Knowledge enhancement-based power grid new energy operation fault discrimination method and system

Abstract

The invention discloses a power grid new energy operation fault judgment method and system based on knowledge enhancement, and relates to the technical field of wind power plant fault diagnosis. The method comprises the following steps: acquiring time sequence data of operation of the new energy equipment of the power grid, and constructing a graph structure according to the time sequence data of operation of the new energy equipment of the power grid to obtain graph data representation; encoding expert knowledge into a hierarchical rule set according to the characteristics of the operation data of the new energy equipment of the power grid, and constructing a knowledge rule base according to the hierarchical rule set; constructing a fault diagnosis model by utilizing a graph neural network with a knowledge enhancement mechanism, and training the fault diagnosis model by taking the hierarchical rule set in the knowledge rule base as a teacher and the graph neural network as a student; and performing anomaly detection on the graph data representation by using the trained fault diagnosis model to obtain a wind power plant fault judgment result. According to the invention, unsupervised accurate identification of complex fault scenes such as gearbox faults and variable pitch system anomalies can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind farm fault diagnosis, and in particular to a power grid new energy operation fault discrimination method and system based on knowledge enhancement. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] With the popularization of intelligent monitoring technology of wind farms, SCADA is widely used in data acquisition of wind turbine operation state. Based on the analysis of multi-parameter time series data, early fault warning and diagnosis of wind turbine can be realized. However, the traditional fault diagnosis method (such as the method based on physical model) highly depends on expert experience, and has problems such as poor adaptability and difficult modeling. In order to introduce intelligent means, the early artificial intelligence method represented by expert system is applied. This method makes logical reasoning according to the rules and prior knowledge summarized by artificial intelligence, which partially reduces the dependence on experience, but still needs to build a complete knowledge base, and it is difficult to deal with unknown faults.

[0004] With the progress of technology, the more pure data-driven method represented by deep learning has gradually become the research mainstream with its strong self-learning and non-linear fitting ability. Common technologies include autoencoder, long short-term memory network and graph neural network, etc. Among them, the unsupervised method such as autoencoder detects anomalies through reconstruction error, but it is sensitive to dynamic working condition changes; the long short-term memory network can capture time dependence, but often ignores the spatial correlation between variables; although the graph neural network can model the topological relationship between variables, most methods assume that the graph structure is static and fixed, which is difficult to adapt to the high dynamic characteristics of wind power data, and has limitations such as complex calculation and weak noise resistance.

[0005] Under the background of continuous expansion of wind farm scale and continuous growth of operation data, the accuracy of fault diagnosis depends seriously on the feature extraction and correlation mining ability of algorithm on high-dimensional, nonlinear and strongly coupled SCADA data. However, the existing methods cannot effectively fuse multi-source information, and cannot well adapt to dynamic changes and consider efficiency and robustness. In addition, most existing artificial intelligence diagnosis schemes often use single model or isolated learning mechanism, lack of multi-level knowledge fusion and adaptive optimization ability, which leads to limited generalization performance in complex fault scenarios, and there is still room for improvement in diagnosis accuracy. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application aims to provide a power grid new energy operation fault discrimination method and system based on knowledge enhancement, which can realize unsupervised accurate identification of complex fault scenarios such as gearbox failure and abnormality of variable pitch system.

[0007] In order to achieve the above object, the present application is realized by the following technical scheme: The first aspect of the present application provides a knowledge-enhanced power grid new energy operation fault discrimination method, comprising the following steps: Obtain the time series data of the operation of the power grid new energy equipment, and construct a graph structure according to the time series data of the operation of the power grid new energy equipment to obtain a graph data representation; According to the characteristics of the power grid new energy equipment operation data, the expert knowledge is encoded into a hierarchical rule set, and a knowledge rule base is constructed according to the hierarchical rule set; A fault diagnosis model is constructed by using a graph neural network with a knowledge enhancement mechanism, the hierarchical rule set in the knowledge rule base is taken as a teacher, and the graph neural network is taken as a student, and the fault diagnosis model is trained; The trained fault diagnosis model is used for abnormal detection of the graph data representation, and a wind farm fault discrimination result is obtained.

[0008] Further, the specific steps of constructing a graph structure according to the time series data of the operation of the power grid new energy equipment are: Design node features according to the physical characteristics of the operation of the power grid new energy equipment; A directed graph is constructed based on the node features using a sliding window method to obtain an adjacency matrix, specifically, the time series data of the operation of the power grid new energy equipment is abstracted into a graph structure, the nodes represent the state of the equipment at the key moment, and the edges represent the causal dependence on the time series; The node features are fused based on an attention mechanism to obtain a dimension-reduced feature matrix.

