Network fine-tuning-based building facade control method, system, controller and storage medium

Abstract

The application relates to the technical field of building control, and discloses a building facade control method and system based on network fine tuning, a controller and a storage medium.The method comprises the following steps: standardizing real-time environment data to obtain an environment state vector; inputting the environment state vector into an initial controller, controlling a facade component and receiving feedback data of a target user; constructing a preference pair in a current period according to the feedback data to optimize the initial controller to obtain a target controller; and inputting an environment state vector of a target area in a next period into the target controller to adjust the facade component.The application improves the efficient management of the facade component by deeply fusing deep reinforcement learning driven by data and lightweight online preference learning technology, and improves the adaptive ability, individualization degree and multi-target real-time optimization efficiency of the equipment.

Description

A method, system, controller, and storage medium for network-based fine-tuning of building facade control. Technical Field

[0001] This invention relates to the field of building control technology, and in particular to a method, system, controller, and computer-readable storage medium for controlling building facades based on network fine-tuning. Background Technology

[0002] As the physical interface between the indoor and outdoor environments, the building facade plays a crucial role in regulating the exchange of energy and matter, such as light, heat, and air. To achieve a balance between energy conservation and indoor environmental comfort, adaptive building facades have emerged. Their core lies in dynamically adjusting the facade's state (such as the angle of shading components and light transmittance) through a control system that responds in real time to changes in indoor and outdoor climate and the needs of residents.

[0003] Existing adaptive facade control systems mainly consist of an environmental perception layer, a control decision layer, and an execution layer. They lack a real-time feedback mechanism for the environment and cannot adapt to long-term changes in the building environment, usage patterns, and equipment status.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a network-based fine-tuning method, system, controller, and computer-readable storage medium for controlling building facades. This aims to address the problem that existing technologies lack adaptive optimization and feedback processes for controlling facade components, thereby increasing the complexity of equipment management and reducing the accuracy of equipment control.

[0006] To achieve the above objectives, the present invention provides a building facade control method based on network fine-tuning, the building facade control method based on network fine-tuning comprising the following steps:

[0007] Acquire real-time environmental data of the target area, and standardize the real-time environmental data to obtain an environmental state vector;

[0008] The environmental state vector is input to the initial controller, and control instructions for each facade component are output. The corresponding facade component is controlled according to each control instruction, and feedback data from the target user is received.

[0009] Based on the feedback data, multiple preference pairs for the current period are constructed, and the initial controller is optimized based on all the preference pairs to obtain the target controller;

[0010] The environmental state vector of the target area in the next cycle is input to the target controller to control all the facade components.

[0011] Optionally, the aforementioned network-based fine-tuning method for controlling building facades, wherein acquiring real-time environmental data of the target area and standardizing the real-time environmental data to obtain an environmental state vector specifically includes:

[0012] Real-time environmental data of the target area is collected at fixed intervals, and a corresponding timestamp is added to each piece of real-time environmental data;

[0013] All abnormal data in the real-time environmental data are removed, all normal data within the preset value range are retained, and all missing values ​​are filled in using linear interpolation to obtain a filtered data sequence.

[0014] Unify all data units and all data magnitude units in the filtered data sequence, and normalize all the data according to each data unit to obtain the corresponding data set;

[0015] Each data set is sequentially concatenated into a one-dimensional array using predefined rules to obtain the environmental state vector for the current period.

[0016] Optionally, the aforementioned network-based fine-tuning method for controlling building facades, wherein inputting the environmental state vector to the initial controller and outputting control commands for each facade component specifically includes:

[0017] For each facade component, the environmental state vector is input into the feature extraction network of the initial controller, and the feature extraction network is used to perform multi-layer feature transformation on the environmental state vector to output a high-dimensional feature vector.

[0018] The high-dimensional feature vector is input into the policy network of the initial controller for prediction, and multiple action probabilities are output.

[0019] The high-dimensional feature vector is input into the value network of the initial controller for expected evaluation, and the expected benefit of the environmental state vector is output.

[0020] Based on all the said action probabilities and the said expected reward, an action probability distribution of the environment state vector is generated, and control instructions for each of the said facade components are generated based on the action probability distribution.

[0021] Optionally, the building facade control method based on network fine-tuning, wherein the step of inputting the environmental state vector into the feature extraction network of the initial controller, and using the feature extraction network to perform multi-layer feature transformation on the environmental state vector to output a high-dimensional feature vector, specifically includes:

[0022] The environmental state vector is input into the feature extraction network of the initial controller, and a linear transformation and nonlinear activation are performed on the environmental state vector to obtain the primary feature vector:

[0023] ;

[0024] in, Represents the primary feature vector. Denotes the first nonlinear activation function. This represents the first weight matrix. Represents the environment state vector. express The bias vector;

[0025] The primary feature vectors are subjected to higher-order cross-processing and abstraction to output high-dimensional feature vectors:

[0026] ;

[0027] in, Represents a high-dimensional feature vector. This represents the second weight matrix. express The bias vector.

