A Multi-Energy Coupled Temperature Field Control Method and System Based on Intelligent Optimization

Abstract

This application discloses a multi-energy coupled temperature field control method and system based on intelligent optimization, belonging to the field of intelligent control technology. In terms of energy efficiency optimization, the method extracts the spatiotemporal features of the temperature field through an ST-CNN spatiotemporal convolutional network and combines it with the NSGA-III algorithm to solve the Pareto front of the multi-objective optimization model, enabling the system to achieve an optimal balance between electrothermal conversion efficiency, coefficient of performance (COP), and human comfort. Secondly, in terms of control accuracy, a 3D convolutional kernel and LSTM module are used to collaboratively analyze the dynamic laws of the temperature field, and CFD turbulence model simulation verification is performed, which can reduce the temperature control error to a small range and significantly improve the uniformity of spatial temperature. Finally, in terms of system intelligence, a near-end strategy optimization algorithm is used to achieve autonomous iterative optimization of control parameters, forming a closed-loop control system of "perception-decision-execution-feedback," reducing or eliminating the need for manual intervention.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to a multi-energy coupled temperature field control method and system based on intelligent optimization. Background Technology

[0002] Multi-energy coupled air conditioning systems typically employ a two-tier architecture of a centralized control platform and zoned execution units. The central control system collects environmental parameters through a network of temperature and pressure sensors, uses PID control algorithms for temperature and humidity regulation, and relies on manual setting of temperature thresholds and operating modes. In large buildings, the system is divided into multiple temperature control zones, each equipped with an independent actuator, achieving basic regulation through preset timing logic or simple feedback mechanisms.

[0003] Multi-split air conditioning systems, as a typical example, connect multiple indoor units to their outdoor units via refrigerant pipes. They rely on electronic expansion valves to regulate refrigerant flow to meet the load demands of different areas. However, this control method lacks precise perception of the spatial distribution of the temperature field.

[0004] Traditional control methods mainly rely on manual experience to set a constant temperature field or on simple temperature control rules to control the temperature field. For example, a normal indoor temperature is kept constant at a certain level, or a simple control measure of cooling during the day and keeping warm at night is adopted in production workshops and warehouses. For higher frequency or more flexible control methods, air conditioning parameters are usually adjusted manually in real time, resulting in high maintenance costs and untimely response of the temperature field. Summary of the Invention

[0005] The main objective of this application is to provide a multi-energy coupled temperature field control method and system based on intelligent optimization, in order to solve the problem that the traditional control methods in the prior art mainly rely on manual experience to set the constant temperature field or to achieve temperature field control according to simple temperature adjustment rules. For higher frequency or more flexible control methods, air conditioning parameters are usually adjusted manually in real time, which leads to high maintenance costs and untimely response of the temperature field.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A multi-energy coupled temperature field control method based on intelligent optimization is disclosed. This method controls the temperature field within a preset area via an air conditioning system. The preset area contains an Internet of Things (IoT) sensor array. The multi-energy coupled temperature field control method includes:

[0008] Step S1: Obtain the original temperature field dataset of the preset area through the IoT sensor array, and transmit the original temperature field dataset to the edge computing node through the MQTT protocol to form an initial three-dimensional temperature field matrix;

[0009] Step S2: Extract the local temperature gradient features of the initial three-dimensional temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial three-dimensional temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network to output a high-dimensional tensor that integrates spatiotemporal features.

[0010] Step S3: Construct a multi-objective optimization model based on electrothermal conversion efficiency, cooling coefficient, and human comfort index, and input the high-dimensional tensor into the multi-objective optimization model. Solve the Pareto front of the multi-objective optimization model using the NSGA-III algorithm to obtain the air conditioning group control parameters.

[0011] Step S4: Send the air conditioning group control parameters to the air conditioning system and collect the equipment response dataset of the air conditioning system;

[0012] Step S5: Construct a CFD simulation environment equipped with a turbulence model based on the preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model. The actual temperature field dataset is obtained through dynamic simulation in the CFD simulation environment.

[0013] Step S6: Obtain the predefined expected temperature field, and perform a difference calculation between the actual temperature field dataset and the expected temperature field dataset to obtain the deviation matrix;

[0014] Step S7: The air conditioning group control parameters are iteratively updated using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thereby obtaining the Pareto optimal solution set of the air conditioning group control parameters;

[0015] Step S8: Convert the Pareto optimal solution set into readable instructions for the air conditioning system and send them to the air conditioning system.

[0016] As a further improvement to this application, step S1 involves acquiring the original temperature field dataset of the preset area through the IoT sensor array, and transmitting the original temperature field dataset to the edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix, including:

[0017] Step S11: Deploy an IoT sensor array based on the preset area according to a hexagonal grid topology, and collect raw temperature data at their respective locations through each temperature sensor in the IoT sensor array;

[0018] Step S12: All raw temperature data are aggregated to the regional gateway of the edge computing node via the Zigbee protocol to form the raw temperature field dataset;

[0019] Step S13: Start the MQTT agent based on the edge computing node to encapsulate the original temperature field dataset into a structured message in JSON format;

[0020] Step S14: The structured message is completed into a continuous temperature field by kriging interpolation through the edge computing node, resulting in an initial three-dimensional temperature field matrix with dimensions N×M×P×1, where N / M / P correspond to the number of discretized grids along the X / Y / Z axes of the preset region, respectively.

[0021] As a further improvement of this application, step S2 involves extracting the local temperature gradient features of the initial three-dimensional temperature field matrix using the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learning the dynamic law of the time dimension of the initial three-dimensional temperature field matrix using the LSTM module of the ST-CNN spatiotemporal convolutional network to output a high-dimensional tensor that fuses spatiotemporal features, including:

[0022] Step S21: Extract features from the initial three-dimensional temperature field matrix using a 3D convolution kernel group of a preset size, and output a feature map with a dimension of N×M×P×16 through the activation function of the ST-CNN spatiotemporal convolutional network.

[0023] Step S22: The feature map is reduced to N / 2×M / 2×P / 2×16 by the 3D max pooling layer of the ST-CNN spatiotemporal convolutional network to obtain 3D convolutional features;

[0024] Step S23: Expand the 3D convolutional features into temporal features according to the time sequence, and capture the temporal dependencies of the temporal features through a bidirectional LSTM structure with a hidden layer dimension of 64, and output the original LSTM output features with a dimension of 2×64×N / 2×M / 2×P / 2.

[0025] Step S24: The original output features of LSTM at all times are weighted and fused through a time attention mechanism to obtain a dynamic weight matrix with dimensions N / 2×M / 2×P / 2×T, where T is the time dimension;

[0026] Step S25: Perform a Hadamard product operation between the dynamic weight matrix and the temporal features to obtain a weighted temporal feature tensor with dimensions of 2×64×N / 2×M / 2×P / 2;

[0027] Step S26: Reduce the number of channels of the weighted temporal feature tensor to 32 by 3D transpose convolution to obtain LSTM features;

[0028] Step S27: The 3D convolutional features and the LSTM features are concatenated by channels, and after normalization, a 48-channel hybrid tensor is obtained.

[0029] Step S28: The hybrid tensor is weighted twice by spatial attention mechanism, and a 32-dimensional feature vector is obtained by global average pooling, which is the high-dimensional tensor.

[0030] As a further improvement to this application, step S3 involves constructing a multi-objective optimization model based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort index. The high-dimensional tensor is then input into the multi-objective optimization model, and the Pareto front of the multi-objective optimization model is solved using the NSGA-III algorithm to obtain the air conditioning group control parameters, including:

[0031] Step S31: Define the objective function of the electrothermal conversion efficiency, construct the objective function of the coefficient of performance (COP) using linear interpolation, and construct the objective function of the human comfort index based on the PMV-PPD standard.

[0032] Step S32: Perform Min-Max normalization on the three objective functions and add preset constraints to each of them;

[0033] Step S33: Define the target space by using the electrothermal conversion efficiency, the cooling coefficient, and the human comfort index as the three coordinate axes of the target space.

[0034] Step S34: Define the feasible region of the target space through all preset constraints, and map the high-dimensional tensor to the target space through linear transformation to obtain a prediction vector based on three objective functions;

[0035] Step S35: Define several random air conditioning group control parameter combinations based on the NSGA-III algorithm, and input all random air conditioning group control parameter combinations into the target space to form several target vectors;

[0036] Step S36: Update all target vectors iteratively through non-dominated sorting, and calculate the fitness value of each target vector based on the predicted vector in each update iteration;

[0037] Step S37: After the preset iteration termination condition is met, output the Pareto front solution set, which includes several combinations of non-dominated air conditioning group control parameters.

[0038] As a further improvement to this application, step S5 involves constructing a CFD simulation environment equipped with a turbulence model based on the preset region, and inputting the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model. The actual temperature field dataset is then obtained through dynamic simulation in the CFD simulation environment, including:

[0039] Step S51: Define a turbulence model based on the geometric features of the preset area and the air supply characteristics of the air conditioning system;

[0040] Step S52: Divide the initial three-dimensional temperature field matrix into a computational domain grid using an unstructured grid, and build a solver environment in the CFD simulation environment;

[0041] Step S53: Perform time-progression calculations through the solver environment and record the actual temperature of each node in the computational domain grid in real time;

[0042] Step S54: Integrate the actual temperatures of all nodes into the actual temperature field dataset.