[0009] Further, the specific steps of encoding expert knowledge into a hierarchical rule set according to the characteristics of the power grid new energy equipment operation data are: Design a knowledge rule hierarchy according to different complexity of fault scenarios; Based on the knowledge rule hierarchy, the expert knowledge related to wind farm anomaly detection is converted into a continuous vector representation according to the characteristics of the power grid new energy equipment operation data, the knowledge rule hierarchy includes a unique alarm layer, a fixed combination layer and a complex combination layer, wherein the unique alarm layer is suitable for simple fault scenarios, the fixed combination layer is suitable for medium complexity faults, and the complex combination layer is suitable for complex faults.

[0010] Further, the specific steps of using the trained fault diagnosis model for abnormal detection of the graph data representation are: The trained fault diagnosis model is used to process the graph data representation to generate a node low-dimensional embedding representation; A restricted Boltzmann machine is used to calculate the anomaly score of the node low-dimensional embedding representation through an energy function; The wind farm fault discrimination result is obtained according to the anomaly score.

[0011] Further, the specific steps of obtaining the wind farm fault discrimination result according to the abnormal score are as follows: Setting an adaptive dynamic abnormality judgment threshold according to historical detection results; Comparing the abnormal score with the dynamic threshold to obtain the wind farm fault discrimination result.

[0012] The second aspect of the application provides a power grid new energy operation fault discrimination system based on knowledge enhancement, comprising: A data acquisition module configured to acquire time series data of power grid new energy equipment operation, and construct a graph structure according to the time series data of power grid new energy equipment operation to obtain graph data representation; A rule base establishment module configured to encode expert knowledge into a hierarchical rule set according to the characteristics of power grid new energy equipment operation data, and construct a knowledge rule base according to the hierarchical rule set; A model construction module configured to construct a fault diagnosis model using a graph neural network with a knowledge enhancement mechanism, use the hierarchical rule set in the knowledge rule base as a teacher and the graph neural network as a student to train the fault diagnosis model; An anomaly detection module configured to use the trained fault diagnosis model to perform anomaly detection on the graph data representation to obtain a wind farm fault discrimination result.

[0013] Further, the data acquisition module is further configured to: Design node features according to physical characteristics of power grid new energy equipment operation; Construct a directed graph based on node features using a sliding window method to obtain an adjacency matrix, specifically, abstract the time series data of power grid new energy equipment operation into a graph structure, with nodes representing the state of the equipment at key moments and edges representing causal dependence on time series; Fuse node features based on an attention mechanism to obtain a reduced feature matrix.

[0014] Further, the rule base establishment module is further configured to: Design knowledge rule levels according to different complexity of fault scenarios; Based on the knowledge rule levels, convert expert knowledge related to wind farm anomaly detection into continuous vector representation according to the characteristics of power grid new energy equipment operation data, the knowledge rule levels include a unique alarm layer, a fixed combination layer and a complex combination layer, wherein the unique alarm layer is suitable for simple fault scenarios, the fixed combination layer is suitable for medium complexity faults, and the complex combination layer is suitable for complex faults.

[0015] Further, the model construction module is further configured to: Use the trained fault diagnosis model to process the graph data representation to generate a node low-dimensional embedding representation; Anomaly scores are calculated by energy function of a restricted Boltzmann machine to represent low-dimensional embedding of nodes; A wind farm fault discrimination result is obtained according to the anomaly scores.

[0016] Further, the anomaly detection module is further configured to: An adaptive dynamic anomaly determination threshold is set according to historical detection results; The anomaly scores are compared with the dynamic threshold to obtain a wind farm fault discrimination result.

[0017] The above one or more technical solutions have the following beneficial effects: The application discloses a power grid new energy operation fault discrimination method and system based on knowledge enhancement, encodes expert knowledge of a wind farm into a propositional logic formula, generates a hierarchical general rule set, including a unique alarm layer, a fixed combination layer and a complex combination layer, to cover fault scenes with different complexities. SCADA (Supervisory Control and Data Acquisition) data of the wind farm are abstracted into a graph structure, nodes are designed based on physical statistical characteristics, and edges are constructed into a directed graph through a sliding window to capture time sequence dependence. Then, the knowledge rule is used as a "teacher", and a GNN (Graph Neural Network) is used as a "student", knowledge is embedded into GNN training by maximizing mutual information, and a regularization term is added to constrain the distance between the output and the knowledge embedding. Finally, an RBM (Restricted Boltzmann Machine) is used as an anomaly detector, a node state score is calculated based on an energy function, and a dynamic threshold adjustment mechanism is combined to realize unsupervised fault discrimination. The application solves the problems of high-dimensional redundancy and label loss of SCADA data, improves the recognition accuracy of complex faults, and improves the reasoning efficiency through model lightening technology, thereby providing reliable support for intelligent operation and maintenance of the wind farm.

