Power grid equipment fault prediction and maintenance scheme generation method

Abstract

The invention provides a power grid equipment fault prediction and maintenance scheme generation method. The method comprises the following steps: acquiring a data set; standardizing the data set to obtain a structured data set; based on the structured data set, defining a quintuple relation structure, and establishing a dynamic knowledge graph; constructing an encoder group; based on the data set and the dynamic knowledge graph, fusing the time sequence feature, the text feature and the graph feature by using an encoder group to obtain a fused feature; based on the fusion feature, obtaining a fault classification probability and a decomposition cost item; according to the fault classification probability and the decomposition cost item, determining a corresponding measure set and a feasible time window set in the dynamic knowledge graph to obtain a search space; and based on the search space, determining a target maintenance scheme so as to solve the problem that when an operation maintenance scheme is formulated by a power grid enterprise at present, systematic consideration on the equipment fault probability, the disposal cost and the resource configuration is often lacked, so that the scheme selection cost is high or execution conflicts exist.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid equipment operation and maintenance, and particularly relates to a power grid equipment fault prediction and maintenance scheme generation method. BACKGROUND

[0002] At present, the power industry is booming, the scale of power grid is continuously expanding, the number of power grid equipment is rapidly increasing, and the structure is increasingly complex. Its operating environment covers different geographical regions and climate conditions, which changes extremely diversely. In this complex scenario, there are various types of power grid equipment faults, and the causes are extremely hidden, which brings great challenges to the safe and stable operation of the power grid. Power grid enterprises urgently need an efficient, accurate and dynamic fault diagnosis and operation decision-making method to ensure the reliability of power supply, reduce economic losses caused by equipment failure, reasonably allocate operation and maintenance resources, and balance operation and maintenance costs and benefits.

[0003] To meet the above needs, the traditional method mainly adopts a fault diagnosis method based on a rule base or expert experience. The rule base is to summarize a large number of known fault phenomena and causes, and to develop a series of fixed rules to judge the equipment fault; the expert experience is to rely on the knowledge and experience of senior experts in the field to manually analyze and diagnose the equipment fault. At the same time, some studies use a single model-based data-driven method to predict faults by analyzing equipment operation data. In terms of operation decision-making, some studies focus on single-target scheduling, such as only considering the shortest equipment maintenance time or the lowest maintenance cost; some studies use a simple cost minimization strategy to develop an operation and maintenance scheme.

[0004] However, the traditional fault diagnosis method based on the rule base or expert experience has obvious drawbacks, has weak generalization ability, is difficult to cope with new emerging fault types not covered in the rule base or experience, and has high maintenance cost with the update of equipment and changes in rules, and cannot adapt to the dynamic changes of equipment operating state. The single model-based data-driven method has certain effect in local scenarios, but performs poorly in multi-source heterogeneous data fusion, and is difficult to effectively integrate data from different equipment and different monitoring systems. In terms of operation decision-making, the current part of the study only focuses on single-target scheduling or simple cost minimization, ignores the multi-objective trade-off demand in the power system, such as the complex relationship between fault probability, disposal cost and resource allocation, and does not fully consider the importance of knowledge structure in decision-making, resulting in problems such as high cost of scheme selection or execution conflict. SUMMARY

[0005] The present application provides a power grid equipment fault prediction and maintenance scheme generation method to solve the technical problem that the existing power grid enterprises often lack systematic consideration of equipment fault probability, disposal cost and resource allocation when developing operation and maintenance schemes, resulting in high cost of scheme selection or execution conflict.

[0006] The application provides a power grid equipment fault prediction and maintenance scheme generation method, comprising:

[0007] Obtaining a data set; the data set comprises operation data, state alarm information and maintenance cost information of power grid equipment;

[0008] Standardizing the data set to obtain a structured data set;

[0009] Based on the structured data set, defining a five-tuple relationship structure and establishing a dynamic knowledge graph; the five-tuple comprises an entity node, a relationship set, a weight matrix, a threshold parameter and a mapping rule;

[0010] Constructing an encoder group; the encoder group comprises a time sequence, a text and a graph encoder;

[0011] Based on the data set and the dynamic knowledge graph, the encoder group is used to obtain time sequence features, text features and graph features;

[0012] The time sequence features, text features and graph features are fused to obtain fused features;

[0013] Based on the fused features, fault classification probabilities and decomposition cost items are obtained; the decomposition cost items comprise materials, labor and power loss;

[0014] According to the fault classification probabilities and the decomposition cost items, corresponding measure sets and feasible time window sets are determined in the dynamic knowledge graph to obtain a search space;

[0015] Based on the search space, a Tchebycheff function is used to determine a target maintenance scheme.

[0016] In some embodiments, the step of standardizing the data set to obtain a structured data set comprises:

[0017] Based on the data set, a standardization algorithm is used to unify the dimensions of different features in the data set;

[0018] Based on the data set, linear interpolation, seasonal quantile backfilling and adjacent backfilling are used for data backfilling processing;

[0019] Based on the data set, the median, MAD robust Z-score and isolated forest anomaly score of the measured data in the data set are used to delete outliers in the measured data;

[0020] The text data in the data set is converted into a calculable vector;

[0021] By aligning the computable vectors with the entity vectors of the knowledge graph, triples and triple confidence are determined to obtain a structured dataset; the triples include: power grid equipment, faults and corresponding measures, and numerical values ​​and standards.

[0022] In some embodiments, the step of converting text data in the dataset into computable vectors includes:

[0023] The text data in the dataset is preprocessed; the preprocessing operations include: text cleaning, segmentation, word segmentation, and annotation.

[0024] The semantics of text fragments in the text data are transformed into computable vectors.

[0025] In some embodiments, the step of determining triples and triple confidence by aligning the computable vector with the knowledge graph entity vector includes:

[0026] By using contrastive learning and triplet loss to optimize model parameters, the semantically relevant text vectors are made closer to the text data, while the semantically unrelated text vectors are made further away from the text data.

[0027] Construct entity vectors for knowledge graphs;

[0028] Entity vectors related to the text fragment are filtered out using the entity vectors in the knowledge graph.

[0029] Calculate the similarity between the computable vector and the entity vector;

[0030] The similarity is fused with semantic evidence to obtain the alignment result;

[0031] Based on the alignment results, triples are extracted and triple confidence scores are calculated.

[0032] In some embodiments, the step of defining a quintuple relation structure and establishing a dynamic knowledge graph based on the structured dataset includes:

[0033] Define the quintuple relation structure and establish an initial knowledge graph;

[0034] Define a joint anomaly event; the joint anomaly event is used to dynamically adjust the weight matrix of the initial knowledge graph when the joint anomaly event is triggered, so that the state of the initial knowledge graph is synchronized with the data corresponding to the joint anomaly event.

[0035] After the initial knowledge graph obtains updated triples, the updated triples are aligned with the initial knowledge graph and an alignment score is calculated; the updated triples were not included in the initial knowledge graph.

