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

Abstract

The application discloses a kind of based on knowledge enhancement's power grid new energy operation fault discrimination method and system, it is related to wind farm fault diagnosis technical field.The method includes steps: obtaining the time series data of power grid new energy equipment operation, and according to the time series data of power grid new energy equipment operation graph structure is constructed, obtains graph data representation;According to the characteristics of power grid new energy equipment operation data expert knowledge is encoded as hierarchical rule set, and according to hierarchical rule set constructs knowledge rule base;With the knowledge enhancement mechanism of graph neural network constructs fault diagnosis model, the hierarchical rule set in knowledge rule base is as teacher, and graph neural network is as student, and the fault diagnosis model is trained;Abnormal detection is carried out to graph data representation using the trained fault diagnosis model, and wind farm fault discrimination result is obtained.The present application can realize the unsupervised accurate identification of complex fault scenarios such as gearbox fault, variable pitch system anomaly.

Description

Technical Field

[0001] This invention relates to the field of wind farm fault diagnosis technology, and in particular to a knowledge-enhanced method and system for identifying faults in the operation of new energy sources in power grids. Background Technology

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

[0003] With the popularization of intelligent monitoring technology in wind farms, SCADA (Supervisory Control and Data Acquisition) is widely used for data acquisition of wind turbine operating status. Based on the analysis of multi-parameter time-series data, early fault warning and diagnosis of wind turbines can be achieved. However, traditional fault diagnosis methods (such as physical model-based methods) heavily rely on expert experience, resulting in poor adaptability and modeling difficulties. To introduce intelligent means, early artificial intelligence methods, represented by expert systems, have been applied. This method relies on manually summarized rules and prior knowledge for logical reasoning, which partially reduces the dependence on experience, but still requires the construction of a complete knowledge base and is difficult to handle unknown faults.

[0004] With technological advancements, purer data-driven methods, such as deep learning, have gradually become the mainstream research approach due to their powerful self-learning and nonlinear fitting capabilities. Common techniques include autoencoders, long short-term memory networks, and graph neural networks. Among these, unsupervised methods like autoencoders detect anomalies through reconstruction errors but are sensitive to dynamic changes in operating conditions; long short-term memory networks can capture temporal dependencies but often ignore spatial relationships between variables; while graph neural networks can model topological relationships between variables, most methods assume a static and fixed graph structure, making it difficult to adapt to the highly dynamic characteristics of wind power data, and they also have limitations such as computational complexity and weak noise resistance.

[0005] With the continuous expansion of wind farm scale and the sustained growth of operational data, the accuracy of fault diagnosis heavily relies on the algorithm's ability to extract features and mine correlations from high-dimensional, nonlinear, and strongly coupled SCADA data. However, existing methods cannot effectively integrate multi-source information, cannot adapt well to dynamic changes, and cannot balance efficiency and robustness. In addition, most existing artificial intelligence diagnostic solutions often use a single model or isolated learning mechanism, lacking multi-level knowledge fusion and adaptive optimization capabilities, resulting in limited generalization performance in complex fault scenarios, and there is still room for improvement in diagnostic accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a knowledge-enhanced method and system for identifying faults in the operation of new energy sources in power grids, which can achieve unsupervised and accurate identification of complex fault scenarios such as gearbox faults and pitch system anomalies.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] The first aspect of this invention provides a knowledge-enhanced method for identifying faults in the operation of new energy sources in power grids, comprising the following steps:

[0009] Obtain time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation;

[0010] Based on the characteristics of the power grid's new energy equipment operation data, expert knowledge is encoded into a hierarchical rule set, and a knowledge rule base is constructed based on the hierarchical rule set;

[0011] A fault diagnosis model is constructed using a graph neural network with a knowledge enhancement mechanism. The hierarchical rule set in the knowledge rule base is used as the teacher and the graph neural network is used as the student to train the fault diagnosis model.

[0012] The trained fault diagnosis model is used to detect anomalies in the graph data representation to obtain the fault discrimination results of the wind farm.

[0013] Furthermore, the specific steps for constructing the graph structure based on the time-series data of the operation of new energy equipment in the power grid are as follows:

[0014] Node characteristics are designed based on the physical characteristics of the operation of new energy equipment in the power grid;

[0015] Based on node features, a directed graph is constructed using the sliding window method to obtain the adjacency matrix. Specifically, the time series data of the operation of new energy equipment in the power grid is abstracted into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series.

[0016] The node features are fused based on the attention mechanism to obtain a dimensionality-reduced feature matrix.

