Wind power plant state optimization decision-making method and system based on graph neural network

Abstract

The invention provides a wind power plant state optimization decision-making method and system based on a graph neural network, and relates to the technical field of artificial intelligence, and the method comprises the steps: firstly constructing a wind power plant state association graph, taking equipment as a node, taking an operation state influence relation between the equipment as an edge, dynamically adjusting the association strength of the edge, and carrying out the node feature propagation through a pre-configured graph neural network, the method comprises the following steps: generating node comprehensive state representation, constructing an evaluation system, evaluating the state of a wind power plant from multiple dimensions, outputting related vectors, and generating an optimization scheme comprising multiple strategies according to an evaluation result. Finally, the scheme is converted into a standardized instruction which is executed by a centralized control system driving node, so that the state of the wind power plant can be comprehensively evaluated, dynamic optimization decision is realized, and the operation efficiency and stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a wind farm state optimization decision-making method and system based on a graph neural network. BACKGROUND

[0002] In the field of wind farm operation management, accurately assessing the state of the wind farm and formulating effective optimization decision-making schemes are key to improving the operation efficiency and stability of the wind farm. Traditional wind farm state assessment methods often focus on independent analysis of individual wind turbine units or local equipment, ignoring the complex operating state influence relationship between devices within the wind farm. For example, for wind turbine units, only their own power generation power, speed and other parameters are considered, without fully considering the mutual influence between adjacent wind turbine units, junction box groups, power collection line sections and step-up transformer units.

[0003] At the same time, existing wind farm optimization decision-making schemes are mostly based on fixed rules or simple statistical models, which are difficult to adapt to the dynamic changes of the operating state of the wind farm. In actual operation, the operating state of each device in the wind farm is affected by various factors such as wind speed, wind direction, equipment aging, etc., resulting in a constantly changing relationship between devices. Traditional optimization decision-making methods cannot capture these changes in real time, so they cannot adjust the optimization strategy in a timely manner, and the overall operation efficiency of the wind farm cannot be effectively improved. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the present application provides a wind farm state optimization decision-making method based on a graph neural network, the method comprising:

[0005] Constructing a wind farm state correlation graph, the nodes corresponding to wind turbine units, junction box groups, power collection line sections and step-up transformer units within the wind farm, and the edges representing the operating state influence relationship between the nodes, the correlation strength of the edges being dynamically adjusted according to the real-time state interaction frequency between the nodes;

[0006] Performing node feature propagation on the wind farm state correlation graph by a pre-configured graph neural network, the graph neural network comprising a graph sampling layer, a graph aggregation layer and a feature mapping layer, the graph sampling layer filtering a core neighborhood node set, the graph aggregation layer correlating and aggregating the state features of the core neighborhood nodes to generate neighborhood correlation features, and the feature mapping layer fusing the node's own state features and the neighborhood correlation features to generate a node comprehensive state representation;

[0007] Constructing a wind farm state assessment system based on the node comprehensive state representation, the wind farm state assessment system comprising a node operating stability assessment module, an inter-node coordinated operation assessment module and an overall output efficiency assessment module, respectively outputting a node stability vector, a node coordination degree vector and an overall efficiency vector;

[0008] generate a wind farm state optimization scheme according to the output result of the wind farm state evaluation system, the wind farm state optimization scheme comprising a node operation mode adjustment strategy, an inter-node collaborative control strategy and a resource scheduling priority strategy;

[0009] convert the wind farm state optimization scheme into a standardized control instruction and send the standardized control instruction to a wind farm centralized control system, and the centralized control system drives the corresponding nodes to perform state adjustment actions.

[0010] In still another aspect, the present application also provides a wind farm state optimization decision system based on a graph neural network, comprising a processor, a machine-readable storage medium, the machine-readable storage medium being connected with the processor, the machine-readable storage medium being used for storing programs, instructions or codes, and the processor being used for executing the programs, instructions or codes in the machine-readable storage medium to realize the above-mentioned method.

[0011] Based on the above aspects, the present application can comprehensively and accurately depict the complex correlation between each device in the wind farm by constructing a wind farm state correlation graph, taking the wind turbine unit, the current box group, the power collection line section and the booster unit in the wind farm as nodes, taking the influence relationship between the nodes as edges, and dynamically adjusting the correlation strength of the edges according to the real-time state interaction frequency. The graph neural network is pre-configured to perform node feature propagation on the wind farm state correlation graph, the core neighborhood node set is filtered through the graph sampling layer, the neighborhood correlation features are generated by the graph aggregation layer for the core neighborhood node state features, and the node comprehensive state representation is generated by the feature mapping layer by fusing the node state features and the neighborhood correlation features. The depth mining and accurate representation of the wind farm state are realized. The wind farm state evaluation system constructed based on the node comprehensive state representation comprises a node operation stability evaluation module, an inter-node collaborative operation evaluation module and an overall output efficiency evaluation module, which can comprehensively evaluate the state of the wind farm from multiple dimensions, output a node stability vector, a node coordination degree vector and an overall efficiency vector, and generate a wind farm state optimization scheme according to the output result of the wind farm state evaluation system. The wind farm state optimization scheme comprises a node operation mode adjustment strategy, an inter-node collaborative control strategy and a resource scheduling priority strategy, which can realize dynamic optimization and adjustment of the wind farm state, effectively improve the overall operation efficiency and stability of the wind farm, and finally convert the wind farm state optimization scheme into a standardized control instruction and send the standardized control instruction to a wind farm centralized control system. The centralized control system drives the corresponding nodes to perform state adjustment actions, ensuring accurate execution of the optimization decision and further improving the operation and management level of the wind farm. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is an execution flow schematic diagram of the wind farm state optimization decision method based on the graph neural network provided by the embodiment of the present application.

[0013] Figure 2is a schematic diagram of exemplary hardware and software components of a wind farm state optimization decision system based on a graph neural network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0014] The present application will be described in detail below with reference to the accompanying drawings of the specification, Figure 1 is a flowchart of a wind farm state optimization decision method based on a graph neural network provided by an embodiment of the present application. The wind farm state optimization decision method based on a graph neural network will be described in detail below.

[0015] Step S110: Construct a wind farm state association graph, nodes correspond to wind turbine units, combiner box groups, power collection line sections and step-up transformer units in the wind farm, edges represent the operating state influence relationship between nodes, and the association strength of the edges is dynamically adjusted according to the real-time state interaction frequency between nodes.

[0016] In the actual operation of the wind farm, the wind turbine units, combiner box groups, power collection line sections and step-up transformer units in the wind farm each bear different functions and have close operating state influence relationships with each other. The wind turbine unit is the core equipment for converting wind energy into electrical energy, and its operating state directly affects the subsequent steps of power collection, transmission and step-up. The combiner box group is responsible for collecting the power output by multiple wind turbine units, and its operating state stability will affect the input power of the power collection line section. The power collection line section is the channel for power transmission, and its operating condition will further affect the work of the step-up transformer unit. The step-up transformer unit steps up the power transmitted by the power collection line section for efficient long-distance transmission.

[0017] In order to comprehensively and accurately depict the relationships between these devices, it is necessary to construct a wind farm state association graph. The wind turbine units, combiner box groups, power collection line sections and step-up transformer units are respectively corresponded to nodes in the graph, and each node represents a specific device entity. The edges are used to represent the operating state influence relationship between nodes, and the above relationship is dynamically changed, and the association strength of the edges is dynamically adjusted according to the real-time state interaction frequency between nodes. For example, when the real-time state interaction between two nodes is frequent, it means that their relationship is close, and the association strength of the edge will be correspondingly improved; on the contrary, if the interaction frequency is low, the association strength of the edge will be reduced. Through the above method of dynamically adjusting the association strength, the actual operating state influence relationship between the devices in the wind farm can be more truly reflected.

[0018] Step S111: Determine the specific device entities of the wind turbine units, combiner box groups, power collection line sections and step-up transformer units in the wind farm, and set each device entity as an independent node of the wind farm state association graph, and give each node a unique device identification code.

[0019] To build the wind farm state correlation graph, the specific equipment entities in the wind farm need to be first identified. For wind turbine units, each wind turbine has its unique model, performance parameters and installation location, and it operates independently but is correlated with each other in the wind farm. The combiner box group is a collection of multiple combiner boxes, and different combiner box groups can be connected with different numbers and types of wind turbine units. The power collection line section is composed of power transmission equipment such as cables, and different power collection line sections can have differences in electrical characteristics such as length, resistance, and capacitance. The step-up transformer unit is a transformer device with a specific transformation ratio and capacity, used to change the voltage level of electric energy.

[0020] These specific equipment entities are set as independent nodes of the wind farm state correlation graph in order to effectively analyze and manage each device individually. Each node is assigned a unique device identification code. For example, a specific coding rule can be used, and the node code of a wind turbine unit starts with "FJ" followed by the serial number of the wind turbine; the node code of a combiner box group starts with "HB"; the node code of a power collection line section starts with "JL"; and the node code of a step-up transformer unit starts with "SB". In this way, each node in the wind farm state correlation graph can be accurately identified and distinguished by the device identification code.

[0021] Step S112: Collect real-time running state data of each node, which includes device current running mode information, state monitoring data and historical state change records.

[0022] The real-time running state data of each node is a key source of information for understanding the operation of the equipment. The device current running mode information reflects the working state of the equipment at a certain time, for example, the wind turbine unit can be in power generation mode, shutdown maintenance mode or standby mode, etc. Different running modes will have different effects on the performance and energy consumption of the equipment. The state monitoring data is the real-time collected equipment running parameters through various sensors. For wind turbine units, these parameters can include speed, power output, blade angle, etc.; for combiner box groups, they can include current, voltage, temperature, etc.; for power collection line sections, they can include line temperature, resistance, transmission power, etc.; for step-up transformer units, they can include transformation ratio, oil temperature, winding temperature, etc. These state monitoring data can reflect the running health status of the equipment in real time.

[0023] The historical state change record is the running state change of the device in the past period of time, which can help analyze the running trend and stability of the device. For example, by checking the historical power output record of the fan unit, the power generation capacity change of the fan unit under different seasons and different wind speed conditions can be known; by analyzing the historical temperature record of the busbar group, whether there is a potential fault hidden danger such as overheating can be judged. Collecting real-time running state data of each node requires installing corresponding sensors on each device and transmitting these data in real time to the data collection center through the data transmission network for storage and analysis.

[0024] Step S113: Analyzing the upstream and downstream relationship of different nodes in the power transmission path, identifying the direct state influence relationship between nodes according to the upstream and downstream relationship, and establishing a direct influence relationship edge. The direct influence relationship edge connects two nodes that exist in a direct power transmission path.