[0043] As a further improvement to this application, step S7 involves iteratively updating the air conditioning group control parameters using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thereby obtaining the Pareto optimal solution set of the air conditioning group control parameters, including:

[0044] Step S71: Construct a dual-head neural network policy model with two independent output heads, namely the parameter control head and the value head;

[0045] Step S72: Periodically acquire the deviation matrix, the L2 norm of the deviation matrix, and the air conditioning group control parameters based on a preset acquisition interval, and integrate them into trajectory data;

[0046] Step S73: Calculate the state value function of the trajectory data using the value head, and calculate the dominance function of the state value function using the generalized dominance estimation method;

[0047] Step S74: Bind trajectory data and advantage function of the same preset acquisition interval into a training sample, and integrate all preset acquisition intervals to obtain a training sample set;

[0048] Step S75, define the pruning objective function of the dual-head neural network strategy model according to equation (1):

[0049] (1);

[0050] in, Let the pruning objective function be... This is the weight set of the dual-head neural network strategy model. The ratio of the action probabilities of the new strategy to those of the old strategy. For the aforementioned advantage function, The pruning threshold is the pruning objective function;

[0051] Step S76: Input the training sample set into the pruning objective function, and iterate and update the dual-head neural network policy model for several epochs using the Adam optimizer;

[0052] Step S77: When the rate of change of the HV index of the Pareto front solution is less than a preset percentage threshold for several consecutive iterations, the iteration is terminated, and the air conditioning group control parameters output by the parameter control head are obtained as the Pareto optimal solution set.

[0053] As a further improvement to this application, step S8, converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending them to the air conditioning system, includes:

[0054] Step S81: Convert the air conditioning group control parameters in the Pareto optimal solution set into physical parameters through linear transformation;

[0055] Step S82: Encode the physical parameters into a data unit structure that conforms to the communication protocol of the air conditioning system;

[0056] Step S83: The data unit structure is encapsulated by the OPC-UA server of the edge computing node and sent to the air conditioning system.

[0057] To achieve the above objectives, this application also provides the following technical solutions:

[0058] A multi-energy coupled temperature field control system based on intelligent optimization, wherein the multi-energy coupled temperature field control system is applied to the multi-energy coupled temperature field control method described above, and the multi-energy coupled temperature field control system includes:

[0059] The initial three-dimensional temperature field matrix acquisition module is used to acquire the original temperature field dataset of the preset area through the Internet of Things sensor array, and transmit the original temperature field dataset to the edge computing node through the MQTT protocol to form the initial three-dimensional temperature field matrix.

[0060] The high-dimensional tensor output module is used to extract the local temperature gradient features of the initial three-dimensional temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial three-dimensional temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network, and output a high-dimensional tensor that integrates spatiotemporal features.

[0061] The air conditioning group control parameter solving module is used to construct a multi-objective optimization model based on electrothermal conversion efficiency, cooling coefficient, and human comfort index, and input the high-dimensional tensor into the multi-objective optimization model, and obtain the air conditioning group control parameters by solving the Pareto front of the multi-objective optimization model through the NSGA-III algorithm;

[0062] The device response dataset acquisition module is used to send the air conditioning group control parameters to the air conditioning system and acquire the device response dataset of the air conditioning system.

[0063] The actual temperature field dataset dynamic simulation module is used to construct a CFD simulation environment equipped with a turbulence model based on the preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model, and obtain the actual temperature field dataset through dynamic simulation in the CFD simulation environment;

[0064] The deviation matrix calculation module is used to obtain a predefined expected temperature field and perform a difference calculation between the actual temperature field dataset and the expected temperature field dataset to obtain a deviation matrix.

[0065] The air conditioning group control parameter optimal solution set acquisition module is used to iteratively update the air conditioning group control parameters through a near-end strategy optimization algorithm to gradually approach the minimum value of the deviation matrix and obtain the Pareto optimal solution set of the air conditioning group control parameters;

[0066] The optimal solution set for air conditioning group control parameters is sent by a module that converts the Pareto optimal solution set into readable instructions for the air conditioning system and sends them to the air conditioning system.

[0067] To achieve the above objectives, this application also provides the following technical solutions:

[0068] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the multi-energy coupled temperature field control method described above.

[0069] To achieve the above objectives, this application also provides the following technical solutions:

[0070] A storage medium storing program instructions that, when executed by a processor, implement the multi-energy coupled temperature field control method described above.

[0071] Beneficial effects:

[0072] This application acquires the original temperature field dataset of a preset area using an IoT sensor array and transmits it to an edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix. Local temperature gradient features of the initial three-dimensional temperature field matrix are extracted using the 3D convolutional kernels of an ST-CNN spatiotemporal convolutional network. The LSTM module of the ST-CNN spatiotemporal convolutional network then learns the dynamic temporal dimension of the initial three-dimensional temperature field matrix, outputting a high-dimensional tensor that integrates spatiotemporal features. A multi-objective optimization model is constructed based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort indices. The high-dimensional tensor is input into the multi-objective optimization model, and the Pareto front of the multi-objective optimization model is solved using the NSGA-III algorithm. The process involves obtaining the air conditioning group control parameters; sending these parameters to the air conditioning system and collecting the equipment response dataset; constructing a CFD simulation environment with a turbulence model based on a preset area; inputting the initial three-dimensional temperature field matrix and equipment response dataset into the turbulence model; dynamically simulating the actual temperature field dataset through the CFD simulation environment; obtaining a predefined expected temperature field; calculating the deviation matrix by difference between the actual temperature field dataset and the expected temperature field dataset; iteratively updating the air conditioning group control parameters using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thus obtaining the Pareto optimal solution set for the air conditioning group control parameters; and converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending them to the air conditioning system. This application optimizes energy efficiency by extracting the spatiotemporal features of the temperature field using an ST-CNN spatiotemporal convolutional network and combining it with the NSGA-III algorithm to solve the Pareto front of a multi-objective optimization model, achieving an optimal balance between electrothermal conversion efficiency, coefficient of performance (COP), and human comfort. Secondly, regarding control accuracy, it employs a 3D convolutional kernel and LSTM module to collaboratively analyze the dynamic laws of the temperature field, coupled with CFD turbulence model simulation verification, reducing temperature control errors to a small range (±0.5℃) and significantly improving spatial temperature uniformity. Finally, in terms of system intelligence, it utilizes a near-end strategy optimization algorithm to achieve autonomous iterative optimization of control parameters, forming a closed-loop control system of "perception-decision-execution-feedback," reducing or eliminating the need for manual intervention. Simultaneously, the deployment of an IoT sensor array shortens the equipment status monitoring response time to the millisecond level, making this application suitable for scenarios with high requirements for temperature control accuracy and energy efficiency, such as large commercial complexes, data centers, production workshops, and warehouses. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the steps of an embodiment of the intelligent optimization-based multi-energy coupling temperature field control method of this application.

[0074] Figure 2 This is a functional block diagram of an embodiment of the intelligent optimization-based multi-energy coupled temperature field control system of this application;

[0075] Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0076] Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0077] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0078] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0079] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same instance, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0080] like Figure 1 As shown, this embodiment provides an example of a multi-energy coupled temperature field control method based on intelligent optimization. In this embodiment, the multi-energy coupled temperature field control method is used to control the temperature field in a preset area through an air conditioning system. The preset area has an Internet of Things sensor array.

[0081] Preferably, multi-energy coupled air conditioning is a high-efficiency system that integrates multiple energy sources (such as electricity, gas, geothermal energy, etc.) with air conditioning technology. Its core lies in improving overall energy efficiency by utilizing multi-energy synergistic optimization (such as combined cooling, heating, and power). Typical technologies include water-cooled multi-split units and intelligent dynamic matching control, which can achieve energy efficiency improvements of over 40%. Compared to traditional multi-split units, multi-energy coupled air conditioning emphasizes coupling across energy networks, such as combining cooling, heating, and clean energy (such as solar energy). Through intelligent control, it enables energy relay in the event of a network outage, making it suitable for large buildings or scenarios requiring high energy efficiency standards.

[0082] Preferably, the preset area is a scenario with high requirements for temperature control accuracy and energy efficiency, such as a large commercial complex, data center, production workshop, or warehouse.

[0083] Specifically, the multi-energy coupled temperature field control method includes the following steps:

[0084] Step S1: Obtain the original temperature field dataset of the preset area through the Internet of Things sensor array, and transmit the original temperature field dataset to the edge computing node through the MQTT protocol to form an initial three-dimensional temperature field matrix.

[0085] Preferably, the IoT sensor array can employ high-precision digital temperature and humidity sensors, deployed according to a spatial gridding strategy, with each grid spaced 3×3m in a preset area. 2 Sensor nodes are installed at the grid vertices; for high-precision scenarios, each 1×1m grid can be used. 2 Sensor nodes are installed at the grid vertices.

[0086] Preferably, each node of the IoT sensor array can integrate a LoRa / NB-IoT dual-mode communication module to achieve low-power wide-area coverage.

[0087] Preferably, edge computing nodes are distributed computing units deployed near data sources or user terminals. They reduce reliance on the cloud through localized processing, achieving low-latency, high-bandwidth, and efficient data processing. Their core functions include real-time data acquisition, edge-side analysis, and rapid decision execution. Typical hardware components include processors (such as multi-core Xeon or ARM chips), storage devices (SSD / memory), and network interfaces (5G / WiFi). The software system supports containerized deployment and lightweight AI inference. In this embodiment, edge computing nodes can significantly optimize the response speed in scenarios such as real-time analysis of factory sensor data and local temperature control decisions.

[0088] Step S2: Extract local temperature gradient features of the initial three-dimensional temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial three-dimensional temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network, and output a high-dimensional tensor that integrates spatiotemporal features.

[0089] Preferably, the 3D convolution kernel can be a 3×3×3 kernel group and a stride of 1×1×1 to extract local gradient features along the X / Y / Z axes of the initial three-dimensional temperature field matrix. Then, the temperature field patterns of different receptive fields are processed in parallel through a 3-layer stacked Inception module. The activation function of the 3D convolution kernel can be LeakyReLU (α=0.3) to suppress the temperature gradient vanishing problem.