[0018] The application solves the problems of high-dimensional redundancy and label loss of SCADA data by fusing wind power physical statistical characteristics and symbolic knowledge. The application divides fault scenes into three levels and constructs a time sequence dependence graph, effectively capturing nonlinear correlation and topological information in the data. The application also realizes efficient adaptive fault detection through knowledge embedding constraints and a dynamic threshold adjustment mechanism.

[0019] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be learned through the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 Flow chart of the power grid new energy operation fault discrimination method based on knowledge enhancement in the embodiment one of the present application. Figure 2 Flow chart of the graph neural network training with knowledge enhancement mechanism in the embodiment one of the present application. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0023] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form, and in addition, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof; The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] Embodiment one: The embodiment one of the present application provides a knowledge enhancement-based power grid new energy operation fault discrimination method. The performance of GNN in wind farm fault detection is improved through a knowledge enhancement mechanism, and unsupervised and high-precision fault identification is realized. According to the characteristics of wind power SCADA data, the expert knowledge is coded into a hierarchical rule set, including a unique alarm layer, a fixed combination layer and a complex combination layer, to cover different complexity fault scenarios. Graph structure modeling is also performed. The SCADA time series data is abstracted into a graph structure. The node features are designed based on physical statistics such as power curve deviation and temperature difference. The edges are constructed into a directed graph through a sliding window to capture the time series dependence, and a graph attention network (GAT) is used to realize feature fusion. Then, a teacher-student training scheme is used. The knowledge rule is used as the "teacher", and the GNN is used as the "student". By adding a knowledge consistency regularization term to the loss function, the distance between the GNN output and the knowledge embedding is constrained. Finally, unsupervised anomaly detection is realized based on RBM. The node state score is calculated through the energy function, and the dynamic threshold is adjusted to optimize the system performance. The embodiment combines physical statistical features and symbolic knowledge, solves the problems of high-dimensional redundancy and label missing of SCADA data, improves the recognition accuracy of complex faults (such as gearbox failure and variable pitch system anomaly), and improves the inference efficiency through model lightweight technology.

[0025] As shown in Figure 1 The embodiment takes a wind farm device in a power grid new energy device as an example for illustration, and specifically includes the following steps: Step 1: Obtain the time series data of the operation of the power grid new energy device, and construct a graph structure according to the time series data of the operation of the power grid new energy device to obtain a graph data representation.

[0026] Step 1.1: Obtain the time series data of the operation of the power grid new energy device.

[0027] In a specific embodiment, the SCADA system continuously collects various operating parameters through sensors installed on the wind turbine to form high-dimensional time series data. The SCADA data specifically includes more than 30 numerical parameters such as temperature, power, and wind speed. The data type is mainly numerical time series.

[0028] Step 1.2: Construct a graph structure according to the time series data of the operation of the power grid new energy device.

[0029] In a specific embodiment, in order to capture the time series dependence and nonlinear association in SCADA data, the embodiment abstracts the time series data of the operation of the power grid new energy device into a graph structure. The node represents the state of the device at the key moment, and the edge represents the causal dependence on the time series.

[0030] Step 1.2.1: Design node features according to the physical characteristics of the new energy equipment operation of the power grid.

[0031] Specifically, node feature design covers three types of physical feature mapping functions that extract key parameters from different dimensions. The physical feature mapping functions include feature curve mapping function, normal behavior model mapping function, and relative temperature difference mapping function.

[0032] Feature curve mapping function : Based on the power-speed characteristic curve of wind turbines, define the power reduction parameter ξ to quantify the power generation efficiency at a unit wind speed. The calculation formula is: (1).

[0033] wherein represents the actual wind speed at the current time, which is the input variable directly monitored by the SCADA system, represents the rated wind speed, i.e., the design wind speed when the wind turbine reaches the rated power. is the cut-in wind speed, which represents the minimum wind speed at which the wind turbine begins to generate power. represents the actual power generation of the wind turbine at the current time. is the rated power, which represents the maximum output power of the wind turbine under design conditions.

[0034] Normal behavior model mapping function : Based on the normal behavior model of key parameters (such as temperature, vibration), calculate the absolute deviation between the predicted value and the actual value. The formula is: (2).

[0035] wherein is the absolute deviation between the predicted value and the actual value. is the main monitoring variable, is the model predicted value, is the true value.

[0036] Relative temperature difference mapping function : Define the temperature difference feature ΔT to capture heat-related failures. The formula is: (3).