[0036] When the alignment score is greater than the score threshold, the updated triple is included in the initial knowledge graph;

[0037] The maintenance cost information is incorporated into the initial knowledge graph using the exponentially weighted moving average method to establish a dynamic knowledge graph.

[0038] In some embodiments, the step of fusing the temporal features, text features, and graph features to obtain fused features includes:

[0039] By using a gating mechanism, the temporal features, text features, and graph features are fused to obtain fused features.

[0040] In some embodiments, after obtaining the fault classification probability and decomposing the cost item based on the fusion features, the method further includes:

[0041] The cross-entropy loss of the fault classification probability, the MAE loss of cost prediction, the graph consistency loss, and the graph Laplacian regularization term are fused using a joint loss function, and the proportion of each loss component is adjusted by weighting.

[0042] In some embodiments, the step of determining the target maintenance plan based on the search space using the Tchebycheff function includes:

[0043] Using total maintenance cost, planned downtime, and fault handling risk score as multiple objective functions, a Deb feasibility rule is introduced to determine a set of maintenance schemes from the search space.

[0044] Based on the set of maintenance plans, the target maintenance plan is determined using the Tchebycheff function.

[0045] In some embodiments, the method further includes:

[0046] Obtain actual maintenance information; the actual maintenance information includes: the actual fault type of the faulty equipment, the actual cost, the handling results, and the equipment recovery rate.

[0047] A deviation analysis is performed on the target maintenance plan and the actual maintenance information to obtain deviation indicators; the deviation indicators include: quantitative cost deviation, classification deviation, and recovery deviation.

[0048] The weights of the dynamic knowledge graph are adjusted based on the aforementioned deviation index.

[0049] In some embodiments, the step of adjusting the weights of the dynamic knowledge graph based on the deviation index includes:

[0050] If the deviation index is greater than the preset deviation threshold, then the threshold parameters, weights, and cost range of the dynamic knowledge graph are updated.

[0051] This application provides a method for predicting power grid equipment faults and generating maintenance plans, including: acquiring a dataset; the dataset includes: operating data, status alarm information, and maintenance cost information of the power grid equipment; standardizing the dataset to obtain a structured dataset; defining a quintuple relation structure based on the structured dataset and establishing a dynamic knowledge graph; the quintuple includes: entity nodes, relation sets, weight matrices, threshold parameters, and mapping rules; constructing an encoder group; the encoder group includes: time-series, text, and graph encoders; based on the dataset and the dynamic knowledge graph, using the encoder group, acquiring time-series features, text features, and graph features; and then... The temporal features, textual features, and graph features are fused to obtain fused features. Based on the fused features, the fault classification probability and decomposed cost items are obtained. The decomposed cost items include: materials, labor, and power outage losses. According to the fault classification probability and the decomposed cost items, the corresponding set of measures and feasible time windows are determined in the dynamic knowledge graph to obtain the search space. Based on the search space, the Tchebycheff function is used to determine the target maintenance plan, so that when power grid companies formulate operation and maintenance plans, they can systematically consider equipment failure probability, handling costs, and resource allocation, thereby reducing the cost of the plan and avoiding the occurrence of execution conflicts. Attached Figure Description

[0052] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the method for generating power grid equipment fault prediction and maintenance plans in this application. Detailed Implementation

[0054] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in 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 in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0055] In some technologies, power grid companies often lack a systematic consideration of equipment failure probability, handling costs, and resource allocation when formulating operation and maintenance plans, leading to excessively high costs or execution conflicts in plan selection. To address this technical problem, this application provides a method for predicting power grid equipment failures and generating maintenance plans. The method for predicting power grid equipment failures and generating maintenance plans is described below:

[0056] This application provides a method for predicting power grid equipment faults and generating maintenance plans, including the following steps:

[0057] First, multi-source data acquisition and consistency processing are performed. Dynamic information such as equipment operation data, status alarms, and maintenance costs are collected from the SCADA system, monitoring terminals, and financial system, and robust dual-channel cleaning is used to remove missing and outlier values. Isolated anomalies are eliminated using Z-scores based on median and MAD, while the isolated forest algorithm is used to identify multidimensional complex anomalies, thereby obtaining high-confidence time series data. For text data, a semantic encoder is used for word segmentation, embedding, and entity recognition, ultimately forming a structured sample library.

[0058] Specifically, SCADA real-time / historical data, protection alarms, maintenance texts, and ERP cost details are aligned as a unified input. Through standardization, backfilling of missing measurements, and anomaly removal, the offset caused by measurement noise and inconsistent sampling is reduced. At the same time, maintenance texts are encoded into semantic vectors, and costs are broken down into three items: materials, labor, and power outage losses. Physical consistency calculations are performed using the system load curve.

[0059] S100: Obtain the dataset; the dataset includes: power grid equipment operation data, status alarm information, and maintenance cost information.

[0060] S200: Standardize the dataset to obtain a structured dataset.

[0061] The step of standardizing the dataset to obtain a structured dataset includes the following sub-steps:

[0062] S210: Based on the dataset, use a standardization algorithm to unify the dimensions of different features in the dataset.

[0063] Specifically, to eliminate the dimensional differences between different features and enable the model to fairly learn the importance of each feature, time-series data needs to be standardized. Simultaneously, to capture the temporal correlation of device states, a fixed-length sliding window sequence needs to be constructed. This patent employs the Z-score standardization algorithm to eliminate the influence of dimensional differences, making different features comparable.

[0064]

[0065] Where, xi,t μ is the measurement value of the i-th original feature at time t. i σ is the mean of feature i. i Let be the standard deviation of feature i. These are the standardized feature values; the source / calculation is obtained from the previous formula. X e Let L be the input window tensor of device e. Let L be the length of the time window. Let d be the feature dimension.

[0066] Some of the data is shown in the table below:

[0067]

[0068] S220: Based on the dataset, perform data backfilling using linear interpolation, seasonal quantile backfilling, and nearest neighbor backfilling.

[0069] Specifically, since SCADA systems may experience data loss due to communication interruptions or sensor malfunctions, directly affecting the integrity of model input, a combination of multiple strategies is required for backfilling: linear interpolation utilizes time continuity, seasonal quantile backfilling considers periodicity, and proximity backfilling references the status of similar equipment. By using dynamic weight fusion, these three methods can ensure data rationality in different data loss scenarios.

[0070]

[0071] in, The characteristic value after backfilling; m i,t For observation indication (value present = 1, missing / abnormal = 0); This is a linear interpolation backfill value; This refers to the seasonal phase quantile backfill value; β1, β2, and β3 are the spatial proximity backfill values; β1, β2, and β3 are the weights of the three backfill strategies.

[0072] To further improve the accuracy of adjacent backfilling, weights should be assigned based on the geographical or electrical distance between equipment. The closer the equipment is, the higher its status relevance and the greater its weight, thus avoiding interference from irrelevant equipment.

[0073]

[0074] Where N(e) is the set of neighboring devices of device e; Let d be the i-th feature value of the neighboring device j. e,j Let w be the geographical / electrical distance between devices e and j. e,j The weights are for neighboring devices. K(r) is the distance kernel function. ε is a small positive number that is excluded from zero.

