Multi-zone electric actuator coordinated control method and system based on carrier communication

Abstract

This invention relates to the field of intelligent control technology and proposes a multi-zone electric actuator coordinated control method and system based on carrier communication. The method includes: collecting carrier communication datasets of electric actuators in multiple zones of a building; using a multi-scale convolutional neural network and a temporal attention weight algorithm for intelligent carrier signal recognition, outputting an actuator control command recognition matrix; based on the carrier communication dataset and the actuator control command recognition matrix, using a hierarchical graph neural network and an asynchronous multi-agent learning algorithm for actuator coordinated control, outputting a hierarchical collaborative control strategy vector; generating a comprehensive coordinated control strategy for the building actuators based on the command recognition matrix and the hierarchical collaborative control strategy vector; and sending control commands to the electric actuators in each zone via power line carrier signals based on the comprehensive coordinated control strategy for the building actuators. This invention achieves intelligent coordinated control among electric actuators in HVAC, water supply and drainage, and fire protection systems.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a multi-zone electric actuator coordinated control method and system based on carrier communication. Background Technology

[0002] Multi-zone electric actuator coordinated control technology based on power line carrier communication is a technology that enables unified scheduling and coordinated control of electric actuators in multiple zones of a building through power line carrier communication. It can utilize existing power line infrastructure for data transmission, avoiding additional wiring costs. Through a unified control strategy, it can achieve coordinated operation of electric actuators in multiple systems such as HVAC, water supply and drainage, and fire protection. It has advantages such as low cost, wide coverage, and convenient deployment, and is an important development direction for modern intelligent building automation control.

[0003] In existing technologies, multi-zone electric actuator control in buildings mainly adopts an independent decentralized control method. Electric actuators for the HVAC, water supply and drainage, and fire protection systems in each zone operate independently according to their local control logic, adjusting based on preset control programs and local sensor feedback. However, when a temperature anomaly occurs in a zone requiring adjustment of the damper actuator, the HVAC system actuators respond independently but cannot coordinate with adjacent zones, leading to energy waste due to excessive cooling or heating in adjacent zones. Furthermore, when the fire protection system needs to urgently activate the smoke exhaust fan actuator, actuators in other systems continue to operate according to their original logic, failing to adjust ventilation paths in a timely manner to meet fire protection needs, resulting in poor fire protection effectiveness and delays in personnel evacuation. Summary of the Invention

[0004] In view of this, the present invention proposes a multi-zone electric actuator coordinated control method and system based on carrier communication. This solves the problems of existing electric actuator coordinated control methods, where when a temperature anomaly occurs in a certain area and the damper actuator needs to be adjusted, the HVAC system actuator responds independently but cannot coordinate with adjacent areas, resulting in excessive cooling or heating in adjacent areas and wasted energy. When the fire protection system needs to urgently start the smoke exhaust fan actuator, the actuators of other systems still operate according to the original logic and fail to adjust the ventilation path in time to meet the fire protection needs, resulting in poor fire protection effect and delay in personnel evacuation.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides a multi-region electric actuator coordinated control method based on carrier communication, comprising the following steps: Collect carrier communication datasets of electric actuators in multiple areas of a building; Based on the aforementioned carrier communication dataset, an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction are used to intelligently identify carrier signals and output an actuator control command identification matrix. Based on the aforementioned carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy are used to perform actuator coordinated control and output a hierarchical cooperative control strategy vector. A comprehensive coordinated control strategy for building actuators is generated based on the actuator control command identification matrix and the hierarchical collaborative control strategy vector. Based on the building actuator integrated coordination control strategy, control commands are sent to the electric actuators in each area via power line carrier signals to control the electric actuators.

[0006] In some embodiments, the step of intelligently identifying carrier signals based on the carrier communication dataset using an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction, and outputting an actuator control command identification matrix, includes: Impedance-aware multi-scale convolutional neural networks are used to extract multi-scale features from carrier communication datasets to obtain multi-scale feature matrices of carrier signals. The time-series attention weighting algorithm for building load prediction is used to adaptively weight the multi-scale feature matrix of the carrier signal to obtain an adaptive weight feature matrix; A domain-adaptive meta-learning algorithm constrained by building physical features is used to perform cross-domain optimization processing on the adaptive weight feature matrix to obtain a domain-adaptive feature matrix, and an actuator control command recognition matrix is ​​generated based on the domain-adaptive feature matrix.

[0007] In some embodiments, the multi-scale feature extraction of the carrier communication dataset using an impedance-aware multi-scale convolutional neural network includes: Multiple convolution kernels of different scales are dynamically constructed based on the power line impedance values ​​in the carrier communication dataset to obtain an adaptive convolution kernel group. The adaptive convolution kernel group is used to perform parallel multi-scale convolution processing on the carrier communication dataset to extract carrier signal feature components in different frequency bands. The amplitude of the carrier signal feature components in different floors is compensated by combining the floor attenuation compensation mechanism, and the compensated carrier signal feature components are fused to obtain a multi-scale feature matrix of the carrier signal.

[0008] In some embodiments, the adaptive weight allocation of the multi-scale feature matrix of the carrier signal using the time-series attention weight algorithm for building load prediction includes: A long short-term memory network is used to perform time-series modeling of the multi-scale feature matrix of the carrier signal to predict the load change trend of each system in the building, thus obtaining a load prediction sequence. Based on the load prediction sequence, a three-level cascaded attention mechanism at the floor, area, and equipment levels is constructed, and the attention weight coefficients of each level are calculated. The weight attenuation parameters of the attention weight coefficients at each level are dynamically adjusted according to the operating status data of the building temperature control system, and the multi-scale feature matrix of the carrier signal is adaptively redistributed to obtain an adaptive weight feature matrix.

[0009] In some embodiments, the domain-adaptive meta-learning algorithm for building physical feature constraints performs cross-domain optimization processing on the adaptive weight feature matrix, including: A building physical feature constraint model is constructed, using building structure parameters and pipeline system layout parameters as physical constraints. A meta-learning network is used to quickly and adaptively train the adaptive weight feature matrix, learning the feature mapping relationship between different building environments under the guidance of the building physical feature constraint model. The network parameters of the meta-learning network are updated through gradient descent optimization to achieve rapid adaptation across building environments, and the adaptive weight feature matrix is ​​converted into a domain adaptive feature matrix.

[0010] In some embodiments, the step of using a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy to perform actuator coordinated control based on the carrier communication dataset and actuator control command recognition matrix, and outputting a hierarchical cooperative control strategy vector, includes: Based on the carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with building three-dimensional spatial constraints is used to construct an actuator coordination topology graph to obtain an actuator spatial topology relation matrix. Based on the actuator space topology matrix, an asynchronous multi-agent learning algorithm with system priority hierarchy is used to perform multi-system coordination decision-making, and a multi-agent coordination decision matrix is ​​obtained. Based on the multi-agent coordination decision matrix, energy consumption constraint optimization is performed to obtain a hierarchical collaborative control strategy vector.

[0011] In some embodiments, the construction of the actuator coordination topology graph using a hierarchical graph neural network with three-dimensional building spatial constraints includes: The building is modeled as a two-layer graph structure consisting of a vertical connection graph and a horizontal connection graph. The vertical connection graph represents the actuator connection relationships between floors, and the horizontal connection graph represents the actuator connection relationships within the same floor. Semantic constraints of the piping system are embedded into the graph node features, and semantic connections between nodes are constructed based on the piping layout of the HVAC, water supply and drainage, and fire protection systems. The graph edge weights are dynamically adjusted based on the real-time power line load status in the carrier communication dataset to obtain the actuator coordination topology graph. The graph structure features of the actuator coordination topology graph are extracted to obtain the actuator spatial topology relation matrix.

[0012] In some embodiments, the use of an asynchronous multi-agent learning algorithm with system priority hierarchy for multi-system coordination decision-making includes: A hierarchical intelligent agent architecture is constructed based on the safety priorities of the fire protection system, water supply and drainage system, and HVAC system, with the intelligent agent of the fire protection system having the highest decision priority. An asynchronous decision-making mechanism is adopted to enable intelligent agents at each level to learn in parallel according to the actuator space topology relation matrix, and the decision results of high-priority intelligent agents are used as constraints for low-priority intelligent agents. The policy network parameters of each intelligent agent are updated through reinforcement learning to obtain the multi-agent coordinated decision matrix.