[0037] wherein and represent any two monitoring parameters in the SCADA data, usually temperature-related variables (such as gearbox oil temperature, generator bearing temperature).

[0038] All features are normalized to form a node feature matrix X ∈ R {N×F} , where N is the number of time points, F is the feature dimension, and R is the real number set.

[0039] Step 1.2.2: Construct a directed graph based on node features using a sliding window method to obtain an adjacency matrix.

[0040] Specifically, first, the SCADA time series data is abstracted into graph nodes, and each node corresponds to a device state feature (such as power, temperature, etc.) at a time point t. To capture the time series dependence, this embodiment uses a sliding window method to construct a directed graph. Set the window size k, if the time difference between nodes i and j satisfies |i-j| ≤ k and i < j, that is, the historical node affects the current node, then in the adjacency matrix A, Ai,j=1, otherwise 0. This structure ensures that the model only depends on past information, which conforms to the physical causality.

[0041] The adjacency matrix A is an N × N binary matrix (where N is the number of nodes, i.e., the number of time series data points), which is used to represent the directed connection between nodes. The neighbor set of each node is defined by the adjacency matrix A, which provides the basis for the operation of GAT. The graph data representation in this embodiment includes the node feature matrix X and the adjacency matrix A.

[0042] Step 1.2.3: Fuse node features based on attention mechanism to obtain a reduced dimension feature matrix.

[0043] Specifically, to overcome the limitations of fixed weights in traditional GCN, this embodiment uses GAT as an encoder to dynamically aggregate neighbor information in the adjacency matrix through learnable weights, more effectively capturing nonlinear relationships in SCADA data. For node i, its feature vector updated after the GAT layer is The calculation formula is: (4).

[0044] wherein, represents the state feature of the wind turbine at node j, which is extracted from the original data through the mapping function formulas 1-3, and is used to represent the operating condition of the device. denotes the neighbor set of node i, . W is a trainable linear transformation weight matrix; used for dimension mapping and transformation of the features of all nodes; is the calculated attention weight, representing the importance of node j to node i; is a nonlinear activation function.

[0045] (5).

[0046] wherein, represents the state feature of the wind turbine at node i, represents the state feature of wind turbine at node k. Parameter k is a temporary index variable used to represent a specific neighbor node in the neighbor set N(i) of node i in mathematical calculation. is the parameter vector of the trainable single-layer feedforward neural network, represents vector concatenation; LeakyReLU is a linear rectifier function with leakage used to introduce nonlinearity. This mechanism enables the model to focus on key neighbor nodes and improves the ability to capture nonlinear correlations. The fused output is a reduced dimension feature matrix H∈R {N×F′} , which provides high representative statistical features for anomaly detection. F' represents the hidden layer dimension of the GAT encoder output, which is usually smaller than the original feature dimension F.

[0047] Step 2: Encode expert knowledge into a hierarchical rule set according to the characteristics of new energy equipment operation data of the power grid, and construct a knowledge rule base according to the hierarchical rule set.

[0048] In a specific embodiment, the present embodiment converts the expert knowledge of wind turbine anomaly detection into a structured form that can be processed by machines. The SCADA data of wind turbines has characteristics such as multi-parameter time dependence and nonlinear correlation. To address these characteristics, the knowledge system is hierarchically divided using propositional logic, and corresponding logical formulas are generated to provide basic input for subsequent graph neural network training. The expert knowledge of wind turbine anomaly detection includes known physical rules and fault modes.

[0049] Step 2.1: Design knowledge rule hierarchy according to different complexity of fault scenarios.

[0050] Specifically, in the present embodiment, the knowledge rule hierarchy includes a unique alarm layer, a fixed combination layer, and a complex combination layer. The unique alarm layer is suitable for simple fault scenarios, the fixed combination layer is suitable for medium complexity faults, and the complex combination layer is suitable for complex faults.

[0051] Knowledge hierarchical design divides expert knowledge into three levels to adapt to fault scenarios of different complexity. Each level is composed of propositions consisting of subject, predicate, and object, as well as logical connectors (such as Λ, V, ) combined into interpretable rules. Table 1 defines the key relationship predicates used in the knowledge expression module. These predicates are designed based on the physical topology of wind farms and monitoring parameters to construct hierarchical rules. By standardizing the functions and examples of predicates, Table 1 ensures the machine readability and scalability of knowledge rules, supporting subsequent knowledge embedding and fault discrimination of graph neural networks.

[0052] Table 1. Relationship predicate definition

[0053] Step 2.1.1: Set up the unique alarm layer.

[0054] Specifically, the unique alarm layer is suitable for simple fault scenarios and can directly identify fault types through a single alarm signal. This embodiment uses simple propositional logic formulas to map faults as direct inference rules. For example, the "gearbox bearing temperature too high" of a wind turbine can be encoded as the following propositional formula: (6).