[0017] Furthermore, the specific steps for encoding expert knowledge into a hierarchical rule set based on the characteristics of power grid renewable energy equipment operation data are as follows:

[0018] The knowledge rule hierarchy is designed based on fault scenarios of varying complexity;

[0019] 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 grid new energy equipment operation data. 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 compound faults.

[0020] Furthermore, the specific steps for anomaly detection using the trained fault diagnosis model on the graph data representation are as follows:

[0021] The trained fault diagnosis model is used to process graph data representations and generate low-dimensional node embedding representations.

[0022] Anomaly scores for low-dimensional embedding representations of nodes are calculated using a restricted Boltzmann machine via an energy function.

[0023] The fault identification results of the wind farm are obtained based on the abnormal scores.

[0024] Furthermore, the specific steps for obtaining wind farm fault identification results based on anomaly scores are as follows:

[0025] An adaptive dynamic anomaly detection threshold is set based on historical detection results;

[0026] The fault identification result of the wind farm is obtained by comparing the anomaly score with the dynamic threshold.

[0027] A second aspect of the present invention provides a knowledge-enhanced power grid renewable energy operation fault identification system, comprising:

[0028] The data acquisition module is configured to acquire time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation;

[0029] The rule base building module is configured to encode expert knowledge into hierarchical rule sets based on the characteristics of the power grid new energy equipment operation data, and to build a knowledge rule base based on the hierarchical rule sets;

[0030] The model building module is configured to build a fault diagnosis model using a graph neural network with a knowledge enhancement mechanism, using a 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.

[0031] The anomaly detection module is configured to use the trained fault diagnosis model to perform anomaly detection on the graph data representation and obtain the fault judgment result of the wind farm.

[0032] Furthermore, the data acquisition module is also configured as follows:

[0033] Node characteristics are designed based on the physical characteristics of the operation of new energy equipment in the power grid;

[0034] Based on node features, a directed graph is constructed using the sliding window method to obtain the adjacency matrix. Specifically, the time series data of the operation of new energy equipment in the power grid is abstracted into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series.

[0035] The node features are fused based on the attention mechanism to obtain a dimensionality-reduced feature matrix.

[0036] Furthermore, the rule base creation module is also configured as follows:

[0037] The knowledge rule hierarchy is designed based on fault scenarios of varying complexity;

[0038] 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 grid new energy equipment operation data. 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 compound faults.

[0039] Furthermore, the model building module is also configured as follows:

[0040] The trained fault diagnosis model is used to process graph data representations and generate low-dimensional node embedding representations.

[0041] Anomaly scores for low-dimensional embedding representations of nodes are calculated using a restricted Boltzmann machine via an energy function.

[0042] The fault identification results of the wind farm are obtained based on the abnormal scores.

[0043] Furthermore, the anomaly detection module is also configured as follows:

[0044] An adaptive dynamic anomaly detection threshold is set based on historical detection results;

[0045] The fault identification result of the wind farm is obtained by comparing the anomaly score with the dynamic threshold.

[0046] The above one or more technical solutions have the following beneficial effects:

[0047] This invention discloses a knowledge-enhanced method and system for fault identification in the operation of renewable energy in power grids. It encodes wind farm expert knowledge into propositional logic formulas, generating a hierarchical general rule set, including a unique alarm layer, a fixed combination layer, and a complex combination layer to cover fault scenarios of varying complexity. The data from the Supervisory Control and Data Acquisition (SCADA) system is abstracted into a graph structure, with nodes designed based on physical statistical features and edges constructed using a sliding window to capture temporal dependencies. Then, using knowledge rules as "teachers" and a Graph Neural Network (GNN) as "students," knowledge is embedded into the GNN training by maximizing mutual information, and a regularization term is added to constrain the distance between the output and the knowledge embedding. Finally, a Restricted Boltzmann Machine (RBM) is used as an anomaly detector, calculating node state scores based on energy functions and combining this with a dynamic threshold adjustment mechanism to achieve unsupervised fault identification. This invention solves the problems of high-dimensional redundancy and missing labels in SCADA data, improves the accuracy of identifying complex faults, and enhances inference efficiency through model lightweighting technology, providing reliable support for intelligent operation and maintenance of wind farms.

[0048] This invention addresses the issues of high-dimensional redundancy and missing labels in SCADA data by integrating wind power physical statistical characteristics with symbolic knowledge. It categorizes fault scenarios into three levels and constructs a temporal dependency graph, effectively capturing nonlinear correlations and topological information within the data. Furthermore, it achieves efficient adaptive fault detection through knowledge embedding constraints and a dynamic threshold adjustment mechanism.