[0013] In some embodiments, the energy consumption constraint optimization process based on the multi-agent coordinated decision matrix includes: A multi-objective optimization function is constructed with communication quality, coordination efficiency, and energy consumption as optimization objectives, and the multi-agent coordination decision matrix is ​​used as the optimization variable. An adaptive particle swarm optimization algorithm is used to solve the multi-objective optimization function, and the gradient of energy consumption change is used as the dominant factor for particle velocity update. Simultaneously, optimization is performed on three time scales: second-level communication scheduling, minute-level load prediction, and hour-level system coordination. The optimal coordination control parameters are obtained through iterative optimization. Based on the optimal coordination control parameters, the multi-agent coordination decision matrix is ​​converted into a hierarchical cooperative control strategy vector.

[0014] On the other hand, the present invention also provides a multi-zone electric actuator coordinated control system based on carrier communication, the system comprising: The data acquisition module is used to collect carrier communication datasets from electric actuators in multiple areas of a building. The signal recognition module is used to intelligently identify carrier signals based on the carrier communication dataset, using an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction, and outputs an actuator control command recognition matrix. The coordination control module is used to perform actuator coordination control based on the carrier communication dataset and actuator control command identification matrix, using a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy, and outputs a hierarchical cooperative control strategy vector. The strategy generation module is used to generate a comprehensive coordinated control strategy for building actuators based on the actuator control instruction identification matrix and the hierarchical collaborative control strategy vector. The control command sending module is used to send control commands to the electric actuators in each area via power line carrier signals based on the building actuator integrated coordination control strategy, so as to control the electric actuators.

[0015] The multi-zone electric actuator coordinated control method and system based on carrier communication of the present invention has the following advantages over the prior art: (1) This application uses impedance-sensing multi-scale convolutional neural network and building load prediction temporal attention weight algorithm to intelligently identify carrier communication dataset, accurately extract actuator control command information contained in carrier signal, and generate actuator control command identification matrix; then, it uses hierarchical graph neural network with building three-dimensional spatial constraints and asynchronous multi-agent learning algorithm with system priority hierarchical to establish cross-regional and cross-system coordination decision mechanism, and uses the parallel collaboration capability of multi-agent learning algorithm and spatial relationship modeling capability of graph neural network to realize intelligent coordination control among electric actuators of HVAC, water supply and drainage and fire protection systems, generate hierarchical collaborative control strategy vector, and realize efficient, intelligent and collaborative control of electric actuators in multiple areas of the building; (2) This application dynamically constructs an adaptive convolution kernel group based on the power line impedance value in the carrier communication dataset, and performs multi-scale feature extraction on the carrier signal by combining the floor attenuation compensation mechanism, accurately capturing the carrier signal feature components in different frequency bands, and obtaining the carrier signal multi-scale feature matrix; then, a long short-term memory network is used to predict the load change trend of each system in the building, and a three-level cascaded attention mechanism at the floor level, area level, and equipment level is constructed, and the weight attenuation parameter is dynamically adjusted according to the operating status of the building temperature control system to achieve adaptive weight allocation of the carrier signal features; further, through the training of the meta-learning network under the guidance of the building physical feature constraint model, the feature mapping relationship between different building environments is learned, and the rapid adaptation across building environments is achieved, thereby improving the accuracy and adaptability of carrier signal identification, and ensuring that the actuator control command can be accurately identified under different building environments and load conditions; (3) This application models the building as a two-layer graph structure of vertical and horizontal connection graphs, embeds the semantic constraints of the pipeline system into the graph node features, dynamically adjusts the graph edge weights in combination with the real-time power line load status in the carrier communication dataset, constructs the actuator coordination topology graph and extracts the spatial topology matrix; then, according to the safety priority of the fire protection system, water supply and drainage system and HVAC system, a hierarchical intelligent agent architecture is constructed, and an asynchronous decision mechanism is adopted to enable the agents at each level to learn in parallel, and the policy network parameters are updated through reinforcement learning to generate a multi-agent coordination decision matrix; further, a multi-objective optimization function with communication quality, coordination efficiency and energy consumption index as optimization objectives is constructed, and an adaptive particle swarm algorithm is used to optimize the three time scales of second-level communication scheduling, minute-level load prediction and hour-level system coordination, thereby realizing the spatial relationship modeling and hierarchical coordination decision between multi-system actuators, and ensuring efficient coordination control and energy consumption optimization under the premise of meeting the system safety priority. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a multi-region electric actuator coordinated control method based on carrier communication according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Existing building actuator control systems generally employ independent control modes, lacking effective coordination mechanisms between systems. When multiple systems within a building, such as HVAC, water supply and drainage, and fire protection, operate simultaneously, conflicts and resource contention easily arise because each system's actuators respond independently to their own control commands. For example, in an emergency, while the fire protection system activates its smoke exhaust fans, the HVAC system continues to operate its air supply equipment according to its predetermined program, causing airflow interference and affecting the effectiveness of fire smoke extraction. Furthermore, traditional control methods cannot dynamically optimize based on the building's real-time load status and spatial layout characteristics, resulting in high overall energy consumption and low coordination efficiency.

[0020] Based on this, in some embodiments, please refer to Figure 1 This invention provides a multi-region electric actuator coordinated control method based on carrier communication, comprising the following steps: S1 collects carrier communication datasets of electric actuators in multiple areas of the building.

[0021] It is understandable that step S1 involves acquiring the operating status data of electric actuators in various areas of the building through power line carrier communication technology. Power line carrier communication utilizes existing power lines as the data transmission medium, achieving full building coverage without the need for additional wiring. The collected dataset includes electrical parameters such as actuator voltage, current, power factor, and operating frequency, as well as management information such as the actuator's system type, installation location, and control command history, providing fundamental data support for coordinated control.

[0022] S2, based on the carrier communication dataset, an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction are used to intelligently identify carrier signals and output an actuator control command identification matrix.

[0023] It is understandable that step S2 addresses the problem of insufficient signal recognition accuracy in carrier communication environments. The impedance-aware multi-scale convolutional neural network can adapt to changes in signal characteristics under different line impedance conditions, extracting multi-dimensional features of the carrier signal through parallel processing of convolutional kernels at multiple scales. The temporal attention weight algorithm dynamically adjusts the attention given to the carrier signal at different times based on the periodic changes in building load, effectively improving the command recognition accuracy in complex electromagnetic environments.

[0024] S3, based on the carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy are used to perform actuator coordinated control and output a hierarchical cooperative control strategy vector.

[0025] It is understandable that step S3 is the core component for achieving cross-system coordinated control. The hierarchical graph neural network models the building as a vertical and horizontal two-layer graph structure, fully considering the three-dimensional spatial distribution characteristics of actuators and the physical connections of the piping system. The asynchronous multi-agent learning algorithm constructs a hierarchical decision-making architecture according to the safety priorities of fire protection, water supply and drainage, and HVAC systems. This ensures that the control requirements of high-priority systems effectively constrain low-priority systems, avoiding conflicts between systems and achieving orderly coordinated control.

[0026] S4. Generate a building actuator integrated coordination control strategy based on the actuator control command identification matrix and the hierarchical collaborative control strategy vector.

[0027] It is understandable that step S4 integrates the identification results and coordination strategies obtained in steps S1-S3 to generate a unified control strategy for the entire building. This strategy includes not only the specific action instructions of each actuator, but also multi-dimensional control parameters such as execution timing, priority level, and energy consumption constraints, ensuring that the actuators of each system can operate in a coordinated manner within a unified framework.

[0028] S5, based on the building actuator integrated coordination control strategy, sends control commands to the electric actuators in each area via power line carrier signals to control the electric actuators.

[0029] It is understandable that step S5 converts the coordinated control strategy into a standard carrier communication protocol format and sends control commands to actuators in each area via power lines. The carrier signal carries complete control information such as actuator identification, action type, and parameter settings. After receiving the command, the actuator executes the corresponding action according to the coordinated strategy, thus realizing distributed coordinated control based on carrier communication.

[0030] Through steps S1-S5, this embodiment effectively solves the cross-system coordination failure problem existing in traditional building electric actuator control systems. This embodiment fully utilizes the convenience of carrier communication and the decision-making advantages of artificial intelligence algorithms to achieve unified coordinated control of actuators in multiple regions and systems, significantly improving the overall coordination and energy efficiency of building equipment operation.