[0055] wherein, represents that the system reports a "temperature anomaly" alarm, is the alarm type, indicating that the temperature monitoring value of the power grid new energy equipment exceeds the normal range, triggering an abnormal alarm, indicates that a fault occurs. is the fault type, indicating that the gearbox of the wind turbine is overheated due to high temperature, which is usually monitored by a temperature sensor.

[0056] This formula indicates that when the system reports a temperature anomaly alarm, it can directly infer the occurrence of the gearbox overheating fault. This formula form is simple and can quickly respond to single parameter anomalies.

[0057] Step 2.1.2: Set up the fixed combination layer.

[0058] Specifically, the fixed combination layer is suitable for moderately complex faults and needs to be identified in combination with the physical topology relationship between devices. This embodiment introduces a relationship predicate to describe the adjacency between wind turbine devices. For example, the "cross-device communication fault" of a wind turbine can be defined as the following formula: (7).

[0059] wherein, indicates a device and are adjacent in the physical topology, indicates a gearbox temperature sensor for monitoring the gearbox temperature, indicates a generator vibration sensor for monitoring the generator vibration state. is the fault type, indicating a coordinated anomaly caused by communication or data inconsistency between sensors, which is common in link interruption in a wind turbine group.

[0060] This formula captures the coordinated anomaly caused by link interruption in a wind turbine group by directly mapping the device location dependency, ensuring that the rule has a clear physical meaning.

[0061] Step 2.1.3: Set up the complex combination layer.

[0062] Specifically, the complex combination layer needs to integrate additional information such as topological relationship, time dependence and equipment chain state for the composite failure of the transmission chain. This embodiment adopts multiple logical connectors to construct composite rules to process complex associations in high-dimensional data. For example, combined with the relative temperature difference feature, the following formula is designed: (8).

[0063] wherein, represents the relative temperature difference, , is a temperature threshold, which is a preset critical value, and an abnormality is triggered when ΔT exceeds the threshold. The threshold is set based on historical data or expert experience, is used to identify the position of the equipment in the transmission chain, and additional information is integrated, is a fault type, which refers to a composite failure of the wind turbine transmission chain (including gearboxes, couplings, etc.) caused by overheating or mechanical stress.

[0064] The formula of this layer can effectively express the complex interaction between devices in the wind turbine, and improve the recognition ability of rare or composite failures.

[0065] Step 2.2: Based on the knowledge rule hierarchy, the expert knowledge related to wind farm anomaly detection is converted into a continuous vector representation according to the characteristics of the power grid new energy equipment operation data.

[0066] In a specific embodiment, this embodiment converts symbolic logic formulas into continuous vector representations, thereby realizing efficient integration with graph neural networks.

[0067] Step 2.2.1: Extract key information from power grid new energy equipment operation data to generate related propositions, and use logical connectors to construct composite formulas according to the generated propositions.

[0068] First, key information is extracted from wind turbine SCADA data to generate three types of propositions: relationship propositions, report propositions and occurrence propositions. Relationship propositions describe the topological association between devices (such as wind turbine adjacency), report propositions capture alarm signals (such as temperature anomalies), and occurrence propositions define fault types (such as gearbox overheating).

[0069] Subsequently, composite formulas are constructed using logical connectors. Formula construction follows an automated process, combining specific instances to form general rules. For example, the specific instances of the complex combination layer formula (8) can be combined into a general formula: (9).

[0070] wherein, represents a logical NOT symbol, represents a logical OR symbol, "vibration anomaly" alarm is not reported, is the alarm type, indicating that the equipment vibration monitoring value is abnormal (such as the vibration of the generator or the gear box exceeding the standard), indicates the gear box sensor, is the vibration amplitude, which is used to quantify the vibration intensity, is the amplitude threshold, which is a preset vibration threshold.

[0071] The formula integrates multiple conditions through disjunction operations to cover complex fault scenarios. To ensure computational efficiency, the formula needs to be converted to conjunctive normal form (CNF) before coding, and further compiled into deterministic decomposable negation normal form (d-DNNF), which supports efficient satisfiability checking and model counting.

[0072] Step 2.2.2: Use a graph neural network encoder to map the graph structure of the composite formula (such as the d-DNNF graph) into a low-dimensional embedding vector.

[0073] Specifically, the encoder training uses a mutual information maximization objective to make the generated embedding vector emb i can preserve the semantic features of the original formula. The embedding vectors corresponding to all formulas constitute the knowledge embedding vector set EMB=emb i This process ensures that symbolic knowledge is converted into continuous vectors aligned with the data feature space, providing a foundation for subsequent teacher-student training schemes.