[0025] In the power transmission path of the wind farm, there is a clear upstream and downstream relationship between different nodes. The fan unit is the source of electricity, and the output power of the fan unit is first collected by the busbar group. The busbar group, as an intermediate link, integrates the power of multiple fan units and transmits it to the collection line section. The collection line section transmits the power to the step-up transformer unit, and after processing by the step-up transformer unit, the power is transmitted to the power grid.

[0026] According to the above upstream and downstream relationship, the direct state influence relationship between nodes can be identified. For example, the running state of the fan unit directly affects the input power of the busbar group. If the fan unit fails or the output power decreases, the power collected by the busbar group will also decrease accordingly. The state of the busbar group directly affects the transmission of the collection line section. If the busbar group has a short circuit fault, it may cause the current of the collection line section to abnormally increase. The state of the collection line section further affects the work of the step-up transformer unit. Overload or short circuit of the collection line section may cause damage to the step-up transformer unit.

[0027] For two nodes that exist in a direct power transmission path, a direct influence relationship edge is established. This edge represents that there is a direct running state influence between the two nodes, and the state change of one node will directly affect the other node. By establishing the direct influence relationship edge, the direct association relationship between devices can be effectively displayed in the wind farm state association graph.

[0028] Step S114: Analyzing the master-slave relationship of different nodes in the control instruction interaction, identifying the indirect state influence relationship between nodes according to the master-slave relationship, establishing an indirect influence relationship edge, and connecting two nodes that exist in a control instruction interaction.

[0029] In addition to the upstream and downstream relationship in the power transmission path, there is also a master-slave relationship in the control instruction interaction between different nodes. A wind farm is usually equipped with a centralized control system, which can send control instructions to each device node to achieve overall control and optimized operation of the wind farm. For example, the centralized control system can send instructions to adjust the speed, blade angle, etc. of the wind turbine unit according to environmental conditions such as wind speed, wind direction, and the demand of the power grid, so as to improve the power generation efficiency of the wind turbine; can send switch control instructions to the combiner box group to achieve reasonable distribution of the power collected by different wind turbine units; can send power regulation instructions to the power collection line section to balance the load of each line; can send transformer ratio adjustment instructions to the step-up transformer unit to meet the voltage requirements of the power grid.

[0030] In the above control instruction interaction, the centralized control system is dominant and is the master node, while each device node is subordinate and is the slave node. According to the above master-slave relationship, the indirect state influence relationship between nodes can be identified. When the centralized control system sends a control instruction to a slave node, the state change of the slave node may indirectly affect other related nodes. For example, the centralized control system sends a command to a wind turbine unit to reduce power output, and the power reduction of the wind turbine unit will cause the input power of the combiner box group connected thereto to decrease, which may affect the operating state of the power collection line section and the step-up transformer unit.

[0031] For two nodes that exist control instruction interaction, an indirect influence relationship edge is established. This edge represents that there is an indirect operating state influence between the two nodes through the control instruction. By establishing the indirect influence relationship edge, the mutual influence relationship between devices in the wind farm state association graph can be comprehensively reflected, not only considering the direct influence of power transmission, but also considering the indirect influence brought by the control instruction.

[0032] Step S115: Based on the real-time state interaction record between nodes, the number of state interactions and the interaction duration corresponding to the direct influence relationship edge in the preset time window are counted, and the association strength of the direct influence relationship edge is dynamically adjusted according to the number of state interactions and the interaction duration.

[0033] Step S1151: Obtain the real-time state interaction record of the two nodes connected by the direct influence relationship edge. The real-time state interaction record includes the start timestamp and the end timestamp of each interaction.

[0034] To dynamically adjust the association strength of the direct influence relationship edge, the real-time state interaction records of the two nodes connected by the direct influence relationship edge need to be obtained first. These records detail the start time and end time of each state interaction between the two nodes. Through analysis of these timestamps, the specific circumstances of the state interaction between the nodes can be obtained. For example, for the direct influence relationship edge between the fan unit and the busbar cabinet group, the real-time state interaction records will show the start time and end time of each time the fan unit outputs power to the busbar cabinet group, as well as the start time and end time of each time the busbar cabinet group feeds back the operating state information to the fan unit.

[0035] Step S1152: Define a preset time window, and count the total number of interaction events in the real-time state interaction records within the preset time window as the state interaction frequency.

[0036] A preset time window is defined, which can be set according to actual needs, for example, it can be one hour, one day, or one week, etc. Within the preset time window, the total number of interaction events in the real-time state interaction records is counted, which is taken as the state interaction frequency. The state interaction frequency reflects the frequency of state interaction between the two nodes within the time period. If the state interaction frequency is high, it means that the connection between the two nodes is close and the mutual influence is large; on the contrary, if the state interaction frequency is low, it means that the association between them is relatively weak.

[0037] Step S1153: Based on the start timestamp and end timestamp of each interaction, the duration of each interaction is calculated, and the total interaction duration is obtained by accumulating the duration of all interactions.

[0038] According to the start timestamp and end timestamp of each interaction, the duration of each interaction can be calculated. For example, for an interaction in which the fan unit outputs power to the busbar cabinet group, the duration of this interaction can be obtained by subtracting the start timestamp from the end timestamp. The total interaction duration is obtained by accumulating the duration of all interactions within the preset time window. The total interaction duration reflects the total length of time of state interaction between the two nodes within the time period. A longer total interaction duration means that the two nodes have more in-depth interaction and influence.

[0039] Step S1154: According to the state interaction frequency, query the preset frequency-strength mapping table to obtain the frequency contribution strength component.

[0040] A frequency-intensity mapping table is preset, which records the intensity components corresponding to different state interaction frequencies. The mapping table is established according to a large amount of historical data and actual experience, and reflects the relationship between the state interaction frequency and the associated intensity. According to the statistical state interaction frequency, the corresponding frequency contribution intensity component is queried in the frequency-intensity mapping table. For example, if the state interaction frequency is in a certain interval, the corresponding intensity component value can be found in the frequency-intensity mapping table, which represents the contribution degree of the state interaction frequency to the associated intensity.

[0041] Step S1155: According to the total interaction duration, query the preset duration-intensity mapping table to obtain the duration contribution intensity component.

[0042] Similarly, a duration-intensity mapping table is preset, which records the intensity components corresponding to different total interaction durations. According to the calculated total interaction duration, the corresponding duration contribution intensity component is queried in the duration-intensity mapping table. The duration contribution intensity component reflects the contribution degree of the total interaction duration to the associated intensity. For example, a longer total interaction duration may correspond to a higher duration contribution intensity component, indicating that long-time interaction between two nodes has a greater impact on the associated intensity between them.

[0043] Step S1156: Feature fusion is performed on the frequency contribution intensity component and the duration contribution intensity component to generate the current associated intensity of the direct influence relationship edge.

[0044] The frequency contribution intensity component and the duration contribution intensity component are feature fused to generate the current associated intensity of the direct influence relationship edge. The feature fusion method can choose a suitable method according to the specific situation, for example, a weighted summation method can be used, and different weights are assigned to the frequency contribution intensity component and the duration contribution intensity component, and then they are added to obtain the current associated intensity. The purpose of this is to comprehensively consider the influence of state interaction frequency and interaction duration on the associated intensity, so that the calculation of the associated intensity more accurately reflects the actual association degree between two nodes.

[0045] Step S1157: Smooth transition processing is performed on the current associated intensity and the historical associated intensity of the direct influence relationship edge, and the associated intensity parameter of the direct influence relationship edge is updated according to the current associated intensity after the smooth transition.

[0046] In order to avoid sharp fluctuations in the correlation strength, the current correlation strength needs to be smoothly transitioned with the historical correlation strength of the directly affecting relationship edge. The smoothing transition can adopt a moving average method, and a more stable correlation strength value is obtained by weighted average of the current correlation strength and the historical correlation strength. According to the current correlation strength after smoothing transition, the correlation strength parameter of the directly affecting relationship edge is updated, so that the change of the correlation strength is more stable and more consistent with the actual operation. At the same time, the updated correlation strength parameter will be used for subsequent node feature propagation and state evaluation and other operations.

[0047] Step S116: Based on the inter-node control instruction response record, the instruction response and timeliness and response accuracy corresponding to the indirectly affecting relationship edge in the preset time window are counted, and the correlation strength of the indirectly affecting relationship edge is dynamically adjusted according to the instruction response and timeliness and response accuracy.

[0048] Step S1161: Obtain the control instruction response record of the two nodes connected by the indirectly affecting relationship edge, and the control instruction response record contains the sending time, receiving time and execution result of each instruction.

[0049] In order to dynamically adjust the correlation strength of the indirectly affecting relationship edge, the control instruction response record of the two nodes connected by the indirectly affecting relationship edge needs to be obtained. These records record in detail the sending time, receiving time and execution result of each control instruction. For example, for the control instruction sent by the centralized control system to the fan unit, the control instruction response record will show the sending time of the instruction, the time when the fan unit receives the instruction, and the power output, speed and other execution results after the fan unit executes the instruction. Through the analysis of these records, the response of the node to the control instruction can be known.

[0050] Step S1162: Draw a preset time window, and count the instruction response and timeliness index in the control instruction response record in the preset time window. The instruction response and timeliness index reflects the time interval from receiving the instruction to starting execution.

[0051] A preset time window is drawn, and the instruction response and timeliness index in the control instruction response record in the preset time window is counted. The instruction response and timeliness index reflects the time interval from receiving the control instruction to starting execution. The shorter the time interval, the more timely the response of the node; the longer the time interval, the longer the delay of the node in responding to the control instruction, which may affect the coordinated operation of the entire wind farm. For example, if the centralized control system sends an instruction to adjust the speed of the fan unit, the fan unit can start adjusting the speed in a short time, which indicates that the instruction response and timeliness is good.

[0052] Step S1163: Statistics the instruction response accuracy index in the control instruction response record within the preset time window. The instruction response accuracy index reflects the degree of compliance of the result of the node executing the instruction with the instruction requirement.

[0053] Within the preset time window, the instruction response accuracy index in the control instruction response record is counted. The index reflects the degree of compliance of the result of the node executing the control instruction with the instruction requirement. For example, the central control system requires the fan unit to adjust the power output to a certain specific value, and the instruction response accuracy index is the closeness of the actual power output of the fan unit to the specific value. If the actual power output is very close to the required power output of the instruction, it means that the instruction response accuracy of the node is high; otherwise, if there is a large deviation, it means that the instruction response accuracy is low.

[0054] Step S1164: According to the instruction response timeliness index, the preset timeliness-intensity mapping table is queried to obtain the timeliness contribution intensity component.

[0055] A timeliness-intensity mapping table is preset, which records the intensity components corresponding to different instruction response timeliness indexes. According to the instruction response timeliness index obtained by counting, the corresponding timeliness contribution intensity component is queried in the timeliness-intensity mapping table. The timeliness contribution intensity component reflects the contribution degree of the instruction response timeliness to the association strength. For example, a shorter instruction response time interval may correspond to a higher timeliness contribution intensity component, indicating that the node response timeliness has a positive role in improving the association strength of the indirect influence relationship edge.