[0031] In some embodiments, based on the above embodiments, step S1 includes: The carrier signal data and device response data of electric actuators in multiple areas of the building during the carrier communication process are collected and organized according to the area and device to obtain the raw carrier communication data.

[0032] For example, carrier communication information of electric actuators in various areas of a building is acquired through a distributed data acquisition device. The carrier signal data includes voltage amplitude. Current phase carrier frequency Signal modulation depth ,in, , Indicates the amplitude of the carrier reference voltage. Indicates the frequency of the modulating signal. Represents a time variable. Indicates the signal modulation depth. Indicates the carrier frequency.

[0033] Equipment response data includes actuator motion feedback. Operating status identifier Power consumption monitoring value ,in, Indicates the area code. This indicates the device number. The raw data acquisition uses a timestamp synchronization mechanism to ensure the timing consistency of data from different areas. The acquisition frequency is set to 1000Hz to meet the Nyquist sampling requirements of the carrier signal.

[0034] The original carrier communication data is time-aligned, noise-filtered, segmented, and encapsulated in a unified format to obtain a carrier communication dataset.

[0035] For example, the time alignment process employs a delay estimation algorithm based on the cross-correlation function to calculate the time offset between data in each region. ,in, , Indicates the first The time offset of each region relative to the reference region Represents a delay variable. For the reference region signal sequence, The signal sequence represents the region to be aligned. Indicates the signal sequence index.

[0036] Noise filtering employs an adaptive Wiener filter, and the filter coefficient update formula is as follows: ,in, For the first The filter weight vector for the next iteration; Indicates the first The filter weight vector for the next iteration; Indicates the step size parameter; Indicates the first The input signal vector for the next iteration; Indicates the first The conjugate of the next iteration error signal; This indicates the iteration number index. After filtering, the signal noise power spectral density is reduced to less than 5% of the original signal.

[0037] Specifically, step size parameter The range of values ​​is In one specific embodiment, in an office building carrier communication environment, when the signal-to-noise ratio is 15dB, setting μ=0.01 can converge the filter to the optimal state within 10 iterations, reducing the signal noise power spectral density to 3% of the original signal.

[0038] Data segmentation and processing are performed according to the carrier communication frame structure, with the length of each data segment set to [length to be specified]. , Indicates the length of the data segment; Indicates the total number of sampling points; Indicates the number of carrier communication frames; This is a rounding down function; the unified format encapsulation adopts the IEEE 802.15.4 standard protocol format, and the packet header includes the source address. Target address Data length Verification code ,in, Indicates the source device address of the data packet. Indicates the destination device address of the data packet. Indicates the length of the data packet payload. This represents the cyclic redundancy check value, ensuring the reliability and integrity of data transmission. After the preprocessing step, the signal-to-noise ratio of the original carrier communication data is improved by approximately 15-20 dB, and the data integrity rate reaches over 99.8%, providing a high-quality input data foundation for carrier signal identification and coordinated control algorithms.

[0039] The data acquisition and preprocessing stage of this embodiment solves the problem of multi-source heterogeneous data fusion in a carrier communication environment. Through precise timing alignment and effective noise suppression, it ensures that artificial intelligence algorithms can be trained and inferred based on a unified and clean dataset, laying a solid data foundation for achieving cross-system coordinated control.

[0040] In some embodiments, based on the above embodiments, step S2 includes: Impedance-aware multi-scale convolutional neural networks are used to extract multi-scale features from carrier communication datasets, resulting in a multi-scale feature matrix of carrier signals.

[0041] For example, the impedance-sensing multi-scale convolutional neural network employs an adaptive convolutional kernel design, dynamically constructing convolutional kernel parameters at multiple scales based on the electric field line impedance value. The formula for calculating the adaptive convolutional kernel size is as follows: ; in, Indicates the scaling factor; Indicates the base core size offset; Indicates the first The kernel size corresponding to each line; For the first The power line impedance value of the line; This is the floor function.

[0042] Specifically, size scaling factor The range of values ​​is Basic core size offset The range of values ​​is In one specific embodiment, in a high-rise residential building, the power line impedance range is set to 50Ω-200Ω, α=1.2, β=3, and the convolution kernel size range is 7×7 to 11×11, effectively extracting carrier signal features of different frequency bands.

[0043] The network structure includes three parallel convolutional branches, each employing... , , The convolution kernel performs multi-scale feature extraction on the carrier signal. The calculation formula for the convolution operation is as follows: ; in, Indicates the first Each scale branch at position The convolution output, Corresponding to , , The convolution kernel; Indicates the position of the input carrier signal. The value, Indicates the first Each scale in position The convolution weights, Indicates the first Bias parameters for each scale Indicates the first Each scale of convolution kernel size.

[0044] The floor attenuation compensation mechanism compensates for the amplitude of the carrier signal at different floors based on the signal propagation distance. The compensation formula is as follows: ; in, Indicates the first Floor in location The compensated eigenvalues; Indicates the first Each scale branch at position The convolution output; Indicates the compensation gain coefficient; Indicates the first The distance from the floor to the reference point; This represents the attenuation constant. Finally, the features at each scale are weighted and fused through a feature fusion layer to obtain the multi-scale feature matrix of the carrier signal. , Indicates the number of feature dimensions. Indicates the length of the time series.

[0045] The time-series attention weighting algorithm for building load prediction is used to adaptively assign weights to the multi-scale feature matrix of the carrier signal to obtain an adaptive weight feature matrix.

[0046] As can be understood, this embodiment first uses a Long Short-Term Memory (LSTM) network to perform time-series modeling of the multi-scale feature matrix of the carrier signal. The hidden state update formula of the LSM network is: ,in, Indicates the first The hidden state vector at each time step, For the first The hidden state vector at each time step; Indicates the first The forget gate output at each time step; Indicates the first Input gate output at each time step; Indicates the first Candidate state vectors at each time step; This represents element-wise multiplication. The formula for calculating the load prediction sequence is: , Indicates the first Load forecast values ​​for each time step; Represents the prediction layer weight matrix; Indicates the first The hidden state vector at each time step; This represents the prediction layer bias vector. Indicates the prediction time step. This represents the Sigmoid activation function.

[0047] A three-layer cascaded attention mechanism is constructed based on the load prediction sequence. The formula for calculating the attention weights of each layer is as follows: ; in, Indicates the first Attention weight of each floor Indicates the first The attentional energy of each floor This represents the floor-level attention weight vector. This represents the floor-level weight matrix. This represents the floor-level bias vector. Indicates the first Load forecast values ​​for each time step Indicates the first The hidden state vector at each time step, It is the hyperbolic tangent activation function; For the first Attention weighting for each floor; Indicates the total number of floors; The function is exponential; the attention weights at the regional and device levels are calculated using a similar method, and the final weight decay parameter is dynamically adjusted based on the operating status of the temperature control system. The adjustment formula is: ; in, Indicates the first Weight decay parameters for time steps; Indicates the initial attenuation parameter. This represents the attenuation adjustment coefficient. Indicates the first The operating status value of the temperature control system at the time step.

[0048] Specifically, initial attenuation parameters The range of values ​​is Attenuation adjustment coefficient The range of values ​​is In one specific embodiment, for the HVAC system of a commercial complex, a configuration is provided. , When the operating status value of the temperature control system is 0.6, the weight decay parameter is adjusted to 0.74 to achieve an adaptive response to load changes.

[0049] The formula for calculating the adaptive weight feature matrix is: ; in, Represents the adaptive weight feature matrix. Indicates the first Number of floors / areas Indicates the first Number of devices in the area Indicates the first Attention weight of each floor Indicates the first regional attention weights Indicates the first Device attention weight, Indicates the first Weight decay parameters for time steps Indicates the first Layer Region 1 The equipment has multi-scale characteristics. This indicates the total number of floors.

[0050] A domain-adaptive meta-learning algorithm constrained by building physical features is used to perform cross-domain optimization processing on the adaptive weight feature matrix to obtain a domain-adaptive feature matrix, and an actuator control command recognition matrix is ​​generated based on the domain-adaptive feature matrix.

[0051] For example, the building physical characteristic constraint model will include building structural parameters Piping system layout parameters As a constraint, among which... Indicates floor height. Indicates the floor area. Indicates the volume of the room. Indicates the pipe diameter. Indicates the length of the pipe. Indicates the branch angle.