[0028] Optionally, the building facade control method based on network fine-tuning, wherein inputting the high-dimensional feature vector into the policy network of the initial controller for prediction and outputting multiple action probabilities specifically includes:

[0029] The high-dimensional feature vector is input into the policy network of the initial controller. The high-dimensional feature vector is then linearly transformed through the fully connected layer of the policy network to map it to a multi-dimensional space corresponding to multiple angles, resulting in a linear transformation.

[0030] ;

[0031] in, Indicates the result of a linear transformation. This represents the third weight matrix. express The bias vector;

[0032] By introducing a nonlinearity into the linear transformation result using a second nonlinear activation function, the action probability corresponding to each angle is obtained:

[0033] ;

[0034] in, Indicates the probability of an action. This represents the second nonlinear activation function.

[0035] Optionally, the aforementioned network-based fine-tuning method for controlling building facades, wherein the step of constructing multiple preference pairs within the current period based on the feedback data, and optimizing the initial controller based on all the preference pairs to obtain the target controller, specifically includes:

[0036] Extract the target user's preference information from the feedback data, and convert the preference information, the environmental state vector of the current period, and the actions corresponding to all the action probabilities into preference pairs;

[0037] The policy loss function is constructed using all the aforementioned preference pairs, and then minimized to optimize the policy network.

[0038] ;

[0039] in, Represents the policy loss function. This represents the fine-tuned strategy parameters. This represents the strategy parameters before fine-tuning. This represents the environment state vector at the feedback time. This represents the action with the highest probability. This indicates that the action has a lower probability than the action itself. This represents the Sigmoid function. This represents the hyperparameter that controls the intensity of fine-tuning. Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability of.

[0040] Optionally, the building facade control method based on network fine-tuning further includes, after constructing a policy loss function using all the preference pairs and optimizing the policy network using the policy loss function:

[0041] For each of the aforementioned facade components, after the movement of the facade component is adjusted, the real-time environmental data of the previous cycle and the current cycle are compared to obtain environmental change data;

[0042] Based on the environmental change data, the thermal comfort item, visual comfort item, and energy consumption item of the facade components of the target user are determined. The thermal comfort item, visual comfort item, and energy consumption item are weighted and fused to obtain a comprehensive reward.

[0043] The overall reward and the expected return are compared to obtain the optimization result of the policy network.

[0044] Furthermore, to achieve the above objectives, the present invention also provides a network-based fine-tuning building facade control system, wherein the network-based fine-tuning building facade control system includes:

[0045] The data acquisition module is used to acquire real-time environmental data of the target area, and to standardize the real-time environmental data to obtain an environmental state vector.

[0046] The equipment control module is used to input the environmental state vector to the initial controller, output control instructions for each facade component, control the corresponding facade component according to each control instruction, and receive feedback data from the target user.

[0047] The strategy optimization module is used to construct multiple preference pairs within the current period based on the feedback data, and optimize the initial controller based on all the preference pairs to obtain the target controller;

[0048] The equipment re-control module is used to input the environmental state vector of the target area in the next cycle to the target controller in order to control all the facade components.

[0049] Furthermore, to achieve the above objectives, the present invention also provides a controller, wherein the controller includes: a memory, a processor, and a network-based fine-tuning building facade control program stored in the memory and executable on the processor, wherein when the network-based fine-tuning building facade control program is executed by the processor, it implements the steps of the network-based fine-tuning building facade control method as described above.

[0050] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a network-based fine-tuning building facade control program, which, when executed by a processor, implements the steps of the network-based fine-tuning building facade control method as described above.

[0051] In this invention, real-time environmental data of the target area is acquired, and the real-time environmental data is standardized to obtain an environmental state vector. The environmental state vector is input to an initial controller, which outputs control instructions for each facade component. The corresponding facade component is controlled according to each control instruction, and feedback data from the target user is received. Based on the feedback data, multiple preference pairs for the current period are constructed, and the initial controller is optimized based on all preference pairs to obtain a target controller. The environmental state vector of the target area in the next period is input to the target controller to control all facade components. This invention improves the efficient management of facade components and enhances the device's adaptability, personalization, and multi-objective real-time optimization efficiency by deeply integrating data-driven deep reinforcement learning with lightweight online preference learning technology. Attached Figure Description

[0052] Figure 1 is a flowchart of a preferred embodiment of the building facade control method based on network fine-tuning of the present invention;

[0053] Figure 2 is a flowchart of the system control of a preferred embodiment of the building facade control method based on network fine-tuning of the present invention;

[0054] Figure 3 is a flowchart of feature transformation of a preferred embodiment of the building facade control method based on network fine-tuning of the present invention;

[0055] Figure 4 is a structural diagram of a preferred embodiment of the building facade control system based on network fine-tuning of the present invention;

[0056] Figure 5 is a structural diagram of a preferred embodiment of the controller of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0058] Currently, common adaptive facade control systems typically employ a hierarchical control architecture, mainly consisting of an environmental perception layer, a control decision layer, and an execution layer. Furthermore, the control strategy of the control decision layer is rigid and inflexible, lacking environmental adaptability. Whether it is a rule-based controller, a PID (Proportional Integration Differentiation) controller, or a Model Predictive Control (MPC), its core parameters (such as rule thresholds, PID gain, and MPC model parameters and weight matrices) are all manually tuned once during the design phase based on specific operating conditions or simulation models.