[0056] Step S1165: According to the instruction response accuracy index, the preset accuracy-intensity mapping table is queried to obtain the accuracy contribution intensity component.

[0057] Similarly, an accuracy-intensity mapping table is preset, which records the intensity components corresponding to different instruction response accuracy indexes. According to the instruction response accuracy index obtained by counting, the corresponding accuracy contribution intensity component is queried in the accuracy-intensity mapping table. The accuracy contribution intensity component reflects the contribution degree of the instruction response accuracy to the association strength. For example, higher instruction response accuracy may correspond to higher accuracy contribution intensity component, indicating that the node accurately executing the instruction has an important influence on the association strength of the indirect influence relationship edge.

[0058] Step S1166: The timeliness contribution intensity component and the accuracy contribution intensity component are fused to generate the current association strength of the indirect influence relationship edge.

[0059] The timeliness contribution strength component and the accuracy contribution strength component are fused to generate the current correlation strength of the indirect influence relationship edge. The fusion manner can be similar to the calculation of the correlation strength of the direct influence relationship edge, and a weighted summation method is adopted. Through the above manner, the influence of the instruction response timeliness and the response accuracy on the correlation strength is comprehensively considered, so that the correlation strength of the indirect influence relationship edge can more accurately reflect the actual correlation degree between two nodes through the control instruction.

[0060] Step S1167: The current correlation strength is smoothly transitioned with the historical correlation strength of the indirect influence relationship edge, and the correlation strength parameter of the indirect influence relationship edge is updated according to the smoothly transitioned current correlation strength.

[0061] In order to avoid the sharp fluctuation of the correlation strength of the indirect influence relationship edge, the current correlation strength is smoothly transitioned with the historical correlation strength of the indirect influence relationship edge. A moving average method can also be adopted to obtain a relatively stable correlation strength value. The correlation strength parameter of the indirect influence relationship edge is updated according to the smoothly transitioned current correlation strength, so that the change of the correlation strength of the indirect influence relationship edge is more stable.

[0062] Step S117: The node device identification code, the real-time running state data, the direct influence relationship edge, the indirect influence relationship edge and the corresponding correlation strength are integrated to form an initial wind farm state correlation graph.

[0063] After the determination of the nodes, the collection of the data, the establishment of the edges and the calculation of the correlation strength are completed, the node device identification code, the real-time running state data, the direct influence relationship edge, the indirect influence relationship edge and the corresponding correlation strength need to be integrated. The node device identification code is used to uniquely identify each node, the real-time running state data reflects the actual running condition of each node, the direct influence relationship edge and the indirect influence relationship edge represent the running state influence relationship between the nodes, and the correlation strength quantifies the strength of the above influence relationship.

[0064] These information are integrated together to form an initial wind farm state correlation graph. The initial wind farm state correlation graph visually displays the running state influence relationship and the correlation strength between the device nodes in the wind farm in a graphical manner. Through the initial wind farm state correlation graph, the mutual relationship between the devices in the wind farm can be quickly obtained.

[0065] Step S118: The node real-time running state data and the interaction record are re-collected according to a preset update period, the correlation strength of each edge is repeatedly adjusted, and the dynamic update and maintenance of the wind farm state correlation graph is realized.

[0066] The operation state of a wind farm is constantly changing, influenced by factors such as wind speed, wind direction, equipment failure, etc. In order to ensure that the wind farm state association graph accurately reflects the actual operation of the wind farm, it is necessary to re-collect node real-time operation state data and interaction records according to the preset update period. The update period can be set according to actual needs, for example, it can be updated every hour, every day, or every week.

[0067] The re-collected data includes real-time operation state data of each node, such as power output and speed of wind turbine units, current and voltage of busbar groups, and real-time state interaction records and control instruction response records between nodes. According to the re-collected data, the process of adjusting the association strength of each edge in steps S115 and S116 is repeated, that is, the association strength of the direct influence relationship edge is adjusted according to the real-time state interaction frequency between nodes, and the association strength of the indirect influence relationship edge is adjusted according to the control instruction response between nodes.

[0068] Through continuous data re-collection and association strength adjustment, dynamic update and maintenance of the wind farm state association graph are realized. Thus, the wind farm state association graph can always reflect the latest operation state influence relationship between each device node in the wind farm.

[0069] Step S120: Perform node feature propagation on the wind farm state association graph through a pre-configured graph neural network, which includes a graph sampling layer, a graph aggregation layer, and a feature mapping layer. The graph sampling layer selects a core neighborhood node set, the graph aggregation layer aggregates the state feature association of the core neighborhood nodes to generate neighborhood association features, and the feature mapping layer fuses the node's own state feature and neighborhood association feature to generate a comprehensive state representation of the node.

[0070] After the wind farm state association graph is constructed, it needs to be propagated through a pre-configured graph neural network to mine the potential relationship and features between nodes. The graph neural network is a neural network specially designed for processing graph structure data, which can effectively capture the dependency relationship between nodes in the graph. The graph neural network includes a graph sampling layer, a graph aggregation layer, and a feature mapping layer, each layer has its specific function, and through the cooperative work of these layers, the propagation and fusion of node features can be realized.

[0071] Step S121: Convert the real-time operation state data of each node in the wind farm state association graph into a node's own state feature vector, and the feature dimension of the node's own state feature vector is adapted to the receiving dimension of the input layer of the graph neural network.

[0072] The real-time operation state data of each node in the wind farm state association graph is multi-dimensional, including current operation mode information, state monitoring data, and historical state change records of the device. In order to input these data into the graph neural network for processing, it is necessary to convert them into a node's own state feature vector.

[0073] First, feature extraction and selection are performed on the real-time running state data to determine which features are the most critical for describing the node state. For example, for a fan unit, the rotational speed, power output, and blade angle can be selected as the key features; for a busbar group, the current, voltage, and temperature can be selected as the key features. Then, these key features are arranged in a certain order to form a node self-state feature vector.

[0074] During the conversion process, it is necessary to ensure that the feature dimension of the node self-state feature vector is compatible with the receiving dimension of the input layer of the graph neural network. If the feature dimensions do not match, it may cause the data to be unable to be normally input into the graph neural network for processing. Therefore, it is necessary to adjust the dimension of the node self-state feature vector according to the requirements of the input layer of the graph neural network, such as through feature dimension reduction or feature expansion to achieve the adaptation of the dimension.

[0075] Step S122: input the node self-state feature vector and the association relationship matrix of the wind farm state association graph into the graph sampling layer, and the association relationship matrix contains the edge type and association strength information between nodes.

[0076] After the conversion of the node self-state feature vector is completed, it is input into the graph sampling layer of the graph neural network together with the association relationship matrix of the wind farm state association graph. The association relationship matrix is a matrix reflecting the edge type and association strength information between nodes. The rows and columns of the matrix correspond to the nodes in the wind farm state association graph, and the elements in the matrix represent the edge type (such as direct influence relationship edge or indirect influence relationship edge) and association strength between two nodes.

[0077] The graph sampling layer needs to select the core neighborhood node set according to the association relationship matrix and the node self-state feature vector. The association relationship matrix provides the connection information and association strength information between nodes, and through these information, the graph sampling layer can determine which nodes are more closely associated with the current node, thereby selecting the core neighborhood nodes.

[0078] Step S123: in the graph sampling layer, for each node, sort all its adjacent nodes in the association relationship matrix according to the association strength from high to low, and select the adjacent nodes with a high ranking and an association strength satisfying a preset threshold to form a core neighborhood node set.

[0079] Step S1231: for the current processing node, extract the identification information and corresponding association strength values of all adjacent nodes related to the node from the association relationship matrix.

[0080] In the graph sampling layer, for the current processing node, the identification information and the corresponding association strength value of all adjacent nodes associated with the node need to be extracted from the association matrix. The association matrix stores the association strength information between all nodes, and by searching the corresponding row or column of the current processing node, the association strength values of all adjacent nodes can be obtained. At the same time, according to the index information of the association matrix, the identification information of the adjacent nodes, such as the device identification code of the node, can be obtained.

[0081] Step S1232: The adjacent nodes are sorted according to the association strength value from high to low to form an adjacent node sorting sequence.

[0082] After obtaining the association strength values of the adjacent nodes, the adjacent nodes are sorted according to the association strength value from high to low. The purpose of sorting is to effectively know which adjacent nodes are more closely associated with the current processing node. Through sorting, the adjacent nodes with higher association strength are arranged in front, and the adjacent nodes with lower association strength are arranged in back, forming an adjacent node sorting sequence.

[0083] Step S1233: Traverse the adjacent node sorting sequence, and sequentially determine whether the association strength value of each adjacent node is greater than or equal to the association strength preset threshold.

[0084] For example, step S12331: Obtain the association strength values of all direct influence relationship edges and indirect influence relationship edges in the wind farm state association graph to form an association strength data set.

[0085] To determine the association strength preset threshold, first, the association strength values of all direct influence relationship edges and indirect influence relationship edges in the wind farm state association graph need to be obtained. These association strength values reflect the association tightness between all nodes in the wind farm. Collecting these association strength values forms an association strength data set. The association strength data set contains the association strength information between all nodes in the wind farm.

[0086] Step S12332: Statistically analyze the association strength data set to calculate the mean and standard deviation of the association strength data set.

[0087] Statistically analyze the association strength data set to calculate the mean and standard deviation. The mean reflects the average level of the association strength, which can be used as a reference value to measure the overall association tightness between nodes. The standard deviation reflects the dispersion of the association strength, that is, the fluctuation of the association strength value relative to the mean. By calculating the mean and standard deviation, the distribution characteristics of the association strength data set can be basically understood.

[0088] Step S12333: Based on the mean and standard deviation, set an initial association strength preset threshold, and the initial association strength initial threshold is the mean minus the standard deviation multiplied by a preset multiple.

[0089] According to the calculated mean and standard deviation, an initial correlation strength preset threshold is set. The initial correlation strength preset threshold is obtained by subtracting the preset multiple of the standard deviation from the mean. The selection of the preset multiple needs to be adjusted according to the actual situation, which determines the strictness of screening the core neighbor nodes. If the preset multiple is larger, the initial correlation strength preset threshold will be lower, and the set of screened core neighbor nodes will be relatively larger; if the preset multiple is smaller, the initial correlation strength preset threshold will be higher, and the set of screened core neighbor nodes will be relatively smaller.

[0090] Step S12334: Collect evaluation data of the set of historical core neighbor nodes, and the evaluation data includes neighbor node coverage and feature propagation accuracy indicators.

