A method and system for AI-based collaborative energy-saving valve cluster control

Abstract

A method and system for AI-based collaborative energy-saving valve cluster control, belonging to the interdisciplinary field of industrial process control and artificial intelligence, comprises the following steps: First, modeling and defining the optimization problem of the valve cluster system; second, constructing an intelligent optimization model; third, training the intelligent optimization model based on federated learning; and fourth, online collaborative control and continuous learning. This is achieved through an energy-saving valve cluster control system, including a valve cluster system modeling module, an intelligent optimization model construction module, an intelligent optimization model federated learning training module, and a collaborative control and continuous learning module. This invention enables direct and dynamic mapping between the valve cluster system state and the optimal valve opening vector, dynamically and selectively integrating key local information of valves, edges, and nodes to achieve precise and adaptive collaboration; significantly reducing the total energy consumption of the system and achieving optimal global energy efficiency and dynamic continuous optimization of the valve cluster system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of industrial process control and artificial intelligence, and relates to an AI-based collaborative energy-saving valve cluster control method and system. Background Technology

[0002] Valves, as key components of fluid control systems, are widely used in major engineering fields such as petrochemicals, power, water conservancy, aerospace, and nuclear industry. Their performance directly affects the safety, efficiency, and energy consumption of the entire system. Valve performance is almost entirely achieved through structural design. An excellent valve structural design means higher safety, lower energy consumption, longer lifespan, and lower total cost, making a fundamental and strategic contribution to industrial equipment and even national economic development.

[0003] In modern industrial production and urban infrastructure such as chemical, pharmaceutical, metallurgical, urban centralized heating and water supply, and HVAC systems in large commercial complexes, there exist complex pipeline clusters consisting of dozens or even hundreds or thousands of valve nodes. These valves undertake critical tasks such as flow distribution, pressure regulation, and temperature control, and their coordinated control performance directly affects the stability of industrial production, product quality, and even the energy consumption of the entire system.

[0004] Valve cluster control has been extensively studied. Yu Xiangyu et al. proposed a supercritical carbon dioxide closed-loop temperature and pressure coupled control system (Chinese Invention Patent CN202011308876.4). However, their method did not consider the impact of coordinated actions among numerous valves on the total energy consumption of the system, lacking a mechanism for global energy efficiency optimization of networked valve clusters. Chen Xun et al. proposed a thermal power peak-shaving and frequency regulation system with thermal storage, along with its thermal power unit and control method (Chinese Invention Patent CN202310464896.8). However, their method only focuses on the control execution between individual valves and does not delve into the coordinated optimization at the valve cluster level. If the valves are in a non-coordinated state, significant energy waste will still occur. Zhang Hong et al. proposed a valve joint control method, equipment, and storage medium for a multi-split air conditioning system (Chinese Invention Patent CN202210945152.3). However, the control logic of their method is based on preset rules or a relatively simple local optimization model, failing to achieve global optimization across systems and units. In summary, existing research indicates that valve cluster control has the following significant shortcomings:

[0005] Focusing solely on system control output and key parameter performance lacks specialized optimization for fine-grained coordination of numerous valves to minimize internal system energy consumption, thus ignoring bottom-up potential for energy efficiency improvement. Control strategy parameters are typically offline or based on fixed models, making it difficult to adapt to uncertainties such as changes in pipeline characteristics over time, lacking self-learning capabilities, and unable to achieve continuous optimization.

[0006] Therefore, in order to overcome the shortcomings of existing technologies, it is necessary to propose an AI-based collaborative energy-saving valve cluster control method and system. Summary of the Invention

[0007] To address the problems and shortcomings of existing technologies, this invention provides an AI-based collaborative energy-saving valve cluster control method and system. Specifically, it is a collaborative energy-saving valve cluster control method and system based on federated learning. By constructing a total energy consumption model of the valve cluster system, with the minimization of the total energy consumption of the valve cluster system as the optimization objective, a graph neural network (GNN) with an attention mechanism is used to design an intelligent optimization model for the valve cluster system. The intelligent optimization model is continuously trained and updated based on federated learning until convergence. Then, through online collaborative control and continuous learning of the intelligent optimization model, an AI collaborative valve cluster control system that takes into account both global energy saving and dynamic collaborative capabilities is finally established, thereby achieving optimal global energy efficiency and dynamic continuous optimization of the valve cluster system.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for controlling energy-saving valve clusters based on AI includes the following steps:

[0010] The first step is to model and define the optimization problem of the valve cluster system. The valve cluster system is constructed as a mathematical model of total energy consumption. Based on the optimization objective of collaborative energy saving in the valve cluster system, and satisfying constraints including node pressure constraints, branch flow constraints, valve opening constraints, core equipment load constraints, and pipeline flow velocity constraints, a mathematical description of the optimization problem with minimizing the total energy consumption of the valve cluster system is completed. Specifically, as follows:

[0011] Step 1.1, Modeling the total energy consumption of the valve cluster system. The direct goal of valve cluster regulation is to reduce the total energy consumption of the core equipment that provides power to the pipeline network. The core equipment includes pumps and fans. Suppose that there are N key control valves in the valve cluster, where the value of N ranges from 15 to 2000. Key control valves refer to valves that have a significant or decisive impact on the total energy consumption of the valve cluster system. The total power consumption model of the valve cluster system is mainly composed of the power consumption of the core equipment that provides power to the pipeline network, as shown in formula (1):

[0012] (1)

[0013] in, This represents the total power consumption of the valve cluster system. The number of core devices in operation. For fluid density, It is the acceleration due to gravity. The head of the m-th core device. Let m be the traffic of the m-th core device. Let m be the operating efficiency of the m-th core device.

[0014] Step 1.2, Optimization Problem Description. Based on the total energy consumption model of the valve cluster system constructed in Step 1.1, the optimization problem is mathematically expressed. By introducing constraints including node pressure constraints, branch flow constraints, valve opening constraints, core equipment load constraints, and pipeline flow velocity constraints, the basic operating requirements of the valve cluster system are met, achieving the optimization objective of minimizing the total energy consumption of the valve cluster system. Specifically:

[0015] Because changes in valve opening alter the hydraulic characteristics of the pipe network, thus affecting the operation of core equipment, the total power consumption of the valve cluster system... It can be represented as an optimization function, which is a valve opening vector consisting of N valve openings. It is expressed as follows:

[0016] (2)

[0017] (3)

[0018] in, To optimize the function; This indicates the valve opening degree, where 0 is fully closed and 1 is fully open. .

