Substation equipment intelligent operation and maintenance decision method based on multi-level confidence decision and security constraint rollback

Abstract

The application discloses a substation equipment intelligent operation and maintenance decision method based on multi-level confidence decision and security constraint rollback, comprising collecting substation equipment multi-source operation data and performing three-level quality grading and differential processing; adopting a deep learning model for fault diagnosis, performing three-way flow according to confidence double thresholds, starting rule and model joint decision in the medium confidence interval, and returning and guiding attention for recalculation when conflicts occur; generating a candidate maintenance scheme, verifying by a safety rule, and implementing positioning-backtracking-equivalent replacement correction on the illegal scheme to output the optimal scheme; collecting execution feedback and evaluating deviation levels, and triggering knowledge graph associated path weight reduction update in high deviation. The application forms a whole-chain closed-loop decision process through four-level judgment-shunt-response nodes, which can effectively improve the automation rate, safety and response speed of substation intelligent operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance decision-making for power equipment, and specifically relates to an intelligent operation and maintenance decision-making method for power equipment based on multi-level confidence judgment and safety constraint backoff. Background Technology

[0002] In the field of power system substation operation and maintenance, with the continuous expansion of the power grid and the surge in the number of equipment, the traditional operation and maintenance model relying on manual experience and regular inspections can no longer meet the requirements of efficiency and reliability. In recent years, artificial intelligence technologies such as deep learning and knowledge graphs have been gradually introduced into the field of substation equipment fault diagnosis. The basic implementation path is as follows: sensors deployed on equipment such as transformers, circuit breakers, and disconnectors collect multi-source operating data such as electrical quantities, thermal quantities, and dissolved gases in oil. After preprocessing, the data is input into a deep learning model for feature extraction and fault classification, outputting diagnostic results and generating maintenance suggestions for operation and maintenance personnel to refer to, thus realizing an initial shift from "manual interpretation" to "model-assisted decision-making".

[0003] However, existing technical solutions generally adopt a linear, unidirectional process of "data input → model processing → result output," which exposes a series of shortcomings in practical engineering applications: First, sensor data is often missing or physically inconsistent due to communication packet loss, equipment drift, etc. Existing solutions lack a hierarchical data quality control mechanism oriented towards the laws of power physics, and directly inputting abnormal data into the model can easily lead to misdiagnosis; Second, the model lacks differentiated processing strategies for low-confidence outputs, and the uniform transfer to manual processing results in low operation and maintenance efficiency; Third, automatically generated maintenance plans lack a safety rule pre-verification step, which cannot ensure that the plan complies with mandatory operating procedures such as "five-prevention" interlocking, and non-compliant plans require full manual rework; Fourth, feedback on the on-site execution effect is only accumulated periodically in batches, and the response to newly emerging error patterns is delayed, making it impossible to achieve rapid closed-loop correction. The above problems seriously restrict the practical application level of substation intelligent operation and maintenance systems. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the above shortcomings, the purpose of this invention is to provide a smart operation and maintenance decision-making method for substation equipment based on multi-level confidence judgment and safety constraint backoff. Through a four-level judgment-diversion-response node, a closed-loop decision-making process is formed, which can effectively improve the automation rate, safety and response speed of substation smart operation and maintenance.

[0005] Technical Solution: To achieve the above objectives, this invention provides an intelligent operation and maintenance decision-making method for substation equipment based on multi-level confidence judgment and safety constraint backoff, including: Step 1: Collect multi-source operating data of power equipment and preprocess it to form structured time-series data; Step 2: The preprocessed data is comprehensively evaluated from three dimensions: completeness, consistency, and timeliness. Based on the evaluation results, the data is divided into three levels: Level I valid, Level II partially abnormal, and Level III severely abnormal. Level II data is imputed and re-verified, and Level III data is physically consistent and reversed before being sent to the subsequent process. If the correction fails, it is transferred to manual processing. Step 3: Use a deep learning model to extract features and classify faults from the effective data output in Step 2, and output the probability distribution of each fault type and its maximum probability value P. max and preliminary diagnostic results; Step 4: Set the high confidence threshold T H and low confidence threshold T L According to P max The interval is divided into three parts: if P max ≥T H Then directly use the model's diagnostic results; if T L ≤P max <T H Then, the rule and model jointly make the decision, and in case of conflict, the decision is fed back to guide the model to recalculate the attention; if P max <T L Then it will be transferred to manual analysis; Step 5: Based on the diagnostic results confirmed in Step 4, generate no fewer than two functionally equivalent candidate maintenance solutions to form a candidate solution set. Step 6: Perform hard safety rule verification on each solution in the candidate solution set. For non-compliant solutions, perform location, backtracking and equivalent replacement correction. Remove solutions that still violate the rules after correction. Compare and score the candidate solution set that passes the verification from three dimensions: number of operation steps, power outage impact and resource matching degree, and sum the scores by weight. Select the solution with the highest comprehensive score as the final recommended solution output. Step 7: Push the solution output in Step 6 to the operations and maintenance personnel for review and execution, and collect feedback data during the execution process simultaneously; Step 8: Compare the actual execution data with the recommended solution, calculate the operation sequence deviation, resource deviation, and effect deviation, and take the maximum value of the three as the comprehensive deviation. Perform graded processing according to the comprehensive deviation level: when the deviation is low, store it in the case library for regular training; when the deviation is medium, perform differentiated attribution analysis; when the deviation is high, immediately trigger the weighted update of the associated paths in the knowledge graph.