[0091] The evaluation data of the set of historical core neighbor nodes is collected, and these data are used to evaluate the quality of the set of core neighbor nodes. The neighbor node coverage reflects the coverage degree of the set of core neighbor nodes to the adjacent nodes of the current processing node, that is, the proportion of the number of adjacent nodes contained in the set of core neighbor nodes to the number of all adjacent nodes. The feature propagation accuracy indicator reflects the accuracy degree of feature propagation through the set of core neighbor nodes, which can be measured by comparing the difference between the propagated features and the actual features, for example.

[0092] Step S12335: Adjust the initial threshold according to the neighbor node coverage in the evaluation data, and if the coverage is lower than the preset coverage benchmark, lower the initial threshold, and if the coverage is higher than the preset coverage upper limit, increase the initial threshold.

[0093] The initial threshold is adjusted according to the neighbor node coverage in the evaluation data. If the neighbor node coverage is lower than the preset coverage benchmark, it means that the set of screened core neighbor nodes may be too narrow and cannot well cover the adjacent nodes of the current processing node, so the initial threshold needs to be lowered to expand the range of the set of core neighbor nodes. If the neighbor node coverage is higher than the preset coverage upper limit, it means that the set of screened core neighbor nodes may be too large and contains some nodes with weak correlation with the current processing node, so the initial threshold needs to be increased to narrow the range of the set of core neighbor nodes.

[0094] Step S12336: Adjust the initial threshold according to the feature propagation accuracy indicator in the evaluation data, and if the accuracy is lower than the preset accuracy benchmark, increase the initial threshold, and if the accuracy is higher than the preset accuracy upper limit, lower the initial threshold.

[0095] Similarly, the initial threshold is adjusted according to the feature propagation accuracy indicator in the evaluation data. If the feature propagation accuracy is lower than the preset accuracy benchmark, it indicates that the selected core neighborhood node set may contain some nodes that are weakly associated with the current processing node, resulting in inaccurate feature propagation. In this case, the initial threshold needs to be increased to filter out more representative core neighborhood nodes. If the feature propagation accuracy is higher than the preset accuracy upper limit, it indicates that the filtering standard may be too strict, and the initial threshold can be appropriately reduced to increase the number of core neighborhood nodes and improve the comprehensiveness of feature propagation.

[0096] Step S12337: The adjusted threshold is taken as the final association strength preset threshold, and the association strength preset threshold and the corresponding adjustment basis are stored.

[0097] The adjusted threshold is taken as the final association strength preset threshold, and the threshold and the corresponding adjustment basis are stored. The adjustment basis is stored to facilitate subsequent analysis and backtracking to understand the reason and process of threshold adjustment. The final association strength preset threshold will be used for subsequent core neighborhood node filtering operations.

[0098] Step S1234: Add the adjacent nodes with an association strength value greater than or equal to the association strength preset threshold to the temporary neighborhood set.

[0099] When traversing the adjacent node sorting sequence, it is determined in turn whether the association strength value of each adjacent node is greater than or equal to the association strength preset threshold. If the association strength value of a certain adjacent node meets the condition, it is added to the temporary neighborhood set. The temporary neighborhood set is a preliminary selected set of adjacent nodes that are strongly associated with the current processing node.

[0100] Step S1235: If the number of nodes in the temporary neighborhood set does not reach the preset neighborhood size upper limit, continue to select the remaining adjacent nodes with the highest association strength value in the sorting sequence to join the temporary neighborhood set until the preset neighborhood size upper limit is reached.

[0101] If the number of nodes in the temporary neighborhood set does not reach the preset neighborhood size upper limit, it indicates that the number of nodes in the set is not enough, and the remaining adjacent nodes with the highest association strength value need to be selected from the adjacent node sorting sequence to join the temporary neighborhood set. The preset neighborhood size upper limit is to control the size of the core neighborhood node set to avoid excessive size leading to increased computational complexity. By continuously selecting the remaining adjacent nodes with the highest association strength value until the preset neighborhood size upper limit is reached, it ensures that the core neighborhood node set has a certain size and can fully reflect the neighborhood characteristics of the current processing node.

[0102] Step S1236: If the number of nodes in the temporary neighborhood set exceeds the upper limit of the preset neighborhood size, the first preset neighborhood size of adjacent nodes with the highest correlation strength value are selected to form the temporary neighborhood set.

[0103] If the number of nodes in the temporary neighborhood set exceeds the upper limit of the preset neighborhood size, it means that the number of nodes in the set is too large and needs to be adjusted. The first preset neighborhood size of adjacent nodes with the highest correlation strength value are selected to form the temporary neighborhood set. In this way, it can be ensured that the nodes in the temporary neighborhood set are strongly associated with the current processing node, and the size of the set is controlled within the preset range.

[0104] Step S1237: The adjacent nodes in the temporary neighborhood set are de-duplicated to ensure that each adjacent node is unique in the set.

[0105] In order to avoid repeated calculation and data redundancy, the adjacent nodes in the temporary neighborhood set are de-duplicated. In actual situations, there may be duplicate adjacent nodes in the temporary neighborhood set due to data errors or other reasons. Through de-duplication, it is ensured that each adjacent node is unique in the set, improving the efficiency and accuracy of subsequent calculations.

[0106] Step S1238: The de-duplicated temporary neighborhood set is determined as the core neighborhood node set of the current processing node.

[0107] After de-duplication, the temporary neighborhood set is determined as the core neighborhood node set of the current processing node. The core neighborhood node set contains adjacent nodes that are strongly associated with the current processing node, and their state features will have a significant impact on the feature propagation and state evaluation of the current processing node. The core neighborhood node set provides input data for the subsequent graph aggregation layer.

[0108] Step S124: The node state feature vector of each neighborhood node in the core neighborhood node set is input into the graph aggregation layer, and the state features of the neighborhood nodes are integrated by the association aggregation algorithm of the graph aggregation layer to generate neighborhood association features. The association aggregation algorithm assigns feature integration weights based on the correlation strength between the neighborhood nodes and the current node.

[0109] After determining the core neighborhood node set of the current processing node, the node state feature vector of each neighborhood node in the core neighborhood node set is input into the graph aggregation layer of the graph neural network. The main function of the graph aggregation layer is to integrate the state features of these neighborhood nodes in association to generate neighborhood association features.

[0110] The graph aggregation layer adopts a correlation aggregation algorithm to realize the integration of features. The correlation aggregation algorithm assigns different feature integration weights to the state features of each neighborhood node according to the correlation strength between the neighborhood node and the current node. The higher the correlation strength of the neighborhood node, the greater the weight of its state feature in the integration process, indicating that the state of the neighborhood node is more important to the current node. For example, for a fan unit node, the state features of the header box group node with higher correlation strength will be given greater weight in the integration.

[0111] Through the correlation aggregation algorithm, the state features of each neighborhood node are integrated according to the assigned weights to generate neighborhood correlation features. The neighborhood correlation features integrate the state information of each neighborhood node in the core neighborhood node set, reflecting the neighborhood environmental features of the current processing node.

[0112] Step S125: input the neighborhood correlation features into the graph attention layer, analyze the correlation closeness between the neighborhood correlation features and the state feature vector of the current node itself through the attention mechanism, assign attention weights to different feature components in the neighborhood correlation features, and generate weighted neighborhood correlation features.

[0113] After obtaining the neighborhood correlation features, they are input into the graph attention layer. The graph attention layer uses the attention mechanism to analyze the correlation closeness between the neighborhood correlation features and the state feature vector of the current node itself. The attention mechanism can automatically learn the importance of different feature components in the neighborhood correlation features to the current node.

[0114] For each feature component in the neighborhood correlation features, the attention mechanism calculates the correlation degree between it and the state feature vector of the current node itself. The higher the correlation closeness of the feature component, the greater the attention weight it can be assigned; the lower the correlation closeness of the feature component, the smaller the attention weight it can be assigned. For example, if a certain feature component in the neighborhood correlation features is closely related to the power output of the current node, the feature component will be assigned a higher attention weight.

[0115] By assigning attention weights to different feature components in the neighborhood correlation features, weighted neighborhood correlation features are generated. The weighted neighborhood correlation features highlight the feature components that are closely related to the current node, making the subsequent feature fusion more targeted and better reflecting the actual state of the current node.

[0116] Step S126: merge the state feature vector of the current node itself and the weighted neighborhood correlation features into a joint feature vector through a feature splicing operation, perform a nonlinear feature transformation on the joint feature vector, and generate a node comprehensive state representation.

[0117] The node self-state feature vector of the current node and the weighted neighborhood association feature are combined into a joint feature vector through a feature splicing operation. The feature splicing operation is to splice two vectors in the dimension, for example, to connect the node self-state feature vector and the weighted neighborhood association feature vector in a certain order to form a longer vector. The joint feature vector contains the state features of the current node itself and the association features of its neighborhood nodes, and the information is more rich.

[0118] The joint feature vector is subjected to nonlinear feature transformation, which is usually realized through a nonlinear activation function. The nonlinear activation function can map the joint feature vector to a new feature space, enhancing the expression ability of the features. Through nonlinear feature transformation, a node comprehensive state representation is generated. The node comprehensive state representation comprehensively reflects the running state of the node more comprehensively and accurately.

[0119] Step S130: Based on the node comprehensive state representation, a wind farm state evaluation system is constructed, which includes a node running stability evaluation module, a node intercoordination running evaluation module and a whole output performance evaluation module, respectively outputting a node stability vector, a node coordination degree vector and a whole performance vector.

[0120] After obtaining the node comprehensive state representation, in order to comprehensively evaluate the running state of the wind farm, a wind farm state evaluation system needs to be constructed based on the node comprehensive state representation. The wind farm state evaluation system includes three evaluation modules, which evaluate the running state of the wind farm from different angles to provide multi-dimensional evaluation results.

[0121] Step S131: A feature sequence representing the running stability of the equipment is extracted from the node comprehensive state representation, and the feature sequence is input into the node running stability evaluation module. The node running stability evaluation module captures the fluctuation trend of the feature sequence through a sliding window feature extraction method, generates a node stability vector, and the node stability vector contains a fluctuation frequency index and a fluctuation amplitude index.

[0122] Step S1311: The feature dimensions related to the running stability of the equipment are selected from the node comprehensive state representation, the feature values corresponding to the selected feature dimensions are arranged in time sequence, and a feature sequence representing the running stability of the equipment is formed.

[0123] The node comprehensive state representation contains multiple feature dimensions, some of which are closely related to the running stability of the equipment. For example, for a wind turbine unit, its power output, speed and other feature dimensions can reflect its running stability; for a busbar box group, the stability of current and voltage is also an important embodiment of the running stability. The feature dimensions related to the running stability of the equipment are selected from the node comprehensive state representation.

[0124] The feature values corresponding to the selected feature dimensions are arranged in time sequence to form a feature sequence representing the running stability of the equipment. The feature sequence records the changes of the running stability features of the equipment within a period of time. For example, arranging the power output values of the fan unit at different times within a day in time sequence obtains a feature sequence representing the running stability of the fan unit.