[0090] Preferably, the dynamic pattern in the time dimension can be obtained by slicing the temperature sequence into 10-second windows, setting the hidden layer dimension of the input LSTM to 128, and capturing the temporal correlation of the temperature field before and after by BiLSTM.

[0091] Step S3: Construct a multi-objective optimization model based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort index. Input the high-dimensional tensor into the multi-objective optimization model and use the NSGA-III algorithm to solve the Pareto front of the multi-objective optimization model to obtain the air conditioning group control parameters.

[0092] Preferably, the electrothermal conversion efficiency can be obtained by establishing a quantitative model of the electrothermal conversion efficiency η, which is calculated as η=Q_heat / P_elec, where Q_heat is the heating capacity and P_elec is the input electrical power. The relationship surface between η and the temperature setpoint and the ambient humidity can be obtained by fitting the experimental data in advance.

[0093] Preferably, a lookup table model for the coefficient of performance (COP) can be constructed using linear interpolation. The input parameters are the evaporation temperature T_evap and the condensation temperature T_cond. The discrete value set of COP is obtained through the performance curve of the air conditioning system, and the interpolation dimension of the compressor frequency is added.

[0094] Preferably, the average thermal sensation index can be calculated and predicted based on the PMV-PPD standard. The input parameters include dry-bulb temperature, relative humidity, air velocity, and average radiant temperature. The Fanger equation is used for real-time calculation to ensure a soft constraint that the absolute value of PMV is ≤0.5.

[0095] Step S4: Send the air conditioning group control parameters to the air conditioning system and collect the equipment response dataset of the air conditioning system.

[0096] Preferably, step S4 can use a bidirectional transmission channel to ensure data transmission efficiency. Downlink control sends group control parameters through the Modbus TCP protocol (port 502) and uses CRC-16 checksum to ensure instruction integrity. Uplink acquisition uses the OPC-UA protocol, port 4840, to read device status data in real time at a sampling frequency of 10Hz.

[0097] Step S5: Construct a CFD simulation environment equipped with a turbulence model based on a preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model. The actual temperature field dataset is obtained through dynamic simulation in the CFD simulation environment.

[0098] Preferably, if dynamic simulation of the preset area is to be performed, the geometric structure of the preset area needs to be generated based on the millimeter-level BIM building information model, and unstructured meshing is performed using ICEM CFD, with the number of meshes controlled between 2 million and 5 million; the air conditioning outlet is set as a velocity inlet, for example, v=0.5-2.5m / s, the wall is set as an insulation boundary, and the doors and windows are set as pressure outlets.

[0099] Preferably, the k-ω SST model can be selected as the turbulence model.

[0100] Step S6: Obtain the predefined expected temperature field, and perform a difference calculation between the actual temperature field dataset and the expected temperature field dataset to obtain the deviation matrix.

[0101] Preferably, the differential calculation requires dimensionality verification to ensure that the measured and expected temperature fields have the same spatial resolution and time step, and the dynamic time warping (DTW) algorithm is used to eliminate sensor sampling time deviation.

[0102] Preferably, the expected temperature field standard is formulated according to GB 50736-2012 "Code for Design of Heating, Ventilation and Air Conditioning of Civil Buildings", setting the temperature range of the activity area (1.1-1.8m height) as 22±1℃ and the allowable deviation of the non-activity area as ±2℃, and the three-dimensional temperature field distribution is generated by spatial interpolation.

[0103] Step S7: The air conditioning group control parameters are iteratively updated using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thereby obtaining the Pareto optimal solution set of the air conditioning group control parameters.

[0104] Step S8: Convert the Pareto optimal solution set into readable instructions for the air conditioning system and send them to the air conditioning system.

[0105] Further, step S1 involves acquiring the original temperature field dataset of a preset area through an IoT sensor array, and transmitting the original temperature field dataset to the edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix. This specifically includes the following steps:

[0106] Step S11: Deploy an IoT sensor array based on a hexagonal grid topology within a preset area, and collect raw temperature data from each temperature sensor in the IoT sensor array at its respective location.

[0107] Preferably, a temperature sensor array can be deployed in a preset area according to a hexagonal grid topology, with each node configured with a digital temperature sensor and the sampling frequency set to 2Hz in accordance with the ISO 7726 standard.

[0108] Step S12: All raw temperature data are aggregated to the regional gateway of the edge computing node via the Zigbee protocol to form the raw temperature field dataset.

[0109] Preferably, the raw temperature data T_raw(x,y,z,t) collected is aggregated to the regional gateway via the Zigbee protocol. The gateway performs data packet timestamp synchronization and removes data points outside the ±5σ (standard deviation) range.

[0110] Preferably, the regional gateway, as an intermediate-level device in the edge computing node architecture, is responsible for aggregating raw temperature data T_raw(x,y,z,t) from the IoT sensor array. Its core functions include data preprocessing (e.g., timestamp synchronization, outlier filtering) and protocol conversion. The edge computing nodes, located at a higher level, undertake complex computational tasks (e.g., constructing a three-dimensional temperature field matrix). Together, they form a three-tiered data processing system of "terminal-gateway-edge".

[0111] Step S13: Start the MQTT agent based on the edge computing node to encapsulate the original temperature field dataset into a structured message in JSON format.

[0112] Preferably, an MQTT broker is started on the edge computing node and QoS=1 is set to ensure at least one transmission. Sensor data is encapsulated in JSON format as a structured message {timestamp: Unix_ms, location: [x,y,z], value: T}, encrypted using TLS, and published to the "temperature / raw" topic. The edge node subscribes to this topic and establishes a data receiving buffer.

[0113] Step S14: The structured message is completed into a continuous temperature field by kriging interpolation through the edge computing node, resulting in an initial three-dimensional temperature field matrix with dimensions N×M×P×1, where N / M / P correspond to the number of discretized grids of the X / Y / Z axes of the preset region, respectively.

[0114] Preferably, every 30 seconds (which needs to be aligned with the LSTM time step below), the edge computing node reconstructs the received discrete data points into a continuous temperature field using the Kriging interpolation algorithm, generating an initial three-dimensional matrix T_init with dimensions N×M×P×1, where N / M / P correspond to the number of discretized grids along the X / Y / Z axes of the preset region, respectively, and the matrix element values ​​are smoothed by Laplace to eliminate interpolation noise.

[0115] Further, in step S2, the local temperature gradient features of the initial three-dimensional temperature field matrix are extracted using the 3D convolutional kernel of the ST-CNN spatiotemporal convolutional network. Then, the dynamic law of the time dimension of the initial three-dimensional temperature field matrix is ​​learned using the LSTM module of the ST-CNN spatiotemporal convolutional network, and a high-dimensional tensor fused with spatiotemporal features is output. Specifically, this includes the following steps:

[0116] Step S21: Extract features from the initial three-dimensional temperature field matrix using a set of 3D convolutional kernels of a preset size, and output a feature map with dimensions N×M×P×16 using the activation function of the ST-CNN spatiotemporal convolutional network.

[0117] Preferably, a group of 16 convolutional kernels (3×3×3) can be used for multi-scale feature extraction, with a stride of (1,1,1). Nonlinearity is introduced through the ReLU activation function, and the output feature map dimension is N×M×P×16.

[0118] Step S22: The feature map is reduced to N / 2×M / 2×P / 2×16 by the 3D max pooling layer of the ST-CNN spatiotemporal convolutional network to obtain 3D convolutional features.

[0119] Step S23: Expand the 3D convolutional features into temporal features according to the time series, and capture the temporal dependencies of the temporal features through a bidirectional LSTM structure with a hidden layer dimension of 64, and output the original LSTM output features with a dimension of 2×64×N / 2×M / 2×P / 2.

[0120] Preferably, the LSTM hidden layer dimension can be set to 64, and the LSTM time step size can be set to 30 seconds, which is aligned with the edge computing node acquisition step size mentioned above.

[0121] Preferably, the original output feature extraction of LSTM can be implemented using the following pseudocode:

[0122] def lstm_temporal_learning(input_tensor):

[0123] # Initialize LSTM cell parameters

[0124] lstm_cells = [LSTMCell(hidden_size=D) for _ in range(H×W)]

[0125] hidden_states = [zeros(D) for _ in range(H×W)]

[0126] cell_states = [zeros(D) for _ in range(H×W)]

[0127] # Time Series Processing

[0128] for t in range(T): # Iterate through the time dimension

[0129] spatial_features = input_tensor[t] # Get the features of the current time slice

[0130] for h in range(H):

[0131] for w in range(W):

[0132] # Calculate spatial location index

[0133] idx = h × W + w

[0134] # LSTM cell computation

[0135] input_gate = sigmoid(W_i × spatial_features[h,w] +U_i × hidden_states[idx])

[0136] forget_gate = sigmoid(W_f × spatial_features[h,w] +U_f × hidden_states[idx])

[0137] output_gate = sigmoid(W_o × spatial_features[h,w] +U_o × hidden_states[idx])

[0138] # Status Update

[0139] cell_states[idx] = forget_gate × cell_states[idx] +input_gate × tanh(W_c × spatial_features[h,w])

[0140] hidden_states[idx] = output_gate × tanh(cell_states[idx])

[0141] # Reorganization Spatial Characteristics

[0142] output_tensor = zeros(H, W, D)

[0143] for h in range(H):

[0144] for w in range(W):

[0145] idx = h × W + w

[0146] output_tensor[h,w] = hidden_states[idx]

[0147] return output_tensor

[0148] Step S24: The original output features of LSTM at all times are weighted and fused through a time attention mechanism to obtain a dynamic weight matrix with dimensions N / 2×M / 2×P / 2×T, where T is the time dimension.

