in the control area makes independent decisions based only on local observation information. The training process utilizes the simulated running data generated by the digital twin as the training environment.

8. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S5, the meta-learning-based policy transfer mechanism includes: constructing a scene encoder to encode building function type, spatial function layout, typical personnel usage patterns, and climate region characteristics into low-dimensional scene embedding vectors; using a model-independent meta-learning algorithm to perform meta-training on simulation task distributions covering multiple building types and climate regions to obtain meta-policy parameters; and when deployed to the target building, performing a small number of gradient updates on the meta-policy network with the scene embedding vector as context conditions to quickly generate personalized control policies adapted to the current scene.

9. The intelligent control method for photovoltaic curtain walls based on digital twin algorithm according to claim 1, characterized in that: In S6, the multi-dimensional safety boundary verification includes: voltage / current safety boundary verification, hot spot risk assessment, building electrical safety verification, and mechanical safety verification; if any safety verification fails, the current control command is rejected and a safety rollback strategy is triggered.

10. A photovoltaic curtain wall intelligent control system based on a digital twin algorithm, used to implement the method described in any one of claims 1 to 9, characterized in that, include: A multi-level sensing network module is used to collect multi-source sensing data and building operation data from the photovoltaic curtain wall system; The digital twin engine module is used to build and run the self-evolving multi-level digital twins, including photovoltaic array twin models, thermal-optical coupling twin models, and component health status twin models, as well as to realize the collaborative operation and online incremental learning of multiple digital twins; The virtual sensor network module is used to extrapolate the local state parameters of all components of the curtain wall based on limited real sensor data. A multi-agent reinforcement learning decision-making module is used to generate multi-objective collaborative optimization control strategies; The meta-learning strategy adaptation module is used to quickly adapt pre-trained meta-policies to the target building usage scenario; The interpretability and security verification module is used to perform physical source interpretation of control decisions and multi-dimensional security boundary verification. The actuator control module is used to send control commands that have passed safety verification to the photovoltaic bracket angle adjustment motor and bypass switch controller, and to receive execution feedback.

Images