[0125] Step S1312: setting the window size and sliding step of the sliding window, the window size is determined according to the change period of the equipment running state, and the sliding step is a preset proportion of the window size.

[0126] In order to capture the fluctuation trend of the feature sequence, the window size and sliding step of the sliding window need to be set. The determination of the window size needs to consider the change period of the equipment running state. For example, for the fan unit, the power output may fluctuate with the change of the wind speed, and the change of the wind speed usually has a certain periodicity, so the window size can be determined according to the change period of the wind speed.

[0127] The sliding step is the distance that the window moves each time, which is usually set as a preset proportion of the window size. The selection of the preset proportion needs to be adjusted according to the actual situation to ensure that the fluctuation trend of the feature sequence can be effectively captured. For example, if the preset proportion is small, the sliding step will be short, the window will move slowly, and the fluctuation trend can be captured in detail; if the preset proportion is large, the sliding step will be long, the window will move fast, and the calculation efficiency can be improved, but some subtle fluctuations may be ignored.

[0128] Step S1313: sliding the sliding window on the feature sequence by the sliding step in turn, and performing statistical analysis on the feature values in each window to calculate the mean and standard deviation of the feature values in the window.

[0129] The sliding window is slid on the feature sequence by the sliding step in turn, and after each sliding, the feature values in the window are statistically analyzed. The mean and standard deviation of the feature values in the window are calculated. The mean reflects the average level of the feature values in the window, and the standard deviation reflects the dispersion degree of the feature values relative to the mean. For example, for the fan unit power output feature values in a window, the calculation of the mean can obtain the average power output in that time period, and the calculation of the standard deviation can obtain the fluctuation of the power output.

[0130] Step S1314: calculating the fluctuation coefficient according to the mean and standard deviation of the feature values in the window, the fluctuation coefficient reflecting the dispersion degree of the feature values in the window.

[0131] The fluctuation coefficient is calculated according to the mean value and the standard deviation of the characteristic values in the window. The fluctuation coefficient is the ratio of the standard deviation to the mean value, which can more intuitively reflect the dispersion degree of the characteristic values in the window. The larger the fluctuation coefficient, the greater the dispersion degree of the characteristic values, and the poorer the running stability of the equipment; the smaller the fluctuation coefficient, the smaller the dispersion degree of the characteristic values, and the better the running stability of the equipment.

[0132] Step S1315: The fluctuation coefficients of all windows are arranged in time sequence to form a fluctuation coefficient sequence, and the overall fluctuation trend of the characteristic sequence is captured through the fluctuation coefficient sequence.

[0133] The fluctuation coefficients of all windows are arranged in time sequence to form a fluctuation coefficient sequence. The fluctuation coefficient sequence records the fluctuation of the characteristic sequence in different time periods. By analyzing the fluctuation coefficient sequence, the overall fluctuation trend of the characteristic sequence can be captured. For example, if the fluctuation coefficient sequence presents a periodic change, it indicates that the running state of the equipment has a periodic fluctuation; if the numerical value of the fluctuation coefficient sequence gradually increases, it indicates that the running stability of the equipment is gradually deteriorating.

[0134] Step S1316: Trend fitting is performed on the fluctuation coefficient sequence to obtain a fluctuation trend curve, and a fluctuation frequency index is extracted from the fluctuation trend curve. The fluctuation frequency index is the number of times that the fluctuation coefficient exceeds a preset fluctuation threshold in a unit time.

[0135] Trend fitting is performed on the fluctuation coefficient sequence, which can use methods such as polynomial fitting and spline fitting. Through trend fitting, a fluctuation trend curve is obtained. The fluctuation trend curve can more effectively show the change trend of the fluctuation coefficient sequence.

[0136] The fluctuation frequency index is extracted from the fluctuation trend curve. The fluctuation frequency index is the number of times that the fluctuation coefficient exceeds a preset fluctuation threshold in a unit time. The preset fluctuation threshold is a pre-set standard for judging whether the fluctuation is abnormal. For example, if the fluctuation coefficient exceeds the preset fluctuation threshold, it indicates that the running state of the equipment has a large fluctuation. The fluctuation frequency index reflects the frequency of the fluctuation of the running state of the equipment. The higher the fluctuation frequency, the poorer the running stability of the equipment.

[0137] Step S1317: A fluctuation amplitude index is extracted from the fluctuation trend curve. The fluctuation amplitude index is the difference between the maximum value and the minimum value in the fluctuation coefficient sequence.

[0138] The fluctuation amplitude index is extracted from the fluctuation trend curve. The fluctuation amplitude index is the difference between the maximum value and the minimum value in the fluctuation coefficient sequence. The fluctuation amplitude reflects the maximum fluctuation range of the running state of the equipment. A larger fluctuation amplitude indicates that the equipment may have a larger unstable situation during operation, which may have an adverse effect on the service life and performance of the equipment.

[0139] Step S1318: Integrate the fluctuation frequency index and the fluctuation amplitude index to generate a node stability vector containing the index name and the corresponding numerical range.

[0140] The fluctuation frequency index and the fluctuation amplitude index are integrated to generate a node stability vector. The node stability vector contains the index name and the corresponding numerical range. For example, the node stability vector can be represented as "fluctuation frequency index: [X1, X2], fluctuation amplitude index: [Y1, Y2]", where [X1, X2] and [Y1, Y2] are the numerical ranges of the fluctuation frequency index and the fluctuation amplitude index, respectively. The node stability vector effectively demonstrates the running stability of the node.

[0141] Step S132: Extract feature pairs representing the interaction characteristics between nodes from the node comprehensive state representation, and input the feature pairs into the inter-node collaborative operation evaluation module. The inter-node collaborative operation evaluation module analyzes the synchronicity and complementarity of the feature pairs through the feature alignment algorithm to generate a node collaboration degree vector, which contains a synchronous response index and a functional complementarity index.

[0142] Feature pairs representing the interaction characteristics between nodes are extracted from the node comprehensive state representation. These feature pairs reflect the interaction relationship between different nodes. For example, for the fan unit and the bus box group nodes, the power output of the fan unit and the input current of the bus box group can form a feature pair; for the collection line section and the step-up transformer unit nodes, the transmission power of the collection line section and the input power of the step-up transformer unit can form a feature pair.

[0143] These feature pairs are input into the inter-node collaborative operation evaluation module. The module uses the feature alignment algorithm to analyze the synchronicity and complementarity of the feature pairs. Synchronicity reflects whether the state changes of different nodes at the same time are consistent. For example, if the power output of the fan unit increases, the input current of the bus box group also increases accordingly, indicating that they have good synchronicity. Complementarity reflects whether the functions of different nodes are complementary to each other. For example, the output of one node can be used as the input of another node, or the functions of two nodes can work together to complete a specific task.

[0144] Through the feature alignment algorithm, the synchronicity and complementarity of the feature pairs are quantitatively analyzed to generate a node collaboration degree vector. The node collaboration degree vector contains a synchronous response index and a functional complementarity index. The synchronous response index reflects the synchronous response ability between nodes, and the functional complementarity index reflects the functional complementarity degree between nodes. The node collaboration degree vector evaluates the collaborative operation of the nodes.

[0145] Step S133: Collect the current environmental perception data of the wind farm, which includes wind speed change trend, wind direction distribution characteristics, and temperature field distribution information.

[0146] The operating state of a wind farm is significantly affected by environmental factors, so it is necessary to collect current environmental perception data of the wind farm. Environmental perception data includes wind speed change trend, wind direction distribution characteristics and temperature field distribution information.

[0147] The wind speed change trend reflects the change of wind speed around the wind farm. Wind speed is a key factor affecting the power generation efficiency of wind turbine units. Different wind speeds will result in different power outputs of wind turbine units. By collecting wind speed change trend data, the change law of wind speed can be known.

[0148] The wind direction distribution characteristics reflect the distribution law of wind direction. Wind direction will affect the windward angle of wind turbine units, and thus affect their power generation efficiency. Knowing the wind direction distribution characteristics can reasonably adjust the installation position and orientation of wind turbine units, and improve the overall power generation efficiency of the wind farm.

[0149] Temperature field distribution information reflects the temperature distribution of different areas in the wind farm. Temperature will affect the performance and life of equipment, for example, excessive temperature may cause equipment to overheat and damage. Collecting temperature field distribution information can timely find abnormal heating of equipment and take appropriate measures to handle it.

[0150] Collecting environmental perception data can be achieved by installing wind speed sensors, wind direction sensors and temperature sensors and other devices in the wind farm. These sensors will transmit the real-time collected data to the data collection center for storage and analysis.

[0151] Step S134: Cross-modal feature fusion is performed between the node comprehensive state representation of all nodes and the environmental perception data to generate the global state feature of the wind farm, and the global state feature of the wind farm is input into the overall output efficiency evaluation module.

[0152] Cross-modal feature fusion is performed between the node comprehensive state representation of all nodes and the environmental perception data. The node comprehensive state representation reflects the operating state of each device node in the wind farm, while the environmental perception data reflects the external environment of the wind farm. Cross-modal feature fusion can integrate different types of data and extract more valuable information.

[0153] When performing cross-modal feature fusion, methods such as feature concatenation and feature weighted summation can be used. For example, the node comprehensive state representation vectors of all nodes and the environmental perception data vectors are concatenated to form a longer vector; or different weights are assigned to the node comprehensive state representation and the environmental perception data, and then weighted summation is performed.

[0154] The wind farm global state feature is generated by cross-modal feature fusion. The wind farm global state feature comprehensively reflects the running state of each device node in the wind farm and the influence of external environmental factors, and more comprehensively reflects the overall running state of the wind farm. The wind farm global state feature is input into the overall output efficiency evaluation module for evaluating the overall output efficiency of the wind farm.

[0155] Step S135: The overall output efficiency evaluation module analyzes the wind farm global state feature through the efficiency mapping function to generate an overall efficiency vector, which includes an output efficiency index and an environmental adaptation index.

[0156] The overall output efficiency evaluation module analyzes the wind farm global state feature through the efficiency mapping function. The efficiency mapping function is a pre-defined function that can map the wind farm global state feature to an efficiency space to generate an overall efficiency vector.

[0157] The overall efficiency vector includes an output efficiency index and an environmental adaptation index. The output efficiency index reflects the efficiency of the wind farm in converting wind energy into electrical energy, which can be measured by calculating the ratio of the actual power generation of the wind farm to the theoretical maximum power generation. The environmental adaptation index reflects the running ability of the wind farm under different environmental conditions, such as power generation efficiency and stability under different wind speed, wind direction and temperature conditions.

[0158] The overall efficiency vector is obtained by analyzing the wind farm global state feature through the efficiency mapping function. The overall efficiency vector evaluates the overall output efficiency of the wind farm.

[0159] Step S136: The node stability vector, the node coordination degree vector and the overall efficiency vector are integrated to construct a wind farm state evaluation system containing multi-dimensional evaluation results.