[0049] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the knowledge-enhanced power grid new energy operation fault identification method in Embodiment 1 of the present invention;

[0052] Figure 2 This is a flowchart of the training process of a graph neural network with a knowledge enhancement mechanism in Embodiment 1 of the present invention. Detailed Implementation

[0053] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. 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 this invention pertains.

[0054] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0056] Example 1:

[0057] This invention provides a knowledge-enhanced method for identifying faults in the operation of renewable energy in power grids. By employing a knowledge enhancement mechanism, the performance of a Generative Neural Network (GNN) in wind farm fault detection is improved, achieving unsupervised and high-precision fault identification. Targeting the characteristics of wind power SCADA data, expert knowledge is encoded into a hierarchical rule set, including a unique alarm layer, a fixed combination layer, and a complex combination layer to cover fault scenarios of varying complexity. Graph structure modeling is also performed, abstracting SCADA time-series data into a graph structure. Node features are designed based on physical statistics such as power curve deviation and temperature differences. Edges are constructed using a sliding window to build a directed graph to capture temporal dependencies, and a Graph Attention Network (GAT) is used for feature fusion. Then, a teacher-student training scheme is adopted, with knowledge rules acting as the "teacher" and the GNN as the "student." A knowledge consistency regularization term is added to the loss function to constrain the distance between the GNN output and the knowledge embedding. Finally, unsupervised anomaly detection is achieved based on an Energy Function (RBM), calculating node state scores using an energy function and optimizing system performance by adjusting dynamic thresholds. This embodiment integrates physical statistical features and symbolic knowledge, solves the problems of high-dimensional redundancy and missing labels in SCADA data, improves the identification accuracy of complex faults (such as gearbox faults and pitch system anomalies), and improves inference efficiency through model lightweighting technology.

[0058] like Figure 1 As shown, this embodiment takes wind farm equipment in the grid's new energy equipment as an example for illustration, and specifically includes the following steps:

[0059] Step 1: Obtain the time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation.

[0060] Step 1.1: Obtain the time-series data of the operation of new energy equipment in the power grid.

[0061] In one specific implementation, the SCADA system continuously collects various operating parameters through sensors installed on the wind turbine, forming high-dimensional time-series data. SCADA data specifically includes more than 30 numerical parameters such as temperature, power, and wind speed, and the data type is primarily numerical time-series.

[0062] Step 1.2: Construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid.

[0063] In one specific implementation, in order to capture the temporal dependencies and nonlinear correlations in SCADA data, this embodiment abstracts the temporal data of the operation of new energy equipment in the power grid into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series.

[0064] Step 1.2.1: Design node characteristics based on the physical characteristics of the operation of new energy equipment in the power grid.

[0065] Specifically, the node feature design encompasses three types of physical feature mapping functions, extracting key parameters from different dimensions. These physical feature mapping functions include feature curve mapping functions, normal behavior model mapping functions, and relative temperature difference mapping functions.

[0066] Characteristic curve mapping function Based on the power-wind speed characteristic curve of wind turbines, a power reduction parameter ξ is defined to quantify the power generation efficiency per unit wind speed. The calculation formula is as follows:

[0067] (1).

[0068] in, This represents the actual wind speed at the current moment, and is an input variable directly monitored by the SCADA system. This indicates the rated wind speed, which is the design wind speed at which the wind turbine reaches its rated power. Cut-in wind speed represents the minimum wind speed at which the wind turbine begins generating electricity. This indicates the actual power generation of the wind turbine at the current moment. Rated power indicates the maximum output power of a wind turbine under design conditions.

[0069] 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:

[0070] (2).

[0071] Where, 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.

[0072] Relative temperature difference mapping function : Define the temperature difference feature ΔT to capture heat-related faults. The formula is:

[0073] (3).

[0074] Where, and represent any two monitoring parameters in the SCADA data, usually temperature-related variables (such as gearbox oil temperature, generator bearing temperature).

[0075] After all features are normalized, a node feature matrix X ∈ R {N×F} is formed, where N is the number of time points, F is the feature dimension, and R is the set of real numbers.

[0076] Step 1.2.2: Construct a directed graph based on the node features using the sliding window method to obtain the adjacency matrix.

[0077] Specifically, first, abstract the SCADA time series data into graph nodes, and each node corresponds to the device state features (such as power, temperature, etc.) at a time point t. To capture the time series dependence, the sliding window method is used in this embodiment 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 Ai,j = 1 in the adjacency matrix A, otherwise it is 0. This structure ensures that the model only depends on past information and conforms to the physical causality law.