[0052] The meta-learning network adopts the MAML (Model-Agnostic Meta-Learning) framework, and the inner loop gradient update formula is: ; in, Indicates the first Update parameters for each task. Indicates the initial network parameters. This represents the learning rate of the inner loop. Indicates the first The training loss function for each task. This indicates the initial network parameters. The gradient operator for partial derivatives; the formula for updating the outer loop parameters is: ; in, This represents the outer loop learning rate. Indicates the first Test loss function for each task.

[0053] The physical constraint loss function is defined as follows: , Represents the physical constraint weight matrix. Represents the adaptive weight feature matrix. Represents the physical constraint vector, and the domain adaptive feature matrix. The calculation formula is: ; in, The meta-learning network mapping function is represented. Represents the adaptive weight feature matrix. The optimized meta-learning network parameters are represented by the actuator control command recognition matrix, which is obtained through mapping via a fully connected layer. , This represents the actuator control command identification matrix. Indicates the instruction recognition weight matrix. This represents the instruction identification bias vector, where each row of the matrix corresponds to the probability distribution of control instructions for an executor.

[0054] In some embodiments, based on the above embodiments, the step of using an impedance-aware multi-scale convolutional neural network to extract multi-scale features from the carrier communication dataset includes: Multiple convolution kernels of different scales are dynamically constructed based on the power line impedance values ​​in the carrier communication dataset to obtain an adaptive convolution kernel group. The adaptive convolution kernel group is used to perform parallel multi-scale convolution processing on the carrier communication dataset to extract carrier signal feature components in different frequency bands. The amplitude of the carrier signal feature components in different floors is compensated by combining the floor attenuation compensation mechanism, and the compensated carrier signal feature components are fused to obtain a multi-scale feature matrix of the carrier signal.

[0055] For example, this embodiment solves the problem of inconsistent feature extraction results caused by fixed convolution kernel parameters under different electrical environments through an adaptive power line impedance adjustment mechanism. The adaptive convolution kernel group construction process dynamically determines the number of convolution kernels based on the statistical characteristics of impedance distribution. , Indicates the number of convolution kernels. Indicates the maximum line impedance. This represents the minimum line impedance.

[0056] Parallel multi-scale convolution processing uses depthwise separable convolution structures to reduce computational complexity. The depthwise convolution operation is... , Indicates the first The passage is in the location The depthwise convolution output; Indicates the first The kernel size is defined by a scale, i.e., the length and width of the kernel, assuming it to be a square kernel; This represents the position index within the convolution kernel along the vertical (i.e., height) direction, with a value range of [value missing]. arrive ; This represents the position index in the horizontal direction (i.e., the width direction) within the convolution kernel, with a value range of... arrive ; Indicates the first The passage is in the location The input signal, Indicates the first The passage is in the location The depthwise convolution weights, Indicates the first Channel depth convolution bias.

[0057] The floor attenuation compensation mechanism is based on the propagation attenuation model of carrier signals in buildings, and the attenuation coefficient is calculated using the following formula: , Indicates the first The attenuation coefficient of the layer, Indicates the reference attenuation coefficient. Indicates the first The distance from the floor to the reference point Indicates the reference distance. The path loss index is represented by the compensated feature fusion, which employs an attention-weighted mechanism. The fusion weight is calculated using the following formula: , Indicates the first Scale feature fusion weights, Represents the fusion attention parameter vector, It is an exponential function. For the first The feature matrix after floor attenuation compensation at each scale For the first The feature matrix after floor attenuation compensation at each scale Represents the total number of scales, multi-scale feature matrix The final representation is .

[0058] In some embodiments, based on the above embodiments, the adaptive weight allocation of the multi-scale feature matrix of the carrier signal using the time-series attention weight algorithm for building load prediction includes: A long short-term memory network is used to perform time-series modeling of the multi-scale feature matrix of the carrier signal to predict the load change trend of each system in the building, thus obtaining a load prediction sequence. Based on the load prediction sequence, a three-level cascaded attention mechanism at the floor, area, and equipment levels is constructed, and the attention weight coefficients of each level are calculated. The weight attenuation parameters of the attention weight coefficients at each level are dynamically adjusted according to the operating status data of the building temperature control system, and the multi-scale feature matrix of the carrier signal is adaptively redistributed to obtain an adaptive weight feature matrix.

[0059] It is understandable that this embodiment solves the problem that traditional fixed weight allocation cannot adapt to dynamic changes in building load. In the temporal modeling process of Long Short-Term Memory networks, the cell state update formula is: ,in, Indicates the first Cell state at time step Indicates the first Cell state at time step Indicates the first Candidate cell states at time steps Indicates the first The output of the forget gate at each time step Indicates the first Input gate output at each time step, This represents element-wise multiplication, and the output gate calculation formula is: , Indicates the output gate status. This represents the output gate weight matrix. For the first The hidden state vector at each time step, Indicates the first Input features at time steps This represents the output gate bias vector.

[0060] The load forecast sequence includes HVAC loads. Water supply and drainage load Fire protection load Three subsequences. The three-layer cascaded attention mechanism adopts a hierarchical calculation strategy, and the device-level attention weight calculation formula is:

[0061] ; in, For the first Device-level attention weights for each device It is an exponential function. The hyperbolic tangent activation function is used. Represents the device-level attention parameter vector. Represents the device-level transformation matrix; Indicates the first The multi-scale feature submatrix of the carrier signal corresponding to each device is derived from the complete multi-scale feature matrix. Extracted from the first Each device-related feature is used to calculate the attention weights for that device; Indicates the first The multi-scale feature sub-matrix of the carrier signal corresponding to each device; This represents the device-level bias vector. Indicates the total number of devices. This indicates the device index.

[0062] The dynamic adjustment of the weight decay parameter takes into account the operating mode of the temperature control system, and the adjustment strategy is as follows: , For the first Adjustment strategy at each time step Indicates the initial attenuation parameter. Indicates the periodic adjustment range. Indicates the load cycle frequency. This indicates a phase shift. The adaptive weight redistribution process achieves the global optimal allocation through multi-level weight products.

[0063] Specifically, the periodic adjustment amplitude The range of values ​​is Load cycle frequency The range of values ​​is Phase shift The range of values ​​is In one specific embodiment, for a hospital building, a configuration is provided. , (i.e., a 30-minute cycle) The simulation of hospital daytime load fluctuations showed that the weight decay parameter varied periodically between 0.72 and 1.08.

[0064] In some embodiments, based on the above embodiments, the domain adaptive meta-learning algorithm for building physical feature constraints performs cross-domain optimization processing on the adaptive weight feature matrix, including: A building physical feature constraint model is constructed, using building structure parameters and pipeline system layout parameters as physical constraints. A meta-learning network is used to quickly and adaptively train the adaptive weight feature matrix, learning the feature mapping relationship between different building environments under the guidance of the building physical feature constraint model. The network parameters of the meta-learning network are updated through gradient descent optimization to achieve rapid adaptation across building environments, and the adaptive weight feature matrix is ​​converted into a domain adaptive feature matrix.

[0065] For example, this embodiment addresses the problem of insufficient model generalization ability caused by differences in feature distribution under different building environments. The physical feature constraint model introduces constraints through the Lagrange multiplier method, and the constraint optimization objective function is:

[0066] ; in, To constrain the optimization objective function, Indicates mission loss. Indicates the first A constrained Lagrange multiplier, Indicates the first A physical constraint function, Indicates the number of constraints. Represents the adaptive weight feature matrix. For building structural parameters, For the pipeline system layout parameters. The rapid adaptive training of the meta-learning network employs second-order gradient optimization, and the Hessian matrix is ​​approximated as follows: , Represents the Hessian matrix; Represents a set of tasks. Indicates the first The loss function for each task is the network parameters. The function; the second-order update formula is .

[0067] Cross-domain feature mapping relationship learning enhances the feature consistency between different building environments through a contrastive learning mechanism. The contrastive loss is calculated as follows: , This represents the similarity calculation function. Indicates positive sample features. Indicates negative sample features. The temperature parameter is represented by the domain adaptive feature matrix, which is further optimized by a domain discriminator. The discriminator loss is... , This represents a domain discriminator. Represents the source domain. Indicates the target domain; Indicates originating from the source domain. The domain-adaptive feature matrix, i.e., the feature representation generated in the source domain building environment; Indicates originating from the target domain A domain-adaptive feature matrix is ​​generated to train a domain discriminator to distinguish feature distributions in different building environments. Finally, cross-building environment feature distribution alignment is achieved through adversarial training.