[0160] The node stability vector, the node coordination degree vector and the overall efficiency vector are integrated to construct a wind farm state evaluation system containing multi-dimensional evaluation results. The node stability vector evaluates the running stability of each node, the node coordination degree vector evaluates the collaborative running state between nodes, and the overall efficiency vector evaluates the overall output efficiency of the wind farm.

[0161] By integrating the three vectors, the wind farm state evaluation system can comprehensively evaluate the running state of the wind farm from multiple angles. For example, by analyzing the node stability vector, the nodes with poor running stability can be found out, and appropriate measures can be taken for optimization; by analyzing the node coordination degree vector, problems in the collaborative running between nodes can be found out, and collaborative control can be performed; by analyzing the overall efficiency vector, the overall output efficiency of the wind farm can be known, and reasonable resource scheduling and optimization strategies can be developed.

[0162] Step S140: generating a wind farm state optimization scheme according to the output result of the wind farm state evaluation system, the wind farm state optimization scheme including a node operation mode adjustment strategy, an inter-node collaborative control strategy, and a resource scheduling priority strategy.

[0163] The output result of the wind farm state evaluation system provides multi-dimensional evaluation information of the wind farm operation state, and according to these information, a wind farm state optimization scheme can be generated. The wind farm state optimization scheme aims to improve the operation efficiency, stability and collaboration of the wind farm, and includes a node operation mode adjustment strategy, an inter-node collaborative control strategy, and a resource scheduling priority strategy.

[0164] Step S141: analyzing the node stability vector output by the wind farm state evaluation system, and identifying nodes with fluctuation frequency indicators and fluctuation amplitude indicators exceeding the preset stability threshold as nodes to be stabilized.

[0165] The node stability vector output by the wind farm state evaluation system is analyzed, and nodes with fluctuation frequency indicators and fluctuation amplitude indicators exceeding the preset stability threshold are identified. The preset stability threshold is a pre-set standard for judging whether the operation stability of the node is qualified. If the fluctuation frequency indicator or the fluctuation amplitude indicator of a node exceeds the preset stability threshold, it means that the operation stability of the node is poor, and needs to be optimized and adjusted, and the node is identified as a node to be stabilized.

[0166] Step S142: analyzing the node coordination degree vector output by the wind farm state evaluation system, and identifying node groups with synchronization response indicators and function complement indicators lower than the preset coordination threshold as node groups to be coordinated.

[0167] The node coordination degree vector output by the wind farm state evaluation system is analyzed, and node groups with synchronization response indicators and function complement indicators lower than the preset coordination threshold are identified. The preset coordination threshold is a pre-set standard for judging whether the collaborative operation between node groups is good. If the synchronization response indicator or the function complement indicator of a node group is lower than the preset coordination threshold, it means that the collaborative operation ability between the node groups is poor, and needs to be coordinated and optimized, and the node group is identified as a node group to be coordinated.

[0168] Step S143: analyzing the overall efficiency vector output by the wind farm state evaluation system, and identifying efficiency bottleneck directions with power efficiency indicators and environmental adaptation indicators not reaching the preset efficiency benchmark.

[0169] The overall performance vector output by the wind farm state evaluation system is analyzed to identify the performance bottleneck direction of the output efficiency index and the environmental adaptation index that does not reach the preset performance benchmark. The preset performance benchmark is a pre-set standard for measuring whether the overall output performance of the wind farm meets the standard. If the output efficiency index or the environmental adaptation index does not reach the preset performance benchmark, it indicates that the wind farm has performance bottlenecks in these aspects, and the specific bottleneck direction needs to be found, such as the performance limitation of some equipment, the influence of environmental factors, etc.

[0170] Step S144: For the node to be stabilized, a node operation mode adjustment strategy is generated in combination with the operation mode characteristics in the node comprehensive state representation, and the node operation mode adjustment strategy includes a target operation mode type and a mode switching trigger condition.

[0171] Step S1441: Obtain the node comprehensive state representation of the node to be stabilized, and extract the operation mode characteristics from the node comprehensive state representation. The operation mode characteristics include the current operation mode identifier and the historical mode switching record.

[0172] For the node to be stabilized, its node comprehensive state representation is obtained, and the operation mode characteristics are extracted therefrom. The operation mode characteristics include the current operation mode identifier and the historical mode switching record. The current operation mode identifier indicates the current operation mode of the node to be stabilized, such as power generation mode, shutdown maintenance mode, etc. The historical mode switching record records the past mode switching of the node to be stabilized, including the time of switching, the reason for switching, etc. By analyzing the operation mode characteristics, the operation mode of the node to be stabilized can be known.

[0173] Step S1442: Query the preset operation mode database to obtain all selectable target operation mode types and the corresponding stability parameter range of each type.

[0174] The preset operation mode database is queried, which stores all selectable target operation mode types and the corresponding stability parameter range of each type. The stability parameter range refers to the reasonable range of the node's operation stability indicators (such as fluctuation frequency indicator, fluctuation amplitude indicator) in this operation mode. For example, for a wind turbine unit, different operation modes may correspond to different power output ranges and speed ranges, which will affect its operation stability. By querying the operation mode database, all selectable target operation mode types and their stability parameter ranges can be known.

[0175] Step S1443: Match the fluctuation frequency indicator and the fluctuation amplitude indicator in the node stability vector of the node to be stabilized with the stability parameter range of each target operation mode type, and select the candidate operation mode type whose fluctuation indicators meet the parameter range.

[0176] The fluctuation frequency index and the fluctuation amplitude index in the node stability vector of the node to be stabilized are matched with the stationary parameter range of each target operation mode type. If the stationary parameter range of a certain target operation mode type can meet the fluctuation frequency index and the fluctuation amplitude index requirements of the node to be stabilized, i.e. the fluctuation index falls within the parameter range, the target operation mode type is taken as a candidate operation mode type. The screened candidate operation mode type is a mode that can improve the operation stability of the node to be stabilized.

[0177] Step S1444: From the candidate operation mode types, the type with the highest adaptability is selected as the target operation mode type according to the mode adaptability scores in the historical mode switching records.

[0178] From the candidate operation mode types, the type with the highest adaptability is selected as the target operation mode type according to the mode adaptability scores in the historical mode switching records. The mode adaptability score is a score evaluated according to the actual operation of the node to be stabilized after switching to different operation modes in the past. The higher the score, the stronger the adaptability of the node to be stabilized to the operation mode. By selecting the target operation mode type with the highest adaptability, the success rate of mode switching can be improved, and the operation stability of the node to be stabilized can be better improved.

[0179] Step S1445: The real-time operation state data of the node to be stabilized is analyzed to determine the precondition parameters required for mode switching, including device load rate, environmental parameter threshold and associated node state.

[0180] The real-time operation state data of the node to be stabilized is analyzed to determine the precondition parameters required for mode switching. The precondition parameters refer to the conditions that must be met by the node to be stabilized before mode switching. These parameters include device load rate, environmental parameter threshold and associated node state. For example, for a fan unit node, the device load rate may need to be within a certain range before switching the operation mode, the environmental wind speed may need to meet a certain threshold requirement, and the operation state of the associated bus box group node also needs to be normal.

[0181] Step S1446: The mode switching trigger condition is set according to the precondition parameters, and the mode switching trigger condition is a logical combination of all precondition parameters being met at the same time.

[0182] The mode switching trigger condition is set according to the precondition parameters. The mode switching trigger condition is a logical combination of all precondition parameters being met at the same time. For example, if the precondition parameters include that the device load rate is between A and B, the environmental wind speed is greater than C, and the associated node state is normal, then the mode switching trigger condition is that the device load rate is between A and B, the environmental wind speed is greater than C, and the associated node state is normal. Only when all precondition parameters are met, the mode switching operation is triggered.

[0183] Step S1447: generating a mode switching transition scheme, the mode switching transition scheme containing a parameter gradient rate in the switching process and an intermediate state holding time length.

[0184] In order to ensure the smoothness of mode switching, a mode switching transition scheme is generated. The mode switching transition scheme contains a parameter gradient rate in the switching process and an intermediate state holding time length. The parameter gradient rate refers to the rate of change of the parameters (such as power output, speed, etc.) of the device during mode switching. A reasonable parameter gradient rate can avoid sharp fluctuations in the device during mode switching. The intermediate state holding time length refers to the time the device stays in the intermediate state during mode switching. A proper intermediate state holding time length can ensure stable transition of the device during switching.

[0185] Step S1448: integrating the target operating mode type, mode switching trigger condition and mode switching transition scheme to generate a node operating mode adjustment strategy.

[0186] The target operating mode type, mode switching trigger condition and mode switching transition scheme are integrated to generate a node operating mode adjustment strategy. The node operating mode adjustment strategy provides specific solutions and guidance for the adjustment of the operating mode of the to-be-stabilized node. By implementing the strategy, the operating stability of the to-be-stabilized node can be improved, and the overall operating efficiency of the wind farm can be improved.

[0187] Step S145: for the to-be-coordinated node group, combining the synchronization response index and the functional complement index in the node coordination degree vector, generating an inter-node collaborative control strategy, the inter-node collaborative control strategy containing an interaction timing adjustment rule and a functional cooperation parameter.

[0188] For the to-be-coordinated node group, combining the synchronization response index and the functional complement index in the node coordination degree vector, an inter-node collaborative control strategy is generated. The inter-node collaborative control strategy aims to improve the collaborative operation ability between the node groups, containing an interaction timing adjustment rule and a functional cooperation parameter.

[0189] The interaction timing adjustment rule is used to adjust the interaction time sequence between nodes. According to the synchronization response index, the synchronization response between nodes is analyzed, and the node pairs with synchronization problems are found out. By adjusting the interaction timing of these node pairs, they can respond at the same time, improving the synchronization between nodes. For example, for a wind turbine unit node and a collector box group node, if their synchronization response index is low, the power output time of the wind turbine unit and the data collection time of the collector box group can be adjusted to make them more synchronized.

[0190] The function matching parameter is determined according to the function complementary index. By analyzing the function characteristics of each node in the to-be-coordinated node group, the parts that can complement each other between the nodes are found out. For example, the output of one node can be used as the input of another node, or the functions of two nodes can work together to complete a specific task. According to these function complementary relationships, function matching parameters such as the ratio of input and output, the time allocation of collaborative work, etc. are determined. By adjusting the function matching parameters, the functions of the nodes can be more complementary, and the overall operation efficiency of the node group can be improved.

[0191] Step S146: For the performance bottleneck direction, a resource scheduling priority strategy is generated by combining the global state characteristics of the wind farm and the environmental perception data. The resource scheduling priority strategy includes the maintenance resource allocation sequence and the upgrade resource allocation proportion.