[0078] The adjacency matrix A is an N×N binary matrix (where N is the number of nodes, that is, the number of time series data points), which is used to represent the directed connection between nodes. Define the neighbor set of each node through the adjacency matrix A, providing a basis for the subsequent operation of GAT. And the graph data representation in this embodiment includes the node feature matrix X and the adjacency matrix A.

[0079] Step 1.2.3: Based on the attention mechanism, fuse the node features to obtain a dimensionality-reduced feature matrix.

[0080] Specifically, to overcome the limitations of fixed weights in traditional GCNs, this embodiment uses GAT as the encoder. By dynamically aggregating neighbor information in the adjacency matrix through learnable weights, it more effectively captures nonlinear correlations in SCADA data. For node i, its feature vector updated after the GAT layer... The calculation formula is:

[0081] (4).

[0082] in, The state characteristics of the wind turbine at node j are extracted from the original data using mapping function formulas 1-3, and are used to characterize the operating status of the equipment. Let i represent the set of neighbors of node i. W is a trainable linear transformation weight matrix; used to perform dimensionality mapping and transformation on the features of all nodes. The attention weights are calculated and represent the importance of node j to node i. It is a non-linear activation function.

[0083] (5).

[0084] in, This represents the state characteristics of the wind turbine at node i. This represents the state characteristics of the wind turbine at node k. The parameter k is a temporary index variable used in mathematical calculations to represent a specific neighbor node in the neighbor set N(i) of node i. For a trainable single-layer feedforward neural network parameter vector, This represents vector concatenation; LeakyReLU is a linear rectified function with leakage, used to introduce nonlinearity. This mechanism allows the model to focus on key neighbor nodes, improving its ability to capture nonlinear associations. The fused output is a dimensionality-reduced feature matrix H∈R. {N×F′} This provides highly 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.

[0085] Step 2: Based on the characteristics of the power grid new energy equipment operation data, encode expert knowledge into a hierarchical rule set, and construct a knowledge rule base based on the hierarchical rule set.

[0086] In one specific implementation, this embodiment transforms the expert knowledge of wind turbine anomaly detection into a machine-processable structured form. Wind turbine SCADA data exhibits characteristics such as multi-parameter temporal dependence and nonlinear correlation. To address these characteristics, propositional logic is used to hierarchically divide the knowledge system and generate corresponding logical formulas, providing the basic input for subsequent graph neural network training. The expert knowledge for wind turbine anomaly detection includes known physical rules and fault modes.

[0087] Step 2.1: Design knowledge rule hierarchies based on fault scenarios of varying complexity.

[0088] Specifically, in this 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 compound faults.

[0089] The hierarchical knowledge design divides expert knowledge into three levels to accommodate fault scenarios of varying complexity. Each level is defined by propositions consisting of subjects, predicates, and objects, as well as logical connectors (such as Λ, V, ...). These are combined into interpretable rules. Table 1 defines the key relation predicates used in the knowledge representation module. These predicates are designed based on the wind farm's physical topology and monitoring parameters to construct hierarchical rules. By standardizing the functions and examples of the predicates, Table 1 ensures the machine readability and scalability of the knowledge rules, supporting subsequent knowledge embedding and fault identification using graph neural networks.

[0090] Table 1. Definition of Relational Predicates

[0091]

[0092] Step 2.1.1: Set a unique alarm layer.

[0093] Specifically, the unique alarm layer is suitable for simple fault scenarios, allowing direct identification of the fault type through a single alarm signal. This embodiment employs a simple propositional logic formula to map faults into direct reasoning rules. For example, the "gearbox bearing temperature too high" fault in a wind turbine can be encoded as the following propositional formula:

[0094] (6).

[0095] in, This indicates that the system has reported a "temperature anomaly" alarm. This is an alarm type, indicating that the temperature monitoring value of the power grid's new energy equipment has exceeded the normal range, triggering an abnormal alarm. This indicates a malfunction has occurred. It refers to a fault type, specifically a fault in the wind turbine gearbox caused by excessive temperature, which is usually detected by a temperature sensor.

[0096] This formula indicates that when the system reports an abnormal temperature alarm, a gearbox overheating fault can be directly inferred. This type of formula is concise and can quickly respond to anomalies in a single parameter.

[0097] Step 2.1.2: Set up a fixed combination layer.