[0068] The carrier signal identification step S2 effectively addresses the technical problems of low signal identification accuracy and insufficient cross-environment generalization ability in carrier communication environments through three key steps: multi-scale feature extraction, temporal attention weight allocation, and domain adaptive optimization. The impedance-aware multi-scale convolutional network can adapt to changes in signal characteristics across different electrical environments, the temporal attention mechanism enables adaptive response to dynamic changes in building load, and the meta-learning algorithm ensures the algorithm's rapid adaptability to different building environments, laying an accurate and reliable signal identification foundation for the generation of coordinated control strategies.

[0069] In some embodiments, based on the above embodiments, step S3 includes: Based on the carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with three-dimensional spatial constraints of the building is used to construct an actuator coordination topology graph, thereby obtaining an actuator spatial topology relation matrix.

[0070] It is understandable that the hierarchical graph neural network adopts the GCN architecture, representing the building actuator network as a two-layer graph structure. The formula for calculating the features of vertically connected graph nodes is: ; in, Indicates the first Layer The vertical connection feature vector of each node. Indicates the first Layer The vertical connection feature vector of each node. Indicates the first Layer vertical connection weight matrix, Represents a node In the set of neighboring nodes in the vertical direction, Represents aggregate functions, Indicates the node index. Indicates the neighbor node index. This represents the layer index of a graph neural network; horizontally connected graphs use the same computational framework but with independent parameters. Gather with neighbors , and The first Layer horizontal connection weight matrix and nodes The set of neighboring nodes in the horizontal direction.

[0071] Semantic constraints in the pipeline system are implemented through node feature embedding, where the initial feature vector of a node is defined as follows: , Indicates the first Each node has a unique thermal code for its system type, corresponding to fire protection, water supply and drainage, and HVAC systems, respectively. Represents three-dimensional position coordinates; Represents the feature vector of pipeline attributes. This indicates the dimension of the pipeline attributes.

[0072] The graph edge weights are dynamically adjusted based on carrier communication data, and the weight update formula is as follows: , Indicates the first Time Node and Edge weights between them Represents a node and The basic edge weights between them Indicates the load difference sensitivity coefficient. Indicates the first Time Node Load status, Indicates the first Time Node Load state, actuator space topology matrix The final layer graph node features are obtained by concatenating them. , The graph neural network represents the first... The vertical connection feature vector of the final layer. The graph neural network represents the first... The horizontal connection feature vectors of the layer This represents the total number of layers in the graph neural network. Represents the total number of nodes. This indicates the dimension of a single-layer feature.

[0073] Specifically, load difference sensitivity coefficient The range of values ​​is In one specific embodiment, a graph neural network model is performed on the office building, and a configuration is established. When the load difference between adjacent nodes is 0.3, the edge weight adjustment coefficient is 0.67, which effectively reflects the impact of load status on the actuator connection strength.

[0074] Based on the actuator space topology matrix, an asynchronous multi-agent learning algorithm with system priority hierarchy is used to perform multi-system coordination decision-making, resulting in a multi-agent coordination decision matrix.

[0075] For example, this embodiment constructs a three-layer hierarchical intelligent agent architecture, with the fire protection system intelligent agent located at the highest priority level. Its policy network adopts a DQN structure, and the action value function update formula is: ; in, The action value function of the fire protection system agent is specifically the expected cumulative reward value for performing a specific action in a given state. Represents the state vector of the fire protection system. Represents the action vector of the fire protection system. This represents the set of network parameters for the fire protection intelligent agent. Indicates network layer index, Indicates the first Layer weight matrix, Indicates the first Layer bias vector, This is the actuator space topology matrix.

[0076] The intelligent agents for water supply and drainage systems and HVAC systems adopt a decision-evaluation architecture, and the policy network outputs a probability distribution. , Indicates the operation of the water supply and drainage system. Indicates the status of the water supply and drainage system. Indicates the constraints of the fire protection system. This represents the network parameters of the water supply and drainage agent strategy. This represents the policy network mapping function.

[0077] The asynchronous decision-making mechanism employs a time-slicing strategy, and the relationship between the decision-making time steps of each agent is as follows: , Indicates the decision-making cycle of the fire protection system. Indicates the decision-making cycle of the water supply and drainage system. This indicates the decision-making cycle of the HVAC system, ensuring that high-priority systems can promptly constrain the decision-making behavior of low-priority systems.

[0078] The policy network parameters are updated through empirical replay and target network mechanisms, and the loss function is: , The loss function represents the agent policy network. Represents the experience pool. Let the experience sample be a quadruple, representing the state, action, immediate reward, and next state, respectively. Indicates the target value. Indicates the discount factor. Represents the target network parameters, multi-agent coordination decision matrix. It is obtained by splicing together the policy outputs of each agent. , These represent the number of intelligent agents in the three systems: fire protection, water supply and drainage, and heating, ventilation, and air conditioning. , , These represent the policy output vectors of the intelligent agents for fire protection, water supply and drainage, and HVAC systems, respectively. This represents the action space dimension of a single intelligent agent.

[0079] Based on the multi-agent coordination decision matrix, energy consumption constraint optimization is performed to obtain a hierarchical collaborative control strategy vector.

[0080] It is understood that this embodiment uses an adaptive particle swarm optimization algorithm to solve a multi-objective optimization problem, and the objective function is designed as follows: ,in, For multi-agent coordination decision matrix, The function representing the communication quality evaluation is calculated using the following formula: , Represents a set of communication links. Indicates the first A multi-agent coordination decision matrix for communication nodes. Indicates the first A multi-agent coordination decision matrix for communication nodes. Represents the numerically stable term; The coordination efficiency function is defined as follows: , Represents a set of coordinated tasks; Indicates the first The completion status of the coordination task is a Boolean function, where the value of the first coordination task is equal to the value of the second coordination task. Returns a true value (1) if the task is successfully completed, otherwise returns a false value (0); Indicates the first The importance weight of each task Indicates an indicator function, Indicates the task index; The total energy consumption function is expressed as follows: , Indicates the first The basic power consumption of each actuator Indicates the energy consumption coefficient of the action intensity. Indicates the number of actuators. These represent the weighting coefficients of the three objectives.

[0081] Specifically, The range of values ​​for are respectively , , In one specific embodiment, for a data center building, a configuration is provided. Corresponding communication quality weights, Corresponding coordination efficiency weight, By assigning corresponding energy consumption weights, a balanced optimization of communication reliability, coordination efficiency, and energy-saving effect can be achieved.

[0082] Specifically, the energy consumption coefficient of motion intensity The range of values ​​is In one specific embodiment, for the energy consumption calculation of the hotel building, a system is set... When the actuator's action intensity is 0.8, the increase in power consumption relative to the base is 3.2%, accurately reflecting the impact of action intensity on system energy consumption.

[0083] The particle velocity update formula is as follows ,in, Indicates the first The particle in the first The velocity vector of the next iteration Indicates the first The particle in the first The velocity vector of the next iteration Indicates inertia weight, These represent the first, second, and third acceleration coefficients, respectively. express Uniformly distributed random numbers Indicates the first The historical best position of each particle Indicates the globally optimal position. Indicates the first The particle in the first The position of the next iteration. Multi-agent coordination decision matrix Find the gradient operator for partial derivatives.

[0084] Specifically, inertia weight The range of values ​​is , The range of values ​​for are respectively , , In one specific embodiment, particle swarm optimization is performed for the shopping mall, and the following settings are configured: To maintain search momentum, Corresponding to individual learning factors, Corresponding to social learning factors, The corresponding energy consumption guidance factor converges to the global optimal solution after 50 iterations, reducing system energy consumption by 15%.

[0085] Multi-timescale optimization is achieved through hierarchical time windows, with second-level communication scheduling windows. Minute-level load forecast window Hourly system coordination window The formula for calculating the adaptive weights at each time scale is: ; in, Indicates the first Weights for each time scale Indicates the first The center time of each scale Indicates the first The time constants at each scale. The optimal coordinated control parameters are obtained through Pareto front selection. , This represents the optimal control parameters. Represents the control parameter vector; This represents a multi-objective optimization function that evaluates the overall performance of control parameters. Represents the feasible parameter space. The transformation function from the decision matrix to the control parameters is represented by the hierarchical collaborative control strategy vector, which is ultimately expressed as... , This indicates the dimension of the control strategy.