[0059] While Model Predictive Control (MPC) can theoretically handle multi-objective optimization, it suffers from high computational complexity in practical applications and static weights. The system lacks the ability to continuously learn from interactive data and evolve itself, making it difficult to achieve stable and excellent personalized performance in long-term, dynamic real-world building environments.

[0060] Based on the above problems, this invention adds a user feedback layer to the above three architectures to realize the process from static setting to dynamic adaptive learning, so as to improve the control efficiency of facade components and improve the overall adaptability and robustness of the system.

[0061] The preferred embodiment of the building facade control method based on network fine-tuning of the present invention, as shown in Figure 1, includes the following steps:

[0062] Step S10: Obtain real-time environmental data of the target area, and standardize the real-time environmental data to obtain an environmental state vector.

[0063] As shown in Figure 2, different environmental data are acquired through multiple sensors in the environmental perception layer, and the multi-source data is further converted into a unified state vector to improve data quality. This provides a basic guarantee for optimizing model performance and can significantly improve data analysis efficiency.

[0064] Specifically, real-time environmental data of the target area is collected at fixed intervals, and a corresponding timestamp is added to each piece of real-time environmental data;

[0065] All abnormal data in the real-time environmental data are removed, all normal data within the preset value range are retained, and all missing values ​​are filled in using linear interpolation to obtain a filtered data sequence.

[0066] Unify all data units and all data magnitude units in the filtered data sequence, and normalize all the data according to each data unit to obtain the corresponding data set;

[0067] Each data set is sequentially concatenated into a one-dimensional array using predefined rules to obtain the environmental state vector for the current period.

[0068] In the embodiments disclosed in this invention, outdoor environmental data, such as total solar radiation intensity, outdoor dry-bulb temperature, relative humidity, and wind speed, are collected by outdoor weather stations installed on the building roof or exterior walls. Indoor environmental data, such as work surface illuminance sensors (installed at 0.8 meters above the ground), eye height illuminance sensors (1.2 meters or 1.7 meters), and indoor temperature and humidity sensors, are collected by indoor environmental sensors arranged in various indoor zones to calculate visual comfort and thermal comfort indices. Finally, the physical state of different exterior facade components (e.g., louver opening angle) can be fed back in real time by facade status sensors.

[0069] Furthermore, these initially acquired real-time environmental data are preprocessed to remove abnormal parameters that are outside the physical range. For some sensors that experience momentary communication failures, there may be multiple missing values. In this case, linear interpolation can be used to fill in these missing values ​​to ensure the continuity of the data stream.

[0070] Furthermore, all processed data are standardized to eliminate differences in physical dimensions (e.g., temperature data has multiple units, so all temperature data units are converted to "degrees Celsius") and magnitudes, thereby eliminating their interference with model training and improving model training efficiency.

[0071] The data standardization process can be achieved through min-max normalization, mapping the raw values ​​of each sensor to the [0, 1] interval. Then, all the preprocessed and standardized data are concatenated into a one-dimensional array in a predefined, fixed order, forming the current environmental state vector. The dimension and order of the environmental state vector remain unchanged after system initialization, ensuring that the deep learning model can stably identify the physical meaning corresponding to each input position. Through this process, the multi-source, heterogeneous raw sensor data is transformed into a clean, unified numerical input (i.e., the environmental state vector) suitable for deep reinforcement learning model decision-making, providing an important data foundation for achieving intelligent and stable control.

[0072] Step S20: Input the environmental state vector into the initial controller, output control instructions for each facade component, control the corresponding facade component according to each control instruction, and receive feedback data from the target user.

[0073] Based on the generated environmental state vector, the execution layer can obtain the device control commands generated after analyzing the environmental state vector. The execution layer consists of motor drivers, relays, etc., and drives the exterior facade mechanisms such as electric sunshade louvers and roller blinds based on the device control commands, changing their physical state (such as rotating to a specified angle), thereby actively adjusting the light and heat entering the room.

[0074] Specifically, for each facade component, the environmental state vector is input into the feature extraction network of the initial controller, and the environmental state vector is transformed into a multi-layer feature vector using the feature extraction network to output a high-dimensional feature vector.

[0075] The high-dimensional feature vector is input into the policy network of the initial controller for prediction, and multiple action probabilities are output.

[0076] The high-dimensional feature vector is input into the value network of the initial controller for expected evaluation, and the expected benefit of the environmental state vector is output.