[0098] Specifically, the fixed combination layer is suitable for medium-complexity faults, which need to be identified in conjunction with the physical topology relationships between devices. This embodiment introduces relational predicates to describe the adjacency between wind turbine devices. For example, the "cross-device communication fault" of a wind turbine can be defined by the following formula:

[0099] (7).

[0100] in, Display device and Adjacent in physical topology This refers to a gearbox temperature sensor, used to monitor the temperature of the gearbox. This refers to a generator vibration sensor, used to monitor the vibration status of a generator. It is a fault type, referring to a coordination anomaly caused by communication or data inconsistency between sensors, which is commonly seen in link interruptions in wind turbine groups.

[0101] This formula captures coordination anomalies caused by link interruptions in wind turbine clusters by directly mapping device location dependencies, ensuring that the rules have clear physical meaning.

[0102] Step 2.1.3: Set up a complex composite layer.

[0103] Specifically, complex combination layers, designed for compound faults such as transmission chain failures, need to integrate additional information such as topological relationships, time dependencies, and equipment chain states. This embodiment employs multiple logical connectors to construct composite rules to handle complex relationships in high-dimensional data. For example, combining the relative temperature difference feature, the following formula is designed:

[0104] (8).

[0105] in, Indicates relative temperature difference. , This is the temperature threshold, a preset critical value. An anomaly is triggered when ΔT exceeds this threshold. The threshold is set based on historical data or expert experience. Used to identify the position of equipment in the drive train and integrate additional information. It refers to a type of fault, specifically a combined fault in the wind turbine's drivetrain (including gearboxes, couplings, etc.) caused by overheating or mechanical stress.

[0106] This formula can effectively express the complex interactions between equipment in a wind turbine and improve the ability to identify rare or complex faults.

[0107] 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 grid new energy equipment operation data.

[0108] In one specific implementation, this embodiment converts symbolic logic formulas into continuous vector representations, thereby achieving efficient integration with graph neural networks.

[0109] Step 2.2.1: Extract key information from the operation data of new energy equipment in the power grid to generate relevant propositions, and construct compound formulas based on the generated propositions using logical connectors.

[0110] First, key information is extracted from the SCADA data of the wind turbine to generate three types of propositions: relational propositions, reporting propositions, and occurrence propositions. Relational propositions describe the topological relationships between devices (such as wind turbine adjacency), reporting propositions capture alarm signals (such as temperature anomalies), and occurrence propositions define the fault type (such as gearbox overheating).

[0111] Subsequently, logical connectors are used to construct compound formulas. Formula construction follows an automated process, forming general rules by merging specific instances. For example, the complex composite layer formula (8) can have its specific instances combined into a general formula:

[0112] (9).

[0113] in, Represents the logical NOT symbol. Represents logical OR symbols, This indicates that no "abnormal vibration" alarm was reported. It is an alarm type, indicating that the equipment vibration monitoring value is abnormal (such as excessive vibration of generator or gearbox). Indicates gearbox sensor, The amplitude of the vibration is used to quantify the vibration intensity. The amplitude threshold is a preset critical value for vibration.

[0114] This formula integrates multiple conditions through disjunction operations to cover complex fault scenarios. To ensure computational efficiency, the formula needs to be converted to conjunctive norm (CNF) form before encoding and further compiled into deterministic decomposable negation norm (d-DNNF) form. This form supports efficient satisfiability checks and model counting.

[0115] Step 2.2.2: Use a graph neural network encoder to map the graph structure of the composite formula (such as a d-DNNF graph) into a low-dimensional embedding vector.

[0116] Specifically, encoder training employs mutual information maximization as its objective, aiming to maximize the generated embedding vector emb. i It 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 transformed into continuous vectors aligned with the data feature space, providing a foundation for subsequent teacher-student training schemes.

[0117] The graph neural network encoder outputs a knowledge rule base, which has the advantage of integrating physical statistical features and symbolic knowledge, thereby improving the interpretability and generalization ability of the model.

[0118] Step 3: 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 the teacher and the graph neural network as the student to train the fault diagnosis model.

[0119] In one specific implementation, this embodiment aims to improve the accuracy and robustness of wind turbine anomaly detection through graph structure modeling and knowledge-enhanced training. Its core lies in transforming SCADA time-series data into a graph structure, then introducing a teacher-student framework to inject physical rule knowledge, achieving a fusion of data-driven and symbolic reasoning. The training process is as follows: Figure 2 As shown.

[0120] To incorporate domain knowledge into the model in the form of soft constraints, this embodiment employs an improved teacher-student training scheme. Its core lies in using knowledge rules as the "teacher" to evaluate the rationality of the output node states of the GNN (student) and thereby guide the training process.