[0075] S230: Based on the dataset, use the median of the measurement data in the dataset, the robust Z-score of MAD, and the outlier score of the isolated forest to remove outliers from the measurement data.

[0076] Specifically, measurement data may contain outliers such as sensor false alarms and interference pulses. Directly inputting these into the model can lead to prediction bias. Therefore, it is necessary to remove outliers through dual-channel detection: robust Z-scores based on median and MAD can resist extreme value interference and are suitable for detecting isolated outliers with single features; isolated forests identify complex outliers through multi-dimensional spatial density differences. Combining the two can improve the outlier recall rate.

[0077]

[0078] Among them, z it The standardized residuals are denoted by MAD. `median(·)` is the median operator. `MAD(·)` is the absolute median difference. it This indicates an anomaly (1 for an anomaly). τ z The threshold is Z / MAD. τ represents the outlier score for isolated forests. iso is the IFORS threshold. I(·) is the indicator function.

[0079] S240: Convert the text data in the dataset into computable vectors. Text data such as maintenance reports and fault work orders contain rich experiential knowledge and need to be converted into machine-understandable vector forms. By mapping the text to a high-dimensional semantic space through a pre-trained semantic encoder, and then calculating the cosine similarity with the knowledge graph entities, the alignment of text information with knowledge graph knowledge can be achieved. In this step, unstructured text such as maintenance reports, fault work orders, and inspection records are converted into high-dimensional vectors (computable vectors), aligned with knowledge graph entities, and triple candidates and confidence scores are generated for dynamic knowledge rewriting and subsequent modeling.

[0080] The step of converting the text data in the dataset into computable vectors includes the following sub-steps:

[0081] S241: Perform preprocessing operations on the text data in the dataset; the preprocessing operations include: text cleaning, segmentation, word segmentation, and annotation operations.

[0082] Specifically, text data often suffers from encoding chaos and non-standard formatting, such as the mixing of units of measurement "megapascal" and "MPa", and semantic fragmentation in long texts. Standardization needs to be achieved through preprocessing and segmentation to provide high-quality input for subsequent semantic encoding.

[0083] Cleaning: unify coding, remove control characters, and standardize units of measurement.

[0084] Segmentation: Segment by Chinese punctuation and common engineering separators, window length M = 1 to 3 sentences, sliding window step size 1.

[0085] Word segmentation / labeling: To assist in subsequent entity recognition and numerical extraction, a domain-specific word segmentation tool is used to segment the fragments, and potential entity boundaries are pre-labeled using the BIO labeling method.

[0086] Output: Sentence / fragment sequence tokens and optional BIO tag / value pairs.

[0087] S242: Convert the semantics of the text fragments in the text data into computable vectors.

[0088] Specifically, in order to transform the semantics of text fragments into computable vector representations, a domain-adaptive semantic encoder is required to achieve accurate encoding of technical terms and ensure that semantically similar text fragment vectors are closer together.

[0089] Encoder: Employs an SBERT model fine-tuned based on a corpus from the power industry domain. Compared to general pre-trained models, the domain-fine-tuned model can better capture subtle differences in specialized semantics such as insulation aging and excessive oil dielectric loss.

[0090] For the preprocessed text segments, the word-level hidden state sequence output by the encoder is aggregated into sentence-level semantic vectors using mean pooling:

[0091] z = f txt (tokens)=MeanPool(H), H=[h1,…,h T ]∈R T×h ;

[0092] Where z is the semantic vector of the text segment. H is the sequence of hidden states of the Transformer. T is the length of the sub-word sequence. h is the dimension of the hidden layer. tokens are the preprocessed word segmentation sequence.

[0093] S250: Align the computable vector with the entity vector of the knowledge graph to determine the triples and triple confidence, and obtain the structured dataset; the triples include: power grid equipment, faults and corresponding measures, values ​​and standards.

[0094] The step of determining triples and triple confidence by aligning the computable vector with the entity vector of the knowledge graph includes the following sub-steps:

[0095] S251: Optimize model parameters using contrastive learning and triplet loss to make the semantically relevant text vectors closer to the text data, and the semantically unrelated text vectors more distant from the text data.

[0096] Specifically, in order to further improve the encoder's semantic understanding of text in the power operation and maintenance field, it is necessary to fine-tune based on the domain corpus, optimize the model parameters through comparative learning and triple loss, so that semantically related text vectors are closer and irrelevant text vectors are more distant.

[0097] To address the requirement that semantically similar text fragments should have similar vectors, InfoNCE loss is used to optimize the discriminative power between positive and negative pairs:

[0098]

[0099] To reinforce the constraint that the distance between the anchor point and the positive sample is less than the distance between the anchor point and the negative sample, a triplet loss is introduced:

[0100] L tri =max(0, m+d(z) a ,z p )-d(z a ,z n )).

[0101] By fusing contrastive learning with triple loss, semantic similarity and distance constraints can be balanced.

[0102] L txt =λ con L con +λ tri L tri ;

[0103] in, These are pairs of words with similar meanings. j λ represents the negative sample vector. sim(·) represents the cosine similarity. τ represents the temperature coefficient. d(·,·) represents the distance. m represents the Triplet interval proposal. con ,λ tri For joint weights.

[0104] Training recommendations: lr = 2e-5 to 5e-5, batch = 64, epochs = 3–10, max_len = 128 / 256, mix precision and gradient accumulation for stable convergence.

[0105] S252: Construct entity vectors for a knowledge graph.

[0106] Specifically, to achieve semantic alignment between text fragments and knowledge graph entities, a prototype vector needs to be constructed for each entity in the graph. This prototype vector serves as a standard representation of the entity's semantics and is used to calculate similarity with the text vector. For entity k in the knowledge graph, the semantic vectors of its name, synonyms, description, and relevant standard clauses are aggregated, and a prototype vector is generated by weighted averaging using IDF.

[0107]

[0108] Among them, e k Let v(·) be the vector of entity k. v(·) is the output vector of the text encoder. syn k A collection of synonyms / aliases. (desc) k This is a fragment describing an entity. u The weighting factor is (IDF).

[0109] S253: Use the knowledge graph entity vectors to filter out entity vectors related to the text fragment.

[0110] Specifically, in order to efficiently filter out entities related to text fragments from the knowledge graph, a two-stage strategy of "initial screening + fine ranking" is adopted to balance recall and precision.

[0111] Based on the word frequency features of text fragments and entity descriptions, the BM25 algorithm is used to quickly recall potential relevant entities. The initial BM25 screening involves using sentence / window and entity description inverted index retrieval to obtain a candidate set C. k .

[0112] The candidate set obtained from the initial screening is further sorted using semantic vector similarity or a cross-encoder:

[0113]

[0114] Among them, Cross-Encoder probabilities (SCE) can be superimposed for fine-grained fusion. Parameter recommendations: TopK = 20~100; BM25 is used for high recall, and vector / cross-coding is used for high precision.