[0086] Specifically, the time constants of the first, second, and third scales. , , The range of values ​​for are respectively Corresponding to seconds, Corresponding to minute level, Corresponding to the hourly level. In one specific embodiment, multi-time-scale optimization is performed on the intelligent building, and settings are configured. , , This achieves an organic combination of second-level communication scheduling, minute-level load prediction, and hour-level system coordination.

[0087] In some embodiments, based on the above embodiments, the construction of the actuator coordination topology graph using a hierarchical graph neural network with three-dimensional building spatial constraints includes: The building is modeled as a two-layer graph structure consisting of a vertical connection graph and a horizontal connection graph. The vertical connection graph represents the actuator connection relationships between floors, and the horizontal connection graph represents the actuator connection relationships within the same floor. Semantic constraints of the piping system are embedded into the graph node features, and semantic connections between nodes are constructed based on the piping layout of the HVAC, water supply and drainage, and fire protection systems. The graph edge weights are dynamically adjusted based on the real-time power line load status in the carrier communication dataset to obtain the actuator coordination topology graph. The graph structure features of the actuator coordination topology graph are extracted to obtain the actuator spatial topology relation matrix.

[0088] For example, this embodiment solves the problem that traditional single-layer graph networks cannot simultaneously represent the horizontal and vertical connectivity relationships of buildings by using a two-layer graph structure modeling. Vertical connectivity graph adjacency matrix. The construction rules are as follows:

[0089] if and ; in, Represents a node The floor where it is located Represents a node The floor where it is located Represents a node System type Represents a node System type: Horizontal connectivity graph adjacency matrix The construction rules are as follows: if ; in, Represents a node Two-dimensional plane coordinates, The Gaussian kernel bandwidth parameter represents the spatial connectivity. The semantic connectivity of the pipeline system employs a learnable attention weighting mechanism; the attention score is calculated as follows:

[0090] ; in, This represents the semantic connectivity learning matrix. Represents a node All neighboring nodes, , They represent the first , The pipeline attribute feature vector of each node, Represents a node The Pipeline attribute feature vectors of neighboring nodes, semantically enhanced node features The updated formula is The real-time edge weight adjustment mechanism dynamically corrects the connection strength based on the time-varying characteristics of the power line load state. The load state normalization calculation formula is as follows:

[0091] ; in, Represents the normalized stable term. Indicates the index of all nodes; Indicates the first Time Node Load status, Indicates the first Time Node The load state; the dynamic edge weight matrix is , Represents static weights. This indicates that the intensity parameters are dynamically adjusted, and the graph structure features are extracted through a cascade of graph convolutional layers.

[0092] In some embodiments, based on the above embodiments, the use of an asynchronous multi-agent learning algorithm with system priority hierarchy for multi-system coordination decision-making includes: A hierarchical intelligent agent architecture is constructed based on the safety priorities of the fire protection system, water supply and drainage system, and HVAC system, with the fire protection system intelligent agent having the highest decision priority. An asynchronous decision-making mechanism is adopted to enable each level of intelligent agent to learn in parallel according to the actuator space topology relation matrix, and the decision results of high-priority intelligent agents are used as constraints for low-priority intelligent agents. The policy network parameters of each intelligent agent are updated through reinforcement learning to obtain the multi-agent coordinated decision matrix.

[0093] It is understandable that this embodiment solves the problem of decision-making conflicts among agents in traditional multi-agent systems by using a hierarchical security priority system. The hierarchical agent architecture adopts a hard constraint mechanism, and the action space constraints of the water supply and drainage system agents are as follows:

[0094] ; in, This indicates the available operational space for the water supply and drainage system under fire safety constraints. This represents the complete operating space of the water supply and drainage system. Represents the constraint function vector. This represents the constraint vector imposed by the fire protection system. This represents the action vector of the water supply and drainage system agent; the HVAC system agent is simultaneously constrained by both the fire protection and water supply and drainage systems, with the following constraints: , Indicates the constraints of the water supply and drainage system. This represents the constraint function of water supply and drainage on HVAC. This indicates that the HVAC system is activated.

[0095] The asynchronous decision-making mechanism uses event-triggered policy updates, while the fire protection system uses synchronous updates. , This indicates that the fire protection system intelligent agent is in the first... The network parameter vector after the next iteration This indicates that the fire protection system intelligent agent is in the first... The network parameter vector after the next iteration Indicates the learning rate of the fire protection system. This indicates the network parameters of the fire protection system. gradient operator, This represents the objective function of the fire protection system. The water supply, drainage, and HVAC systems employ an asynchronous update mechanism, with the update condition being... or , Indicates the threshold for parameter variation. Indicates a forced asynchronous update cycle. This indicates the number of steps since the last update.

[0096] Policy network training employs a PER (Prioritized Experience Replay) mechanism, where experience priority is calculated as follows: , Indicates the first Prioritizing experience Indicates timing difference error. Indicates the priority offset. This represents the priority index parameter, with importance sampling weights being... , Indicates the size of the experience pool. The importance sampling index is represented by the multi-agent coordinated decision matrix, which is obtained by weighted fusion of the policies of each agent.

[0097] In some embodiments, based on the above embodiments, the energy consumption constraint optimization process based on the multi-agent coordinated decision matrix includes: A multi-objective optimization function is constructed with communication quality, coordination efficiency, and energy consumption as optimization objectives, and the multi-agent coordination decision matrix is ​​used as the optimization variable. An adaptive particle swarm optimization algorithm is used to solve the multi-objective optimization function, and the gradient of energy consumption change is used as the dominant factor for particle velocity update. Simultaneously, optimization is performed on three time scales: second-level communication scheduling, minute-level load prediction, and hour-level system coordination. The optimal coordination control parameters are obtained through iterative optimization. Based on the optimal coordination control parameters, the multi-agent coordination decision matrix is ​​converted into a hierarchical cooperative control strategy vector.

[0098] It is understood that the optimization algorithm in this embodiment solves the problem that single-objective optimization cannot balance multiple aspects of system performance through multi-objective Pareto optimal search. The multi-objective optimization function adopts the weighted Chebyshev method, and the objective function is reconstructed as follows: , Represents the weight vector. Indicates the first One objective function, Indicates the first The ideal point for each goal.

[0099] The gradient of energy consumption change is calculated using the finite difference method, and the gradient estimate is: , Indicates the finite difference step size. Indicates the first A unit vector of dimension, the energy-oriented term for particle velocity updates is: , This represents the energy consumption optimization intensity coefficient.

[0100] Multi-timescale optimization employs a decomposition and coordination strategy, including second-level subproblems, minute-level subproblems, and hour-level subproblems. The consistency of solutions at each timescale is guaranteed through Lagrange dual decomposition.

[0101] The optimal coordinated control parameters are determined through a multi-objective evolutionary search, and the Pareto front update rule is as follows: , For the first The Pareto front solution set after the next iteration. Indicates the first The Pareto front solution set after the next iteration. This represents the newly generated solution set. Represents a candidate solution vector. Representing Pareto dominance, the hierarchical collaborative control strategy vector is obtained through a linear combination of the optimal solutions. , Indicates the final Pareto front. Indicates the first The weights of each Pareto solution Indicates the first One optimal control parameter This represents the Pareto solution index.

[0102] In some embodiments, based on the above embodiments, step S4 includes: The actuator control command identification matrix and the hierarchical collaborative control strategy vector are matched and fused to obtain a comprehensive control decision matrix.

[0103] It is understandable that the matching and fusion processing employs a tensor fusion algorithm based on semantic alignment. First, it performs dimensional matching on the actuator control command identification matrix and the hierarchical collaborative control strategy vector. The actuator control command identification matrix is... OK The column is a real matrix, and the hierarchical collaborative control strategy vector is... A real vector with 1 row and 1 column. Indicates the total number of actuators. Indicates the number of control instruction types. This indicates the number of levels in the hierarchical strategy.