[0121] Step 3.1: Generate knowledge signals.

[0122] Specifically, the knowledge rules (Equations 6-8) are transformed into discriminant functions applicable to node state vectors. Each rule R corresponds to a function. Its output value represents the node state. The degree to which the rule is met.

[0123] Step 3.2: Design the loss function.

[0124] In one specific implementation, the total loss This is used in 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: data reconstruction loss ensures that the model can learn the normal operating patterns in SCADA data, while knowledge consistency loss embeds expert-summarized physical rules and prior knowledge of faults into the model as soft constraints. It directly optimizes the weight parameters of the graph neural network (GNN) itself and is the foundation for building the model's capabilities.

[0125] Specifically, the total loss function It consists of data reconstruction loss and knowledge consistency loss:

[0126] (10).

[0127] Where λ is a hyperparameter used to balance the weights of the two losses; The reconstruction error is based on self-supervised learning of the input data. Defined as the difference between the confidence score and the expected value obtained after applying all knowledge rules to the embedding H of the GNN output node, using binary cross-entropy loss:

[0128] (11).

[0129] This mechanism allows GNNs to learn data features while their outputs are subject to soft constraints from physical rules, thereby efficiently embedding symbolic knowledge into the data-driven model and enhancing the model's generalization ability in scenarios where labels are scarce.

[0130] Step 4: Use the trained fault diagnosis model to perform anomaly detection on the graph data representation to obtain the fault judgment results of the wind farm.

[0131] In one specific implementation, this embodiment uses a pre-trained GNN model to achieve anomaly detection and performance optimization of wind turbine units. The key innovation is the application of knowledge-enhanced features to unsupervised detection, improving the generalization ability to unknown anomalies.

[0132] Step 4.1: Use the trained fault diagnosis model to process the graph data representation and generate low-dimensional embedding representations of nodes.

[0133] Step 4.2: Use a restricted Boltzmann machine to calculate the anomaly score of the low-dimensional embedding representation of the nodes through an energy function.

[0134] Specifically, anomaly detection employs a Restricted Boltzmann Machine (RBM), calculating anomaly scores using an energy function. The RBM uses a contrastive divergence algorithm for unsupervised training, learning parameters on normal data to ensure lower energy values ​​for normal nodes and higher energy values ​​for anomaly nodes. The energy value calculation formula is as follows:

[0135] (12).

[0136] Where z is the node embedding vector, h is the hidden layer variable, and a, b, ... These are the RBM parameters. Node embeddings are optimized using DGI (Deep Graph Infomax) in KE-GNN, which improves fault detection capability by maximizing the local-global mutual information fusion graph structure features. Anomaly scores are based on energy values. Set a dynamic threshold to determine anomalies.

[0137] Step 4.3: Obtain the wind farm fault identification result based on the anomaly score.

[0138] Step 4.3.1: Set an adaptive dynamic anomaly detection threshold based on historical detection results.

[0139] Specifically, this embodiment addresses the non-stationarity of wind farm operation due to sudden wind speed changes and power regulation, where traditional fixed threshold methods are prone to generating numerous false alarms. Therefore, an adaptive threshold update algorithm is designed. This algorithm dynamically adjusts the anomaly detection threshold based on historical detection results within a sliding window. Its update rules are as follows:

[0140] (13).

[0141] in, Let t be the anomaly detection threshold. Let η be the anomaly detection threshold at time t+1, and η be the learning rate. This is used as the anomaly detection loss, for adaptive optimization of the threshold θ during the online application phase after the model is deployed. This mechanism adapts to environmental changes through online learning, reducing the false positive rate.

[0142] Step 4.3.2: Compare the anomaly score with the dynamic threshold to obtain the wind farm fault identification result.

[0143] Distillation loss in this embodiment Used in the compression phase before model deployment. After the teacher model is trained, a lighter model is needed to meet the stringent computational efficiency requirements of wind farm field equipment. At this point, distillation loss guides a lightweight student model to mimic the output behavior of the teacher model. It optimizes the parameters of the student model, aiming to significantly improve the model's inference speed with minimal loss of accuracy, making it more suitable for real-time monitoring.

[0144] To meet real-time requirements, a knowledge distillation technique 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). This is based on distillation loss. The training objective function is as follows:

[0145] (14).

[0146] in, For cross-entropy loss, For KL divergence loss, and The output distributions for students and teachers are respectively. To balance the hyperparameters, the weights of the two losses in the overall objective function are adjusted. This scheme significantly improves the model's inference speed while maintaining the accuracy of the original model, making it more suitable for real-time online monitoring and fault early warning of wind farms.