[0115] S254: Calculate the similarity between the computable vector and the entity vector. S255: Fuse the similarity with semantic evidence to obtain the alignment result.

[0116] Specifically, a single similarity metric may be affected by differences in textual expression, leading to alignment deviations. Therefore, it is necessary to integrate multiple semantic evidences to calculate a comprehensive alignment score, thereby improving the reliability of entity alignment. The integrated score is as follows:

[0117] score(u,k)=λ cos ·sim(z u ,e k )+λ ce ·s ce (u,k)+λ sp ·r uk ;

[0118] The threshold and write-back are:

[0119] match(u,k)=I(score(u,k)≥γ), γ∈[0,1];

[0120] Where u is the text segment index, and sim(·) is the cosine similarity. ce (u,k) represents the matching probability of the Cross-Encoder, r uk The prior on the graph; the source / calculation is to calculate the shortest path using A and normalize it. λ cos , λ ce , λ sp γ is the weighting coefficient. γ is the write-back threshold.

[0121] S256: Based on the alignment results, extract triples and calculate triple confidence.

[0122] Specifically, the alignment scores output by the model may exhibit distributional shifts, and the same text fragment may match multiple semantically similar entities. Therefore, confidence calibration and non-maximum suppression are needed to optimize the reliability and uniqueness of the results. Temperature calibration or:

[0123] or

[0124] Where a and b are Platt calibration parameters; source / calculation is the minimum log loss fit on the validation set. T is the temperature; source / calculation is the cross-validation selection.

[0125] Among them, non-maximum suppression (NMS) is adopted: if multiple entities belong to the same superclass and overlap, the one with the highest s is retained to avoid repeated write-back.

[0126] Specifically, by aligning text with entities, structured entity-relation triples are extracted, and numerical information in the text is anchored to the corresponding feature entities, thus supplementing the knowledge graph with incremental relations and attributes.

[0127] Triple generation: Text-to-entity conversion: (equipment / fault / measure / standard / cost item) is determined by match(u,k)=1. Text-to-value conversion: Regular expression + unit table identifies (index name, value, unit); the value is anchored to the corresponding feature entity.

[0128] Confidence level (numerical standard for equipment failure measures):

[0129]

[0130] Where, p triple For triple confidence; source / calculation is the minimum / geometric mean of the relevant edge matching confidence. When p triple ≥γ kgIf the result is 0.75, the triple is written into step S300, along with the source field. This step, through text semantic parsing and entity alignment, achieves the transformation from unstructured text to structured knowledge, providing high-quality incremental data for the dynamic updating of the knowledge graph and supporting intelligent reasoning in subsequent operation and maintenance solutions.

[0131] The method further includes:

[0132] Operation and maintenance costs need to be broken down into physically interpretable components. Among these, power outage losses are directly related to outage duration and system load, and need to be calculated based on load curve integration: the higher the load during a period, the greater the outage loss. Therefore, it is necessary to quantify the power outage cost for different time windows through discrete summation to provide a basis for subsequent scheme optimization. To accurately calculate the economic loss of power outages at different times, an integral model is constructed based on the system load curve and unit loss coefficient.

[0133]

[0134] Among them, c mat For material costs, c lab For labor costs. out Costs incurred due to power outages. sys (t u λ represents the system load power at the sampling time. pen t0 is the outage loss factor per unit power (yuan / kWh). t0 is the outage start time. τ is the planned outage duration. Δt is the discrete time step. U is the number of integration steps.

[0135] S300: Based on the structured dataset, a quintuple relationship structure is defined to establish a dynamic knowledge graph. The quintuple includes: entity nodes, relationship sets, weight matrices, threshold parameters, and mapping rules. An initial knowledge graph is then established based on the structured dataset. Subsequently, a time-driven mechanism is introduced to automatically trigger attribute and threshold updates for corresponding nodes, keeping the graph state synchronized with the field data. Semantic matching is performed between text descriptions and entity relationships. Semantic similarity is calculated using SBERT encoding, and graph edge weights are updated to achieve dynamic fusion of unstructured knowledge. Finally, operating costs, the effectiveness of treatment measures, and equipment health status are written back into the graph nodes to achieve closed-loop knowledge updates.

[0136] Specifically, by constructing a core dynamic knowledge graph, the graph structure G serves as the central hub for structural priors and consistency constraints. The parameterized evolution of the threshold set triggered by joint abnormal events and the "fault-feature" edge weights triggers adaptive updates of the threshold parameters. New operational knowledge is aligned based on textual semantic similarity, and the cost range is dynamically corrected using an exponentially weighted moving average method.

[0137] The steps of defining a quintuple relation structure and establishing a dynamic knowledge graph based on the structured dataset include the following sub-steps:

[0138] S310: Define the quintuple relation structure and establish the initial knowledge graph.

[0139] Specifically, to clearly depict the entity relationships in the field of power grid equipment operation and maintenance, it is necessary to mathematically define the core components of the graph, including entity nodes, relationship sets, weight matrices, threshold parameters, and mapping rules, so that the graph can store both static knowledge and dynamic parameters.

[0140]

[0141] Where V is the set of nodes (equipment, faults, characteristics, measures, costs); Let A be the set of relations; let O be the set of threshold parameters; and let Π be the set of mapping rules.

[0142] S320: Define a joint anomaly event; the joint anomaly event is used to dynamically adjust the weight matrix of the initial knowledge graph when the joint anomaly event is established, so that the state of the initial knowledge graph is synchronized with the data corresponding to the joint anomaly event.

[0143] Specifically, a single feature anomaly may be a normal fluctuation, while a combined anomaly of multiple features is more likely to indicate a fault. Therefore, a combined anomaly event needs to be defined as the trigger condition for map updates. When an anomaly is triggered, the feature threshold and fault-feature relationship weights are dynamically adjusted to make the map more closely reflect the current state of the equipment.

[0144]

[0145] Among them, E t For joint anomaly event determination. T oil,t This refers to the oil temperature. To accelerate oil temperature rise. H t θ represents relative humidity. T θ r θ H This corresponds to the threshold. α is the threshold smoothing coefficient. Δ T This is the rollback amount. F,feat Let η be the fault-feature edge. η is the edge weight learning rate. δ is the direction and magnitude of the most recently disposed residual.

[0146] S330: After the initial knowledge graph obtains the updated triplet, the updated triplet is aligned with the initial knowledge graph and an alignment score is calculated; the updated triplet is not included in the initial knowledge graph; S340: When the alignment score is greater than the score threshold, the updated triplet is included in the initial knowledge graph.

[0147] Specifically, maintenance texts may contain new knowledge not yet included in the knowledge graph. This new knowledge often exists in text form and is not included in the existing knowledge graph, causing the graph's knowledge to lag behind actual operational needs. To integrate this new knowledge into the graph, the text fragments need to be aligned with the graph entities. An alignment score is calculated by combining the semantic similarity between the text and entities and the relevance of the shortest path in the graph. When the score exceeds a threshold, the new triple is written into the graph, achieving automatic knowledge completion.