[0149] Preferably, in the LSTM module of ST-CNN, the temporal attention mechanism achieves feature fusion by calculating the importance weights of features at each time step. First, a linear transformation is performed on the LSTM hidden state to obtain the query vector (Q), key vector (K), and value vector (V). Then, the weight distribution along the temporal dimension is calculated using scaled dot product attention. Finally, the weighted features are residually concatenated with the original LSTM output. The specific pseudocode is as follows:

[0150] def temporal_attention(lstm_output):

[0151] # Input shape: [batch_size, time_steps, hidden_dim]

[0152] Q = dense_layer(lstm_output) # Query vector projection

[0153] K = dense_layer(lstm_output) # Key vector projection

[0154] V = dense_layer(lstm_output) # Value vector projection

[0155] # Calculate attention score

[0156] attention_scores = tf.matmul(Q, K, transpose_b=True) / sqrt(hidden_dim)

[0157] attention_weights = softmax(attention_scores, axis=-1)

[0158] # Feature-weighted fusion

[0159] weighted_features = tf.matmul(attention_weights, V)

[0160] return lstm_output + weighted_features # Residual connection

[0161] Step S25: Perform Hadamard product operation on the dynamic weight matrix and the temporal features to obtain a weighted temporal feature tensor with dimensions of 2×64×N / 2×M / 2×P / 2.

[0162] Preferably, the Hadamard product operation is represented by F_weighted=W_dynamic⊙F_LSTM, where F_weighted is the weighted temporal feature tensor, W_dynamic is the dynamic weight matrix, F_LSTM is the weighted temporal feature tensor, and ⊙ represents element-wise multiplication.

[0163] Preferably, max pooling with a window size of 3 and a step size of 1 is performed on the weighted temporal feature tensor along the time dimension, compressing the output dimension to 2×64×N / 2×M / 2×P / 2.

[0164] Step S26: Reduce the number of channels of the weighted temporal feature tensor to 32 by 3D transpose convolution to obtain LSTM features.

[0165] Preferably, the pseudocode for obtaining LSTM features is as follows:

[0166] # Pseudocode Implementation

[0167] def dynamic_feature_fusion(conv_3d, lstm_out):

[0168] # Calculation of dynamic weights for spatial features

[0169] spatial_weights = softmax(conv1d(conv_3d.mean(dim=2,3), kernel_size=3)) # Attention along the time dimension

[0170] # Calculation of dynamic weights for time features

[0171] temporal_weights = sigmoid(linear(lstm_out)) # Gating based on LSTM hidden states

[0172] # Feature Alignment and Weighted Fusion

[0173] expanded_lstm = lstm_out.unsqueeze(2).unsqueeze(3) # Expand to [B,T,1,1,D]

[0174] weighted_conv = conv_3d × spatial_weights.unsqueeze(-1) # Spatial feature weighting

[0175] weighted_lstm = expanded_lstm × temporal_weights.unsqueeze(-1) # Time feature weighting

[0176] # Cross-modal stitching and dimensionality reduction

[0177] fused_tensor = concat([weighted_conv, weighted_lstm], dim=-1) # Concatenate channels

[0178] return layer_norm(fused_tensor) # Normalize the output

[0179] Step S27: Channel concatenation of 3D convolutional features and LSTM features is performed, and after normalization, a 48-channel hybrid tensor is obtained.

[0180] Step S28: The hybrid tensor is weighted twice by spatial attention mechanism, and the output of global average pooling is used to obtain a 32-dimensional feature vector, which is the high-dimensional tensor.

[0181] Preferably, the pseudocode for the double-weighted method is as follows:

[0182] def spatial_attention_pooling(input_tensor, output_dims=32):

[0183] # Channel Compression

[0184] compressed = Conv2D(filters=input_tensor.shape[-1] / / 4, kernel_size=1)(input_tensor)

[0185] compressed = ReLU()(compressed)

[0186] # Spatial Attention Generation

[0187] avg_pool = tf.reduce_mean(compressed, axis=-1, keepdims=True)

[0188] max_pool = tf.reduce_max(compressed, axis=-1, keepdims=True)

[0189] concat = tf.concat([avg_pool, max_pool], axis=-1)

[0190] attention = Conv2D(filters=1, kernel_size=7, padding='same',activation='sigmoid')(concat)

[0191] # Feature weighting

[0192] weighted = Multiply()([input_tensor, attention])

[0193] # Global pooling output

[0194] output = GlobalAveragePooling2D()(weighted)

[0195] output = Dense(output_dims)(output) # Ensure the output is 32-dimensional

[0196] return output

[0197] Further, in step S3, a multi-objective optimization model is constructed based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort index. A high-dimensional tensor is input into the multi-objective optimization model, and the Pareto front of the multi-objective optimization model is solved using the NSGA-III algorithm to obtain the air conditioning group control parameters. Specifically, this includes the following steps:

[0198] Step S31: Define the objective function for electrothermal conversion efficiency, construct the objective function for the coefficient of performance (COP) using linear interpolation, and construct the objective function for the human comfort index based on the PMV-PPD standard.

[0199] Step S32: Perform Min-Max normalization on the three objective functions and add preset constraints to each.

[0200] Preferably, preset constraints include, for example, hard constraints (temperature 18-26℃, humidity 40%-60%) and soft constraints (COP≥3.0).

[0201] Step S33: Define the target space by using the electrothermal conversion efficiency, the coefficient of performance (COP), and the human comfort index as the three coordinate axes of the target space.

[0202] Preferably, the electrothermal conversion efficiency, the coefficient of performance (COP), and the human comfort index can be used as the X, Y, and Z axes of the target space, respectively.

[0203] Step S34: Define the feasible region of the target space through all preset constraints, and map the high-dimensional tensor to the target space through linear transformation to obtain the prediction vector based on three objective functions.

[0204] Preferably, linear transformations can be implemented using PyTorch code:

[0205] import torch

[0206] import torch.nn as nn

[0207] class MultiObjectiveMapper(nn.Module):

[0208] def __init__(self, input_dims):

[0209] super().__init__()

[0210] self.flatten = nn.Flatten()

[0211] total_features = 1

[0212] for dim in input_dims[1:]: # Ignore batch dimension

[0213] total_features ×= dim

[0214] # Mapping layer of three objective functions

[0215] self.eta_layer = nn.Linear(total_features, 1)

[0216] self.cop_layer = nn.Linear(total_features, 1)

[0217] self.pmv_layer = nn.Linear(total_features, 1)

[0218] def forward(self, x):

[0219] x_flat = self.flatten(x)

[0220] eta = torch.sigmoid(self.eta_layer(x_flat)) # Electrothermal efficiency [0,1]

[0221] cop = torch.relu(self.cop_layer(x_flat)) # Coefficient of performance (COP) ≥ 0

[0222] pmv = torch.tanh(self.pmv_layer(x_flat)) × 2.5 # PMV[-2.5,2.5]

[0223] return torch.cat([eta, cop, pmv], dim=1)

[0224] Step S35: Define several random air conditioning group control parameter combinations based on the NSGA-III algorithm, and input all random air conditioning group control parameter combinations into the target space to form several target vectors.

[0225] Preferably, an initial population of size N=100 can be generated, and combinations of air conditioning group control parameters (such as set temperature, wind speed, and air supply angle) can be randomly generated, ensuring that the parameters are within the constraints.

[0226] Step S36: Update all target vectors iteratively through non-dominated sorting, and calculate the fitness value of each target vector based on the prediction vector in each update iteration.

[0227] Preferably, updating all target vectors iteratively through non-dominated sorting can be implemented using the following pseudocode:

[0228] Input:

[0229] - Population size N, maximum number of iterations T

[0230] - Reference point set Z (generated using a uniform distribution based on the number of targets M=3)

[0231] - The high-dimensional tensor X output by ST-CNN is used as the initial solution.

[0232] Output:

[0233] - Pareto optimal solution set (air conditioning group control parameters)

[0234] Begin:

[0235] Initialize population P0 = GenerateInitialPopulation(X) # Generate initial parameters from high-dimensional tensors

[0236] for t = 1 to T do

[0237] Q_t = GeneticOperation(P_t-1) # Crossover mutation

[0238] R_t = P_t-1 ∪ Q_t # Merge parent and child generations

[0239] F = FastNonDominatedSort(R_t) # Fast Non-Dominated Sort

[0240] P_t = Φ

[0241] i = 1

[0242] while |P_t| + |F_i| ≤ N do

[0243] P_t = P_t ∪ F_i # Fill the new population layer by layer

[0244] i = i + 1

[0245] end while

[0246] if |P_t| < N then

[0247] NormalizeObjectives(F_i) # Normalize the target values

[0248] GenerateReferencePoints(Z, M) # Generate 3D reference points

[0249] AssociateSolutions(F_i, Z) # Associated solutions and reference points

[0250] NichingSelection(F_i, Z, N-|P_t|) # Niche selection based on reference point

[0251] end if

[0252] UpdateIdealPoint(R_t) # Update the ideal point

[0253] end for

[0254] return NonDominatedSolutions(P_t) # Return to the Pareto frontier

[0255] End

[0256] Step S37: After the preset iteration termination condition is met, output the Pareto front solution set, which includes several combinations of non-dominated air conditioning group control parameters.

[0257] Preferably, the population is iteratively updated (maximum number of iterations T=50) through non-dominated sorting, crowding calculation and reference point selection of NSGA-III, and finally the Pareto front solution set is output, which contains multiple non-dominated air conditioning group control parameter combinations, that is, the preset iteration termination condition is the maximum number of iterations T=50.