[0147] Example 2:

[0148] Embodiment 2 of the present invention provides a knowledge-enhanced power grid renewable energy operation fault identification system, comprising:

[0149] The data acquisition module is configured to acquire time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation;

[0150] The data acquisition module is also configured as follows:

[0151] Node characteristics are designed based on the physical characteristics of the operation of new energy equipment in the power grid;

[0152] Based on node features, a directed graph is constructed using the sliding window method to obtain the adjacency matrix. Specifically, the time series data of the operation of new energy equipment in the power grid is abstracted into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series.

[0153] The node features are fused based on the attention mechanism to obtain a dimensionality-reduced feature matrix.

[0154] The rule base building module is configured to encode expert knowledge into hierarchical rule sets based on the characteristics of the power grid new energy equipment operation data, and to build a knowledge rule base based on the hierarchical rule sets;

[0155] The rule base creation module is also configured as follows:

[0156] The knowledge rule hierarchy is designed based on fault scenarios of varying complexity;

[0157] 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 grid new energy equipment operation data. 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 compound faults.

[0158] The model building module is configured to build a fault diagnosis model using a graph neural network with a knowledge enhancement mechanism, using a 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.

[0159] The model building module is also configured as follows:

[0160] The trained fault diagnosis model is used to process graph data representations and generate low-dimensional node embedding representations.

[0161] Anomaly scores for low-dimensional embedding representations of nodes are calculated using a restricted Boltzmann machine via an energy function.

[0162] The fault identification results of the wind farm are obtained based on the abnormal scores.

[0163] The anomaly detection module is configured to use the trained fault diagnosis model to perform anomaly detection on the graph data representation and obtain the fault judgment result of the wind farm.

[0164] The anomaly detection module is also configured as follows:

[0165] An adaptive dynamic anomaly detection threshold is set based on historical detection results;

[0166] The fault identification result of the wind farm is obtained by comparing the anomaly score with the dynamic threshold.

[0167] The steps and methods involved in the above embodiment two correspond to those in embodiment one. For specific implementation details, please refer to the relevant description section of embodiment one.

[0168] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0169] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A 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 flow or function according to the embodiments of this application is 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 through 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 via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0170] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A knowledge-enhanced method for identifying faults in the operation of new energy sources in power grids, characterized in that: Includes the following steps: Obtain time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation; Based on the characteristics of the power grid's new energy equipment operation data, expert knowledge is encoded into a hierarchical rule set, and a knowledge rule base is constructed based on the hierarchical rule set; A fault diagnosis model is constructed using a graph neural network with a knowledge enhancement mechanism. The hierarchical rule set in the knowledge rule base is used as the teacher and the graph neural network is used as the student to train the fault diagnosis model. An improved teacher-student training scheme is adopted, in which knowledge rules are used as teachers to evaluate the rationality of the output node state of GNN and guide the training process accordingly. Generate knowledge signals: Transform knowledge rules into discriminant functions that can be applied to node state vectors; each rule R corresponds to a function. Its output value represents the node state. The degree to which the rule is met; Design the loss function: Total Loss Used in the initial training phase of the model; the loss function consists of two parts: data reconstruction loss ensures the model can learn the normal operating patterns in SCADA data, while knowledge consistency loss embeds expert-summarized physical rules and prior fault knowledge into the model as soft constraints; it directly optimizes the weight parameters of the graph neural network itself, which is the foundation for building model capabilities; specifically, the total loss function It consists of data reconstruction loss and knowledge consistency loss: Where λ is a hyperparameter used to balance the weights of the two losses; The reconstruction error is based on self-supervised learning of the input data. Defined as the difference between the confidence score and the expected value obtained after applying all knowledge rules to the embedding H of the GNN output node, using binary cross-entropy loss: ； The trained fault diagnosis model is used to detect anomalies in the graph data representation to obtain the fault discrimination results of the wind farm.

2. The knowledge-enhanced power grid renewable energy operation fault identification method as described in claim 1, characterized in that, The specific steps for constructing the graph structure based on the time-series data of the operation of new energy equipment in the power grid are as follows: Node characteristics are designed based on the physical characteristics of the operation of new energy equipment in the power grid; Based on node features, a directed graph is constructed using the sliding window method to obtain the adjacency matrix. Specifically, the time series data of the operation of new energy equipment in the power grid is abstracted into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series. The node features are fused based on the attention mechanism to obtain a dimensionality-reduced feature matrix.