[0148] score(u,v)=λ cos ·sim(z u ,e v )+λ sp ·r uv ;

[0149] Where score(u,v) is the overall matching score λ between the text fragment and the entity. cos , λ sp These are the two weighting coefficients. u e is the vector of the text segment u. v r is a vector for entity v. uv The shortest path relevance is γ (0-1). The write-back rule is to write a new triple and update Π when the score ≥ γ. γ is the write-back threshold.

[0150] S350: Using the exponentially weighted moving average method, the maintenance cost information is included in the initial knowledge graph to establish a dynamic knowledge graph.

[0151] Specifically, the corresponding maintenance cost range for a fault will change with market prices and labor rates, and needs to be dynamically updated based on the most recent actual costs. Using an exponentially weighted moving average method retains historical experience while incorporating new data, making the cost range more closely reflect current realities and providing constraints for subsequent cost forecasts.

[0152] [l y ,u y ]=β[l y ,u y ]+(1-β)[min(C y ),max(C y )).

[0153] In graph inference, the embedding features of nodes need to be consistent with the graph structure. For example, insulation aging and excessive oil dielectric loss are strongly correlated nodes, and their embedding features should be similar; while insulation aging and mechanical jamming are weakly correlated nodes, and their features should differ significantly. Without structural constraints, the node embeddings generated by the model may deviate from the graph relationships. For instance, generating features for mechanical jamming that are similar to those for insulation aging could lead to inference errors. Therefore, a graph Laplacian matrix is ​​introduced as a regularization term. By using the difference between the degree matrix and the adjacency matrix, the structural relationships between nodes are quantified, constraining the distribution of node embeddings and ensuring that nodes with similar structures have similar features, thus guaranteeing the consistency of graph inference.

[0154] L = DA, D = diag(∑ j A ij );

[0155] Among them, [l y ,u y [ ] represents the cost range for fault category y. C y This represents the total actual cost set for this type of failure in the near future. β is the smoothing coefficient. L is the graph Laplace matrix. D is the degree matrix.

[0156] For example, a collaborative fusion of three models—LSTM, semantic encoder, and graph neural network—is used to simultaneously predict equipment failure probability distribution and maintenance costs. The LSTM model extracts evolutionary features from standardized time-series sequences, the SBERT encoder generates textual semantic feature vectors, and the graph encoder generates node embeddings based on the graph structure. A gating mechanism is used to fuse the three feature representations, enabling dynamic weight allocation. The output consists of a joint head module that outputs the failure classification probability and the decomposed cost term, respectively.

[0157] Specifically, by performing joint optimization of failure probability and cost, multiple models work together to jointly predict failure probability and operation and maintenance costs. This involves using LSTM to extract the temporal evolution features of equipment status, SBERT to transform the semantic features of the text, a graph encoder to capture structural features in the graph, and a gating mechanism to adaptively fuse these three types of features to generate a unified representation. Finally, graph consistency constraints and graph Laplacian regularization are added to the loss function to ensure that the prediction results conform to both data patterns and domain knowledge constraints.

[0158] S400: Construct an encoder group; the encoder group includes: time-series, text, and graph encoders; S500: Based on the dataset and the dynamic knowledge graph, use the encoder group to obtain time-series features, text features, and graph features.

[0159] Specifically, the operating status of equipment exhibits strong temporal correlation, requiring the use of Long Short-Term Memory (LSTM) networks to capture this dynamic characteristic. LSTM, through input gates, forget gates, and output gates, controls the flow of information, solving the gradient vanishing problem of traditional RNNs. It can effectively learn trend changes within long time windows and output feature vectors containing temporal evolution patterns.

[0160] i t =σ(W i x t +U i h t-1 +b i ),f t =σ(W f x t +U f h t-1 +b f )

[0161]

[0162] Where, x t Let W be the input vector at time t. * U * b * These are trainable parameters. σ(·) is the Sigmoid activation. tanh(·) is the hyperbolic tangent activation. t f t o t For input / forget / output gates. t ,h t This refers to the cellular state and the hidden state. seq This is a temporal representation.

[0163] Specifically, textual data such as maintenance reports and fault work orders contain a wealth of experiential information that cannot be obtained from time-series data. For example, severe sludge deposition was found during the last maintenance. This type of information is often an important precursor to a fault, but it needs to be converted into machine-understandable semantic features. Through a semantic encoder, the technical expressions in the text can be transformed into fixed-dimensional vectors, allowing the model to capture the fault correlations behind the text. At the same time, the "fault-feature-measure" relationships stored in the knowledge graph are domain prior knowledge, which can provide structural constraints for model reasoning and avoid overfitting where the model relies solely on the data. Through a graph encoder, the node relationships in the graph can be transformed into node embedded features, allowing the model to combine domain knowledge during reasoning, improving the accuracy and interpretability of predictions.

[0164] h txt =f txt (report e ), h kg =f kg (G,A);

[0165] Among them, report e This is the maintenance report text for equipment e. txt For semantic encoders. kg For graph encoders. h txt h kg This is the corresponding representation vector.

[0166] In graph encoders, GraphSAGE's core advantage lies in its ability to generate node embeddings by aggregating the features of a node's neighbors, without relying on the entire graph data. This makes it suitable for handling dynamically updated knowledge graphs. In power grid operation and maintenance graphs, the failure risk of a device node depends not only on its own attributes but also on its neighboring nodes, such as associated fault and feature nodes. GraphSAGE generates the current layer's embedding by averaging the node's own previous-layer embedding with the previous-layer embeddings of its neighbors, and then applying the result to an activation function. This captures the local structural information of nodes and allows newly added nodes to quickly generate embeddings through neighbor aggregation. This solves the problem of insufficient generalization ability of traditional graph embedding methods for new nodes, making graph encoders more adaptable to the needs of dynamic graphs.

[0167]

[0168] Retrieve the layer output of the device node:

[0169]

[0170] in, Let N(v) be the embedding vector of node v at layer k. Let N(v) represent the set of neighboring nodes of node v. Let K represent the layer number of the graph encoder.

[0171] S600: The time-series features, text features, and graph features are fused to obtain fused features.

[0172] The step of fusing the temporal features, text features, and graph features to obtain the fused features includes the following sub-steps:

[0173] S610: Using a gating mechanism, the temporal features, text features, and graph features are fused to obtain fused features.

[0174] Specifically, the importance of temporal features, textual features, and graph features varies across different fault scenarios: in sudden fault scenarios, temporal features most directly reflect the fault state and are of the highest importance; in rare fault scenarios, textual features are often the key basis and require close attention; in common fault scenarios, graph features provide stable prior constraints, helping to improve prediction accuracy. Simply concatenating the three types of features does not allow the model to dynamically adjust the weights of each feature according to the scenario, potentially leading to unimportant features interfering with the prediction results. Therefore, a gating mechanism is needed. A learnable gating vector adaptively allocates weights based on the characteristics of the input data—for example, in a scenario of sudden oil temperature rise, the gating vector assigns high weights to temporal features and low weights to textual and graph features; in rare fault scenarios, the gating vector increases the weight of textual features, thus making the fused features more closely match the current scenario and improving the model's adaptability and prediction accuracy.