[0192] For the performance bottleneck direction identified by the wind farm state evaluation system, a resource scheduling priority strategy needs to be generated by combining the global state characteristics of the wind farm and the environmental perception data. The global state characteristics of the wind farm comprehensively consider the running state of each node and the influence of environmental factors, and the environmental perception data provides environmental information such as wind speed, wind direction, and temperature.

[0193] The maintenance resource allocation sequence is an important part of the resource scheduling priority strategy. According to the global state characteristics of the wind farm, the running state and potential fault risk of each device node are analyzed. For those nodes with poor running state and high fault risk, maintenance resources are allocated preferentially. For example, for wind turbine nodes that often fluctuate and have large fluctuation amplitudes, maintenance personnel should be arranged for inspection and repair in priority. At the same time, environmental perception data is considered, such as in harsh environmental conditions (such as strong wind, high temperature, etc.), some devices are more prone to failure, and the maintenance priority of these devices should be increased accordingly.

[0194] The upgrade resource allocation proportion is also a key content of the resource scheduling priority strategy. According to the performance bottleneck direction, the device nodes that need to be upgraded and the key direction of upgrading are determined. For example, if it is found that the overall output efficiency of the wind farm is low due to the performance limitation of some wind turbine units, more upgrade resources can be allocated to these wind turbine units. At the same time, combined with the environmental perception data, for those device nodes that are greatly affected by environmental factors and can better adapt to the environment after upgrading, the allocation proportion of upgrade resources is increased. For example, in areas with large wind speed changes, the control system of the wind turbine unit is upgraded to enable it to better track wind speed changes and improve power generation efficiency.

[0195] By reasonably determining the maintenance resource allocation sequence and the upgrade resource allocation proportion, the resource scheduling priority strategy is generated. This strategy can optimize the resource allocation of the wind farm, improve the overall performance of the wind farm, and solve the performance bottleneck problem.

[0196] Step S147: The integration node runs the mode adjustment strategy, the inter-node collaborative control strategy, and the resource scheduling priority strategy to form a wind farm state optimization scheme covering multi-dimensional optimization requirements.

[0197] The node operation mode adjustment strategy, the inter-node collaborative control strategy, and the resource scheduling priority strategy are integrated to form a wind farm state optimization scheme covering multi-dimensional optimization requirements. These three strategies respectively optimize the operation state of the wind farm from three different angles of node operation stability, inter-node collaborative operation, and overall performance improvement.

[0198] The node operation mode adjustment strategy adjusts the operation mode of the to-be-stabilized node with poor operation stability to improve its stability. The inter-node collaborative control strategy adjusts the interaction timing and functional coordination parameters of the to-be-coordinated node group with poor collaborative operation capability to improve the inter-node collaboration. The resource scheduling priority strategy targets the performance bottleneck direction of the wind farm and allocates maintenance resources and upgrade resources to improve the overall performance of the wind farm.

[0199] Integrating these three strategies together can comprehensively solve the problems existing in the operation of the wind farm. For example, for a wind farm with low overall performance due to unstable operation of wind turbine units and poor inter-node collaboration, the wind farm state optimization scheme can simultaneously implement the node operation mode adjustment strategy to stabilize the operation of wind turbine units, implement the inter-node collaborative control strategy to improve the inter-node collaborative operation capability, and implement the resource scheduling priority strategy to reasonably allocate resources and improve the overall performance.

[0200] The wind farm state optimization scheme provides a comprehensive and systematic optimization framework for the operation management of the wind farm. By implementing the wind farm state optimization scheme, the operation efficiency, stability, and collaboration of the wind farm can be improved, and the sustainable development of the wind farm can be realized.

[0201] Step S150: The wind farm state optimization scheme is converted into standardized control instructions and sent to the wind farm centralized control system, and the centralized control system drives the corresponding nodes to perform state adjustment actions.

[0202] After generating the wind farm state optimization scheme, it needs to be converted into standardized control instructions. The standardized control instruction is the instruction format that the wind farm centralized control system can recognize and execute. The node operation mode adjustment strategy, the inter-node collaborative control strategy, and the resource scheduling priority strategy in the wind farm state optimization scheme all need to be converted into specific control instructions.

[0203] For the node operation mode adjustment strategy, the target operation mode type, mode switching trigger condition and mode switching transition scheme are converted into corresponding control instructions. For example, for a fan unit node, the target operation mode type is converted into the target parameter setting in the control instruction, the mode switching trigger condition is converted into the trigger condition judgment logic in the control instruction, and the mode switching transition scheme is converted into the parameter adjustment step in the control instruction.

[0204] For the inter-node cooperative control strategy, the interaction timing adjustment rule and functional cooperation parameter are converted into control instructions. For example, for the cooperative control of a fan unit node and a bus box group node, the interaction timing adjustment rule is converted into the time synchronization control logic in the control instruction, and the functional cooperation parameter is converted into the input / output ratio setting in the control instruction.

[0205] For the resource scheduling priority strategy, the maintenance resource allocation order and the upgrade resource deployment ratio are converted into control instructions. For example, the maintenance resource allocation order is converted into the maintenance task scheduling order in the control instruction, and the upgrade resource deployment ratio is converted into the upgrade task allocation ratio in the control instruction.

[0206] The converted standardized control instructions are sent to the wind farm control system. The wind farm control system is the core control center of the wind farm, responsible for receiving and executing control instructions. After receiving the standardized control instructions, the control system drives the corresponding nodes to perform state adjustment actions according to the content of the instructions. For example, for the control instruction of the node operation mode adjustment strategy, the control system will monitor the real-time operation state of the stable node, and when the mode switching trigger condition is met, the operation mode of the node will be adjusted according to the mode switching transition scheme. For the control instruction of the inter-node cooperative control strategy, the control system will coordinate the operation of each node in the node group to be coordinated, and control according to the interaction timing adjustment rule and the functional cooperation parameter. For the control instruction of the resource scheduling priority strategy, the control system will arrange maintenance personnel and upgrade equipment, and allocate resources according to the maintenance resource allocation order and the upgrade resource deployment ratio.

[0207] By converting the wind farm state optimization scheme into standardized control instructions and sending them to the wind farm control system, the implementation of the wind farm state optimization scheme is realized, and the wind farm can operate according to the optimization scheme to improve the overall performance of the wind farm.

[0208] Figure 2 A schematic diagram of exemplary hardware and software components of a wind farm state optimization decision system 100 based on a graph neural network that can implement the idea of the present application is shown. For example, the processor 120 can be used in the wind farm state optimization decision system 100 based on a graph neural network, and is used to perform the functions in the present application.

[0209] The wind farm state optimization decision system 100 based on the graph neural network can be a general server or a special-purpose server, both of which can be used to implement the wind farm state optimization decision method based on the graph neural network of the present application. The present application only shows one server, but for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0210] For example, the wind farm state optimization decision system 100 based on the graph neural network can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Illustratively, the wind farm state optimization decision system 100 based on the graph neural network can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The wind farm state optimization decision system 100 based on the graph neural network also includes an I / O interface 150 between the computer and other input / output devices.

[0211] For the sake of illustration, only one processor is described in the wind farm state optimization decision system 100 based on the graph neural network. However, it should be noted that the wind farm state optimization decision system 100 based on the graph neural network in the present application can also include multiple processors, so the steps performed by one processor described in the present application can also be jointly performed or individually performed by multiple processors. For example, if the processor of the wind farm state optimization decision system 100 based on the graph neural network performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or individually performed in one processor. For example, a first processor performs step A, a second processor performs step B, or a first processor and a second processor jointly perform steps A and B.

[0212] In addition, the present application also provides a readable storage medium, in which computer executable instructions are pre-set, and when a processor executes the computer executable instructions, the wind farm state optimization decision method based on the graph neural network is realized.

[0213] It should be noted that, in order to simplify the description of the present application disclosed and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A wind farm state optimization decision-making method based on graph neural networks, characterized in that, The method includes: Construct a wind farm status association diagram. Nodes correspond to wind turbine units, combiner boxes, collector line sections and step-up transformer units within the wind farm. Edges represent the influence relationship between the operating status of nodes. The association strength of the edges is dynamically adjusted according to the frequency of real-time status interaction between nodes. The pre-configured graph neural network performs node feature propagation on the wind farm state association graph. The graph neural network includes a graph sampling layer, a graph aggregation layer, and a feature mapping layer. The graph sampling layer filters the core neighborhood node set. The graph aggregation layer associates and aggregates the state features of the core neighborhood nodes to generate neighborhood association features. The feature mapping layer fuses the node's own state features and neighborhood association features to generate a comprehensive state representation of the node. A wind farm status assessment system is constructed based on the comprehensive state representation of nodes. The wind farm status assessment system includes a node operation stability assessment module, a node inter-node collaborative operation assessment module, and an overall power output efficiency assessment module, which respectively output node stability vector, node collaboration vector, and overall efficiency vector. Generate wind farm condition optimization schemes based on the output results of the wind farm condition assessment system; The wind farm status optimization scheme is converted into standardized control commands and sent to the wind farm centralized control system, which then drives the corresponding nodes to perform status adjustment actions.

2. The wind farm state optimization decision-making method based on graph neural networks according to claim 1, characterized in that, The construction of the wind farm state association diagram includes: Identify the specific equipment entities of wind turbine units, combiner box groups, collector line sections and step-up transformer units within the wind farm, set each equipment entity as an independent node in the wind farm status association diagram, and assign a unique equipment identification code to each node; Collect real-time operating status data from each node. The real-time operating status data includes the current operating mode information of the equipment, status monitoring data, and historical status change records. Analyze the upstream and downstream relationships of different nodes in the power transmission path, identify the direct state influence relationship between nodes based on the upstream and downstream relationship, establish the direct influence relationship edge, and connect two nodes with a direct power transmission path; Analyze the master-slave relationship of different nodes in the control command interaction, identify the indirect state influence relationship between nodes based on the master-slave relationship, establish indirect influence relationship edges, and connect two nodes that have control command interaction; Based on real-time state interaction records between nodes, the number of state interactions and the duration of interactions directly affecting the relationship edges within a preset time window are counted, and the association strength of the directly affecting relationship edges is dynamically adjusted according to the number of state interactions and the duration of interactions. Based on the control command response records between nodes, the timeliness and accuracy of the command response corresponding to the indirect influence relationship edge within a preset time window are statistically analyzed, and the association strength of the indirect influence relationship edge is dynamically adjusted according to the timeliness and accuracy of the command response. By integrating node device identification codes, real-time operating status data, direct and indirect influence relationship edges, and their corresponding association strengths, an initial wind farm status association diagram is formed. The system re-collects real-time operating status data and interaction records of nodes according to the preset update cycle, and repeatedly adjusts the correlation strength of each side to achieve dynamic updating and maintenance of the wind farm status correlation diagram.