[0019] The optimization objective of a valve cluster system is to find an optimal valve opening vector while satisfying the requirements of all branch flow rates, node pressures, core equipment loads, and pipeline velocities. This minimizes the total power consumption of the valve cluster system, i.e., the optimization function. The value of is minimized; the specific optimization problem is as follows:

[0020] (4)

[0021] The following constraints must be met:

[0022] Node pressure constraint: The actual pressure on each node must be greater than the minimum pressure required by that node; otherwise, the node will not receive sufficient throughput, as follows:

[0023] (5)

[0024] Branch flow constraint: It is essential to ensure that the actual flow rate of each branch equals the required flow rate for that branch. This is a fundamental requirement for the normal operation of a valve cluster system, as follows:

[0025] (6)

[0026] Valve opening constraints: The opening degree of each valve must be within physical limits to ensure valve safety. 0 represents fully closed, and 1 represents fully open, as follows:

[0027] (7)

[0028] Core equipment load constraints: The actual power consumption of core equipment must not exceed its maximum allowable load to ensure the safety of core equipment, as follows:

[0029] (8)

[0030] Pipeline velocity constraints: The actual flow velocity within each pipe should be lower than the maximum permissible flow velocity for that pipe to prevent water hammer, corrosion, and noise, and to ensure pipeline safety, as follows:

[0031] (9)

[0032] in, For nodes The actual pressure at the location; For nodes The minimum pressure required at the point; j represents the node number, and J represents the total number of pipeline nodes; branch road The actual flow rate at the location; branch road The required flow rate; c represents the branch number, and C represents the total number of branches in the pipeline network; For the first The opening degree of each valve; L represents the valve number, and L represents the total number of valves in the pipeline network; The actual power consumption of the m-th core device; This represents the maximum allowable load of the core device; m represents the core device number, and M represents the total number of core devices. Let q be the actual flow velocity in the q-th pipe; The maximum allowable flow velocity is q in the q-th pipe; q represents the pipe number, and Q represents the number of pipes in the network.

[0033] The second step is to construct an intelligent optimization model. Based on the optimization problem established in the first step, the input and output of the intelligent optimization model are designed, and a graph neural network (GNN) with an attention mechanism is used to realize the state of the valve cluster system and the optimal valve opening vector. The direct and dynamic mapping completes the construction of the intelligent optimization model. Specifically:

[0034] Step 2.1: Design the input and output of the intelligent optimization model. Based on the optimization problem established in Step 1, design the input and output of an intelligent optimization model to directly output the system state that enables optimization to achieve the desired optimization. Minimized ,in The optimal valve opening vector is obtained by solving the optimization problem established in the first step, as follows:

[0035] The input feature vector of the intelligent optimization model Including branch demand flow based on the first step Based on the node pressure in the first step Information includes the outlet pressure and head of core equipment, real-time pressure and flow readings of valves in the pipeline network, ambient temperature, and the number of core equipment units currently in operation.

[0036] The intelligent optimization model outputs an action vector. This directly addresses the valve opening vector, composed of N valve openings, in the first step of the optimization problem. Using the Sigmoid activation function, its value range is naturally constrained to the interval [0,1].

[0037] Step 2.2: Design the network structure of the intelligent optimization model. Based on the input and output of the intelligent optimization model designed in Step 2.1, a graph neural network (GNN) is used as the backbone, and an attention mechanism is introduced to cope with various dynamic operating conditions of the valve cluster system. This results in a hybrid network structure with a graph neural network as the backbone and an attention mechanism integrated, thus obtaining the network structure of the intelligent optimization model; specifically as follows:

[0038] First, the valve cluster and its connected pipe network are abstracted into a directed graph G:

[0039] (10)

[0040] in, It is a set of nodes, including water source and heat source nodes, branch nodes, and valve nodes; It is a set of edges (pipes).

[0041] In the constructed directed graph G, features are defined for each node and edge in the directed graph. These features are directly derived from the input feature vector (system state) in step 2.1. Defined node characteristics Includes node type, current pressure value, demand pressure, current temperature value, and edge characteristics. This includes pipe length, diameter, current flow rate, total required flow rate, and ambient temperature. If a valve is installed on the edge, the valve is treated as a special edge feature, which includes the valve type, current opening degree, and historical operating status of the valve.

[0042] Subsequently, the features of nodes and edges are input into a graph neural network (GNN). Through message passing mechanism, messages are propagated, aggregated, and updated to accurately learn the spatial correlation between node pressure and pipeline flow, accurately capture the physical characteristics of pressure propagation and hydraulic linkage in the pipeline network, and thus calculate the globally coordinated valve action and the valve opening achieved.

[0043] Meanwhile, to address various dynamic operating conditions of the valve cluster system, an attention mechanism is introduced. This is achieved by calculating the attention coefficient. This allows the intelligent optimization model to dynamically track the valves, edges, and nodes most relevant to the current optimization objective by assigning weights to different nodes and edges under different operating conditions. The attention coefficient calculation formula is as follows:

[0044] (11)

[0045] (12)

[0046] in, Attention score; For nodes The set of all neighboring nodes in a directed graph. For nodes In a directed graph, at a certain neighbor node; For nodes And its original scores in all its neighboring nodes in the directed graph; For activation functions; This is the attention vector; They are nodes and its neighboring nodes The input feature vector; This is the trainable weight matrix for the nodes; For nodes The input feature vector is processed by a trainable weight matrix The vector after linear transformation, For nodes The input feature vector is processed by a trainable weight matrix The vector after linear transformation; Let be the trainable weight matrix for the opposite edge; For concatenation operators; For nodes Summing the sum of any neighboring nodes in a directed graph.

[0047] Finally, the message aggregation in the message passing mechanism is transformed into a weighted summation, enabling the intelligent optimization model to dynamically and selectively integrate key local information from valves, edges, and nodes, achieving precise and adaptive collaboration.

[0048] (13)

[0049] in, For nodes The input feature vector after weighted summation.

[0050] The third step involves training the intelligent optimization model using federated learning. Based on the intelligent optimization model built in the second step, a loss function is designed using the federated learning framework to minimize the model. This process is repeated iteratively until convergence, ensuring the intelligent optimization model can be continuously updated and dynamically optimized in actual operation. Specifically:

[0051] Step 3.1, Training Objective and Parameter Definition. Based on the intelligent optimization model constructed in Step 2, train a high-quality global model of the valve cluster system. ,in The parameters are global model parameters, and their training objective is to output a valve opening command that minimizes global energy consumption and satisfies all constraints based on the system state. Simultaneously, k edge nodes are defined to participate in the training, each node k possessing its own private dataset. Define the training rounds for federated learning, using t=0,1,2,...,T to represent the training rounds. Here, k ranges from 3 to 500, and T ranges from 20 to 100.

[0052] Step 3.2, Multi-terminal Initialization. Based on the global model parameters obtained in Step 3.1. On the cloud platform, global model parameters are randomly initialized. And the latest global model parameters of the cloud platform. The parameters are distributed to all defined edge nodes to implement global model parameters for each edge node. Synchronous initialization.