[0006] Furthermore, the above eight steps connect the four originally independent stages of data governance, model decision-making, security verification, and feedback learning through a "judgment-diversion-correction-response" logic. This allows the system to automatically enter the corresponding processing branch when facing four typical risks: data anomalies, low model confidence, security rule conflicts, and execution deviations. There are logical dependencies among the four stages: step two ensures the input quality of step three; the conflict feedback in step four depends on the feature extraction layer structure of step three; the equivalent substitution in step six depends on the multi-candidate generation in step five; and the immediate weight reduction in step eight directly affects the verification logic of the next step two. The overall architecture is interconnected, enabling this invention to significantly improve the system's automation rate, security, and closed-loop response speed without changing the deep learning model itself.

[0007] Furthermore, in step two, the completeness assessment uses the missing rate as an indicator: R miss =(N miss / N total )×100% Where N tota l represents the total number of fields for the preset measurement points, N miss This represents the number of missing fields. The conformity assessment uses the violation rate as an indicator: R violate =(N violate / N rule )×100% Where N rule N is the preset total number of physical constraint rules. violate The number of rules violated includes three-phase current balance constraints, voltage-power consistency constraints, and switching state and electrical quantity consistency constraints. Timeliness assessment uses data delay metrics: D delay =t now -t latest Where t now t represents the current system time. latestThe data is timestamped to reflect the latest valid data, and is combined with sampling breakpoint detection and out-of-order detection. Completeness, consistency, and timeliness are three independent dimensions for measuring the usability of power data. Traditional solutions typically only perform simple filtering based on a single dimension, such as removing only null values. This weighting, however, uses three calculable quantitative indicators—missing data rate, violation rate, and latency—to transform data quality assessment from qualitative judgment to quantitative grading. The pre-set physical constraints in the consistency assessment, such as three-phase current balance, voltage-power consistency, and switch status and current consistency, are unique physical laws of the power system. These differ from general data quality checks, making data grading truly geared towards power business scenarios and providing an objective and reproducible benchmark for subsequent differentiated processing.

[0008] Furthermore, in step two, the physical consistency reverse correction of Level III data involves: cross-validating severely abnormal data using redundant measurement data from associated devices; attempting to correct contradictory values ​​based on the power system topology; and re-performing consistency checks after correction. If three consecutive corrections fail, an alarm is triggered, and the data for that time window is removed from the automatic diagnostic queue. For severely abnormal data, instead of simply discarding it or transferring it to manual intervention, this method leverages the inherent topological redundancy of the power system. Electrical quantity constraints exist between adjacent devices (such as Kirchhoff's laws). Redundant measurements from associated devices are used to cross-validate and reverse-calculate contradictory values. This differs from common data interpolation or anomaly removal methods, enabling the maximum recovery of severely abnormal data and improving data utilization without relying on manual intervention.

[0009] Furthermore, in step four, the specific method for joint rule and model decision-making is as follows: the current feature vector is simultaneously input into a preset expert rule base for independent decision-making. If the rule decision result is consistent with the model decision result, the result is adopted; if they are inconsistent, the conflict information is encoded and sent back to the feature extraction layer in step three, triggering the model to recalculate the attention weights for the device dimension pointed to by the rule, generating a new confidence score, and then comparing it with the threshold. The recalculation is performed a maximum of two times. If no consensus can be reached, the process is downgraded to manual processing. This invention differs from the traditional serial filtering of model first and then rule or the parallel integration of model results and rule results. When two inference paradigms disagree, it does not simply perform voting or weighted averaging, but instead encodes the device dimension pointed to by the rule decision result and sends it back to the feature extraction layer, triggering the model to perform targeted attention recalculation, so that the model focuses on the disputed device dimension during secondary inference.