[0104] Dimensional matching is achieved through a linear transformation. The transformation process involves multiplying the actuator control instruction recognition matrix by the alignment weight matrix and then adding the alignment bias matrix to obtain the dimension-aligned instruction matrix. The alignment weight matrix is ​​as follows: OK The column is a real matrix, and the alignment bias matrix is... OK A column of real numbers.

[0105] The semantic alignment process is implemented through executor function encoding. The function encoding adopts a combination of one-hot encoding and continuous embedding. The encoding process is to... After one-hot encoding the function type of the actuator, it is combined with the first... The vector concatenation operation is performed on the consecutive embedding vectors of the actuator to obtain the _th _ ... The function encoding vector of the executor. Tensor fusion adopts a multimodal attention mechanism, and the attention weight is calculated as the _th _th_ ... The query vector of the executor is multiplied by the transpose of the key vector of the j-th layer policy, divided by the square root of the key vector dimension, and then exponentially multiplied. Finally, it is divided by the sum of the exponential operations of the corresponding calculation results of all policy layers to obtain the result of the j-th layer policy. The actuator and the first The fusion attention weights of the layered strategies. The comprehensive control decision matrix is ​​calculated by summing over all hierarchical strategies, with each term being the sum of the fusion attention weight multiplied by the corresponding element of the dimension-aligned instruction matrix and the corresponding element of the hierarchical collaborative control strategy vector, yielding the [previous term]. The integrated control decision vector of each actuator achieves semantic-level deep fusion of instruction recognition results and collaborative control strategies through this fusion mechanism.

[0106] Based on the comprehensive control decision matrix, priority sorting and conflict resolution are performed to obtain the comprehensive coordination control strategy for building actuators.

[0107] For example, the priority ranking adopts a multi-criteria decision-making algorithm based on the analytic hierarchy process (AHP), constructing a priority judgment matrix of G rows and G columns, where G represents the number of priority judgment criteria, and the elements of the judgment matrix are calculated as criteria. Importance weight divided by criteria Importance weights represent the criteria Relative Criteria The importance ratio.

[0108] Consistency testing is performed by calculating a consistency index. The consistency index equals the largest eigenvalue of the judgment matrix minus the number of priority judgment criteria, then divided by the number of priority judgment criteria minus one. A judgment matrix is ​​considered to have satisfactory consistency when the consistency index divided by the random consistency index is less than 0.1. The executor priority score is calculated by summing over all priority judgment criteria, with each criterion representing the i-th priority criterion. The weight of the first criterion is multiplied by the weight of the second criterion. The actuator in the first Quality scoring under the first criterion. Conflict resolution is handled using an optimization algorithm based on constraint satisfaction. Conflict detection is achieved through a resource occupancy matrix, and resource conflict is determined when the first criterion is met. The actuator in the first When the sum of the utilization rates of all resources in the resource set required at any given time is greater than 1, the indicator function outputs 1 to indicate that there is a resource conflict; otherwise, it outputs 0 to indicate that there is no conflict.

[0109] Conflict resolution employs a time window adjustment strategy. The adjusted execution time equals the original execution time plus a time offset. The time offset is determined by minimizing the conflict cost function, which is a double summation over all actuators and all time points. Each term is the absolute value of the conflict penalty weight multiplied by the conflict indicator function, plus the delay penalty weight multiplied by the time offset. The building actuator integrated coordination control strategy is ultimately represented as a set of final coordinated control strategies containing actuator numbers, adjusted execution times, and integrated control decisions. Each element in the set is a triple, including the actuator index. The adjusted execution time of the actuator and the comprehensive control decision vector of the actuator.

[0110] In some embodiments, based on the above embodiments, step S5 includes: The integrated coordination control strategy for building actuators is processed by carrier coding to obtain the carrier control frame sequence corresponding to the electric actuators in each area.

[0111] It is understandable that the carrier coding processing adopts a hierarchical coding algorithm based on the building electric actuator control command standard. First, the integrated coordinated control strategy of the building actuators is classified and coded according to the actuator function type. The classification coding process involves... The function type of each executor is hashed using a hash function and then XORed with the current timestamp to obtain the result. The type encoding of each executor is used, where a hash function is used to map the function type to a numerical value, a timestamp represents the current time stamp, and an XOR operation ensures the uniqueness of the encoding.

[0112] The parameterized encoding of control commands employs a hybrid radix encoding method. The encoding process involves summing all dimensions from zero to the number of control decision dimensions minus one, with each term representing the first radix. The actuator Dimensional control decision value multiplied by the first dimensional encoding cardinality The power of 1, yielding the 1st power. The parameterized encoding vector of each actuator, where the control decision value represents the decision value of that dimension, the encoding cardinality represents the base of that dimension, and the number of control decision dimensions represents the total dimension of the decision vector.

[0113] Time synchronization coding is achieved through phase differential modulation of the carrier signal, and the phase difference is calculated as the... After adjusting the execution time of each actuator, subtract the reference time base, divide by the time resolution, and then calculate the arctangent function value to obtain the first value. The phase differential synchronization encoding of each actuator is used, where the reference time base represents the unified time reference point of the system, and the time resolution represents the smallest unit of precision for time measurement.

[0114] The carrier control frame sequence is constructed according to the building carrier communication protocol standard. The frame structure includes five fields: frame header, actuator address, control command, parameter data, and checksum. The length of a single frame is calculated as the sum of the length of the frame header, the address field, the command field, the data field, and the checksum field. The carrier control frame sequence corresponding to the electric actuator in each area is represented as the [number missing]. The region contains the collection of all actuator control frames.

[0115] The carrier control frame sequence is modulated into a power carrier signal according to a preset communication timing, and control commands are sent to the electric actuators in each area through the power line.

[0116] For example, the modulation process employs a hybrid modulation method combining orthogonal frequency division multiplexing and spread spectrum modulation. First, a serial-to-parallel conversion of the carrier control frame is performed. This conversion divides the carrier control frame sequence into several parallel subcarrier data streams. The subcarrier allocation process involves... The carrier control frame sequence of the region is based on the subcarrier index. Subtract one, multiply by the number of parallel data streams, and add one to the subcarrier index. Divide the range by multiplying it by the number of parallel data streams to obtain the first... Each subcarrier carries a data segment.

[0117] Orthogonal frequency division multiplexing modulation is implemented using fast Fourier transform. The modulation process involves summing all subcarriers from zero to the total number of subcarriers minus one, with each term being the nth term. Multiplying the modulation symbol of the nth subcarrier by the imaginary unit of the natural constant, by twice pi, by the subcarrier index, by the time-domain sampling index, by the total number of subcarriers raised to the power of the sum of its parts, and then by the square root of the total number of subcarriers, yields the nth subcarrier. The orthogonal frequency division multiplexing modulated signal at time t, where the modulation symbol represents the digital information carried by the subcarrier, the imaginary unit represents the basic unit of the imaginary part of the complex number, and the time-domain sampling index represents the sampling point position in the time domain.

[0118] Spread spectrum modulation is achieved through a pseudo-random sequence. The spread spectrum sequence is generated using a linear feedback shift register. The generating polynomial is a quadratic plus a cubic plus a fifth plus a seventh of a delay operator. The spread spectrum signal is calculated by multiplying the orthogonal frequency division multiplexing modulated signal by the value of the pseudo-random sequence at the time-domain sampling index modulo the length of the pseudo-random sequence, thus obtaining the spread spectrum signal.

[0119] The preset communication timing scheduling adopts a time-division multiple access mechanism. The time slot allocation process is to add the frame start time to the region index and multiply it by the time slot duration to obtain the first time slot. The time slot start time of the region, where the frame start time represents the start time of the communication frame, and the time slot duration represents the length of communication time allocated to each region.

[0120] Finally, the power line carrier signal is coupled into the power line through a power amplifier. The carrier power control process involves multiplying the maximum carrier power by the square of the spread spectrum signal amplitude minus the natural constant, and dividing by the square power of the power normalization parameter to obtain the th . The carrier power at any given time, where the maximum carrier power represents the upper limit of the system's allowed power, and the power normalization parameter represents the normalization coefficient of power control. Control commands are sent to the electric actuators in each area via power line carrier signals to achieve unified and coordinated control of distributed multi-area electric actuators.