[0258] Further, in step S5, a CFD simulation environment equipped with a turbulence model is constructed based on a preset region, and the initial three-dimensional temperature field matrix and device response dataset are input into the turbulence model. The actual temperature field dataset is obtained through dynamic simulation in the CFD simulation environment. Specifically, this includes the following steps:

[0259] Step S51: Define a turbulence model based on the geometric features of the preset area and the air supply characteristics of the air conditioning system.

[0260] Preferably, a k-ε or k-ω SST turbulence model is selected based on the geometric characteristics of the preset area (such as floor height and spatial layout) and the characteristics of the air conditioning system (such as air supply method and airflow organization). The closure coefficient is calibrated through experimental data, and the turbulence intensity is dynamically adjusted according to the wind speed parameters in the equipment response dataset to ensure that the model can accurately capture the mixing process of the air conditioning outlet air and the indoor air.

[0261] Step S52: Divide the initial three-dimensional temperature field matrix into a computational domain mesh using an unstructured mesh, and build a solver environment in the CFD simulation environment.

[0262] Preferably, based on the spatial coordinate system of the initial three-dimensional temperature field matrix, an unstructured mesh generation method is used to generate the computational domain mesh, wherein the near-wall region is subjected to y+<1 densification processing, and key areas such as air conditioning outlets, return air inlets, and personnel activity areas are set as boundary conditions, wherein the outlet velocity field is calculated based on the fan frequency parameters in the equipment response dataset.

[0263] Step S53: Perform time-progression calculations through the solver environment and record the actual temperature of each node in the computational domain grid in real time.

[0264] Preferably, a solver environment is built in OpenFOAM or ANSYS Fluent, the coupled calculation of the energy equation and the turbulence model is enabled, the time step is set to 0.1 seconds to match the dynamic changes of the temperature field, and the PISO algorithm is used to handle the pressure-velocity coupling. The initial three-dimensional temperature field matrix is ​​used as the initial condition, and the operating parameters in the equipment response dataset (such as compressor power and valve opening) are used as transient boundary conditions.

[0265] Step S54: Integrate the actual temperatures of all nodes into an actual temperature field dataset.

[0266] Preferably, time-progression calculations are performed using a solver, and the temperature, velocity, and pressure field data of each node in the computational domain are recorded in real time. The calculation is terminated when the simulation time reaches the typical response cycle of the air conditioning system (e.g., 5 minutes). The simulation results are then reorganized according to the same spatial dimension of the initial three-dimensional temperature field matrix, and the actual temperature field dataset containing timestamps is output.

[0267] Further, in step S7, the air conditioning group control parameters are iteratively updated using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thereby obtaining the Pareto optimal solution set of the air conditioning group control parameters. This specifically includes the following steps:

[0268] Step S71: Construct a dual-head neural network policy model with two independent output heads, namely the parameter control head and the value head.

[0269] Preferably, the dual-head neural network strategy model includes a shared feature extraction layer (3 fully connected layers, 128 neurons per layer, ReLU activation) and two independent output heads. The parameter control head outputs the mean vector of the air conditioning group control parameters (temperature setting, fan speed, and airflow angle), while the value head outputs the state value function estimate for dominance function calculation.

[0270] Preferably, a tanh activation function needs to be applied to the parameter control head to map the parameter values ​​to the [-1,1] interval, and then convert them to the actual physical range through a linear transformation (e.g., temperature 16-26℃ corresponds to [-1,1]→[0.5,1.0]).

[0271] Step S72: Periodically acquire the deviation matrix, the L2 norm of the deviation matrix, and the air conditioning group control parameters based on the preset acquisition interval, and integrate them into trajectory data.

[0272] Preferably, the current strategy is executed in a CFD simulation environment to collect trajectory data containing state (deviation matrix), action (group control parameters), and reward (deviation matrix norm). Each episode samples 50 steps to correspond to 25 seconds of physical time.

[0273] Step S73: Calculate the state value function of the trajectory data using the value head, and calculate the dominance function of the state value function using the generalized dominance estimation method.

[0274] The generalized advantage estimation (GAE) method is adopted, with λ=0.95 and discount factor γ=0.99, and the advantage function A_t is calculated through the state value function output by the value head.

[0275] Step S74: The trajectory data and the advantage function of the same preset acquisition interval are bound into a training sample, and a training sample set is obtained by integrating all preset acquisition intervals.

[0276] Step S75, define the pruning objective function of the dual-head neural network strategy model according to equation (1):

[0277] (1).

[0278] in, To prune the objective function, This is the weight set of the dual-head neural network strategy model. The ratio of the action probabilities of the new strategy to those of the old strategy. For the dominant function, The pruning threshold is the pruning objective function.

[0279] Preferably, the trimming threshold is typically between 0.1 and 0.3.

[0280] Preferably, the pruning mechanism of the objective function is when Exceeding When the target value is within the specified range, the target value is forcibly truncated to the boundary value. At the same time, the smaller value between the original target and the truncated target is taken to ensure that the policy update does not deviate too far from the old policy.

[0281] Step S76: Input the training sample set into the pruning objective function, and iterate and update the dual-head neural network policy model for several epochs using the Adam optimizer.

[0282] Preferably, the policy network can be updated using the Adam optimizer every 10 epochs, with gradients computed using a mini-batch of 32 samples per epoch, and a KL divergence constraint (KL threshold 0.01) applied.

[0283] Step S77: When the rate of change of the HV index of the Pareto front solution is less than the preset percentage threshold for several consecutive iterations, the iteration is terminated, and the air conditioning group control parameters output by the parameter control head are obtained as the Pareto optimal solution set.

[0284] Preferably, iteration termination is triggered when the rate of change of the Hypervolume (HV) index of the Pareto front solution is less than 1% for five consecutive iterations, or when the KL divergence remains in the range [0.01, 0.05] and the variance of the dominance function decreases to less than 10% of its initial value, or when the moving average fluctuation of the strategy loss and value loss does not exceed ±5% for three consecutive epochs. For higher data accuracy, it can be set to terminate only when all three conditions are met simultaneously. The pseudocode is as follows:

[0285] # Pseudocode implementation example

[0286] def should_terminate(pareto_front, ppo_metrics):

[0287] # Convergence check

[0288] hv_change = calc_hypervolume_change(pareto_front[-5:])

[0289] kl_stable = check_kl_divergence(ppo_metrics.kl_history)

[0290] # Constraint Satisfaction Check

[0291] temp_valid = all(16 <= x.temp <= 26 for x in pareto_front[-1])

[0292] energy_comfort = calc_dominated_solutions(pareto_front,

[0293] energy_thresh=0.9,

[0294] pmv_range=(-0.5,0.5))

[0295] # PPO Indicator Check

[0296] clip_violation = ppo_metrics.clip_ratio > 0.95

[0297] advantage_stable = np.mean(ppo_metrics.advantage_std) < 0.1

[0298] return (hv_change < 0.01 and kl_stable and temp_valid

[0299] and energy_comfort >= 0.8 and not clip_violation

[0300] and advantage_stable)

[0301] Further, step S8, converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending it to the air conditioning system, specifically includes the following steps:

[0302] Step S81: Convert the air conditioning group control parameters in the Pareto optimal solution set into physical parameters through linear transformation.

[0303] Preferably, the multidimensional parameter vector (including temperature setting, wind speed, air supply angle, etc.) in the Pareto optimal solution set can be converted into physical quantities through a predefined linear transformation matrix. The temperature setting value is constrained to the range of 16-26℃ by the tanh activation function, the wind speed parameter is mapped to a percentage value of 0-100%, and the air supply angle is converted into an absolute angle value of 0-360°.

[0304] Step S82: Encode the physical parameters into a data unit structure that conforms to the air conditioning system communication protocol.

[0305] Preferably, control command frames can be constructed according to the communication protocol supported by the air conditioning system (such as BACnet or Modbus), and the physical parameter values ​​can be encoded into the data format specified by the protocol, including the device address (such as air conditioning unit ID), parameter type (such as temperature setpoint), parameter value (such as 23.5℃) and check bit, to ensure that the command conforms to the data unit structure of the ISO 16484-5 standard.

[0306] Step S83: The data unit structure is encapsulated by the OPC-UA server of the edge computing node and sent to the air conditioning system.

[0307] Preferably, the encapsulated instructions can be pushed to the air conditioning group control system through the OPC-UA server of the edge computing node. The publish-subscribe model is used to ensure the real-time performance of the instructions. The transmission delay of the temperature control instructions needs to be controlled within 500ms. At the same time, the timestamp of the instruction sending and the response time of the air conditioning system are recorded for subsequent strategy verification.

[0308] Preferably, after the command is sent, the actual operating parameters of the air conditioning system (such as the air outlet temperature and the compressor frequency) are continuously monitored, and the feedback data is sent back to the edge computing node through the MQTT protocol. The data is compared with the expected parameters to generate a control error signal, which will serve as the input constraint for the next round of Pareto front optimization.