[0010] Furthermore, in step six, the hard safety rule verification involves checking each action to see if the maintenance plan violates the preset safety rule library. This preset safety rule library includes at least five-prevention interlocking logic rules, live-line safety distance rules, and dispatching operation procedure rules. The location, backtracking, and equivalent replacement correction involves marking the violating action and its position in the operation sequence, tracing back the operation purpose of the action, and selecting a non-violating alternative action from the equivalent operation mapping table based on the topological connection relationship and functional redundancy of the power grid equipment. Each plan is allowed only one correction opportunity. Safety constraints are placed after plan generation and before optimal output, and a precise "location-backtracking-replacement" repair mechanism is designed. Traditional plans can only be reworked and regenerated after a violation is discovered. However, this invention, by tracing the operation purpose of the violating action and utilizing the functional redundancy relationship in the power equipment topology, such as the functional mutual backup between the main circuit breaker and the bypass circuit breaker, performs equivalent replacement, minimizing the modification to the violating plan.

[0011] Furthermore, in step six, each scheme S j The formula for calculating the number of operation steps is:

[0012] Where N j For scheme S j Total number of operation steps The minimum number of steps in the candidate solution set; The power outage impact score is calculated according to the following rules: when At that time, Score (j) outage =100; when And U j When Score > 0 (j) outage =0; when And U j When >0, ; Among them U j For scheme S j The number of users affected by the power outage is calculated in real time through power supply path analysis of the current power grid topology. The minimum number of users affected in the candidate solution set; The formula for calculating resource matching score is:

[0013] Where M represents scheme S j Total number of resource items required, s iLet s be the satisfaction level of the i-th resource. i The value is the ratio of the number of resources currently available to the number required by the plan, and s i ≤1. All three dimensions adopt relative comparisons between candidate solutions rather than absolute benchmarks. The operation step score is based on the minimum number of steps in the candidate set, the power outage impact score is based on the minimum number of affected users in the candidate set, and the resource matching score is calculated as the arithmetic mean of the satisfaction of each resource item. The relative comparison strategy of this invention avoids the engineering problem that the theoretical optimal value cannot be obtained in advance, and makes the score based entirely on the currently available real-time data, ensuring the complete feasibility of the method in engineering.

[0014] Furthermore, in step six, each scheme S j The formula for calculating the overall score is:

[0015] Where α+β+γ=1, the weighting coefficients are pre-set according to business priorities; if there are ties for the highest comprehensive score, the one with the higher power outage impact score is selected first. Through the flexible configuration of the weighting coefficients, the system can adapt to business priorities in different scenarios, such as safety production priority, efficiency priority, or resource-constrained priority, thus enhancing the scenario adaptability of the method.

[0016] Furthermore, in step eight, the operation sequence deviation is calculated using the edit distance ratio:

[0017] Where Seq actual Seq is the actual operation sequence. rec Recommended operation sequence; Resource deviation is calculated using the material variation rate: Devres=(N rec / N diff )×100% Where N rec N represents the total number of spare parts and tools in the recommended solution. diff This represents the number of differences between the actual materials used and the recommended materials. The effect deviation is calculated using the average relative deviation:

[0018] Where K is the number of evaluation state variables, and V actual,i V is the actual recovered measurement value of the i-th state variable. rec,i R is the expected recovery value of the i-th state variable. iThe normal fluctuation range of the i-th state variable is defined as follows: Operation sequence deviation is calculated using the edit distance ratio, which quantifies the degree of structured difference between two variable-length sequences, rather than simply judging whether they are consistent; resource deviation is calculated using the difference rate, which assesses the degree of deviation in resource consumption from a quantitative perspective; and effect deviation is calculated using the average relative deviation, which normalizes the deviation through the normal fluctuation range, allowing state variables of different dimensions to be weighted and integrated, thus avoiding the one-sidedness of evaluation by a single indicator.

[0019] Furthermore, in step eight, the overall deviation Dev all =max(Dev seq Dev res Dev eff When Dev all ≤10% is considered low deviation, when Dev all A deviation of >40% or the occurrence of a safety alarm during execution is considered high deviation, while deviations in between are considered medium deviation. The two threshold levels of 10% and 40% divide the deviation space into three levels, corresponding to different handling strategies. This allows the system to execute differentiated closed-loop feedback based on the degree of risk, avoiding overly sensitive and frequent interventions while enabling timely responses to high-risk deviations.