[0077] Based on all the said action probabilities and the said expected reward, an action probability distribution of the environment state vector is generated, and control instructions for each of the said facade components are generated based on the action probability distribution.

[0078] In the embodiments disclosed in this invention, a deep reinforcement learning (DRL) controller replaces the traditional rule table, PID, or MPC controller. Furthermore, an Actor-Critic architecture (where Actor corresponds to the executor (policy network) and Critic corresponds to the evaluator (value network)) is adopted at the controller level. As shown in Figure 3, the environment state vector is passed through a deep neural network containing two fully connected layers (e.g., 256 neurons per layer, using the ReLU activation function) to ultimately obtain a high-dimensional feature vector.

[0079] Furthermore, the environmental state vector is input into the feature extraction network of the initial controller, and a linear transformation and nonlinear activation are performed on the environmental state vector to obtain a primary feature vector:

[0080] ;

[0081] in, Represents the primary feature vector. Denotes the first nonlinear activation function. This represents the first weight matrix. Represents the environment state vector. express The bias vector;

[0082] The primary feature vectors are subjected to higher-order cross-processing and abstraction to output high-dimensional feature vectors:

[0083] ;

[0084] in, Represents a high-dimensional feature vector. This represents the second weight matrix. express The bias vector.

[0085] In the embodiments disclosed in this invention, the environmental state vector is converted into a primary feature vector in the first layer; The weight matrix is ​​[256, 9]. It is a 256-dimensional bias vector, which undergoes a linear transformation and nonlinear activation at the first layer. Each row vector can be viewed as a "focus pattern" of the original 9-dimensional state, used to calculate the initial weighted combination relationship between different sensor data (for example, simultaneously focusing on "outdoor radiation" and "indoor temperature" with high weights to initially capture the solar heat gain effect); the ReLU function introduces nonlinearity, enabling the model to learn more complex patterns.

[0086] Furthermore, in the second layer, the primary feature vectors are transformed into high-dimensional feature vectors; The weight matrix is ​​[256, 256]. It is another 256-dimensional bias vector; based on the primary feature vector, this layer performs higher-order cross-linking and abstraction, and can learn more complex deep environmental state patterns that are crucial to the final decision, such as the interaction between primary feature A (characterizing radiative heat gain) and primary feature B (characterizing ventilation effect).

[0087] In the above-disclosed embodiments, high-dimensional feature vectors, as shared and information-rich abstract representations, are fed into the subsequent policy network and value network respectively. The policy network executes the probability distribution of five discrete shading angle actions, while the value network evaluates the long-term value of the current state based on the high-dimensional feature vectors to provide a data foundation for subsequent model optimization.

[0088] Furthermore, the high-dimensional feature vector is input into the policy network of the initial controller, and a linear transformation is performed on the high-dimensional feature vector through the fully connected layer of the policy network to map the high-dimensional feature vector to a multi-dimensional space corresponding to multiple angles, thereby obtaining the linear transformation result:

[0089] ;

[0090] in, Indicates the result of a linear transformation. This represents the third weight matrix. express The bias vector;

[0091] By introducing a nonlinearity into the linear transformation result using a second nonlinear activation function, the action probability corresponding to each angle is obtained:

[0092] ;

[0093] in, Indicates the probability of an action. This represents the second nonlinear activation function.

[0094] In the embodiments disclosed in this invention, Given a weight matrix of [5, 256], during the linear transformation process, the 256-dimensional high-dimensional feature vectors are linearly combined and mapped to a 5-dimensional space corresponding to 5 selectable shading angles (0°, 15°, 30°, 45°, 60°). Each row in the algorithm (representing a 256-dimensional vector) is responsible for learning how to identify patterns of preference for a particular action from high-dimensional feature vectors.

[0095] Furthermore, the results of the linear transformation are applied... By introducing nonlinearity into the function, the model can learn and express more complex nonlinear relationships between features and action values, making it more powerful than a simple linear mapping.

[0096] Step S30: Based on the feedback data, construct multiple preference pairs for the current period, and optimize the initial controller based on all the preference pairs to obtain the target controller.

[0097] The feedback layer, as a key layer of this system, generates user-related comfort feedback based on the control results of the aforementioned devices, thereby enabling a closed-loop optimization process for subsequent control strategies. For example, the collected subjective feelings must be set with specific, measurable indicators (thermal comfort: temperature too high or too low; lighting environment: glare or no glare). Feedback commands will only relay this information without distinguishing the specific cause. After this feedback information reaches the control system, the control system will find the most suitable facade control action to achieve the goal (e.g., making glare non-glare) based on these indicators, and increase the probability of the DRL controller outputting this action under these environmental conditions.

[0098] Specifically, the target user's preference information is extracted from the feedback data, and the preference information, the environmental state vector of the current period, and the actions corresponding to all the action probabilities are converted into preference pairs;

[0099] The policy loss function is constructed using all the aforementioned preference pairs, and then minimized to optimize the policy network.