[0121] In some embodiments, the present invention also provides a multi-region electric actuator coordinated control system based on carrier communication, the system comprising: The data acquisition module is used to collect carrier communication datasets from electric actuators in multiple areas of a building. The signal recognition module is used to intelligently identify carrier signals based on the carrier communication dataset, using an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction, and outputs an actuator control command recognition matrix. The coordination control module is used to perform actuator coordination control based on the carrier communication dataset and actuator control command identification matrix, using a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy, and outputs a hierarchical cooperative control strategy vector. The strategy generation module is used to generate a comprehensive coordinated control strategy for building actuators based on the actuator control instruction identification matrix and the hierarchical collaborative control strategy vector. The control command sending module is used to send control commands to the electric actuators in each area via power line carrier signals based on the building actuator integrated coordination control strategy, so as to control the electric actuators.

[0122] This embodiment effectively solves the cross-system coordination failure problem existing in traditional building electric actuator control systems by proposing a multi-zone electric actuator coordination control system based on carrier communication. This embodiment fully utilizes the convenience of carrier communication and the decision-making advantages of artificial intelligence algorithms to achieve unified coordination control of actuators in multiple zones and systems, significantly improving the overall coordination and energy efficiency of building equipment operation.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-region electric actuator coordinated control method based on carrier communication, characterized in that, Includes the following steps: Collect carrier communication datasets of electric actuators in multiple areas of a building; Based on the aforementioned carrier communication dataset, an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction are used to intelligently identify carrier signals and output an actuator control command identification matrix. Based on the aforementioned carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy are used to perform actuator coordinated control and output a hierarchical cooperative control strategy vector. A comprehensive coordinated control strategy for building actuators is generated based on the actuator control command identification matrix and the hierarchical collaborative control strategy vector. Based on the building actuator integrated coordination control strategy, control commands are sent to the electric actuators in each area via power line carrier signals to control the electric actuators.

2. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 1, characterized in that, Based on the carrier communication dataset, the method employs an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction to intelligently identify carrier signals and output an actuator control command identification matrix, including: Impedance-aware multi-scale convolutional neural networks are used to extract multi-scale features from carrier communication datasets to obtain multi-scale feature matrices of carrier signals. The time-series attention weighting algorithm for building load prediction is used to adaptively weight the multi-scale feature matrix of the carrier signal to obtain an adaptive weight feature matrix; A domain-adaptive meta-learning algorithm constrained by building physical features is used to perform cross-domain optimization processing on the adaptive weight feature matrix to obtain a domain-adaptive feature matrix, and an actuator control command recognition matrix is ​​generated based on the domain-adaptive feature matrix.

3. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 2, characterized in that, The method of using an impedance-aware multi-scale convolutional neural network to extract multi-scale features from a carrier communication dataset includes: Multiple convolution kernels of different scales are dynamically constructed based on the power line impedance values ​​in the carrier communication dataset to obtain an adaptive convolution kernel group. The adaptive convolution kernel group is used to perform parallel multi-scale convolution processing on the carrier communication dataset to extract carrier signal feature components in different frequency bands. The amplitude of the carrier signal feature components in different floors is compensated by combining the floor attenuation compensation mechanism, and the compensated carrier signal feature components are fused to obtain a multi-scale feature matrix of the carrier signal.

4. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 2, characterized in that, The time-series attention weighting algorithm using building load prediction adaptively assigns weights to the multi-scale feature matrix of the carrier signal, including: A long short-term memory network is used to perform time-series modeling of the multi-scale feature matrix of the carrier signal to predict the load change trend of each system in the building, thus obtaining a load prediction sequence. Based on the load prediction sequence, a three-level cascaded attention mechanism at the floor, area, and equipment levels is constructed, and the attention weight coefficients of each level are calculated. The weight attenuation parameters of the attention weight coefficients at each level are dynamically adjusted according to the operating status data of the building temperature control system, and the multi-scale feature matrix of the carrier signal is adaptively redistributed to obtain an adaptive weight feature matrix.

5. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 2, characterized in that, The domain-adaptive meta-learning algorithm for building physical feature constraints performs cross-domain optimization processing on the adaptive weight feature matrix, including: A building physical feature constraint model is constructed, using building structure parameters and pipeline system layout parameters as physical constraints. A meta-learning network is used to quickly and adaptively train the adaptive weight feature matrix, learning the feature mapping relationship between different building environments under the guidance of the building physical feature constraint model. The network parameters of the meta-learning network are updated through gradient descent optimization to achieve rapid adaptation across building environments, and the adaptive weight feature matrix is ​​converted into a domain adaptive feature matrix.

6. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 1, characterized in that, The actuator coordinated control is performed using a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy, based on the carrier communication dataset and actuator control command recognition matrix. The output is a hierarchical coordinated control strategy vector, including: Based on the carrier communication dataset and actuator control command recognition matrix, a hierarchical graph neural network with building three-dimensional spatial constraints is used to construct an actuator coordination topology graph to obtain an actuator spatial topology relation matrix. Based on the actuator space topology matrix, an asynchronous multi-agent learning algorithm with system priority hierarchy is used to perform multi-system coordination decision-making, and a multi-agent coordination decision matrix is ​​obtained. Based on the multi-agent coordination decision matrix, energy consumption constraint optimization is performed to obtain a hierarchical collaborative control strategy vector.

7. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 6, characterized in that, The hierarchical graph neural network using three-dimensional building spatial constraints to construct the actuator coordination topology graph includes: The building is modeled as a two-layer graph structure consisting of a vertical connection graph and a horizontal connection graph. The vertical connection graph represents the actuator connection relationships between floors, and the horizontal connection graph represents the actuator connection relationships within the same floor. Semantic constraints of the piping system are embedded into the graph node features, and semantic connections between nodes are constructed based on the piping layout of the HVAC, water supply and drainage, and fire protection systems. The graph edge weights are dynamically adjusted based on the real-time power line load status in the carrier communication dataset to obtain the actuator coordination topology graph. The graph structure features of the actuator coordination topology graph are extracted to obtain the actuator spatial topology relation matrix.

8. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 6, characterized in that, The asynchronous multi-agent learning algorithm employing system priority hierarchy for multi-system coordination decision-making includes: A hierarchical intelligent agent architecture is constructed based on the safety priorities of the fire protection system, water supply and drainage system, and HVAC system, with the intelligent agent of the fire protection system having the highest decision priority. An asynchronous decision-making mechanism is adopted to enable intelligent agents at each level to learn in parallel according to the actuator space topology relation matrix, and the decision results of high-priority intelligent agents are used as constraints for low-priority intelligent agents. The policy network parameters of each intelligent agent are updated through reinforcement learning to obtain the multi-agent coordinated decision matrix.

9. The multi-region electric actuator coordinated control method based on carrier communication as described in claim 6, characterized in that, The energy consumption constraint optimization process based on the multi-agent coordinated decision matrix includes: A multi-objective optimization function is constructed with communication quality, coordination efficiency, and energy consumption as optimization objectives, and the multi-agent coordination decision matrix is ​​used as the optimization variable. An adaptive particle swarm optimization algorithm is used to solve the multi-objective optimization function, and the gradient of energy consumption change is used as the dominant factor for particle velocity update. Simultaneously, optimization is performed on three time scales: second-level communication scheduling, minute-level load prediction, and hour-level system coordination. The optimal coordination control parameters are obtained through iterative optimization. Based on the optimal coordination control parameters, the multi-agent coordination decision matrix is ​​converted into a hierarchical cooperative control strategy vector.

10. A multi-zone electric actuator coordinated control system based on carrier communication, used to execute the multi-zone electric actuator coordinated control method based on carrier communication as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to collect carrier communication datasets from electric actuators in multiple areas of a building. The signal recognition module is used to intelligently identify carrier signals based on the carrier communication dataset, using an impedance-aware multi-scale convolutional neural network and a time-series attention weight algorithm for building load prediction, and outputs an actuator control command recognition matrix. The coordination control module is used to perform actuator coordination control based on the carrier communication dataset and actuator control command identification matrix, using a hierarchical graph neural network with building three-dimensional spatial constraints and an asynchronous multi-agent learning algorithm with system priority hierarchy, and outputs a hierarchical cooperative control strategy vector. The strategy generation module is used to generate a comprehensive coordinated control strategy for building actuators based on the actuator control instruction identification matrix and the hierarchical collaborative control strategy vector. The control command sending module is used to send control commands to the electric actuators in each area via power line carrier signals based on the building actuator integrated coordination control strategy, so as to control the electric actuators.

Images