[0020] Furthermore, in step eight, the specific method for triggering the emergency update process of the knowledge graph in case of high deviation is as follows: extract the equipment nodes, fault modes and recommended maintenance paths involved in the current case, multiply the recommendation weight of the triple associated path in the knowledge graph by the attenuation coefficient, and generate an early warning notification to push to the management personnel. The differential attribution analysis for the mean deviation is as follows: examine Dev separately. seq Dev res Dev eff Three deviation indicators, if Dev seq >max(Dev res Dev eff And Dev seq If the percentage is ≤40%, it is attributed to changes in on-site conditions and only recorded for record-keeping; if Dev res >Dev seq or Dev eff >Dev seqIf the bias is high, the error is attributed to inaccurate model recommendations, and the sample is labeled as a difficult sample and stored in the optimization pool. For high bias, immediate weight reduction is used instead of delayed retraining, allowing the system to immediately correct the association weights in the knowledge graph after an error occurs, achieving rapid suppression of similar error patterns. This differs from the traditional approach's delayed response mode, which requires accumulating a large number of samples before periodic retraining. For medium bias, the relative magnitude of three bias indicators is used for attribution and source tracing, clearly distinguishing between changes in on-site conditions and inaccurate model recommendations. The former is only recorded for record-keeping, while the latter is stored in the optimization pool, avoiding the resource waste and overfitting risk of retraining the model indiscriminately for all biases.

[0021] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. The present invention is a substation intelligent operation and maintenance decision-making method based on multi-level confidence judgment and safety constraint backoff. Through four-level judgment-diversion-feedback nodes, typical risks such as data anomalies, low model confidence, safety violations, and execution deviations are incorporated into the structured processing path, which greatly reduces the proportion of manual intervention and improves the system automation rate. 2. The present invention is based on a multi-level confidence judgment and safety constraint backoff intelligent operation and maintenance decision-making method for substation equipment. Each link is based on deterministic information that can be obtained in real time or pre-configured during the operation of the power system. It does not rely on any external theoretical optimal value that cannot be obtained in advance, and has strong engineering feasibility. 3. The present invention is a substation intelligent operation and maintenance decision-making method based on multi-level confidence judgment and safety constraint backoff. By triggering the knowledge graph association path weight reduction in real time with high deviation, it can quickly suppress error patterns and make up for the response lag defect of traditional solutions that only perform periodic retraining. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the steps of the intelligent operation and maintenance decision-making method for substation equipment based on multi-level confidence judgment and safety constraint backoff as described in this invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] Example 1

[0025] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0026] This embodiment uses the #1 main transformer of a 220kV substation of a power supply company as the application object, and takes an abnormal rise in transformer oil temperature as a typical scenario to illustrate the method of the present invention. The substation is equipped with transformer oil temperature sensors, winding temperature sensors, load rate monitoring terminals, an online dissolved gas chromatograph for oil, and a circuit breaker / disconnector position status acquisition terminal. All data is aggregated to the data preprocessing unit via the station control layer communication network. The steps are as follows: Figure 1 As shown.

[0027] I. Specific Implementation of Step 1 Each sensor synchronously collects operational data with a sampling period of Ts=1 second. The collected data types include: three-phase voltage, three-phase current, top oil temperature, winding temperature, concentrations of seven gas components in the oil (H2, CH4, C2H2, C2H4, C2H6, CO, CO2), tap changer position, and cooler status. The preprocessing unit performs timestamp alignment, median filtering for noise reduction, and min-max normalization on the above data to form structured time-series samples.

[0028] II. Specific Implementation of Step 2 The preprocessed data enters the data quality assessment module. This module conducts a comprehensive evaluation from three dimensions: completeness, consistency, and timeliness.

[0029] Regarding completeness: Based on the pre-set standard measurement point list (24 fields in total) in the transformer equipment ledger, the current data vector is checked field by field. If the number of missing fields in the current sample is 1 (a certain phase voltage is missing), then according to formula R... miss =(N miss / N total ) × 100% to calculate the missing rate: R miss =1 / 24×100%≈4.17%.