[0074] The graph neural network encoder outputs a knowledge rule base, which has the advantage of integrating physical statistical features and symbolic knowledge, improving the model's explainability and generalization ability.

[0075] Step 3: Use a graph neural network with a knowledge enhancement mechanism to build a fault diagnosis model, using the hierarchical rule set in the knowledge rule base as the teacher and the graph neural network as the student to train the fault diagnosis model.

[0076] In one specific embodiment, the present embodiment aims to improve the accuracy and robustness of wind turbine abnormality detection through graph structure modeling and knowledge enhancement training. The core is to convert SCADA time series data into a graph structure, introduce a teacher-student framework to inject physical rule knowledge, realize the fusion of data-driven and symbolic reasoning, and the training process is as shown in Figure 2 .

[0077] To integrate domain knowledge into the model in the form of soft constraints, the present embodiment uses an improved teacher-student training scheme. The core is to use knowledge rules as "teachers" to evaluate the rationality of the output node state of GNN (student), and guide the training process accordingly.

[0078] Step 3.1: Generate knowledge signals.

[0079] Specifically, the knowledge rules (equations 6-8) are converted into discriminant functions that can be applied to the node state vector. Each rule R corresponds to a function whose output value represents the degree to which the node state matches the rule.

[0080] Step 3.2: Design the loss function.

[0081] In a specific embodiment, the total loss is used for the initial training phase of the model. The goal of this phase is to establish a high-performance baseline model (i.e., the teacher model). The loss function consists of two parts: the data reconstruction loss ensures that the model can learn the normal operation mode in the SCADA data, and the knowledge consistency loss embeds the physical rules and fault prior knowledge summarized by experts into the model in the form of soft constraints. It directly optimizes the weight parameters of the graph neural network (GNN) itself and is the foundation of model capability construction.

[0082] Specifically, the total loss function is composed of the data reconstruction loss and the knowledge consistency loss: (10).

[0083] where λ is a hyperparameter used to balance the weights of the two losses; is the reconstruction error based on self-supervised learning of input data; is defined as the difference between the confidence obtained by applying all knowledge rules to the GNN output node embedding H and the expected value, using binary cross-entropy loss: (11).

[0084] This mechanism allows the GNN to learn data features while its output is subject to soft constraints from physical rules, effectively embedding symbolic knowledge into data-driven models and enhancing the model's generalization ability in label-scarce scenarios.

[0085] Step 4: Use the trained fault diagnosis model to perform anomaly detection on the graph data representation to obtain the wind farm fault discrimination result.

[0086] In a specific embodiment, the present embodiment realizes wind turbine anomaly detection and performance optimization based on the trained GNN model. The key innovation is to use knowledge-enhanced features for unsupervised detection, improving the generalization ability for unknown anomalies.

[0087] Step 4.1: Use the trained fault diagnosis model to process the graph data representation to generate a low-dimensional embedding representation of the nodes.

[0088] Step 4.2: Calculate the anomaly score of the node low-dimensional embedding representation by the energy function using the restricted Boltzmann machine.

[0089] Specifically, the anomaly detection uses the restricted Boltzmann machine to calculate the anomaly score by the energy function. The RBM is unsupervised trained using the contrastive divergence algorithm to learn the parameters from normal data, so that the energy value of the normal node is low and the energy value of the abnormal node is high. The energy value calculation formula is: (12).

[0090] Where z is the node embedding vector, h is the hidden layer variable, a, b, are RBM parameters. The node embedding is optimized by DGI (Deep Graph Infomax) in KE-GNN, which integrates the graph structure features by maximizing the local-global mutual information, and improves the fault discrimination ability. The anomaly score is based on the energy value , and a dynamic threshold is set to determine the anomaly.

[0091] Step 4.3: Obtain the wind farm fault discrimination result according to the anomaly score.

[0092] Step 4.3.1: Set an adaptive dynamic anomaly determination threshold according to the historical detection results.

[0093] Specifically, the embodiment is aimed at the non-stationarity of the operating condition of the wind farm caused by sudden change of wind speed, power regulation, etc. The traditional fixed threshold method is prone to produce a large number of false positives. Therefore, a threshold adaptive updating algorithm is designed. The algorithm dynamically adjusts the anomaly determination threshold based on the historical detection results in the sliding window, and the updating rule is as follows: (13).

[0094] Where, is the anomaly determination threshold at time t, is the anomaly determination threshold at time t+1, η is the learning rate, is the anomaly detection loss, which is used for adaptive optimization of the threshold θ in the online application stage after the model goes online. This mechanism adapts to environmental changes through online learning, reducing the false positive rate.