[0053] Step 3.3, Intelligent Optimization Model Training. Based on the private dataset obtained in Step 3.1. And the latest global model parameters of the cloud platform obtained in step 3.2 The intelligent optimization model is trained as follows:

[0054] Each edge node k is... As the initial point, use its private dataset. Multiple training rounds are performed. The private dataset contains... 1 sample, of which The value range is 50 to 20000, and the format is as follows:

[0055] (14)

[0056] in, These are system state features, derived from the input feature vector (system state) in step 2.1. Under different working conditions different; For various system state characteristics The corresponding optimal valve opening.

[0057] Step 3.4: Design the loss function to minimize the loss. This is based on the private training dataset obtained in Step 3.3. Design a loss function that minimizes the loss function. :

[0058] (15)

[0059] (16)

[0060] (17)

[0061] (18)

[0062] (19)

[0063] (20)

[0064] in, For fitting loss, For pressure-constrained losses, For flow constraint loss, For power consumption loss, For regularization terms, The pressure constraint loss coefficient, The flow constraint loss coefficient, The power loss factor is... The coefficient of the regularization term. The valve opening predicted by the model. Let be the square of the Euclidean norm. For private datasets The system state characteristics With various system state characteristics The summation is performed on the samples corresponding to the optimal valve opening.

[0065] Step 3.5: Define the optimization objective of the intelligent optimization model. This is based on minimizing the loss function obtained in Step 3.4. The optimization objective of the intelligent optimization model is to determine the optimal set of global model parameters for each edge node. This makes the private dataset of the k-th edge node... During the calculation, we obtained The value of reaches its minimum, as shown in formula (21):

[0066] (twenty one)

[0067] Step 3.6, Global model parameter update. Based on the minimized loss function obtained in Step 3.4. Stochastic gradient descent is used to perform the steps in step 3.5. The update process will continue multiple times until it is manually interrupted.

[0068] (twenty two)

[0069] in, The updated global model parameters for each edge node , The global model parameters of each edge node before this update , The learning rate is a parameter, and its value range is... , This is the gradient operator.

[0070] Meanwhile, based on the latest global model parameters of the cloud platform trained in step 3.3... This yields the parameter update amounts between the current optimal global model for each edge node after multiple updates and the latest global model of the trained cloud platform. :

[0071] (twenty three)

[0072] Step 3.7, Global Model Parameter Upload and Aggregation. Based on the parameter update amount obtained in Step 3.6. Each edge node uploads its updates to the cloud platform. After collecting all updates from participating edge nodes, the cloud platform executes a federated averaging algorithm.

[0073] (twenty four)

[0074] (25)

[0075] in, A randomly selected subset of edge nodes;

[0076] Finally, the cloud platform updates its latest global model parameters after training. :

[0077] (26)

[0078] in, This is the global learning rate, with a value ranging from 0.1 to 1.0.

[0079] Step 3.8, Intelligent Optimization Model Iteration Loop and Convergence Judgment. Based on the updated global model parameters in Step 3.7. This is used as the starting point for the next round of training, and the iteration loop continues, repeating steps 3.3 to 3.8, until one of the following convergence conditions is met. Then, the loop exits, and the globally model of the valve cluster system that has been trained and converged is obtained. This enables iterative looping and convergence judgment of the intelligent optimization model. The convergence condition is:

[0080] (1) The performance of the global model no longer improves significantly;

[0081] (2) Reach the maximum number of training rounds T;

[0082] (3) The change in global model parameters is lower than the threshold, where the threshold range is... .

[0083] The fourth step is online collaborative control and continuous learning. This involves using the global model of the valve cluster system that has converged during the training in the third step. The system acquires status data and generates the optimal valve opening vector command for all valves at the current moment. It then issues and executes this optimal valve opening vector command. Finally, an intelligent optimization model performs performance evaluation and continuous learning to achieve global energy efficiency optimization and dynamic continuous optimization of the valve cluster system. Details are as follows:

[0084] Step 4.1, Status Acquisition. The valve cluster system acquires real-time status data of the entire pipeline network through a sensor network, constructing the input feature vector for the current moment. .

[0085] Step 4.2, Instruction Generation. Based on the input feature vector of the valve cluster system acquired in Step 4.1 at the current moment, this vector is input into the deployed global model of the valve cluster system that has been trained and converged in the third step. Subsequently, the global model of the valve cluster system performs a forward propagation calculation, outputting the optimal valve opening vector command for all N valves at the current moment. :

[0086] (27)

[0087] Step 4.3, Instruction Issuance and Execution. Instructions based on the optimal valve opening vector obtained in Step 4.2. It then sends the information to the valve cluster system to drive the valve actuators to operate.

[0088] Step 4.4, Performance Evaluation and Continuous Learning. Based on the valve actuator actions driven in Step 4.3, the total power consumption of the valve cluster system after execution is evaluated. Continuous monitoring and performance evaluation are required.

[0089] If the actual energy-saving effect is good or the expected target is achieved, then the overall energy efficiency of the valve cluster system is optimized.

[0090] If the actual energy-saving effect is not good or fails to meet the expected target, a new round of intelligent optimization model training and updating will be triggered, so that the intelligent optimization model can continuously learn from the operation and achieve dynamic and continuous optimization of the energy efficiency of the valve cluster system.

[0091] An AI-based collaborative energy-saving valve cluster control system is disclosed. This system implements the aforementioned AI-based collaborative energy-saving valve cluster control method. The energy-saving valve cluster control system includes a valve cluster system modeling module, an intelligent optimization model construction module, an intelligent optimization model federated learning training module, and a collaborative control and continuous learning module. Specifically:

[0092] The valve cluster system modeling module consists of pumps, fans, and valves in the pipeline network. It is used to construct the total energy consumption model of the valve cluster system and is the theoretical basis of the entire energy-saving valve cluster control system.

[0093] The intelligent optimization model construction module is connected to the valve cluster system modeling module and consists of a cloud platform center. It is used to construct intelligent optimization models to achieve optimization solutions for the valve cluster system state and valve opening. The direct and dynamic mapping is the model cluster of the entire energy-saving valve cluster control system.

[0094] The intelligent optimization model federated learning training module is connected to the intelligent optimization model construction module and the valve cluster system modeling module. It consists of a cloud platform center and distributed model training edge nodes. It is used for intelligent optimization model training and updating, so as to realize the continuous updating and dynamic optimization of the intelligent optimization model in actual operation. It is the training center of the entire energy-saving valve cluster control system.

[0095] The collaborative control and continuous learning module is connected to the intelligent optimization model federated learning training module, the intelligent optimization model construction module, and the valve cluster system modeling module. It consists of a high-performance server, a sensor network, and valve actuators. It is used to generate, issue, and execute the optimal valve opening vector command for all valves at the current moment, enabling the intelligent optimization model to perform performance evaluation and continuous learning, and achieve global energy efficiency optimization and dynamic continuous optimization of the valve cluster system. It is the learning and optimization center of the entire energy-saving valve cluster control system.