[0030] Regarding consistency: The module's preset physical constraint rule set includes four rules: three-phase current balance constraint, voltage-power consistency constraint, temperature-load correlation constraint, and circuit breaker state-current consistency constraint. Each rule is validated individually. Assuming that in the current sample, the three-phase current satisfies the balance constraint, the voltage-power consistency constraint, the temperature-load correlation is valid, and the circuit breaker state is consistent with the current, then N... violate =0, according to formula R violate =(N violate / N rule ) × 100% to calculate the violation rate: R violate =0 / 4×100%=0%.

[0031] Regarding timeliness: The module detects the delay D between the current data timestamp and the system time. delay =0.5 seconds, less than Ts (1 second), there are no breakpoints or out-of-order sequences.

[0032] The data quality assessment module divides the current data records into three levels according to the following criteria: if and only when R miss =0%, R violate =0%, D delay ≤Ts and there are no breaks or out-of-order conditions, it is determined as Class I valid and directly sent to Step 3; when R miss ≤15% or R violate ≤20% or D delay ≤3Ts, it is determined as Class II partial anomaly; when R miss >15% or R violate >20% or D delay >3Ts, it is determined as Class III severe anomaly.

[0033] Comprehensive determination: R miss =4.17%≤15%, R violate =0%, D delay =0.5<Ts. According to the above classification criteria, this sample meets the Class I determination conditions and is marked as Class II partial abnormal data.

[0034] Class II data starts the Gaussian process regression imputation sub-process. The module takes the historical data of the transformer at the same moment in the latest 7 days as the training set, takes winding temperature, load rate and ambient temperature as covariates, establishes a GPR model to perform predictive imputation on the missing voltage field. After the imputation is completed, the assessment is re-executed, and all three indicators meet the Class I criteria (R miss =0%, R violate =0%, D delay ≤Ts), marked as "imputed" and sent to Step 3.

[0035] III. Specific implementation of Step 3 and Step 4 In Step 3, the deep learning model integrating multi-scale Transformer and GAT encodes the input feature vector and outputs the probability distribution of each fault type. After calculation, P max =0.87, and the corresponding fault type is "abnormal temperature rise of transformer oil".

[0036] In Step 4, the preset thresholds T H =0.92, T L =0.65. Current P max =0.87, which satisfies T L ≤P max <T H , and enters the rule-model joint inference in the medium confidence interval.

[0037] There is a rule in the expert rule base: "If T oilIf the temperature is >85°C, the load rate is >80%, and the total hydrocarbons in the oil chromatogram are not excessive, then it is determined to be overheating due to load. The rule base judgment result is "overheating due to load", which is consistent with the model judgment of "abnormal increase in oil temperature". Therefore, this diagnostic result is adopted to proceed to step 5.

[0038] IV. Specific Implementation of Steps 5 and 6 In step 5, the scheme generation model adopts the BeamSearch decoding strategy (bundle width B=3) and outputs the following 3 candidate maintenance schemes, forming a candidate set S={S1,S2,S3}:

[0039] Step 6 first performs a hard safety rule check on each of the three schemes. None of the three schemes violated the five-prevention interlocking logic, live safety distance, or dispatching operation procedures. Next, a relative comparison and scoring is performed.

[0040] In terms of the number of operation steps, the minimum number of steps among the candidate solutions is 3 (that is, S1 and S3 are both 3 steps). S1 and S3 get 100 points, and S2 (5 steps) gets 3 / 5×100=60 points.

[0041] Regarding the impact of the power outage, according to the revised scoring rules: due to min(U k ) = 0 (i.e., S2 achieves zero power outage), S2 gets 100 points, S1 and S3 are due to Uⱼ = 1200 > 0 and min(U k If ) = 0, all scores are 0.

[0042] Regarding resource matching, the system integrates with the ERP inventory and scheduling system for real-time queries: 2 cooler spare parts in stock (sufficient), 3 maintenance team members (on duty), sufficient operation shift workers, and 1 inspection electrician on duty. All resource items in the three solutions are met, according to the formula. The calculation shows that the resource matching degree of all three schemes is 100 points.

[0043] With weights α=0.3, β=0.4, and γ=0.3, calculate the comprehensive score according to formula (9): S1: 0.3×100+0.4×0+0.3×100=60 S2: 0.3×60+0.4×100+0.3×100=88 S3: 0.3×100+0.4×0+0.3×100=60 S2 (load transfer path) was selected as the final recommended solution. Although this solution has the most operation steps, it achieves zero power outage impact and complies with the principle of prioritizing safe production.