[0100] ;

[0101] in, Represents the policy loss function. This represents the fine-tuned strategy parameters. This represents the strategy parameters before fine-tuning. This represents the environment state vector at the feedback time. This represents the action with the highest probability. This indicates that the action has a lower probability than the action itself. This represents the Sigmoid function. This represents the hyperparameter that controls the intensity of fine-tuning. Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability of.

[0102] In this process, the feedback data input by the user is uploaded to the controller to fine-tune the neural network.

[0103] Specifically, the preference information submitted by users through the feedback layer (such as "comfortable", "overheating", and manually set angle) is transformed into standardized preference pairs. The preference pairs include the environmental state vector at the time of feedback, the preferred action (e.g., the optimal cooling angle calculated by the system when the user feels "overheating"), and the non-preferred action (e.g., the angle currently selected by the controller that causes user dissatisfaction). These preference samples are stored in the controller's local database with timestamps.

[0104] For both non-optimal and optimal actions, the system will provide an additional confirmation process to prevent users from accidentally adjusting the device, which could reduce the accuracy of the model after training.

[0105] Furthermore, after accumulating a certain number of new preference samples, the system initiates an online fine-tuning process and uses a direct preference optimization method to fine-tune the policy network. Utilizing the stored preference samples, the system minimizes the policy loss function through a gradient descent algorithm, updating only the policy parameters of the Actor network. This process directly increases the output probability of the policy network for preferred actions while reducing the probability of for non-preferred actions, thereby gradually aligning the control policy with the individual user's true preferences.

[0106] The invention introduces a "feedback layer" and an "online fine-tuning mechanism based on DPO (Direct Preference Optimization)," constructing a complete personalized learning loop. When users submit feedback such as "too hot" or "glare" through the terminal, or make manual adjustments, the system not only triggers an immediate response to alleviate discomfort, but more importantly, it transforms this feedback, along with a complete environmental snapshot, into structured preference samples. Subsequently, these samples are used to fine-tune the DRL policy network through a direct preference optimization algorithm. This process optimizes directly at the policy probability distribution level, making the network more inclined to choose user-preferred actions (such as a lower shading angle to enhance ventilation and cooling) when encountering similar environmental conditions in the future, while suppressing actions that lead to dissatisfaction. This is equivalent to "branding" a personalized imprint on a general-optimized DRL policy. Compared with existing technologies, this invention achieves continuous and automatic alignment between control policies and end-user subjective feelings for the first time. It can dynamically capture and adapt to the differentiated needs of different users, and even the same user at different times and under different tasks, thereby significantly reducing user complaints and manual interventions caused by environmental discomfort, and improving user satisfaction and experience in buildings from the source.

[0107] Furthermore, for each of the facade components, after the movement of the facade component is adjusted, the real-time environmental data of the previous cycle and the current cycle are compared to obtain environmental change data;

[0108] Based on the environmental change data, the thermal comfort item, visual comfort item, and energy consumption item of the facade components of the target user are determined. The thermal comfort item, visual comfort item, and energy consumption item are weighted and fused to obtain a comprehensive reward.

[0109] The overall reward and the expected return are compared to obtain the optimization result of the policy network.

[0110] In this process, the controller is optimized in real time based on user feedback. For the current optimization process, based on the embodiments disclosed in this invention, the controller periodically uses the collected environmental state vector, the adjusted actions and the comprehensive reward to generate a new state for the next cycle. With the goal of maximizing long-term cumulative reward, the model parameters of the policy network and the value network are updated through a preset algorithm. This process enables the controller to automatically learn the optimal control strategy in complex environments without the need for manual setting of rules or precise physical models.

[0111] In this process, after applying the current action to the real building physics environment, the environment naturally evolves to a new state after one control cycle (e.g., 5 minutes), which is the environmental data measured by the system through sensors in the next cycle. This state is essential data for training the DRL controller, and is used to calculate the key values ​​of the environmental state vector in the next cycle under the influence of the current action.

[0112] Step S40: Input the environmental state vector of the target area in the next cycle into the target controller to control all the facade components.

[0113] In each control cycle, after the controller receives the environmental state vector, the policy network samples the action with the highest probability (i.e., the target angle) and sends it to the execution layer to adjust the facade components. The system calculates a comprehensive reward (i.e., the expected benefit generated by the value network) based on the environmental changes after the action is executed. This comprehensive reward is composed of a weighted sum of thermal comfort, visual comfort, and energy consumption items. All items have been normalized, and the weights can be dynamically adjusted according to the season or policy to improve the model's adaptive optimization capability and enhance the user experience.

[0114] This invention utilizes deep reinforcement learning as its core decision-maker, essentially a machine learning system capable of directly learning strategies from data through trial-and-error interaction. After deployment, the DRL controller continuously adjusts its internal neural network parameters based on real-time environmental conditions and multi-objective reward feedback, thereby dynamically optimizing the control strategy. This means the system no longer relies on potentially outdated empirical models from the design phase, but instead possesses online learning and adaptive capabilities. For example, when new obstructions around a building cause changes in sunlight conditions, the DRL controller can automatically explore and learn new, more suitable shading strategies through reward signals (such as penalties for insufficient illumination), whereas traditional systems require manual readjustment. This data-driven learning mechanism fundamentally solves the performance degradation problem caused by environmental time-varying factors and model mismatch in existing systems, significantly improving the overall adaptability and robustness of the system in different seasons, weather conditions, and long-term operation.