[0096] The beneficial effects of this invention are as follows:

[0097] (1) This invention constructs a mathematical model of the total energy consumption of a valve cluster system, takes minimizing the total energy consumption of the valve cluster system as the optimization objective, and uses a graph neural network (GNN) with an attention mechanism to design an intelligent optimization model for the valve cluster system. The intelligent optimization model is constructed, and the state of the valve cluster system and the optimal valve opening vector are realized. The direct and dynamic mapping can dynamically and differentiately integrate key local information of valves, edges and nodes, achieving precise and adaptive collaboration.

[0098] (2) Based on federated learning, the present invention continuously trains and updates the intelligent optimization model until it converges. Then, through online collaborative control and continuous learning of the intelligent optimization model, an AI collaborative valve cluster control system that takes into account both global energy saving and dynamic collaborative capabilities is finally established, which significantly reduces the total energy consumption of the system and thus achieves the optimal global energy efficiency and dynamic continuous optimization of the valve cluster system. Attached Figure Description

[0099] Figure 1 This is a schematic diagram of the process of an AI-based collaborative energy-saving valve cluster control method according to the present invention.

[0100] Figure 2 This is a schematic diagram of an AI-based collaborative energy-saving valve cluster control system according to the present invention. Detailed Implementation

[0101] The invention will be further illustrated below with a specific implementation example. This embodiment describes a large-scale centralized district heating system in a city. The system includes one large heat source plant, one relay pump station, 20 heating substations, and several building complexes, comprising dozens of key control valves, forming a valve cluster. The goal is to achieve coordinated energy conservation in the large-scale centralized heating system.

[0102] A method for controlling energy-saving valve clusters based on AI includes the following steps:

[0103] The first step is to model and define the optimization problem of the valve cluster system. The valve cluster system is constructed as a mathematical model of total energy consumption. Based on the optimization objective of collaborative energy saving in the valve cluster system, and satisfying constraints including node pressure constraints, branch flow constraints, valve opening constraints, core equipment load constraints, and pipeline flow velocity constraints, a mathematical description of the optimization problem with minimizing the total energy consumption of the valve cluster system is completed. Specifically, as follows:

[0104] Step 1.1, Modeling the total energy consumption of the valve cluster system. The direct goal of valve cluster regulation is to reduce the total energy consumption of the core equipment that provides power to the pipeline network. The core equipment includes pumps and fans. Suppose that there are N key control valves in the valve cluster, where the value of N ranges from 15 to 2000. Key control valves refer to valves that have a significant or decisive impact on the total energy consumption of the valve cluster system. The total power consumption model of the valve cluster system is mainly composed of the power consumption of the core equipment that provides power to the pipeline network, as shown in formula (1):

[0105] (1)

[0106] in, This represents the total power consumption of the valve cluster system. The number of core devices in operation. For fluid density, It is the acceleration due to gravity. The head of the m-th core device. Let m be the traffic of the m-th core device. Let m be the operating efficiency of the m-th core device.

[0107] Step 1.2, Optimization Problem Description. Based on the total energy consumption model of the valve cluster system constructed in Step 1.1, the optimization problem is mathematically expressed. By introducing constraints including node pressure constraints, branch flow constraints, valve opening constraints, core equipment load constraints, and pipeline flow velocity constraints, the basic operating requirements of the valve cluster system are met, achieving the optimization objective of minimizing the total energy consumption of the valve cluster system. Specifically:

[0108] Because changes in valve opening alter the hydraulic characteristics of the pipe network, thus affecting the operation of core equipment, the total power consumption of the valve cluster system... It can be represented as an optimization function, which is a valve opening vector consisting of N valve openings. It is expressed as follows:

[0109] (2)

[0110] (3)

[0111] in, To optimize the function; This indicates the valve opening degree, where 0 is fully closed and 1 is fully open. .

[0112] The optimization objective of a valve cluster system is to find an optimal valve opening vector while satisfying the requirements of all branch flow rates, node pressures, core equipment loads, and pipeline velocities. This minimizes the total power consumption of the valve cluster system, i.e., the optimization function. The value of is minimized. The specific optimization problem is as follows:

[0113] (4)

[0114] The following constraints must be met:

[0115] Node pressure constraint: The actual pressure on each node must be greater than the minimum pressure required by that node; otherwise, the node will not receive sufficient throughput, as follows:

[0116] (5)

[0117] Branch flow constraint: It is essential to ensure that the actual flow rate of each branch equals the required flow rate for that branch. This is a fundamental requirement for the normal operation of a valve cluster system, as follows:

[0118] (6)

[0119] Valve opening constraints: The opening degree of each valve must be within physical limits to ensure valve safety. 0 represents fully closed, and 1 represents fully open, as follows:

[0120] (7)

[0121] Core equipment load constraints: The actual power consumption of core equipment must not exceed its maximum allowable load to ensure the safety of core equipment, as follows:

[0122] (8)

[0123] Pipeline velocity constraints: The actual flow velocity within each pipe should be lower than the maximum permissible flow velocity for that pipe to prevent water hammer, corrosion, and noise, and to ensure pipeline safety, as follows:

[0124] (9)

[0125] in, For nodes The actual pressure at the location; For nodes The minimum pressure required at the point; j represents the node number, and J represents the total number of pipeline nodes; branch road The actual flow rate at the location; branch road The required flow rate; c represents the branch number, and C represents the total number of branches in the pipeline network; For the first The opening degree of each valve; L represents the valve number, and L represents the total number of valves in the pipeline network; The actual power consumption of the m-th core device; This represents the maximum allowable load of the core device; m represents the core device number, and M represents the total number of core devices. Let q be the actual flow velocity in the q-th pipe; The maximum allowable flow velocity is q in the q-th pipe; q represents the pipe number, and Q represents the number of pipes in the network.

[0126] In this embodiment, It includes one heat source plant node, one relay pump station node, 20 heating station nodes, and 20 pipeline junction nodes; For each heating station, there is a corresponding user branch; These are key control valves, including 2 core regulating valves (heat source, pump station outlet), 20 branch regulating valves (inlet of each heating station), and 2 network balancing valves (used to optimize hydraulic conditions). This includes one main circulation pump located at the heat source plant and one booster pump located at the relay pumping station in the middle of the pipeline network; This is used to connect all the main and branch pipelines to the aforementioned nodes. Other data can be collected in real time through the sensor network deployed in the system.

[0127] The second step is to construct an intelligent optimization model. Based on the optimization problem established in the first step, the input and output of the intelligent optimization model are designed, and a graph neural network (GNN) with an attention mechanism is used to realize the state of the valve cluster system and the optimal valve opening vector. The direct and dynamic mapping completes the construction of the intelligent optimization model. Specifically:

[0128] Step 2.1: Design the input and output of the intelligent optimization model. Based on the optimization problem established in Step 1, design the input and output of an intelligent optimization model to directly output the system state that enables optimization to achieve the desired optimization. Minimized ,in The optimal valve opening vector is obtained by solving the optimization problem established in the first step, as follows:

[0129] The input feature vector of the intelligent optimization model Including branch demand flow based on the first step Based on the node pressure in the first step Information includes the outlet pressure and head of core equipment, real-time pressure and flow readings of valves in the pipeline network, ambient temperature, and the number of core equipment units currently in operation.