[0309] This embodiment acquires the original temperature field dataset of a preset area through an IoT sensor array and transmits it to an edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix. Local temperature gradient features of the initial three-dimensional temperature field matrix are extracted using the 3D convolutional kernels of the ST-CNN spatiotemporal convolutional network. The LSTM module of the ST-CNN spatiotemporal convolutional network then learns the dynamic temporal dimension of the initial three-dimensional temperature field matrix, outputting a high-dimensional tensor that integrates spatiotemporal features. A multi-objective optimization model is constructed based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort indices. The high-dimensional tensor is input into the multi-objective optimization model, and the Pareto front of the multi-objective optimization model is solved using the NSGA-III algorithm. The process involves obtaining the air conditioning group control parameters; sending these parameters to the air conditioning system and collecting the equipment response dataset; constructing a CFD simulation environment with a turbulence model based on a preset area; inputting the initial three-dimensional temperature field matrix and equipment response dataset into the turbulence model; dynamically simulating the actual temperature field dataset through the CFD simulation environment; obtaining a predefined expected temperature field; calculating the deviation matrix by difference between the actual temperature field dataset and the expected temperature field dataset; iteratively updating the air conditioning group control parameters using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thus obtaining the Pareto optimal solution set for the air conditioning group control parameters; and converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending them to the air conditioning system. In terms of energy efficiency optimization, this embodiment extracts the spatiotemporal features of the temperature field through the ST-CNN spatiotemporal convolutional network and solves the Pareto front of the multi-objective optimization model using the NSGA-III algorithm, enabling the system to achieve an optimal balance between electrothermal conversion efficiency, coefficient of performance (COP), and human comfort. Secondly, in terms of control accuracy, a 3D convolutional kernel and LSTM module are used to collaboratively analyze the dynamic laws of the temperature field, and CFD turbulence model simulation verification is performed, which can reduce the temperature control error to a small range (±0.5℃), significantly improving the uniformity of spatial temperature. Finally, in terms of system intelligence, the near-end strategy optimization algorithm realizes the autonomous iterative optimization of control parameters, forming a closed-loop control system of "perception-decision-execution-feedback", reducing or eliminating the need for manual intervention. At the same time, the deployment of IoT sensor arrays can shorten the equipment status monitoring response time to the millisecond level, making this embodiment suitable for scenarios with high requirements for temperature control accuracy and energy efficiency, such as large commercial complexes, data centers, production workshops, and warehouses.

[0310] like Figure 2 As shown, this embodiment provides an example of a multi-energy coupled temperature field control system based on intelligent optimization. In this embodiment, the multi-energy coupled temperature field control system is applied to the multi-energy coupled temperature field control method as described in the above embodiment.

[0311] Specifically, the multi-energy coupled temperature field control system includes an initial three-dimensional temperature field matrix acquisition module 1, a high-dimensional tensor output module 2, an air conditioning group control parameter solving module 3, an equipment response dataset acquisition module 4, an actual temperature field dataset dynamic simulation module 5, a deviation matrix calculation module 6, an air conditioning group control parameter optimal solution set acquisition module 7, and an air conditioning group control parameter optimal solution set transmission module 8, which are sequentially electrically or signalally connected.

[0312] The module 1 for acquiring the initial 3D temperature field matrix is ​​used to acquire the original temperature field dataset of a preset area through an IoT sensor array, and transmit the original temperature field dataset to the edge computing node via the MQTT protocol to form the initial 3D temperature field matrix. The module 2 for high-dimensional tensor output is used to extract the local temperature gradient features of the initial 3D temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial 3D temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network to output a high-dimensional tensor that integrates spatiotemporal features. The module 3 for solving the air conditioning group control parameters is used to construct a multi-objective optimization model based on electrothermal conversion efficiency, cooling coefficient, and human comfort index, and input the high-dimensional tensor into the multi-objective optimization model. The Pareto front of the multi-objective optimization model is solved by the NSGA-III algorithm to obtain the air conditioning group control parameters. The module 3 for acquiring the equipment response dataset is also included. The data collection module 4 is used to send the air conditioning group control parameters to the air conditioning system and collect the equipment response dataset of the air conditioning system; the actual temperature field dataset dynamic simulation module 5 is used to construct a CFD simulation environment equipped with a turbulence model based on a preset area, and input the initial three-dimensional temperature field matrix and the equipment response dataset into the turbulence model, and obtain the actual temperature field dataset through dynamic simulation in the CFD simulation environment; the deviation matrix calculation module 6 is used to obtain the predefined expected temperature field, and calculate the deviation matrix by difference between the actual temperature field dataset and the expected temperature field dataset; the air conditioning group control parameter optimal solution set acquisition module 7 is used to iteratively update the air conditioning group control parameters through a near-end strategy optimization algorithm to gradually approach the minimum value of the deviation matrix, and obtain the Pareto optimal solution set of the air conditioning group control parameters; the air conditioning group control parameter optimal solution set sending module 8 is used to convert the Pareto optimal solution set into readable instructions of the air conditioning system and send it to the air conditioning system.

[0313] Furthermore, the initial three-dimensional temperature field matrix acquisition module 1 specifically includes a first initial three-dimensional temperature field matrix acquisition unit, a second initial three-dimensional temperature field matrix acquisition unit, a third initial three-dimensional temperature field matrix acquisition unit, and a fourth initial three-dimensional temperature field matrix acquisition unit that are electrically or signal-connected in sequence; the fourth initial three-dimensional temperature field matrix acquisition unit is electrically or signal-connected to the high-dimensional tensor output module 2.

[0314] The system comprises four main components: a first initial three-dimensional temperature field matrix acquisition unit, a second initial three-dimensional temperature field matrix acquisition unit, and a third initial three-dimensional temperature field matrix acquisition unit. The first initial three-dimensional temperature field matrix acquisition unit is used to deploy an IoT sensor array based on a hexagonal grid topology within a preset region, and to collect raw temperature data from each temperature sensor in the IoT sensor array at its respective location. The second initial three-dimensional temperature field matrix acquisition unit is used to aggregate all raw temperature data to the regional gateway of the edge computing node via the Zigbee protocol, forming a raw temperature field dataset. The third initial three-dimensional temperature field matrix acquisition unit is used to start an MQTT agent based on the edge computing node, and to encapsulate the raw temperature field dataset into a structured message in JSON format. The fourth initial three-dimensional temperature field matrix acquisition unit is used to perform Kriging interpolation on the structured message through the edge computing node to complete it into a continuous temperature field, resulting in an initial three-dimensional temperature field matrix with dimensions N×M×P×1, where N / M / P correspond to the number of discretized grids along the X / Y / Z axes of the preset region, respectively.

[0315] Furthermore, the high-dimensional tensor output module 2 specifically includes a first high-dimensional tensor output unit, a second high-dimensional tensor output unit, a third high-dimensional tensor output unit, a fourth high-dimensional tensor output unit, a fifth high-dimensional tensor output unit, a sixth high-dimensional tensor output unit, a seventh high-dimensional tensor output unit, and an eighth high-dimensional tensor output unit that are electrically or signal-connected in sequence; the first high-dimensional tensor output unit is electrically or signal-connected to the fourth initial three-dimensional temperature field matrix acquisition unit, and the eighth high-dimensional tensor output unit is electrically or signal-connected to the air conditioning group control parameter solving module 3.

[0316] The system comprises four high-dimensional tensor output units: the first high-dimensional tensor output unit extracts features from the initial three-dimensional temperature field matrix using a pre-defined group of 3D convolutional kernels, and outputs a feature map of dimension N×M×P×16 through the activation function of the ST-CNN spatiotemporal convolutional network; the second high-dimensional tensor output unit reduces the dimension of the feature map to N / 2×M / 2×P / 2×16 through the 3D max pooling layer of the ST-CNN spatiotemporal convolutional network, obtaining 3D convolutional features; the third high-dimensional tensor output unit unfolds the 3D convolutional features into temporal features according to the time sequence, and captures the temporal dependencies of the temporal features through a bidirectional LSTM structure with a hidden layer dimension of 64, outputting the original LSTM output features of dimension 2×64×N / 2×M / 2×P / 2; and the fourth high-dimensional tensor output unit performs weighted fusion through a temporal attention mechanism. The original LSTM output features at all time points are used to obtain a dynamic weight matrix of dimension N / 2×M / 2×P / 2×T, where T is the time dimension. The fifth high-dimensional tensor output unit is used to perform a Hadamard product operation on the dynamic weight matrix and the temporal features to obtain a weighted temporal feature tensor of dimension 2×64×N / 2×M / 2×P / 2. The sixth high-dimensional tensor output unit is used to reduce the number of channels of the weighted temporal feature tensor to 32 through 3D transposed convolution to obtain LSTM features. The seventh high-dimensional tensor output unit is used to concatenate the 3D convolution features and LSTM features through channels, and after normalization, a 48-channel hybrid tensor is obtained. The eighth high-dimensional tensor output unit is used to perform secondary weighting on the hybrid tensor through a spatial attention mechanism, and output a 32-dimensional feature vector through global average pooling, which is the high-dimensional tensor.

[0317] Furthermore, the air conditioning group control parameter solving module 3 specifically includes a first air conditioning group control parameter solving unit, a second air conditioning group control parameter solving unit, a third air conditioning group control parameter solving unit, a fourth air conditioning group control parameter solving unit, a fifth air conditioning group control parameter solving unit, a sixth air conditioning group control parameter solving unit, and a seventh air conditioning group control parameter solving unit that are electrically or signal-connected in sequence; the first air conditioning group control parameter solving unit is electrically or signal-connected to the eighth high-dimensional tensor output unit, and the seventh air conditioning group control parameter solving unit is electrically or signal-connected to the equipment response dataset acquisition module 4.

[0318] The system comprises four main components: a first air conditioning group control parameter solving unit, a second air conditioning group control parameter solving unit, and a third air conditioning group control parameter solving unit. The first unit defines the objective function for electrothermal conversion efficiency, the objective function for the coefficient of performance (COP) using linear interpolation, and the fourth unit for the human comfort index based on the PMV-PPD standard. The third unit uses electrothermal conversion efficiency, COP, and the human comfort index as the three coordinate axes of the objective space to define the objective space. The fourth unit defines the feasible region of the objective space using all preset constraints and applies the high-dimensional tensor to the target space. A linear transformation is applied to the target space to obtain a prediction vector based on three objective functions. The fifth air conditioning group control parameter solving unit is used to define several random air conditioning group control parameter combinations based on the NSGA-III algorithm, and input all random air conditioning group control parameter combinations into the target space to form several target vectors. The sixth air conditioning group control parameter solving unit is used to update all target vectors iteratively through non-dominated sorting, and calculate the fitness value of each target vector based on the prediction vector in each update iteration. The seventh air conditioning group control parameter solving unit is used to output the Pareto front solution set including several non-dominated air conditioning group control parameter combinations after the preset iteration termination condition is reached.