[0115] This invention improves the efficient management of facade components by deeply integrating data-driven deep reinforcement learning with lightweight online preference learning technology, thereby enhancing the equipment's adaptability, personalization, and multi-objective real-time optimization efficiency. By deeply integrating data-driven deep reinforcement learning with lightweight online preference learning technology, it not only effectively overcomes the core bottlenecks of existing adaptive facade control systems in terms of adaptability, personalization, multi-objective real-time optimization, and long-term evolution, but also achieves efficient and reliable deployment and operation in engineering. This provides a practical technical solution for building a truly intelligent, comfortable, energy-saving, and human-centered next-generation building environment control system.

[0116] Furthermore, as shown in Figure 4, based on the above-mentioned network-based fine-tuning method for controlling building facades, the present invention also provides a network-based fine-tuning control system for building facades, wherein the network-based fine-tuning control system includes:

[0117] Data acquisition module 51 is used to acquire real-time environmental data of the target area, and to standardize the real-time environmental data to obtain an environmental state vector.

[0118] The equipment control module 52 is used to input the environmental state vector to the initial controller, output control instructions for each facade component, control the corresponding facade component according to each control instruction, and receive feedback data from the target user.

[0119] The strategy optimization module 53 is used to construct multiple preference pairs within the current period based on the feedback data, and optimize the initial controller based on all the preference pairs to obtain the target controller;

[0120] The equipment re-control module 54 is used to input the environmental state vector of the target area in the next cycle to the target controller in order to control all the facade components.

[0121] Furthermore, as shown in Figure 5, based on the above-described network-based fine-tuning method and system for controlling building facades, the present invention also provides a controller, which includes a processor 10, a memory 20, and a display 30. Figure 5 only shows some components of the controller; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0122] In some embodiments, the memory 20 may be an internal storage unit of the controller, such as the controller's hard drive or memory. In other embodiments, the memory 20 may be an external storage device of the controller, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the controller. Further, the memory 20 may include both the controller's internal storage unit and an external storage device. The memory 20 is used to store application software and various types of data installed on the controller, such as the program code installed on the controller. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a network-based fine-tuning building facade control program 40, which can be executed by the processor 10 to implement the network-based fine-tuning building facade control method of this application.

[0123] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the network-based fine-tuning building facade control method.

[0124] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information from the controller and to display a visual user interface. Components of the controller communicate with each other via a system bus.

[0125] In one embodiment, when processor 10 executes the network-based fine-tuning building facade control program 40 in memory 20, the following steps are performed:

[0126] Acquire real-time environmental data of the target area, and standardize the real-time environmental data to obtain an environmental state vector;

[0127] The environmental state vector is input to the initial controller, and control instructions for each facade component are output. The corresponding facade component is controlled according to each control instruction, and feedback data from the target user is received.

[0128] Based on the feedback data, multiple preference pairs for the current period are constructed, and the initial controller is optimized based on all the preference pairs to obtain the target controller;

[0129] The environmental state vector of the target area in the next cycle is input to the target controller to control all the facade components.

[0130] Specifically, the step of acquiring real-time environmental data of the target area and standardizing the real-time environmental data to obtain an environmental state vector includes:

[0131] Real-time environmental data of the target area is collected at fixed intervals, and a corresponding timestamp is added to each piece of real-time environmental data;

[0132] All abnormal data in the real-time environmental data are removed, all normal data within the preset value range are retained, and all missing values ​​are filled in using linear interpolation to obtain a filtered data sequence.

[0133] Unify all data units and all data magnitude units in the filtered data sequence, and normalize all the data according to each data unit to obtain the corresponding data set;

[0134] Each data set is sequentially concatenated into a one-dimensional array using predefined rules to obtain the environmental state vector for the current period.

[0135] Specifically, the step of inputting the environmental state vector into the initial controller and outputting control commands for each facade component includes:

[0136] For each facade component, the environmental state vector is input into the feature extraction network of the initial controller, and the feature extraction network is used to perform multi-layer feature transformation on the environmental state vector to output a high-dimensional feature vector.

[0137] The high-dimensional feature vector is input into the policy network of the initial controller for prediction, and multiple action probabilities are output.

[0138] The high-dimensional feature vector is input into the value network of the initial controller for expected evaluation, and the expected benefit of the environmental state vector is output.

[0139] Based on all the said action probabilities and the said expected reward, an action probability distribution of the environment state vector is generated, and control instructions for each of the said facade components are generated based on the action probability distribution.