[0175] g=σ(W g [h seq h txt h kg ]+b g ), h=g⊙h seq +(1-g)⊙(W2h txt +W3h kg );

[0176] Where [·; ·] represents vector concatenation operations; source / computation represents tensor operations; W g W2, W3, b g is the trainable parameter; source / calculation is obtained through training optimization; g is the gate vector (0-1); source / calculation is calculated using this formula; h is the fused joint representation; source / calculation is calculated using this formula.

[0177] S700: Based on the fusion features, obtain the fault classification probability and decomposed cost items; the decomposed cost items include: materials, labor, and power outage losses.

[0178] After obtaining the fault classification probability and decomposing the cost item based on the fused features, the method further includes the following steps:

[0179] S710: The cross-entropy loss of the fault classification probability, the MAE loss of cost prediction, the graph consistency loss, and the graph Laplacian regularization term are fused using a joint loss function, and the proportion of each loss component is adjusted by weighting.

[0180] Specifically, to simultaneously identify fault types and estimate costs, the model needs to design a joint output head: the classification head outputs the probability distribution of various fault types via softmax, and the regression head outputs three costs: materials, labor, and power outage losses. Power outage losses must be calculated based on the system load curve to ensure consistency with the physical meaning of actual power outage windows, avoiding cost predictions that deviate from the real-world scenario.

[0181]

[0182] Among them, g cls g reg Output headers for classification and regression. This represents the probability distribution of fault categories; the source / calculation is the softmax output. The costs are estimated to include three items: materials, labor, and power outage losses. To predict the length of the outage window, the remaining variables are the same as those in the cost decomposition and outage loss calculation steps.

[0183] The method further includes:

[0184] To ensure the prediction results align with domain knowledge, the joint loss function integrates the cross-entropy loss for fault classification, the MAE loss for cost prediction, the graph consistency loss, and the graph Laplacian regularization term, adjusting the importance of each component through weights. A graph consistency constraint needs to be added to the loss function: a penalty is applied when the predicted cost exceeds the defined range of the graph. Simultaneously, graph Laplacian regularization is introduced to constrain the consistency between the fused features and the graph structure. Ultimately, the joint loss function achieves simultaneous optimization of classification accuracy and cost prediction precision.

[0185]

[0186] Among them, l y u y The upper and lower bounds of the cost range for category y. The total cost is predicted. H is the embedding matrix of samples or nodes within a batch. L is the graph Laplacian matrix. y is the true fault label. λ1, λ2, and λ3 are the joint loss weights. CE / MAE is the cross-entropy / mean absolute error.

[0187] For example, the prediction results are linked to a knowledge graph structure, and multi-objective optimization is performed using the NSGA-II non-dominated ranking evolutionary algorithm to generate an executable and optimal maintenance plan. First, based on the predicted fault type, the corresponding set of measures and feasible time windows are retrieved from the knowledge graph to form a search space. Using total maintenance cost, planned downtime, and fault handling risk scores as multi-objective functions, Deb feasibility rules are introduced to handle practical constraints such as resource availability, standard compliance, and scheduling conflicts. After non-dominated ranking, the Pareto front is selected, and the optimal solution is chosen using the Tchebycheff function.

[0188] S800: Based on the fault classification probability and the decomposition cost item, determine the corresponding set of measures and the set of feasible time windows in the dynamic knowledge graph to obtain the search space.

[0189] Specifically, the multi-objective approach is transformed into a single-objective function. The comprehensive performance of each solution is quantified by weighted summation, with the weights dynamically adjusted according to the operation and maintenance strategy. Compliance and resource constraints are also incorporated to ultimately select the optimal and feasible solution.

[0190]

[0191] Resources available to satisfy mandatory provisions:

[0192] Std(r) satisfies the mandatory clause, Res(r) ≤ available resources;

[0193] Where r represents the candidate disposal measure. This is a forecast of the total cost of the project. The downtime duration of the plan. ρ(r) is the risk score. w c w τ w risk For multi-objective weights. R y Let be the set of available measures for category y. Std(r) is the set of applicable standard clauses for the solution. Res(r) is the resource requirement of the solution. ★ This is the optimal solution.

[0194] To avoid subjectivity in risk scoring, an auditable decomposition formula is provided: severity, probability of occurrence, equipment importance, and execution risk are linearly synthesized after quantification. The weights are learned by experts or a validation set to ensure interpretability and reproducibility.

[0195] ρ(r)=w S S y +w P P y +w K κ e +w M μ r ;

[0196] Among them, S y Severity; P y κ represents the probability of occurrence. e For equipment importance; μ r Risks associated with the implementation of measures; S w P w K w M The sum is 1.

[0197] Specifically, the selection of outage windows directly impacts outage losses and grid operational safety. For instance, a one-hour outage during peak daytime industrial electricity consumption can result in losses of up to 150,000 yuan, while a one-hour outage during off-peak hours in the early morning only incurs losses of 50,000 yuan. Therefore, it is necessary to optimize the outage window based on the system load curve. Simultaneously, the selection of outage windows must avoid other maintenance schedules. For example, if a substation has already scheduled line maintenance between 2-4 AM, selecting the same time period for transformer maintenance would lead to a complete substation outage. Therefore, it is necessary to detect outage window conflicts and impose penalties on conflicting windows. By constructing an objective function that includes outage losses and conflict penalties, the optimal outage window with the minimum loss and no conflicts can be selected from the alternative windows. If all outage windows conflict, the window with the minimum conflict penalty is selected, ensuring that the maintenance schedule minimizes losses while guaranteeing grid safety.

[0198]

[0199] Where W represents the set of alternative shutdown time windows. χ(W) m (This is a penalty for scheduling conflicts.) conf Conflict weights. This represents the power outage loss corresponding to the time window. W ★ To determine the optimal time window, when resources are scarce, the resource priority of each solution needs to be calculated: the solution with lower resource requirements and higher equipment urgency has higher priority, ensuring that limited resources are prioritized for critical equipment and efficient solutions, thereby improving overall operation and maintenance efficiency.

[0200] To ensure physical consistency in measuring power outage losses, integration or equal-interval discrete summation is performed on the selected time window based on the system load curve. Higher load periods result in greater outage losses; when electricity prices / compensation coefficients are uncertain, scenario expectations or risk measurements can be performed.

[0201]

[0202] Among them, P sys (t) represents the load curve; λ pen Δt is the unit power loss coefficient; Δt is the sampling interval.

[0203] When resources are scarce, a preliminary screening based on resource priority is used to prioritize key and resource-saving solutions into the feasible set. Then, a fine-tuning using a multi-objective function can significantly reduce the size of MILP and improve the feasibility rate.

[0204]

[0205] Where: req k (r) / avail k κ represents the resource requirements / availability of solution r; e For equipment importance; w K The weights are determined by the number of π(r). Schemes with higher π(r) are preferred.

[0206] S900: Based on the search space, the target maintenance plan is determined using the Tchebycheff function.