[0130] The intelligent optimization model outputs an action vector. This directly addresses the valve opening vector, composed of N valve openings, in the first step of the optimization problem. Using the Sigmoid activation function, its value range is naturally constrained to the interval [0,1].

[0131] Step 2.2: Design the network structure of the intelligent optimization model. Based on the input and output of the intelligent optimization model designed in Step 2.1, a graph neural network (GNN) is used as the backbone, and an attention mechanism is introduced to cope with various dynamic operating conditions of the valve cluster system. This results in a hybrid network structure with a graph neural network as the backbone and an attention mechanism integrated, thus obtaining the network structure of the intelligent optimization model; specifically as follows:

[0132] First, the valve cluster and its connected pipe network are abstracted into a directed graph G:

[0133] (10)

[0134] in, It is a set of nodes, including water source and heat source nodes, branch nodes, and valve nodes; It is a set of edges (pipes).

[0135] In the constructed directed graph G, features are defined for each node and edge in the directed graph. These features are directly derived from the input feature vector (system state) in step 2.1. Defined node characteristics Includes node type, current pressure value, demand pressure, current temperature value, and edge characteristics. This includes pipe length, diameter, current flow rate, total required flow rate, and ambient temperature. If a valve is installed on the edge, the valve is treated as a special edge feature, which includes the valve type, current opening degree, and historical operating status of the valve.

[0136] Subsequently, the features of nodes and edges are input into a graph neural network (GNN). Through message passing mechanism, messages are propagated, aggregated, and updated to accurately learn the spatial correlation between node pressure and pipeline flow, accurately capture the physical characteristics of pressure propagation and hydraulic linkage in the pipeline network, and thus calculate the globally coordinated valve action and the valve opening achieved.

[0137] Meanwhile, to address various dynamic operating conditions of the valve cluster system, an attention mechanism is introduced. This is achieved by calculating the attention coefficient. This allows the intelligent optimization model to dynamically track the valves, edges, and nodes most relevant to the current optimization objective by assigning weights to different nodes and edges under different operating conditions. The attention coefficient calculation formula is as follows:

[0138] (11)

[0139] (12)

[0140] in, Attention score; For nodes The set of all neighboring nodes in a directed graph. For nodes In a directed graph, at a certain neighbor node; For nodes And its original scores in all its neighboring nodes in the directed graph; For activation functions; This is the attention vector; They are nodes and its neighboring nodes The input feature vector; This is the trainable weight matrix for the nodes; For nodes The input feature vector is processed by a trainable weight matrix The vector after linear transformation, For nodes The input feature vector is processed by a trainable weight matrix The vector after linear transformation; Let be the trainable weight matrix for the opposite edge; For concatenation operators; For nodes Summing the sum of any neighboring nodes in a directed graph.

[0141] Finally, the message aggregation in the message passing mechanism is transformed into a weighted summation, enabling the intelligent optimization model to dynamically and selectively integrate key local information from valves, edges, and nodes, achieving precise and adaptive collaboration.

[0142] (13)

[0143] in, For nodes The input feature vector after weighted summation.

[0144] The third step involves training the intelligent optimization model using federated learning. Based on the intelligent optimization model built in the second step, a loss function is designed using the federated learning framework to minimize the model. This process is repeated iteratively until convergence, ensuring the intelligent optimization model can be continuously updated and dynamically optimized in actual operation. Specifically:

[0145] Step 3.1, Training Objectives and Parameter Definitions: Based on the intelligent optimization model constructed in Step 2, train a high-quality global model of the valve cluster system. ,in The parameters are global model parameters, and their training objective is to output a valve opening command that minimizes global energy consumption and satisfies all constraints based on the system state. Simultaneously, k edge nodes are defined to participate in the training, each node k possessing its own private dataset. Define the training rounds for federated learning, using t=0,1,2,...,T to represent the training rounds. Here, k ranges from 3 to 500, and T ranges from 20 to 100.

[0146] Step 3.2, Multi-terminal Initialization. Based on the global model parameters obtained in Step 3.1. On the cloud platform, global model parameters are randomly initialized. And the latest global model parameters of the cloud platform. The parameters are distributed to all defined edge nodes to implement global model parameters for each edge node. Synchronous initialization.

[0147] Step 3.3, Intelligent Optimization Model Training. Based on the private dataset obtained in Step 3.1. And the latest global model parameters of the cloud platform obtained in step 3.2 The intelligent optimization model is trained as follows:

[0148] Each edge node k is... As the initial point, use its private dataset. Multiple training rounds are performed. The private dataset contains... 1 sample, of which The value range is 50 to 20000, and the format is as follows:

[0149] (14)

[0150] in, These are system state features, derived from the input feature vector (system state) in step 2.1. Under different working conditions different; For various system state characteristics The corresponding optimal valve opening.

[0151] Step 3.4: Design the loss function to minimize the loss. This is based on the private training dataset obtained in Step 3.3. Design a loss function that minimizes the loss function. :

[0152] (15)

[0153] (16)

[0154] (17)

[0155] (18)

[0156] (19)

[0157] (20)

[0158] in, For fitting loss, For pressure-constrained losses, For flow constraint loss, For power consumption loss, For regularization terms, The pressure constraint loss coefficient, The flow constraint loss coefficient, The power loss factor is... The coefficient of the regularization term. The valve opening predicted by the model. Let be the square of the Euclidean norm. For private datasets The system state characteristics With various system state characteristics The summation is performed on the samples corresponding to the optimal valve opening.

[0159] Step 3.5: Define the optimization objective of the intelligent optimization model. This is based on minimizing the loss function obtained in Step 3.4. The optimization objective of the intelligent optimization model is to determine the optimal set of global model parameters for each edge node. This makes the private dataset of the k-th edge node... During the calculation, we obtained The value of reaches its minimum, as shown in formula (21):

[0160] (twenty one)

[0161] Step 3.6, Global model parameter update. Based on the minimized loss function obtained in Step 3.4. Stochastic gradient descent is used to perform the steps in step 3.5. The update process will continue multiple times until it is manually interrupted.

[0162] (twenty two)

[0163] in, The updated global model parameters for each edge node , The global model parameters of each edge node before this update , The learning rate is a parameter, and its value range is... , This is the gradient operator.

[0164] Meanwhile, based on the latest global model parameters of the cloud platform trained in step 3.3... This yields the parameter update amounts between the current optimal global model for each edge node after multiple updates and the latest global model of the trained cloud platform. :

[0165] (twenty three)

[0166] Step 3.7, Global Model Parameter Upload and Aggregation. Based on the parameter update amount obtained in Step 3.6. Each edge node uploads its updates to the cloud platform. After collecting all updates from participating edge nodes, the cloud platform executes a federated averaging algorithm.