3. The wind farm state optimization decision-making method based on graph neural networks according to claim 1, characterized in that, The process involves propagating node features on a wind farm state association graph using a pre-configured graph neural network. The graph neural network includes a graph sampling layer, a graph aggregation layer, and a feature mapping layer. The graph sampling layer filters a core neighborhood node set. The graph aggregation layer aggregates the state features of the core neighborhood nodes to generate neighborhood association features. The feature mapping layer fuses the node's own state features with the neighborhood association features to generate a comprehensive node state representation, including: The real-time operating status data of each node in the wind farm state association graph is converted into the node's own state feature vector. The feature dimension of the node's own state feature vector is adapted to the receiving dimension of the graph neural network input layer. The node's own state feature vector and the association matrix of the wind farm state association graph are input into the graph sampling layer. The association matrix contains information on the edge type and association strength between each node. In the graph sampling layer, for each node, all its neighboring nodes are sorted from high to low according to the association strength in the association matrix, and the neighboring nodes that are ranked first and whose association strength meets the preset threshold are selected to form the core neighborhood node set. The state feature vectors of each neighboring node in the core neighborhood node set are input into the graph aggregation layer. The state features of the neighboring nodes are integrated through the association aggregation algorithm of the graph aggregation layer to generate neighborhood association features. The association aggregation algorithm assigns feature integration weights based on the association strength between the neighboring node and the current node. The neighborhood association features are input into the graph attention layer. The attention mechanism is used to analyze the degree of correlation between the neighborhood association features and the current node's own state feature vector. Attention weights are assigned to different feature components in the neighborhood association features to generate weighted neighborhood association features. The node's own state feature vector and the weighted neighborhood association features are merged into a joint feature vector through feature concatenation. The joint feature vector is then subjected to nonlinear feature transformation to generate a comprehensive state representation of the node.

4. The wind farm state optimization decision-making method based on graph neural networks according to claim 1, characterized in that, The wind farm state assessment system, constructed based on the comprehensive state representation of nodes, includes a node operation stability assessment module, an inter-node collaborative operation assessment module, and an overall power output efficiency assessment module. These modules respectively output node stability vectors, node collaboration vectors, and overall efficiency vectors, including: The feature sequence representing the stable operation of the equipment is extracted from the node comprehensive status characterization. The feature sequence is input into the node operation stability assessment module. The node operation stability assessment module captures the fluctuation trend of the feature sequence through the sliding window feature extraction method and generates a node stability vector. The node stability vector includes fluctuation frequency index and fluctuation amplitude index. Feature pairs representing the interaction characteristics between nodes are extracted from the node comprehensive state representation. The feature pairs are input into the node collaborative operation evaluation module. The node collaborative operation evaluation module analyzes the synchronization and complementarity of the feature pairs through the feature alignment algorithm and generates a node collaboration degree vector. The node collaboration degree vector includes synchronization response index and functional complementarity index. Collect current environmental sensing data of the wind farm, which includes wind speed change trends, wind direction distribution characteristics, and temperature field distribution information. The node integrated state representation of all nodes is fused with environmental perception data across modal features to generate global state features of the wind farm. The global state features of the wind farm are then input into the overall power output efficiency evaluation module. The overall power output efficiency assessment module analyzes the global state characteristics of the wind farm through the efficiency mapping function and generates an overall efficiency vector, which includes power output efficiency indicators and environmental adaptability indicators. By integrating node stability vector, node coordination vector, and overall performance vector, a wind farm status assessment system containing multi-dimensional evaluation results is constructed.

5. The wind farm state optimization decision-making method based on graph neural networks according to claim 1, characterized in that, The step of generating a wind farm condition optimization scheme based on the output results of the wind farm condition assessment system includes: The node stability vector output by the wind farm state assessment system is analyzed, and nodes whose fluctuation frequency and fluctuation amplitude exceed the preset stability threshold are identified as nodes to be stabilized. The node coordination vector output by the wind farm status assessment system is analyzed, and node groups whose synchronization response index and functional complementarity index are lower than the preset coordination threshold are identified as node groups to be coordinated. Analyze the overall efficiency vector output by the wind farm condition assessment system to identify the bottleneck directions in which the power output efficiency index and environmental adaptability index fail to meet the preset efficiency benchmark. For nodes that need to be stabilized, a node operation mode adjustment strategy is generated based on the operation mode characteristics in the node's comprehensive state representation. The node operation mode adjustment strategy includes the target operation mode type and the mode switching trigger condition. For the group of nodes to be coordinated, a collaborative control strategy between nodes is generated by combining the synchronization response index and functional complementarity index in the node collaboration degree vector. The collaborative control strategy between nodes includes interaction timing adjustment rules and functional coordination parameters. To address performance bottlenecks, a resource scheduling priority strategy is generated by combining the global status characteristics of the wind farm and environmental perception data. This strategy includes maintaining the order of resource allocation and upgrading the proportion of resource deployment. By integrating node operation mode adjustment strategies, inter-node collaborative control strategies, and resource scheduling priority strategies, a wind farm state optimization scheme covering multi-dimensional optimization needs is formed.

6. The wind farm state optimization decision-making method based on graph neural networks according to claim 2, characterized in that, The method involves recording real-time state interactions between nodes, counting the number of state interactions and their duration within a preset time window that directly affect the relationship edges, and dynamically adjusting the association strength of the directly affecting relationship edges based on the number of state interactions and their duration. Obtain the real-time state interaction record of two nodes that directly affect the connection of the relationship edge. The real-time state interaction record includes the start timestamp and end timestamp of each interaction. Define a preset time window and count the total number of interaction events in the real-time status interaction record within the preset time window as the number of status interactions. Based on the start and end timestamps of each interaction, calculate the duration of each interaction, and sum the durations of all interactions to obtain the total duration of the interaction. The number of state interactions is used to query a preset number-intensity mapping table to obtain the number contribution intensity component. The duration of the interaction is used to query the preset duration-intensity mapping table to obtain the duration contribution intensity component; The frequency contribution intensity component and the duration contribution intensity component are fused to generate the current association strength that directly affects the relationship edge; The current association strength is smoothly transitioned to the historical association strength of the directly affected relationship edges, and the association strength parameters of the directly affected relationship edges are updated based on the current association strength after the smooth transition.

7. The wind farm state optimization decision-making method based on graph neural networks according to claim 3, characterized in that, In the graph sampling layer, for each node, all its neighboring nodes are sorted from high to low according to the association strength in the association matrix. The neighboring nodes that rank highest and whose association strength meets a preset threshold are selected to form a core neighborhood node set, including: For the current processing node, extract the identification information and corresponding association strength values ​​of all neighboring nodes related to this node from the association matrix; The adjacent nodes are sorted from high to low according to their association strength value, forming an adjacent node sorting sequence; Traverse the sorted sequence of adjacent nodes and determine in turn whether the association strength value of each adjacent node is greater than or equal to the preset association strength threshold. Add adjacent nodes with association strength values ​​greater than or equal to the preset association strength threshold to a temporary neighborhood set; If the number of nodes in the temporary neighborhood set does not reach the preset neighborhood size limit, the remaining adjacent nodes with the highest association strength value in the sorting sequence will continue to be added to the temporary neighborhood set until the preset neighborhood size limit is reached. If the number of nodes in the temporary neighborhood set exceeds the preset neighborhood size limit, then the number of adjacent nodes above the preset neighborhood size limit will be truncated according to the association strength value from high to low to form the temporary neighborhood set. The adjacent nodes in the temporary neighborhood set are deduplicated to ensure that each adjacent node is unique in the set; The deduplicated temporary neighborhood set is determined as the core neighborhood node set of the current processing node.

8. The wind farm state optimization decision-making method based on graph neural networks according to claim 4, characterized in that, The feature sequence characterizing the stable operation of the equipment is extracted from the node's comprehensive state characterization. This feature sequence is then input into the node operational stability assessment module. The module uses a sliding window feature extraction method to capture the fluctuation trend of the feature sequence and generates a node stability vector, including: Feature dimensions related to equipment operational stability are selected from the comprehensive state characterization of nodes. The feature values ​​corresponding to the selected feature dimensions are arranged in chronological order to form a feature sequence characterizing equipment operational stability. Set the window size and sliding step size of the sliding window. The window size is determined according to the change cycle of the device's operating status, and the sliding step size is a preset ratio of the window size. The sliding window is slid sequentially on the feature sequence according to the sliding step size. Statistical analysis is performed on the feature values ​​in each window to calculate the mean and standard deviation of the feature values ​​in the window. The volatility coefficient is calculated based on the mean and standard deviation of the feature values ​​within the window. The volatility coefficient reflects the degree of dispersion of the feature values ​​within the window. Arrange the volatility coefficients of all windows in chronological order to form a volatility coefficient sequence, and capture the overall volatility trend of the feature sequence through the volatility coefficient sequence; The volatility coefficient sequence is fitted with a trend to obtain a volatility trend curve. The volatility frequency index is extracted from the volatility trend curve. The volatility frequency index is the number of times the volatility coefficient exceeds the preset volatility threshold per unit time. The volatility amplitude index is extracted from the volatility trend curve. The volatility amplitude index is the difference between the maximum and minimum values ​​in the volatility coefficient sequence. Integrate the volatility frequency indicator and volatility amplitude indicator to generate a node stability vector that includes the indicator name and the corresponding numerical range.

9. The wind farm state optimization decision-making method based on graph neural networks according to claim 5, characterized in that, For nodes to be stabilized, a node operation mode adjustment strategy is generated based on the operation mode characteristics in their comprehensive node state representation. This strategy includes the target operation mode type and mode switching trigger conditions, including: Obtain the comprehensive state representation of the node to be stabilized, and extract the operation mode features from the comprehensive state representation. The operation mode features include the current operation mode identifier and historical mode switching records. Query the preset operating mode database to obtain all available target operating mode types and the corresponding stability parameter range for each type; The fluctuation frequency and fluctuation amplitude indices in the node stability vector of the node to be stabilized are matched with the stability parameter range of each target operating mode type, and candidate operating mode types whose fluctuation indices meet the parameter range are selected. From the candidate operating mode types, the type with the highest fitness score in the historical mode switching record is selected as the target operating mode type. Analyze the real-time operating status data of the nodes to be stabilized to determine the prerequisite parameters required for mode switching. The prerequisite parameters include equipment load rate, environmental parameter thresholds and associated node status. The mode switching trigger condition is set according to the precondition parameters. The mode switching trigger condition is a logical combination that satisfies all the precondition parameters at the same time. A mode switching transition scheme is generated, which includes the parameter change rate and the duration of intermediate state maintenance during the switching process; Integrate the target operating mode type, mode switching trigger conditions, and mode switching transition scheme to generate a node operating mode adjustment strategy.

10. A wind farm state optimization decision system based on graph neural networks, characterized in that, The system includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the wind farm state optimization decision-making method based on graph neural networks as described in any one of claims 1-9.

Images