[0207] The step of determining the target maintenance plan based on the search space using the Tchebycheff function includes the following sub-steps:

[0208] S910: Using total maintenance cost, planned downtime, and fault handling risk score as multiple objective functions, introduce the Deb feasibility rule to determine the maintenance scheme set within the search space; S920: Based on the maintenance scheme set, use the Tchebycheff function to determine the target maintenance scheme.

[0209] Specifically, each individual x corresponds to a combination of measures and time windows (r, W). To maintain comparability, the three objectives are directly derived from business metrics: total cost includes materials, labor, and power outage losses; duration is derived from standard operating procedures (SOPs); and risk scores are calculated using the auditable decomposition method. NSGA-II performs non-dominated ranking in the unweighted objective space to avoid artificial weighting masking the Pareto structure.

[0210]

[0211] Among them, ERP / model output (materials, labor) and EMS load integral (power outage loss). Source: Standard Operating Procedure (SOP) working hours × Condition Correction Factor. Risk score ρ(r) is calculated as follows; R y W and W represent the set of measures for map retrieval and the set of alternative time windows for peak avoidance / scheduling generation, respectively.

[0212] For example, to avoid subjectivity in risk assessment, the risk score is decomposed into four auditable indicators, which are then normalized and weighted. The weights can be set by experts or learned from historical data, and can be configured according to different scenarios.

[0213] ρ(r)=w S S y +w P P y +w K K e +w M μ r w S +w P +w K +w M =1;

[0214] Among them, S y The severity of the fault; P y K represents the probability of occurrence. e For equipment importance; μ rTo mitigate execution risks.

[0215] For example, NSGA-II natively handles unconstrained problems. Here, the Deb feasibility rule prioritizes feasible solutions, while defining a total constraint violation quantity φ(x) to rank them in peer comparisons and use it in the penalty function. Three types of constraints—standard compliance, resource availability, and scheduling conflicts—are all quantified into φ(x), ensuring both project executability and preserving search diversity. Standard resource conflicts:

[0216]

[0217] Where std_ok(r)∈{0,1} is the clause check; req k avail k χ(W)∈{0,1} is the time window and the overlap of work orders / resources under construction.

[0218] The Deb rules include: feasible takes precedence over infeasible; the two feasible ratios are Pareto dominant; and the two infeasible ratios are φ(x). The alternative is a penalized objective F. i (x)=f i (x)+λφ(x), where λ is a large constant.

[0219] For example, power outage losses are obtained by integrating the system load curve and loss coefficient within a candidate time window. The engineering implementation uses discrete summation, and the load can be taken from historical or predicted curves and can be extended with scenario expectations or CVaR risk, thereby explicitly incorporating uncertainty into the target assessment.

[0220]

[0221] Among them, P sys (t u ) represents EMS / dispatch load; λ pen Δt is the compensation / electricity price coefficient; Δt is the sampling interval.

[0222] For example, NSGA-II constructs hierarchies using Pareto dominance and then uses crowding distance to characterize the density of solutions in each layer, prioritizing the preservation of boundaries and sparse regions to prevent premature convergence. The following formula gives the standard calculation of crowding distance, summed over the normalized intervals of each objective. Dominance is defined as when x1 < x2. and f j (x1) <f j (x2). The boundary individuals are:

[0223] Boundary individual CD = ∞.

[0224] For example, since the decision involves discrete measures and discrete time windows, a classification encoding and discrete operators are used. Selection is done using a tournament; crossover is performed on each site using uniform crossover; mutation is achieved through random replacement of neighboring candidate sets; the repair operator automatically replaces the default with the nearest feasible time window / low-resource measure, significantly improving the feasibility rate. The encoding is for the individual x = (r, W), selection is a binary tournament; uniform crossover is:

[0225]

[0226] The mutation is based on probability p. m Replace r or W with the neighbors, where the neighbor sampling follows a soft minimization distribution:

[0227]

[0228] The fix is ​​to replace φ(x)>0 with:

[0229]

[0230] If there is no feasible time window, then in R y Internal selection consumes less resources / terms passed r.

[0231] For example, when selecting the final solution from the Pareto front, the Tchebycheff scalarization method is used to select the optimal solution based on business weights; simultaneously, knee points are extracted as alternatives to balance the trade-offs among multiple objectives. Termination occurs when gen ≥ G. max or ΔHV<∈ HV (Near K generation). Tchebycheff's point selection is:

[0232]

[0233] The Pareto solution with the smallest score is selected as the first choice, and the other two are alternatives.

[0234] For example, the system collects information such as actual fault type, actual cost, handling results, and equipment recovery rate, and performs deviation analysis with the model prediction results to quantify three types of indicators: cost deviation, classification deviation, and recovery deviation. When the deviation exceeds a threshold, it triggers the update of threshold parameters, edge weights, and cost intervals in the knowledge graph; at the same time, based on sample deviation and data time decay, it reconstructs the training sample weights and performs lightweight retraining of the model.

[0235] Specifically, by constructing a fully closed-loop mechanism, actual operational data is compared with predicted results to quantify deviations. These deviations drive incremental updates of the graph parameters and model parameters, enabling the system to continuously evolve with accumulated experience. Simultaneously, data distribution drift is monitored through the PSI index, triggering lightweight model retraining when drift exceeds limits. All updates are verified in a shadow environment before hot-switching, and the system automatically degrades to safe mode during SCADA outages, ultimately ensuring long-term stable system operation.

[0236] The method further includes the following steps:

[0237] S1000: Obtain actual maintenance information; the actual maintenance information includes: the actual fault type of the faulty equipment, the actual cost, the handling result and the equipment recovery rate; S1100: Perform deviation analysis on the target maintenance plan and the actual maintenance information to obtain deviation indicators; the deviation indicators include: quantitative cost deviation, classification deviation and recovery deviation.

[0238] Specifically, to assess the gap between predictions and reality, three core biases need to be quantified: cost bias, classification bias, and recovery bias. The larger the bias, the more adjustments are needed to the existing knowledge or model parameters.

[0239]

[0240] Among them, c act This refers to actual material / labor / power outage losses. The model predicts the cost. Δc is the cost deviation vector. y is the actual fault category. This represents the predicted probability distribution. Δp is the classification error indicator. recover_rate is the recovery rate after treatment. ∈ rec This indicates a state of non-recovery.

[0241] S1200: Adjust the weights of the dynamic knowledge graph based on the deviation index.

[0242] The step of adjusting the weights of the dynamic knowledge graph based on the deviation index includes the following sub-steps:

[0243] S1210: If the deviation index is greater than the preset deviation threshold, then update the threshold parameters, weights and cost range of the dynamic knowledge graph.

[0244] Specifically, the cost ranges in the graph are dynamically corrected based on actual cost data to ensure that the ranges reflect the current market conditions; the weights of graph relationships are adjusted according to the success rate of fault handling, with relationships with more successful cases given higher weights; and the warning thresholds are updated based on actual outliers of features such as oil temperature to make the thresholds more closely match the current state of the equipment. Based on the difference in the updated distribution, when the PSI exceeds the threshold, it indicates a significant distribution drift, requiring a model update to adapt to the new distribution.