[0319] Furthermore, the actual temperature field dataset dynamic simulation module 5 specifically includes a first actual temperature field dataset dynamic simulation unit, a second actual temperature field dataset dynamic simulation unit, a third actual temperature field dataset dynamic simulation unit, and a fourth actual temperature field dataset dynamic simulation unit that are electrically or signal-connected in sequence; the first actual temperature field dataset dynamic simulation unit is electrically or signal-connected to the equipment response dataset acquisition module 4; and the fourth actual temperature field dataset dynamic simulation unit is electrically or signal-connected to the deviation matrix calculation module 6.

[0320] The first actual temperature field dataset dynamic simulation unit is used to define a turbulence model based on the geometric features of a preset region and the air supply characteristics of the air conditioning system; the second actual temperature field dataset dynamic simulation unit is used to divide the initial three-dimensional temperature field matrix into a computational domain grid through an unstructured grid and build a solver environment in the CFD simulation environment; the third actual temperature field dataset dynamic simulation unit is used to perform time-progression calculations through the solver environment and record the actual temperature of each node in the computational domain grid in real time; the fourth actual temperature field dataset dynamic simulation unit is used to integrate the actual temperatures of all nodes into an actual temperature field dataset.

[0321] Furthermore, the optimal solution set acquisition module 7 for air conditioning group control parameters specifically includes a first optimal solution set acquisition unit for air conditioning group control parameters, a second optimal solution set acquisition unit for air conditioning group control parameters, a third optimal solution set acquisition unit for air conditioning group control parameters, a fourth optimal solution set acquisition unit for air conditioning group control parameters, a fifth optimal solution set acquisition unit for air conditioning group control parameters, a sixth optimal solution set acquisition unit for air conditioning group control parameters, and a seventh optimal solution set acquisition unit for air conditioning group control parameters, which are electrically or signal-connected to each other. The first optimal solution set acquisition unit for air conditioning group control parameters is electrically or signal-connected to the deviation matrix calculation module 6, and the seventh optimal solution set acquisition unit for air conditioning group control parameters is electrically or signal-connected to the optimal solution set transmission module 8 for air conditioning group control parameters.

[0322] Among them, the first air conditioning group control parameter optimal solution set acquisition unit is used to construct a dual-head neural network strategy model with two independent output heads, namely the parameter control head and the value head.

[0323] The second air conditioning group control parameter optimal solution set acquisition unit is used to periodically acquire the deviation matrix, the L2 norm of the deviation matrix, and the air conditioning group control parameters based on a preset acquisition interval, and integrate them into trajectory data.

[0324] The third air conditioning group control parameter optimal solution set acquisition unit is used to calculate the state value function of the trajectory data through the value head, and to calculate the dominance function of the state value function through the generalized dominance estimation method.

[0325] The fourth air conditioning group control parameter optimal solution set acquisition unit is used to bind the trajectory data and advantage function of the same preset acquisition interval into a training sample, and integrate them based on all preset acquisition intervals to obtain a training sample set.

[0326] The fifth air conditioning group control parameter optimal solution set acquisition unit is used to define the pruning objective function of the dual-head neural network strategy model according to equation (1):

[0327] (1).

[0328] in, To prune the objective function, This is the weight set of the dual-head neural network strategy model. The ratio of the action probabilities of the new strategy to those of the old strategy. For the dominant function, The pruning threshold is the pruning objective function.

[0329] The sixth unit for obtaining the optimal solution set of air conditioning group control parameters is used to input the training sample set into the pruning objective function and to iteratively update the dual-head neural network policy model for several epochs through the Adam optimizer.

[0330] The seventh air conditioning group control parameter optimal solution set acquisition unit is used to terminate the iteration when the rate of change of the HV index of the Pareto front solution is less than a preset percentage threshold for several consecutive iterations, and the air conditioning group control parameters output by the parameter control head are the Pareto optimal solution set.

[0331] Furthermore, the optimal solution set transmission module 8 for air conditioning group control parameters specifically includes a first optimal solution set transmission unit for air conditioning group control parameters, a second optimal solution set transmission unit for air conditioning group control parameters, and a third optimal solution set transmission unit for air conditioning group control parameters, which are electrically or signal-connected in sequence; the first optimal solution set transmission unit for air conditioning group control parameters is electrically or signal-connected to the seventh optimal solution set acquisition unit for air conditioning group control parameters.

[0332] The first air conditioning group control parameter optimal solution set sending unit is used to convert the air conditioning group control parameters in the Pareto optimal solution set into physical parameters through linear transformation; the second air conditioning group control parameter optimal solution set sending unit is used to encode the physical parameters into a data unit structure that conforms to the air conditioning system communication protocol; and the third air conditioning group control parameter optimal solution set sending unit is used to encapsulate the data unit structure through the OPC-UA server of the edge computing node and send it to the air conditioning system.

[0333] It should be noted that this embodiment is a functional module embodiment based on the above method embodiment. For the preferred, extended, limited, exemplified and principle explanation parts of this embodiment, please refer to the above embodiment. This embodiment will not repeat them.

[0334] This embodiment acquires the original temperature field dataset of a preset area through an IoT sensor array and transmits it to an edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix. Local temperature gradient features of the initial three-dimensional temperature field matrix are extracted using the 3D convolutional kernels of the ST-CNN spatiotemporal convolutional network. The LSTM module of the ST-CNN spatiotemporal convolutional network then learns the dynamic temporal dimension of the initial three-dimensional temperature field matrix, outputting a high-dimensional tensor that integrates spatiotemporal features. A multi-objective optimization model is constructed based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort indices. The high-dimensional tensor is input into the multi-objective optimization model, and the Pareto front of the multi-objective optimization model is solved using the NSGA-III algorithm. The process involves obtaining the air conditioning group control parameters; sending these parameters to the air conditioning system and collecting the equipment response dataset; constructing a CFD simulation environment with a turbulence model based on a preset area; inputting the initial three-dimensional temperature field matrix and equipment response dataset into the turbulence model; dynamically simulating the actual temperature field dataset through the CFD simulation environment; obtaining a predefined expected temperature field; calculating the deviation matrix by difference between the actual temperature field dataset and the expected temperature field dataset; iteratively updating the air conditioning group control parameters using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thus obtaining the Pareto optimal solution set for the air conditioning group control parameters; and converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending them to the air conditioning system. In terms of energy efficiency optimization, this embodiment extracts the spatiotemporal features of the temperature field through the ST-CNN spatiotemporal convolutional network and solves the Pareto front of the multi-objective optimization model using the NSGA-III algorithm, enabling the system to achieve an optimal balance between electrothermal conversion efficiency, coefficient of performance (COP), and human comfort. Secondly, in terms of control accuracy, a 3D convolutional kernel and LSTM module are used to collaboratively analyze the dynamic laws of the temperature field, and CFD turbulence model simulation verification is performed, which can reduce the temperature control error to a small range (±0.5℃), significantly improving the uniformity of spatial temperature. Finally, in terms of system intelligence, the near-end strategy optimization algorithm realizes the autonomous iterative optimization of control parameters, forming a closed-loop control system of "perception-decision-execution-feedback", reducing or eliminating the need for manual intervention. At the same time, the deployment of IoT sensor arrays can shorten the equipment status monitoring response time to the millisecond level, making this embodiment suitable for scenarios with high requirements for temperature control accuracy and energy efficiency, such as large commercial complexes, data centers, production workshops, and warehouses.

[0335] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 3 As shown, the electronic device 9 includes a processor 91 and a memory 92 coupled to the processor 91.

[0336] The memory 92 stores program instructions for implementing the intelligently optimized multi-energy coupled temperature field control method of any of the above embodiments.

[0337] The processor 91 is used to execute program instructions stored in the memory 92 for intelligent optimization-based multi-energy coupled temperature field control.

[0338] The processor 91 can also be referred to as a CPU (Central Processing Unit). The processor 91 may be an integrated circuit chip with signal processing capabilities. The processor 91 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0339] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 The storage medium 10 in this embodiment stores program instructions 101 capable of implementing all the above methods. These program instructions 101 can be stored in the storage medium as a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0340] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0341] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-energy coupled temperature field control method based on intelligent optimization, wherein the multi-energy coupled temperature field control method is used to control the temperature field within a preset area through an air conditioning system, the preset area having an Internet of Things sensor array, characterized in that, The multi-energy coupled temperature field control method includes: Step S1: Obtain the original temperature field dataset of the preset area through the IoT sensor array, and transmit the original temperature field dataset to the edge computing node through the MQTT protocol to form an initial three-dimensional temperature field matrix; Step S2: Extract the local temperature gradient features of the initial three-dimensional temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial three-dimensional temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network to output a high-dimensional tensor that integrates spatiotemporal features. Step S3: Construct a multi-objective optimization model based on electrothermal conversion efficiency, cooling coefficient, and human comfort index, and input the high-dimensional tensor into the multi-objective optimization model. Solve the Pareto front of the multi-objective optimization model using the NSGA-III algorithm to obtain the air conditioning group control parameters. Step S4: Send the air conditioning group control parameters to the air conditioning system and collect the equipment response dataset of the air conditioning system; Step S5: Construct a CFD simulation environment equipped with a turbulence model based on the preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model. The actual temperature field dataset is obtained through dynamic simulation in the CFD simulation environment. Step S6: Obtain the predefined expected temperature field, and perform a difference calculation between the actual temperature field dataset and the expected temperature field dataset to obtain the deviation matrix; Step S7: The air conditioning group control parameters are iteratively updated using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, thereby obtaining the Pareto optimal solution set of the air conditioning group control parameters; Step S8: Convert the Pareto optimal solution set into readable instructions for the air conditioning system and send them to the air conditioning system.

2. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S1: Obtain the original temperature field dataset of the preset area through the IoT sensor array, and transmit the original temperature field dataset to the edge computing node via the MQTT protocol to form an initial three-dimensional temperature field matrix, including: Step S11: Deploy an IoT sensor array based on the preset area according to a hexagonal grid topology, and collect raw temperature data at their respective locations through each temperature sensor in the IoT sensor array; Step S12: All raw temperature data are aggregated to the regional gateway of the edge computing node via the Zigbee protocol to form the raw temperature field dataset; Step S13: Start the MQTT agent based on the edge computing node to encapsulate the original temperature field dataset into a structured message in JSON format; Step S14: The structured message is completed into a continuous temperature field by kriging interpolation through the edge computing node, resulting in an initial three-dimensional temperature field matrix with dimensions N×M×P×1, where N / M / P correspond to the number of discretized grids along the X / Y / Z axes of the preset region, respectively.

3. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S2: Extract local temperature gradient features of the initial three-dimensional temperature field matrix using the 3D convolutional kernels of the ST-CNN spatiotemporal convolutional network. Then, learn the dynamic laws of the temporal dimension of the initial three-dimensional temperature field matrix using the LSTM module of the ST-CNN spatiotemporal convolutional network, and output a high-dimensional tensor that fuses spatiotemporal features, including: Step S21: Extract features from the initial three-dimensional temperature field matrix using a 3D convolution kernel group of a preset size, and output a feature map with a dimension of N×M×P×16 through the activation function of the ST-CNN spatiotemporal convolutional network. Step S22: The feature map is reduced to N / 2×M / 2×P / 2×16 by the 3D max pooling layer of the ST-CNN spatiotemporal convolutional network to obtain 3D convolutional features; Step S23: Expand the 3D convolutional features into temporal features according to the time sequence, and capture the temporal dependencies of the temporal features through a bidirectional LSTM structure with a hidden layer dimension of 64, and output the original LSTM output features with a dimension of 2×64×N / 2×M / 2×P / 2. Step S24: The original output features of LSTM at all times are weighted and fused through a time attention mechanism to obtain a dynamic weight matrix with dimensions N / 2×M / 2×P / 2×T, where T is the time dimension; Step S25: Perform a Hadamard product operation between the dynamic weight matrix and the temporal features to obtain a weighted temporal feature tensor with dimensions of 2×64×N / 2×M / 2×P / 2; Step S26: Reduce the number of channels of the weighted temporal feature tensor to 32 by 3D transpose convolution to obtain LSTM features; Step S27: The 3D convolutional features and the LSTM features are concatenated by channels, and after normalization, a 48-channel hybrid tensor is obtained. Step S28: The hybrid tensor is weighted twice by spatial attention mechanism, and a 32-dimensional feature vector is obtained by global average pooling, which is the high-dimensional tensor.

4. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S3: Construct a multi-objective optimization model based on electrothermal conversion efficiency, coefficient of performance (COP), and human comfort index. Input the high-dimensional tensor into the multi-objective optimization model and solve the Pareto front of the multi-objective optimization model using the NSGA-III algorithm to obtain the air conditioning group control parameters, including: Step S31: Define the objective function of the electrothermal conversion efficiency, construct the objective function of the coefficient of performance (COP) using linear interpolation, and construct the objective function of the human comfort index based on the PMV-PPD standard. Step S32: Perform Min-Max normalization on the three objective functions and add preset constraints to each of them; Step S33: Define the target space by using the electrothermal conversion efficiency, the cooling coefficient, and the human comfort index as the three coordinate axes of the target space. Step S34: Define the feasible region of the target space through all preset constraints, and map the high-dimensional tensor to the target space through linear transformation to obtain a prediction vector based on three objective functions; Step S35: Define several random air conditioning group control parameter combinations based on the NSGA-III algorithm, and input all random air conditioning group control parameter combinations into the target space to form several target vectors; Step S36: Update all target vectors iteratively through non-dominated sorting, and calculate the fitness value of each target vector based on the predicted vector in each update iteration; Step S37: After the preset iteration termination condition is met, output the Pareto front solution set, which includes several combinations of non-dominated air conditioning group control parameters.

5. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S5: Construct a CFD simulation environment equipped with a turbulence model based on the preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model. The actual temperature field dataset is obtained through dynamic simulation in the CFD simulation environment, including: Step S51: Define a turbulence model based on the geometric features of the preset area and the air supply characteristics of the air conditioning system; Step S52: Divide the initial three-dimensional temperature field matrix into a computational domain grid using an unstructured grid, and build a solver environment in the CFD simulation environment; Step S53: Perform time-progression calculations through the solver environment and record the actual temperature of each node in the computational domain grid in real time; Step S54: Integrate the actual temperatures of all nodes into the actual temperature field dataset.

6. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S7: Iteratively update the air conditioning group control parameters using a near-end strategy optimization algorithm to gradually approximate the minimum value of the deviation matrix, obtaining the Pareto optimal solution set of the air conditioning group control parameters, including: Step S71: Construct a dual-head neural network policy model with two independent output heads, namely the parameter control head and the value head; Step S72: Periodically acquire the deviation matrix, the L2 norm of the deviation matrix, and the air conditioning group control parameters based on a preset acquisition interval, and integrate them into trajectory data; Step S73: Calculate the state value function of the trajectory data using the value head, and calculate the dominance function of the state value function using the generalized dominance estimation method; Step S74: Bind trajectory data and advantage function of the same preset acquisition interval into a training sample, and integrate all preset acquisition intervals to obtain a training sample set; Step S75, define the pruning objective function of the dual-head neural network strategy model according to equation (1): （1）； in, Let the pruning objective function be... This is the weight set of the dual-head neural network strategy model. The ratio of the action probabilities of the new strategy to those of the old strategy. For the aforementioned advantage function, The pruning threshold is the pruning objective function. Step S76: Input the training sample set into the pruning objective function, and iterate and update the dual-head neural network policy model for several epochs using the Adam optimizer; Step S77: When the rate of change of the HV index of the Pareto front solution is less than a preset percentage threshold for several consecutive iterations, the iteration is terminated, and the air conditioning group control parameters output by the parameter control head are obtained as the Pareto optimal solution set.

7. The multi-energy coupled temperature field control method according to claim 1, characterized in that, Step S8, converting the Pareto optimal solution set into readable instructions for the air conditioning system and sending them to the air conditioning system, includes: Step S81: Convert the air conditioning group control parameters in the Pareto optimal solution set into physical parameters through linear transformation; Step S82: Encode the physical parameters into a data unit structure that conforms to the communication protocol of the air conditioning system; Step S83: The data unit structure is encapsulated by the OPC-UA server of the edge computing node and sent to the air conditioning system.

8. A multi-energy coupled temperature field control system based on intelligent optimization, wherein the multi-energy coupled temperature field control system is applied to the multi-energy coupled temperature field control method as described in any one of claims 1 to 7, characterized in that, The multi-energy coupled temperature field control system includes: The initial three-dimensional temperature field matrix acquisition module is used to acquire the original temperature field dataset of the preset area through the Internet of Things sensor array, and transmit the original temperature field dataset to the edge computing node through the MQTT protocol to form the initial three-dimensional temperature field matrix. The high-dimensional tensor output module is used to extract the local temperature gradient features of the initial three-dimensional temperature field matrix through the 3D convolution kernel of the ST-CNN spatiotemporal convolutional network, and then learn the dynamic law of the time dimension of the initial three-dimensional temperature field matrix through the LSTM module of the ST-CNN spatiotemporal convolutional network, and output a high-dimensional tensor that integrates spatiotemporal features. The air conditioning group control parameter solving module is used to construct a multi-objective optimization model based on electrothermal conversion efficiency, cooling coefficient, and human comfort index, and input the high-dimensional tensor into the multi-objective optimization model, and obtain the air conditioning group control parameters by solving the Pareto front of the multi-objective optimization model through the NSGA-III algorithm; The device response dataset acquisition module is used to send the air conditioning group control parameters to the air conditioning system and acquire the device response dataset of the air conditioning system. The actual temperature field dataset dynamic simulation module is used to construct a CFD simulation environment equipped with a turbulence model based on the preset region, and input the initial three-dimensional temperature field matrix and the device response dataset into the turbulence model, and obtain the actual temperature field dataset through dynamic simulation in the CFD simulation environment; The deviation matrix calculation module is used to obtain a predefined expected temperature field and perform a difference calculation between the actual temperature field dataset and the expected temperature field dataset to obtain a deviation matrix. The air conditioning group control parameter optimal solution set acquisition module is used to iteratively update the air conditioning group control parameters through a near-end strategy optimization algorithm to gradually approach the minimum value of the deviation matrix and obtain the Pareto optimal solution set of the air conditioning group control parameters; The optimal solution set for air conditioning group control parameters is sent by a module that converts the Pareto optimal solution set into readable instructions for the air conditioning system and sends them to the air conditioning system.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the multi-energy coupled temperature field control method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores program instructions, which, when executed by a processor, enable the implementation of the multi-energy coupled temperature field control method as described in any one of claims 1 to 7.

Images