[0095] Step 4.3.2: Compare the anomaly score with the dynamic threshold to obtain the wind farm fault discrimination result.

[0096] The distillation loss in the embodiment is The compression stage before model deployment. After the teacher model training is completed, in order to meet the harsh requirements of the on-site equipment of the wind power plant on the calculation efficiency, a lighter model needs to be obtained. At this time, the role of the distillation loss is to guide a lightweight student model to imitate the output behavior of the teacher model. The object it optimizes is the parameters of the student model, and the goal is to significantly improve the inference speed of the model without losing accuracy as much as possible, so that it is more suitable for real-time monitoring.

[0097] To meet the real-time requirement, the knowledge distillation technology is used to compress the model. The output of the pre-trained large model (teacher) is used as a supervision signal to train a lightweight model (student). Based on the distillation loss The training objective function is as follows: (14).

[0098] Wherein, is the cross-entropy loss, is the KL divergence loss, and are the output distributions of the student and the teacher respectively, is a balance hyperparameter used to adjust the weight of the two losses in the total objective function. This scheme substantially improves the inference speed of the model while basically maintaining the accuracy of the original model, making it more suitable for real-time online monitoring and fault warning of wind power plants.

[0099] Embodiment two: The embodiment two of the application provides a power grid new energy operation fault discrimination system based on knowledge enhancement, comprising: A data acquisition module is configured to acquire time series data of power grid new energy equipment operation, and construct a graph structure according to the time series data of power grid new energy equipment operation to obtain graph data representation; The data acquisition module is further configured to: Design node features according to the physical characteristics of the power grid new energy equipment operation; Construct a directed graph based on the node features using a sliding window method to obtain an adjacency matrix. Specifically, the time series data of the power grid new energy equipment operation is abstracted into a graph structure, the nodes represent the state of the equipment at the key moment, and the edges represent the causal dependence on the time series; Fusion of node features based on attention mechanism to obtain a reduced feature matrix.

[0100] A rule base establishment module is configured to encode expert knowledge into a hierarchical rule set according to the characteristics of the power grid new energy equipment operation data, and construct a knowledge rule base according to the hierarchical rule set; The rule base establishment module is further configured to: Design knowledge rule levels according to different complexity of fault scenarios; Based on the knowledge rule level, the expert knowledge related to the wind farm anomaly detection is converted into a continuous vector representation according to the characteristics of the new energy equipment operation data of the power grid, and the knowledge rule level includes a unique alarm layer, a fixed combination layer and a complex combination layer, wherein the unique alarm layer is suitable for a simple fault scenario, the fixed combination layer is suitable for a medium complexity fault, and the complex combination layer is suitable for a composite fault.

[0101] The model construction module is configured to construct a fault diagnosis model by using a graph neural network with a knowledge enhancement mechanism, take the hierarchical rule set in the knowledge rule base as a teacher and the graph neural network as a student, and train the fault diagnosis model. The model construction module is further configured to: process the graph data representation by using the trained fault diagnosis model to generate a node low-dimensional embedding representation; adopt a restricted Boltzmann machine to calculate an anomaly score of the node low-dimensional embedding representation through an energy function; obtain a wind farm fault discrimination result according to the anomaly score.

[0102] The anomaly detection module is configured to perform anomaly detection on the graph data representation by using the trained fault diagnosis model to obtain a wind farm fault discrimination result.

[0103] The anomaly detection module is further configured to: set an adaptive dynamic anomaly judgment threshold according to a historical detection result; compare the anomaly score with the dynamic threshold to obtain the wind farm fault discrimination result.

[0104] The steps and method embodiments of the above embodiment two correspond to the method embodiment one, and the specific implementation can refer to the related description part of embodiment one.

[0105] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application. In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network or other programmable device. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired or wireless manner. The computer readable storage medium can be any available medium accessible by the computer or a data processing device such as a server, data center and the like integrated with one or more available media. The available medium can be a magnetic medium, an optical medium or a semiconductor medium and the like. The above description is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A knowledge-enhanced power grid new energy operation fault discrimination method, characterized in that, The method comprises the following steps: obtain time series data of grid new energy equipment operation, and construct a graph structure according to the time series data of grid new energy equipment operation to obtain a graph data representation; encode expert knowledge into a hierarchical rule set according to the characteristics of the grid new energy equipment operation data, and construct a knowledge rule base according to the hierarchical rule set; construct a fault diagnosis model by using a graph neural network with a knowledge enhancement mechanism, use the hierarchical rule set in the knowledge rule base as a teacher and the graph neural network as a student to train the fault diagnosis model; use the trained fault diagnosis model to perform anomaly detection on the graph data representation to obtain a wind farm fault discrimination result.