[0245]

[0246] in, q represents the proportion of recent samples in bin b. b This represents the proportion of the data in the b-th bin during the baseline period. B is the number of bins. PSI is the stability index.

[0247] For example, to avoid performance degradation caused by updated model or graph parameters, the new parameters need to be validated in a shadow environment. The new parameters must simultaneously meet the following criteria: reduced cost prediction error, reduced false alarm rate, and increased early warning lead time. Only after all three indicators are met can the parameters be hot-swapped to the production environment; otherwise, automatic rollback is required to ensure system stability. The early warning lead time is:

[0248] ΔMAE<-η c ΔFPR<-η p Δ lead time > η a ;

[0249] Where ΔMAE is the difference in cost MAE between the old and new models. ΔFPR is the change in false alarm rate. ΔLead Amount is the change in warning lead time. η c η p η a To improve the threshold.

[0250] This application provides a method for predicting power grid equipment faults and generating maintenance plans to address the problem of power grid operation and maintenance optimization. By comprehensively considering multi-modal data features such as power, voltage, temperature, and vibration during the operation of power grid equipment, a multi-channel feature extraction network is constructed, integrating static structural characteristics and dynamic temporal behavior to achieve joint modeling of information from different dimensions. Unlike traditional shallow models based on single time-series signals, this method introduces an attention mechanism and graph structure representation, significantly improving the model's ability to recognize equipment operating status and the accuracy of anomaly identification.

[0251] This application proposes a causal response modeling framework based on LSTM and a fuzzy neural inference system. This framework can capture the evolutionary patterns of faults while using fuzzy rules to express implicit logical relationships under complex operating conditions. Compared to existing models that rely on hard discrimination or shallow rule logic, this application has significant advantages in interpretability and stability when dealing with abrupt changes and uncertain conditions.

[0252] The above detailed embodiments further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A method for predicting power grid equipment faults and generating maintenance plans, characterized in that, include: Obtain the dataset; The dataset includes: power grid equipment operation data, status alarm information, and maintenance cost information; The dataset is then standardized to obtain a structured dataset; Based on the structured dataset, a quintuple relation structure is defined, and a dynamic knowledge graph is established; the quintuple includes: entity node, relation set, weight matrix, threshold parameter, and mapping rule; Construct an encoder group; the encoder group includes: time-series, text, and graph encoders; Based on the dataset and the dynamic knowledge graph, the encoder group is used to obtain temporal features, text features, and graph features; The temporal features, text features, and graph features are fused to obtain the fused features; Based on the fusion features, the fault classification probability and decomposed cost items are obtained; the decomposed cost items include: materials, labor, and power outage losses; Based on the fault classification probability and the decomposition cost item, the corresponding set of measures and feasible time windows are determined in the dynamic knowledge graph to obtain the search space; Based on the search space, the target maintenance plan is determined using the Tchebycheff function.

2. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 1, characterized in that, The step of standardizing the dataset to obtain a structured dataset includes: Based on the dataset, a standardization algorithm is used to unify the dimensions of different features in the dataset; Based on the dataset, data backfilling was performed using linear interpolation, seasonal quantile backfilling, and nearest neighbor backfilling. Based on the dataset, outliers in the measurement data are removed using the median of the measurement data in the dataset, the robust Z-score of MAD, and the outlier score of isolated forest. Convert the text data in the dataset into computable vectors; By aligning the computable vectors with the entity vectors of the knowledge graph, triples and triple confidence are determined to obtain a structured dataset; the triples include: power grid equipment, faults and corresponding measures, and numerical values ​​and standards.

3. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 2, characterized in that, The step of converting the text data in the dataset into computable vectors includes: The text data in the dataset is preprocessed; the preprocessing operations include: text cleaning, segmentation, word segmentation, and annotation. The semantics of text fragments in the text data are transformed into computable vectors.

4. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 2, characterized in that, The step of determining triples and triple confidence by aligning the computable vector with the entity vector of the knowledge graph includes: By using contrastive learning and triplet loss to optimize model parameters, the semantically relevant text vectors are made closer to the text data, while the semantically unrelated text vectors are made further away from the text data. Construct entity vectors for knowledge graphs; Entity vectors related to the text fragment are filtered out using the entity vectors in the knowledge graph. Calculate the similarity between the computable vector and the entity vector; The similarity is fused with semantic evidence to obtain the alignment result; Based on the alignment results, triples are extracted and triple confidence scores are calculated.

5. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 2, characterized in that, The steps of defining a quintuple relation structure and establishing a dynamic knowledge graph based on the structured dataset include: Define the quintuple relation structure and establish an initial knowledge graph; Define a joint anomaly event; the joint anomaly event is used to dynamically adjust the weight matrix of the initial knowledge graph when the joint anomaly event is triggered, so that the state of the initial knowledge graph is synchronized with the data corresponding to the joint anomaly event. After the initial knowledge graph obtains updated triples, the updated triples are aligned with the initial knowledge graph and an alignment score is calculated; the updated triples were not included in the initial knowledge graph. When the alignment score is greater than the score threshold, the updated triple is included in the initial knowledge graph; The maintenance cost information is incorporated into the initial knowledge graph using the exponentially weighted moving average method to establish a dynamic knowledge graph.

6. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 1, characterized in that, The step of fusing the temporal features, text features, and graph features to obtain the fused features includes: By using a gating mechanism, the temporal features, text features, and graph features are fused to obtain fused features.

7. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 1, characterized in that, After obtaining the fault classification probability and decomposing the cost item based on the fused features, the method further includes: The cross-entropy loss of the fault classification probability, the MAE loss of cost prediction, the graph consistency loss, and the graph Laplacian regularization term are fused using a joint loss function, and the proportion of each loss component is adjusted by weighting.

8. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 1, characterized in that, The step of determining the target maintenance plan based on the search space and using the Tchebycheff function includes: Using total maintenance cost, planned downtime, and fault handling risk score as multiple objective functions, a Deb feasibility rule is introduced to determine a set of maintenance schemes from the search space. Based on the set of maintenance plans, the target maintenance plan is determined using the Tchebycheff function.

9. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 1, characterized in that, The method further includes: Obtain actual maintenance information; the actual maintenance information includes: the actual fault type of the faulty equipment, the actual cost, the handling results, and the equipment recovery rate. A deviation analysis is performed on the target maintenance plan and the actual maintenance information to obtain deviation indicators; the deviation indicators include: quantitative cost deviation, classification deviation, and recovery deviation. The weights of the dynamic knowledge graph are adjusted based on the aforementioned deviation index.

10. The method for predicting and generating maintenance plans for power grid equipment faults according to claim 9, characterized in that, The step of adjusting the weights of the dynamic knowledge graph based on the deviation index includes: If the deviation index is greater than the preset deviation threshold, then the threshold parameters, weights, and cost range of the dynamic knowledge graph are updated.

Images