[0167] (twenty four)

[0168] (25)

[0169] in, This is a randomly selected subset of edge nodes.

[0170] Finally, the cloud platform updates its latest global model parameters after training. :

[0171] (26)

[0172] in, This is the global learning rate, with a value ranging from 0.1 to 1.0.

[0173] Step 3.8, Intelligent Optimization Model Iteration Loop and Convergence Judgment. Based on the updated global model parameters in Step 3.7. This is used as the starting point for the next round of training, and the iteration loop continues, repeating steps 3.3 to 3.8, until one of the following convergence conditions is met. Then, the loop exits, and the globally model of the valve cluster system that has been trained and converged is obtained. This enables iterative looping and convergence judgment of the intelligent optimization model. The convergence condition is:

[0174] (1) The performance of the global model no longer improves significantly;

[0175] (2) Reach the maximum number of training rounds T;

[0176] (3) The change in global model parameters is lower than the threshold, where the threshold range is... .

[0177] In this embodiment, the 24 valves and their associated sensors and data are divided into 5 edge nodes, including the heat source plant area (V1), the relay pump station and the northern pipeline network (V2), the central pipeline network (V3), the southern pipeline network (V4), and the remote loop balance zone (V5). Therefore, k=5. T=50. , =10, =5, =1, =10 -4 , , Ultimately, the intelligent optimization model for the large-scale centralized heating system exited the iteration loop after reaching the maximum number of training rounds.

[0178] The fourth step is online collaborative control and continuous learning. This involves using the global model of the valve cluster system that has converged during the training in the third step. The system acquires status data and generates the optimal valve opening vector command for all valves at the current moment. It then issues and executes this optimal valve opening vector command. Finally, an intelligent optimization model performs performance evaluation and continuous learning to achieve global energy efficiency optimization and dynamic continuous optimization of the valve cluster system. Details are as follows:

[0179] Step 4.1, Status Acquisition. The valve cluster system acquires real-time status data of the entire pipeline network through a sensor network, constructing the input feature vector for the current moment. .

[0180] Step 4.2, Instruction Generation. Based on the input feature vector of the valve cluster system acquired in Step 4.1 at the current moment, this vector is input into the deployed global model of the valve cluster system that has been trained and converged in the third step. Subsequently, the global model of the valve cluster system performs a forward propagation calculation, outputting the optimal valve opening vector command for all N valves at the current moment. :

[0181] (27)

[0182] Step 4.3, Instruction Issuance and Execution. Instructions based on the optimal valve opening vector obtained in Step 4.2. It then sends the information to the valve cluster system to drive the valve actuators to operate.

[0183] Step 4.4, Performance Evaluation and Continuous Learning. Based on the valve actuator actions driven in Step 4.3, the total power consumption of the valve cluster system after execution is evaluated. Continuous monitoring and performance evaluation are required.

[0184] If the actual energy-saving effect is good or the expected target is achieved, then the overall energy efficiency of the valve cluster system is optimized.

[0185] If the actual energy-saving effect is not good or fails to meet the expected target, a new round of intelligent optimization model training and updating will be triggered, so that the intelligent optimization model can continuously learn from the operation and achieve dynamic and continuous optimization of the energy efficiency of the valve cluster system.

[0186] In this embodiment, a comparison of core energy-saving data before and after adopting AI collaborative control is shown in Table 1. The expected energy-saving target is: under the premise of ensuring the heating quality of all heating stations, the total power consumption (pump energy consumption) of the system in one year should be reduced by 15% compared with that before adopting AI collaborative control. After adopting the AI ​​collaborative control of this invention, the total power consumption in one year is reduced by 18%, the average power consumption per unit area in one year is reduced by 17.8%, the equivalent standard coal emission reduction is 380t, the carbon dioxide emission reduction is 990t, and the hydraulic imbalance of the pipeline network is improved from high to excellent. Obviously, the expected energy-saving target has been achieved, significantly reducing the overall energy consumption of a large-scale centralized regional heating system in a city, and realizing the optimal global energy efficiency and dynamic continuous optimization of the valve cluster system.

[0187] Table 1: Comparison of core energy-saving data before and after using AI model collaborative control in the embodiments

[0188]

[0189] An AI-based collaborative energy-saving valve cluster control system is disclosed. This system implements the aforementioned AI-based collaborative energy-saving valve cluster control method. The energy-saving valve cluster control system includes a valve cluster system modeling module, an intelligent optimization model construction module, an intelligent optimization model federated learning training module, and a collaborative control and continuous learning module. Specifically:

[0190] The valve cluster system modeling module consists of pumps, fans, and valves in the pipeline network. It is used to construct the total energy consumption model of the valve cluster system and is the theoretical basis of the entire energy-saving valve cluster control system.

[0191] The intelligent optimization model construction module is connected to the valve cluster system modeling module and consists of a cloud platform center. It is used to construct intelligent optimization models to achieve optimization solutions for the valve cluster system state and valve opening. The direct and dynamic mapping is the model cluster of the entire energy-saving valve cluster control system.

[0192] The intelligent optimization model federated learning training module is connected to the intelligent optimization model construction module and the valve cluster system modeling module. It consists of a cloud platform center and distributed model training edge nodes. It is used for intelligent optimization model training and updating, so as to realize the continuous updating and dynamic optimization of the intelligent optimization model in actual operation. It is the training center of the entire energy-saving valve cluster control system.

[0193] The collaborative control and continuous learning module is connected to the intelligent optimization model federated learning training module, the intelligent optimization model construction module, and the valve cluster system modeling module. It consists of a high-performance server, a sensor network, and valve actuators. It is used to generate, issue, and execute the optimal valve opening vector command for all valves at the current moment, enabling the intelligent optimization model to perform performance evaluation and continuous learning, and achieve global energy efficiency optimization and dynamic continuous optimization of the valve cluster system. It is the learning and optimization center of the entire energy-saving valve cluster control system.