3. The knowledge-enhanced power grid renewable energy operation fault identification method as described in claim 1, characterized in that, The specific steps for encoding expert knowledge into a hierarchical rule set based on the characteristics of power grid renewable energy equipment operation data are as follows: The knowledge rule hierarchy is designed based on fault scenarios of varying complexity; 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 grid new energy equipment operation data. 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 compound faults.

4. The knowledge-enhanced power grid renewable energy operation fault identification method as described in claim 1, characterized in that, The specific steps for anomaly detection using the trained fault diagnosis model on graph data representation are as follows: The trained fault diagnosis model is used to process graph data representations and generate low-dimensional node embedding representations. Anomaly scores for low-dimensional embedding representations of nodes are calculated using a restricted Boltzmann machine via an energy function. The fault identification results of the wind farm are obtained based on the abnormal scores.

5. The knowledge-enhanced power grid renewable energy operation fault identification method as described in claim 4, characterized in that, The specific steps for obtaining wind farm fault identification results based on anomaly scores are as follows: An adaptive dynamic anomaly detection threshold is set based on historical detection results; The fault identification result of the wind farm is obtained by comparing the anomaly score with the dynamic threshold.

6. A knowledge-enhanced power grid renewable energy operation fault identification system, characterized in that, include: The data acquisition module is configured to acquire time-series data of the operation of new energy equipment in the power grid, and construct a graph structure based on the time-series data of the operation of new energy equipment in the power grid to obtain a graph data representation; The rule base building module is configured to encode expert knowledge into hierarchical rule sets based on the characteristics of the power grid new energy equipment operation data, and to build a knowledge rule base based on the hierarchical rule sets; The model building module is configured to build a fault diagnosis model using a graph neural network with a knowledge enhancement mechanism, using a 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. An improved teacher-student training scheme is adopted, in which knowledge rules are used as teachers to evaluate the rationality of the output node state of GNN and guide the training process accordingly. Generate knowledge signals: Transform knowledge rules into discriminant functions that can be applied to node state vectors; each rule R corresponds to a function. Its output value represents the node state. The degree to which the rule is met; Design the loss function: Total Loss Used in the initial training phase of the model; the loss function consists of two parts: data reconstruction loss ensures the model can learn the normal operating patterns in SCADA data, while knowledge consistency loss embeds expert-summarized physical rules and prior fault knowledge into the model as soft constraints; it directly optimizes the weight parameters of the graph neural network itself, which is the foundation for building model capabilities; specifically, the total loss function It consists of data reconstruction loss and knowledge consistency loss: Where λ is a hyperparameter used to balance the weights of the two losses; The reconstruction error is based on self-supervised learning of the input data. Defined as the difference between the confidence score and the expected value obtained after applying all knowledge rules to the embedding H of the GNN output node, using binary cross-entropy loss: ； The anomaly detection module is configured to use the trained fault diagnosis model to perform anomaly detection on the graph data representation and obtain the fault judgment result of the wind farm.

7. The knowledge-enhanced power grid new energy operation fault identification system as described in claim 6, characterized in that, The data acquisition module is also configured as follows: Node characteristics are designed based on the physical characteristics of the operation of new energy equipment in the power grid; Based on node features, a directed graph is constructed using the sliding window method to obtain the adjacency matrix. Specifically, the time series data of the operation of new energy equipment in the power grid is abstracted into a graph structure, where nodes represent the state of the equipment at critical moments and edges represent the causal dependencies in the time series. The node features are fused based on the attention mechanism to obtain a dimensionality-reduced feature matrix.

8. The knowledge-enhanced power grid renewable energy operation fault identification system as described in claim 6, characterized in that, The rule base creation module is also configured as follows: The knowledge rule hierarchy is designed based on fault scenarios of varying complexity; 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 grid new energy equipment operation data. 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 compound faults.

9. The knowledge-enhanced power grid renewable energy operation fault identification system as described in claim 6, characterized in that, The model building module is also configured as follows: The trained fault diagnosis model is used to process graph data representations and generate low-dimensional node embedding representations. Anomaly scores for low-dimensional embedding representations of nodes are calculated using a restricted Boltzmann machine via an energy function. The fault identification results of the wind farm are obtained based on the abnormal scores.

10. The knowledge-enhanced power grid renewable energy operation fault identification system as described in claim 9, characterized in that, The anomaly detection module is also configured as follows: An adaptive dynamic anomaly detection threshold is set based on historical detection results; The fault identification result of the wind farm is obtained by comparing the anomaly score with the dynamic threshold.

Images