[0140] Specifically, the step of inputting the environmental state vector into the feature extraction network of the initial controller, and using the feature extraction network to perform multi-layer feature transformation on the environmental state vector to output a high-dimensional feature vector includes:

[0141] The environmental state vector is input into the feature extraction network of the initial controller, and a linear transformation and nonlinear activation are performed on the environmental state vector to obtain the primary feature vector:

[0142] ;

[0143] in, Represents the primary feature vector. Denotes the first nonlinear activation function. This represents the first weight matrix. Represents the environment state vector. express The bias vector;

[0144] The primary feature vectors are subjected to higher-order cross-processing and abstraction to output high-dimensional feature vectors:

[0145] ;

[0146] in, Represents a high-dimensional feature vector. This represents the second weight matrix. express The bias vector.

[0147] Specifically, the step of inputting the high-dimensional feature vector into the policy network of the initial controller for prediction and outputting multiple action probabilities includes:

[0148] The high-dimensional feature vector is input into the policy network of the initial controller. The high-dimensional feature vector is then linearly transformed through the fully connected layer of the policy network to map it to a multi-dimensional space corresponding to multiple angles, resulting in a linear transformation.

[0149] ;

[0150] in, Indicates the result of a linear transformation. This represents the third weight matrix. express The bias vector;

[0151] By introducing a nonlinearity into the linear transformation result using a second nonlinear activation function, the action probability corresponding to each angle is obtained:

[0152] ;

[0153] in, Indicates the probability of an action. This represents the second nonlinear activation function.

[0154] Specifically, the step of constructing multiple preference pairs within the current period based on the feedback data, and optimizing the initial controller based on all the preference pairs to obtain the target controller includes:

[0155] Extract the target user's preference information from the feedback data, and convert the preference information, the environmental state vector of the current period, and the actions corresponding to all the action probabilities into preference pairs;

[0156] The policy loss function is constructed using all the aforementioned preference pairs, and then minimized to optimize the policy network.

[0157] ;

[0158] in, Represents the policy loss function. This represents the fine-tuned strategy parameters. This represents the strategy parameters before fine-tuning. This represents the environment state vector at the feedback time. This represents the action with the highest probability. This indicates that the action has a lower probability than the action itself. This represents the Sigmoid function. This represents the hyperparameter that controls the intensity of fine-tuning. Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability of.

[0159] The process of constructing a policy loss function using all the preference pairs and optimizing the policy network using the policy loss function further includes:

[0160] For each of the aforementioned facade components, after the movement of the facade component is adjusted, the real-time environmental data of the previous cycle and the current cycle are compared to obtain environmental change data;

[0161] Based on the environmental change data, the thermal comfort item, visual comfort item, and energy consumption item of the facade components of the target user are determined. The thermal comfort item, visual comfort item, and energy consumption item are weighted and fused to obtain a comprehensive reward.

[0162] The overall reward and the expected return are compared to obtain the optimization result of the policy network.

[0163] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a network-based fine-tuning building facade control program, which, when executed by a processor, implements the steps of the network-based fine-tuning building facade control method as described above.

[0164] In summary, this invention provides a network-based fine-tuning method and related equipment for controlling building facades. The method includes: acquiring real-time environmental data of a target area; standardizing the real-time environmental data to obtain an environmental state vector; inputting the environmental state vector into an initial controller, outputting control instructions for each facade component, controlling the corresponding facade component according to each control instruction, and receiving feedback data from a target user; constructing multiple preference pairs within the current period based on the feedback data, optimizing the initial controller based on all preference pairs to obtain a target controller; and inputting the environmental state vector of the target area in the next period into the target controller to control all facade components. This invention improves the efficient management of facade components and enhances the device's adaptability, personalization, and multi-objective real-time optimization efficiency by deeply integrating data-driven deep reinforcement learning with lightweight online preference learning technology.

[0165] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or controller that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or controller. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or controller that includes that element.

[0166] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0167] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for controlling building facades based on network fine-tuning, characterized in that, The network-based fine-tuning method for controlling building facades includes: acquiring real-time environmental data of a target area; standardizing the real-time environmental data to obtain an environmental state vector; inputting the environmental state vector into an initial controller, outputting control commands for each facade device, controlling the corresponding facade device according to each control command, and receiving feedback data from the target user; specifically, inputting the environmental state vector into the initial controller and outputting control commands for each facade device includes: for each facade device, inputting the environmental state vector into the feature extraction network of the initial controller, using the feature extraction network to perform multi-layer feature transformation on the environmental state vector, and outputting a high-dimensional feature vector; inputting the high-dimensional feature vector into the policy network of the initial controller for prediction, and outputting multiple action probabilities; and inputting the high-dimensional feature vector into the value network of the initial controller for prediction. The evaluation process involves: assessing and outputting the expected reward of the environmental state vector; generating an action probability distribution of the environmental state vector based on all action probabilities and the expected reward, and generating control commands for each of the facade devices based on the action probability distribution; constructing multiple preference pairs within the current period based on the feedback data, and optimizing the initial controller based on all preference pairs to obtain the target controller; specifically, constructing multiple preference pairs within the current period based on the feedback data, and optimizing the initial controller based on all preference pairs to obtain the target controller includes: extracting the target user's preference information from the feedback data, and converting the preference information, the environmental state vector of the current period, and the actions corresponding to all action probabilities into preference pairs; constructing a policy loss function using all preference pairs, and minimizing the policy loss function to optimize the policy network. ;in, Represents the policy loss function. This represents the fine-tuned strategy parameters. This represents the strategy parameters before fine-tuning. This represents the environment state vector at the feedback time. This represents the action with the highest probability. This indicates that the action has a lower probability than the action itself. This represents the Sigmoid function. This represents the hyperparameter that controls the intensity of fine-tuning. Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability, Indicates use At that time Under the condition of selection The probability; input the environmental state vector of the target area in the next cycle to the target controller to adjust all the facade equipment.