[0194] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. An AI-based collaborative energy-saving valve cluster control method, characterized in that, AI-based collaborative energy-saving valve cluster control method The method comprises the following steps: Step 1, valve cluster system modeling and optimization problem determination; the valve cluster system is constructed as a mathematical model related to total energy consumption, and based on the optimization goal of valve cluster system collaborative energy saving, the constraint conditions including node pressure constraint, branch flow constraint, valve opening constraint, core device load constraint and pipeline flow velocity constraint are met, and the description of the optimization problem with the minimum total energy consumption of the valve cluster system is completed; Step 2, constructing an intelligent optimization model; Based on the optimization problem established in the first step, by designing the input and output of the intelligent optimization model, using the graph neural network GNN with attention mechanism, the direct and dynamic mapping of the valve cluster system state and the optimal valve opening vector is realized, and the construction of the intelligent optimization model is completed . Third step, training the intelligent optimization model based on federated learning; based on the intelligent optimization model constructed in the second step, the minimum loss function is designed through the federated learning framework, the intelligent optimization model is trained and updated, and multiple iteration cycles are performed until convergence, and the global model of the valve cluster system is obtained ; Specifically: Step 3.1, training target and parameter definition; based on the intelligent optimization model constructed in the second step, a high-quality global model of the valve cluster system is trained wherein is the global model parameter, and the training target is to output the valve opening command that minimizes the global energy consumption and meets all constraints according to the system state; meanwhile, k edge nodes are defined to participate in the training, and each node k has its private data set ; the federated learning training round is defined, and t=0, 1, 2,..., T represents the training round; wherein, the value range of k is 3-500, and the value range of T is 20-100; Step 3.2, multi-terminal initialization; global model parameters obtained based on step 3.1 , random initialization of global model parameters in the cloud platform , and the current latest global model parameters of the cloud platform are sent to all defined edge nodes to achieve the synchronous initialization of the global model parameters of each edge node ​ Step 3.3, intelligent optimization model training; based on the private data set obtained in step 3.1 and the current global model parameters of the cloud platform obtained in step 3.2 , intelligent optimization model training is performed to obtain a private data set ; Step 3.4, design minimization loss function; based on the post-training private dataset obtained in step 3.3 , design minimization loss function ; Step 3.5, defining the optimization objective of the intelligent optimization model; based on the minimized loss function obtained in step 3.4 , determining the optimization objective of the intelligent optimization model, that is, finding a set of current optimal global model parameters of each edge node , so that when the parameter is calculated on the private data set of the kth edge node , the value of reaches the minimum, as shown in formula (21): （21）； Step 3.6, global model parameter update; based on the minimized loss function obtained in step 3.4 perform multiple updates of the parameters in step 3.5 using the stochastic gradient descent method until manual interruption: ​ （22）； wherein, are the global model parameters of each edge node before this update , are the global model parameters of each edge node before this update , is the learning rate of the parameter, and the value range is , is the gradient operator; Meanwhile, based on the step 3.3, the current latest global model parameters of the cloud platform after training are obtained , the parameter update amount between the current optimal global model of each edge node after multiple updates and the current latest global model of the cloud platform after training is obtained : （23）； Step 3.7, global model parameter uploading and aggregation; based on the parameter update amount obtained in step 3.6 Each edge node uploads it to the cloud platform; after the cloud platform collects the updates of all participating edge nodes, it performs the federated averaging algorithm: （24）； （25）； wherein, is a randomly selected subset of edge nodes; Finally, the cloud platform updates its current latest global model parameters after training : （26）； wherein, is a global learning rate; Step 3.

8. Intelligent optimization model iteration loop and convergence judgment; update the global model parameters based on step 3.7 , as the starting point for the next round of training, continue the iteration loop, repeat steps 3.3 to 3.8 until one of the following convergence conditions is met, exit the loop, and get the trained global model of the valve cluster system , realize intelligent optimization model iteration loop and convergence judgment; Fourthly, online collaborative control and continuous learning; based on the third step, the global model of the valve cluster system is trained to converge The optimal valve opening vector command of all valves at the current time is output by state acquisition and instruction generation, and the optimal valve opening vector command is issued and executed. Finally, the intelligent optimization model performs performance evaluation and continuous learning to achieve global energy efficiency optimization and dynamic continuous optimization of the valve cluster system.

2. The AI-based collaborative energy-saving valve cluster control method according to claim 1, characterized in that, The first step is specifically: Step 1.1, valve cluster system total energy consumption modeling; the direct goal of valve cluster regulation is to reduce the total energy consumption of core devices that provide power for the pipe network; let the valve cluster have N key control valves, and the key control valve refers to the valve that has a significant or decisive influence on the total energy consumption of the valve cluster system; the total power consumption model of the valve cluster system is mainly composed of the power consumption of the core devices that provide power for the pipe network, as shown in formula (1): （1）； wherein, is the total power consumption of the valve cluster system; is the number of core devices in operation, is the fluid density, is the gravitational acceleration, is the head of the mth core device, is the flow of the mth core device, is the operating efficiency of the mth core device; Step 1.2, optimization problem description; based on the total energy consumption model of the valve cluster system constructed in step 1.1, the optimization problem is mathematically expressed; by introducing constraint conditions including node pressure constraint, branch flow constraint, valve opening constraint, core device load constraint and pipeline flow velocity constraint, the basic operation requirements of the valve cluster system are met, and the optimization goal of minimizing the total energy consumption of the valve cluster system is achieved; specifically: Total power consumption of a valve cluster system The valve opening vector consisting of N valve openings is represented as an optimization function is represented as follows: （2）； （3）； wherein is an optimization function; is a valve opening, 0 is fully closed, 1 is fully open, wherein ; The optimization goal of the valve cluster system is to find an optimal valve opening vector under the premise of meeting all branch flow, node pressure, core device load and pipeline flow rate , so that the total power consumption of the valve cluster system is minimized, that is, the value of the optimization function is minimized; the specific optimization problem is as follows: （4）。 3. The AI-based collaborative energy-saving valve cluster control method according to claim 2, characterized in that, The constraint condition is: Node pressure constraint: ensure that the actual pressure of each node is greater than the minimum pressure required by the node, as follows: （5）； Branch flow constraint: ensure that the actual flow of each branch is equal to the required flow of the branch, as follows: （6）； Valve opening constraint: the opening of each valve is within the physical limit range, where 0 is fully closed and 1 is fully open, as follows: （7）； Core device load constraint: the actual power consumption of the core device cannot exceed its maximum allowable load, as follows: （8）； Pipeline flow velocity constraint: the actual flow velocity in each pipeline should be lower than the maximum allowable flow velocity of the pipeline, as follows: （9）； in, For nodes The actual pressure at the location; For nodes The minimum pressure required at the point; j represents the node number, and J represents the total number of pipeline nodes; branch road The actual flow rate at the location; branch road The required flow rate; c represents the branch number, and C represents the total number of branches in the pipeline network; For the first The opening degree of each valve; L represents the valve number, and L represents the total number of valves in the pipeline network; The actual power consumption of the m-th core device; This represents the maximum allowable load of the core device; m represents the core device number, and M represents the total number of core devices. Let q be the actual flow velocity in the q-th pipe; The maximum allowable flow velocity is q in the q-th pipe; q represents the pipe number, and Q represents the number of pipes in the network.