2. The knowledge enhancement-based power grid new energy operation fault judgment method according to claim 1, characterized in that, The specific steps of constructing a graph structure according to the time series data of grid new energy equipment operation are as follows: design node features according to the physical characteristics of grid new energy equipment operation; construct a directed graph based on the node features by using a sliding window method to obtain an adjacency matrix, specifically, abstract the time series data of grid new energy equipment operation into a graph structure, nodes represent the state of the equipment at key moments, and edges represent the causal dependence on the time series; fuse the node features based on an attention mechanism to obtain a dimension-reduced feature matrix. 3.The knowledge-enhanced power grid new energy operation fault identification method of claim 1, wherein, The specific steps of encoding expert knowledge into a hierarchical rule set according to the characteristics of the grid new energy equipment operation data are as follows: design knowledge rule levels according to different complexity of fault scenarios; based on the knowledge rule levels, convert expert knowledge related to wind farm anomaly detection into a continuous vector representation according to the characteristics of the grid new energy equipment operation data, the knowledge rule levels include a unique alarm layer, a fixed combination layer and a complex combination layer, wherein the unique alarm layer is suitable for simple fault scenarios, the fixed combination layer is suitable for medium complexity faults, and the complex combination layer is suitable for complex faults. 4.The knowledge-enhanced power grid new energy operation fault identification method of claim 1, wherein, The specific steps of using the trained fault diagnosis model to perform anomaly detection on the graph data representation are as follows: use the trained fault diagnosis model to process the graph data representation to generate a node low-dimensional embedding representation; use a restricted Boltzmann machine to calculate an anomaly score of the node low-dimensional embedding representation through an energy function; obtain a wind farm fault discrimination result according to the anomaly score.

5. The knowledge enhancement-based power grid new energy operation fault identification method of claim 4, wherein, The specific steps of obtaining a wind farm fault discrimination result according to the anomaly score are as follows: set an adaptive dynamic anomaly judgment threshold according to historical detection results; compare the anomaly score with the dynamic threshold to obtain a wind farm fault discrimination result.

6. The knowledge-based enhanced power grid new energy operation fault discrimination system, characterized in that, It comprises: a data acquisition module configured to obtain time series data of grid new energy equipment operation, and construct a graph structure according to the time series data of grid new energy equipment operation to obtain a graph data representation; a rule base establishment module configured to encode expert knowledge into a hierarchical rule set according to the characteristics of the grid new energy equipment operation data, and construct a knowledge rule base according to the hierarchical rule set; a model construction module configured to construct a fault diagnosis model by using a graph neural network with a knowledge enhancement mechanism, use the hierarchical rule set in the knowledge rule base as a teacher and the graph neural network as a student to train the fault diagnosis model; an anomaly detection module configured to use the trained fault diagnosis model to perform anomaly detection on the graph data representation to obtain a wind farm fault discrimination result.

7. The knowledge enhancement based power grid new energy operation fault diagnosis system according to claim 6, characterized in that, The data acquisition module is further configured to: The node features are designed according to the physical characteristics of the operation of the grid new energy equipment; A directed graph is constructed based on the node features by using a sliding window method to obtain an adjacency matrix, specifically, time series data of the operation of the grid new energy equipment is abstracted into a graph structure, a node represents the state of the equipment at a key moment, and an edge represents the causal dependence on the time series; The node features are fused based on an attention mechanism to obtain a reduced feature matrix.

8. The knowledge enhancement based power grid new energy operation fault diagnosis system of claim 6, wherein, The rule base establishment module is further configured to: Design a knowledge rule level according to different complexity of fault scenarios; Based on the knowledge rule level, the expert knowledge related to the wind farm anomaly detection is converted into a continuous vector representation according to the characteristics of the grid new energy equipment operation data, the knowledge rule level includes a unique alarm layer, a fixed combination layer and a complex combination layer, wherein the unique alarm layer is suitable for simple fault scenarios, the fixed combination layer is suitable for medium complexity faults, and the complex combination layer is suitable for complex faults.

9. The knowledge enhancement based power grid new energy operation fault diagnosis system of claim 6, wherein, The model construction module is further configured to: Process the graph data representation by using the trained fault diagnosis model to generate a node low-dimensional embedding representation; An abnormal score of the node low-dimensional embedding representation is calculated by using a restricted Boltzmann machine through an energy function; The wind farm fault discrimination result is obtained according to the abnormal score.

10. The knowledge enhancement based power grid new energy operation fault discrimination system of claim 9, wherein, The anomaly detection module is further configured to: Set an adaptive dynamic anomaly judgment threshold according to the historical detection result; Compare the abnormal score with the dynamic threshold to obtain the wind farm fault discrimination result.

Images