2. The building facade control method based on network fine-tuning according to claim 1, characterized in that, The process of acquiring real-time environmental data of the target area and standardizing the real-time environmental data to obtain an environmental state vector specifically includes: collecting real-time environmental data of the target area at fixed intervals and adding a corresponding timestamp to each piece of real-time environmental data; removing all abnormal data from the real-time environmental data, retaining all normal data within a preset value range, and filling in all missing values ​​using linear interpolation to obtain a filtered data sequence; unifying all data units and magnitude units in the filtered data sequence, and normalizing all the data according to each data unit to obtain a corresponding data set; and sequentially concatenating each data set into a one-dimensional array according to predefined rules to obtain the environmental state vector for the current period.

3. The building facade control method based on network fine-tuning according to claim 1, characterized in that, The step of inputting the environmental state vector into the feature extraction network of the initial controller, and using the feature extraction network to perform multi-layer feature transformation on the environmental state vector to output a high-dimensional feature vector, specifically includes: inputting the environmental state vector into the feature extraction network of the initial controller, performing linear transformation and nonlinear activation on the environmental state vector to obtain a primary feature vector; ;in, Represents the primary feature vector. Denotes the first nonlinear activation function. This represents the first weight matrix. Represents the environment state vector. express The bias vector; the primary feature vector is subjected to higher-order cross processing and abstraction processing to output a high-dimensional feature vector: ;in, Represents a high-dimensional feature vector. This represents the second weight matrix. express The bias vector.

4. The building facade control method based on network fine-tuning according to claim 1, characterized in that, The step of inputting the high-dimensional feature vector into the policy network of the initial controller for prediction and outputting multiple action probabilities specifically includes: inputting the high-dimensional feature vector into the policy network of the initial controller, and performing a linear transformation on the high-dimensional feature vector through the fully connected layer of the policy network to map the high-dimensional feature vector to a multi-dimensional space corresponding to multiple angles, thereby obtaining the linear transformation result. ;in, Indicates the result of a linear transformation. This represents the third weight matrix. express The bias vector; by using a second nonlinear activation function to introduce nonlinearity into the linear transformation result, the action probability corresponding to each angle is obtained: ;in, Indicates the probability of an action. This represents the second nonlinear activation function.

5. The building facade control method based on network fine-tuning according to claim 1, characterized in that, The process of constructing a policy loss function using all the aforementioned preference pairs and optimizing the policy network using the policy loss function further includes: for each of the facade devices, after the action of the facade device is adjusted, comparing the real-time environmental data of the previous period and the current period to obtain environmental change data; determining the target user's thermal comfort item, visual comfort item, and the facade device's energy consumption item based on the environmental change data; weighting and fusing the thermal comfort item, visual comfort item, and energy consumption item to obtain a comprehensive reward; and comparing the comprehensive reward with the expected return to obtain the optimization result of the policy network.

6. A building facade control system based on network fine-tuning, characterized in that, The network-based fine-tuning building facade control system is used to implement the network-based fine-tuning building facade control method as described in any one of claims 1-5. The network-based fine-tuning building facade control system includes: a data acquisition module, used to acquire real-time environmental data of a target area, standardize the real-time environmental data to obtain an environmental state vector; a device control module, used to input the environmental state vector to an initial controller, output control commands for each facade device, control the corresponding facade device according to each control command, and receive feedback data from the target user; a strategy optimization module, used to construct multiple preference pairs within the current period based on the feedback data, optimize the initial controller based on all preference pairs to obtain a target controller; and a device re-control module, used to input the environmental state vector of the target area in the next period to the target controller to adjust all facade devices.

7. A controller, characterized in that, The controller includes: a memory, a processor, and a network-based fine-tuning building facade control program stored in the memory and executable on the processor. When the network-based fine-tuning building facade control program is executed by the processor, it implements the steps of the network-based fine-tuning building facade control method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a network-based fine-tuning building facade control program, which, when executed by a processor, implements the steps of the network-based fine-tuning building facade control method as described in any one of claims 1-5.

Images