4. The AI-based collaborative energy-saving valve cluster control method according to claim 3, characterized in that, The second step is specifically: Step 2.1: Design the input and output of the intelligent optimization model; Based on the optimization problem established in Step 1, design the input and output of an intelligent optimization model to directly output the system state that enables optimization to achieve the desired optimization. Minimized ,in The optimal valve opening vector is obtained by solving the optimization problem established in the first step, as follows: input feature vector of the intelligent optimization model including branch demand flow in the first step , node pressure in the first step , outlet pressure and lift of core equipment, real-time pressure and flow readings of valves in the pipe network, and environmental temperature, and the number of currently running core equipment information the output action vector of the intelligent optimization model directly for the valve opening vector consisting of N valve openings in the first step optimization problem using a sigmoid activation function, naturally constraining its value range to the interval [0, 1] Step 2.2, design the network structure of the intelligent optimization model; based on the input and output of the intelligent optimization model designed in step 2.1, adopt graph neural network GNN as the main part, and introduce attention mechanism to cope with various dynamic conditions of the valve cluster system, obtain a hybrid network structure taking graph neural network as the main part and integrating attention mechanism, and obtain the network structure of the intelligent optimization model; specifically as follows: Firstly, the valve cluster and the connected pipe network are abstracted as a directed graph G: （10）； wherein, is a set of nodes, including water source and heat source nodes, branch nodes, and valve nodes; is a set of edges; In the constructed directed graph G, for each node and edge in the directed graph, features are defined, which are directly derived from the input feature vectors in step 2.1 ; defined node features including node type, current pressure value, demand pressure, current temperature value, edge features including pipe length, diameter, current flow value, total demand flow, ambient temperature; Subsequently, the features of nodes and edges are input into the graph neural network GNN, and message propagation, aggregation and update are performed through the message passing mechanism, to learn the spatial correlation between node pressure and pipeline flow, capture the physical properties of pressure propagation and hydraulic linkage in the pipe network, and calculate the globally coordinated valve action and the valve opening achieved thereby; Meanwhile, in order to deal with various dynamic conditions of the valve cluster system, an attention mechanism is introduced; by calculating the attention coefficient The weights of different nodes and edges under different working conditions are allocated, and the intelligent optimization model dynamically tracks the valves, edges and nodes most relevant to the current optimization goal. Finally, the message aggregation in the message passing mechanism is changed to weighted summation, so that the intelligent optimization model can dynamically and differently integrate the key local information of valves, edges and nodes, and realize accurate and adaptive collaboration: （13）； wherein, is a node weighted sum of the input feature vectors; is an activation function; is a node a set of all neighbor nodes in the directed graph; is a node vector after a linear transformation of the input feature vectors of the node by a trainable weight matrix 5. The AI-based collaborative energy-saving valve cluster control method according to claim 4, characterized in that, In step 2.2, the attention coefficient calculation formula is as follows: （11）； （12）； wherein, is an attention score; is a node a set of all neighbor nodes of in the directed graph, is a node a certain neighbor node of in the directed graph; is a node and its original score of all neighbor nodes of in the directed graph; is an activation function; is an attention vector; are input feature vectors of node and its neighbor nodes respectively; is a trainable weight matrix for nodes, is an input feature vector of node after a linear transformation by the trainable weight matrix ; is an input feature vector of node after a linear transformation by the trainable weight matrix ; is a trainable weight matrix for edges; is a concatenation operator; is a sum over all neighbor nodes of node in the directed graph.

6. The AI-based collaborative energy-saving valve cluster control method according to claim 1, characterized in that, In the third step: The step 3.3 is: Each edge node k performs the following steps: As a starting point, it uses its private dataset for multiple rounds of training; the private dataset contains samples, where ranges from 50 to 20,000 and has the following form: （14）； wherein, is a system state feature derived from the input feature vector in step 2.1 , the optimal valve opening under different working conditions are different; are various system state features the corresponding valve optimal opening; In step 3.4, the loss function is minimized is: （15）； （16）； （17）； （18）； （19）； （20）； wherein, is a pressure constraint loss, is a flow constraint loss, is a power consumption loss, is a power consumption loss, is a regularization term, is a pressure constraint loss coefficient, is a flow constraint loss coefficient, is a power consumption loss coefficient, is a regularization term coefficient, is a model predicted valve opening, is a squared Euclidean norm, is a sum over samples of the private dataset consisting of the system state features and the optimal valve opening corresponding to the system state features and the optimal valve opening corresponding to the system state features In the step 3.7, the global learning rate is in the range of 0.1-1.0; The convergence condition of the step 3.8 is: (1) the global model performance no longer improves significantly; (2) the maximum training round T is reached; (3) the variation of the global model parameters is lower than a threshold, wherein the threshold range is .

7. The AI-based collaborative energy-saving valve cluster control method according to claim 6, characterized in that, The fourth step is specifically: Step 4.1, state acquisition; the valve cluster system acquires the state data of the entire pipe network in real time through the sensor network, and constructs the input feature vector at the current time ; Step 4.2, instruction generation; based on the input feature vector of the current time of the valve cluster system collected in step 4.1, it is input into the deployed third step training converged valve cluster system global model , and then the valve cluster system global model is calculated once forward propagation, outputting the optimal valve opening degree vector instruction of all N valves at the current time : （27）； Step 4.3, instruction issuing and execution; based on the optimal valve opening vector instruction obtained in step 4.2 , which is issued to the valve cluster system to drive the valve actuator to act. Step 4.4, performance evaluation and continuous learning; Based on the action of the valve actuator driven by step 4.3, the total power consumption of the valve cluster system after execution Continuous monitoring is performed, and performance evaluation is performed.

8. The AI-based collaborative energy-saving valve cluster control method according to claim 7, characterized in that, In the step 4.4: If the actual energy-saving effect is good or reaches the expected target, the global energy efficiency of the valve cluster system is optimized; If the actual energy-saving effect is not good or does not reach the expected target, a new intelligent optimization model training and update is triggered, so that the intelligent optimization model can continuously learn from operation and realize dynamic and continuous optimization of the energy efficiency of the valve cluster system.

9. An AI-based collaborative energy-saving valve cluster control system, characterized in that, The AI collaborative energy-saving valve cluster control method of any one of claims 1-8 is realized by the energy-saving valve cluster control system, which comprises a valve cluster system modeling module, an intelligent optimization model construction module, an intelligent optimization model federated learning training module, and a collaborative control and continuous learning module; specifically: The valve cluster system modeling module is composed of pumps, fans and valves in the pipe network, and is used to construct a total energy consumption model of the valve cluster system; The intelligent optimization model construction module accesses the valve cluster system modeling module, and is used for constructing an intelligent optimization model, realizing direct and dynamic mapping of valve cluster system state and valve opening optimization solution ​ The intelligent optimization model federated learning training module accesses the intelligent optimization model construction module and the valve cluster system modeling module, and is used for intelligent optimization model training and update, so as to continuously update and dynamically optimize the intelligent optimization model in actual operation; The collaborative control and continuous learning module accesses the intelligent optimization model federated learning training module, the intelligent optimization model construction module and the valve cluster system modeling module, and is used to generate, issue and execute the optimal valve opening degree vector instruction of all valves at the current time, so that the intelligent optimization model can perform performance evaluation and continuous learning, and realize global energy efficiency optimization and dynamic and continuous optimization of the valve cluster system.

Images