[0044] V. Specific Implementation of Steps 7 and 8 In step 7, plan S2 is pushed to the operations and maintenance personnel. After review and confirmation, the personnel execute the plan, collecting the actual sequence Seq during execution. actual With the recommended sequence Seq rec Consistent with the actual materials used, the equipment recovered to the following state after execution: oil temperature dropped from 92°C to 78°C (expected 75°C), and winding hot spot temperature dropped from 105°C to 85°C (expected 82°C).

[0045] In step 8, the deviation of the operation sequence is calculated using the following three formulas: the edit distance ratio, the resource deviation using the material difference rate, and the effect deviation using the average relative deviation. Dev seq =0% (the sequence is completely identical); Dev res =0% (materials are perfectly matched); Dev eff =1 / 2×(|78-75| / R oil +|85-82| / R winding ) × 100%. Take R oil =40°C (normal fluctuation range 60°C~100°C), R winding =50°C (normal fluctuation range 60°C~110°C), calculated Dev eff ≈7.5%.

[0046] Determined based on overall deviation, Dev all =max(0%, 0%, 7.5%)=7.5%≤10%, which is considered low bias. This case is stored as a positive sample in the case library for subsequent regular incremental training.

[0047] Example 2

[0048] Based on Example 1, this example illustrates the reverse correction process for Level III severe anomaly data in step two.

[0049] Suppose that in another sampling, the measured value of phase B current on the low-voltage side of main transformer #1 deviates significantly from the three-phase balance constraint (the normal value is about 300A, the measured value is 450A, while the phase B current calculated from the phase A and C currents on the high-voltage side according to the transformation ratio is about 290A). After evaluation in three dimensions in step 2, it is determined to be a level III severe abnormal data, and the reverse correction subprocess is initiated.

[0050] The module retrieves a query from the power grid topology database and finds another redundant current transformer (with the same accuracy class and an independent sampling circuit) installed at the feeder switch connected to phase B on the low-voltage side of main transformer #1. Its current measurement value is read as 295A. The module substitutes the high-voltage side phase A current, phase C current, and this redundant measurement value into Kirchhoff's current law for ampere-turn balance calculation, which passes. The redundant measurement value of 295A is used to correct the original contradictory value of 450A. After correction, a consistency check (three-phase current balance check and voltage-power consistency check) is performed again, both of which pass. The module is then marked "corrected" and proceeds to step 3.

[0051] If the verification fails three times in a row using different associated devices, an alarm is triggered, the data in that time window is marked as "pending manual verification", temporarily removed from the automatic diagnosis queue, and a pending work order is generated and pushed to the manual review queue of the operation and maintenance personnel, waiting for manual confirmation before being added back.

[0052] Example 3

[0053] Based on Example 1, this example illustrates the specific process of verifying the hard security rules and correcting the violation scheme in step six.

[0054] Suppose that in a certain fault scenario, among the three candidate solutions output by the solution generation model, one of the operation actions of S2 is "suspending the grounding wire while the #1 main transformer is energized", which triggers the energized safety distance rule verification.

[0055] Step 6 first performs the location-backtracking-replacement operation: Location: Mark "suspended grounding wire" as a violation, located in step 4 of the sequence; Retrospective analysis reveals that the purpose of this action was to "ensure the maintenance area is safely grounded" (the voltage testing operation had already been completed in the preceding steps). Replacement: Search the equivalent operation mapping table and select the alternative action "after verifying that there is no voltage, hang the grounding wire" that satisfies the same operation purpose and does not violate regulations to replace the original action; Re-verification: Perform a hard verification on the replaced solution. If all actions pass, the corrected solution will be included in the scoring.

[0056] Example 4

[0057] Based on Example 1, this example illustrates the specific process of triggering knowledge graph weight reduction in step eight when there is high bias.

[0058] Suppose that in another failure, the recommended solution suggested following the S2 path, but due to temporary maintenance of the #2 main transformer, the load could not be transferred, and the maintenance personnel temporarily switched to the S1 path. After execution, Dev was calculated. seq =50% (due to the complete change of the operation sequence), Devres =0%, Dev eff =5%, Overall Deviation all =50%>40%, judged as high deviation.

[0059] The system extracts the ternary association information for this case (#1 main transformer, abnormal oil temperature rise, shutdown and maintenance path after load transfer). It retrieves the original recommended weight W=0.85 for this fault mode from the knowledge graph, multiplies it by the attenuation coefficient λ=0.3, and reduces the weight to W'=0.255. Simultaneously, an early warning notification is generated and pushed to management personnel. When the system encounters the #1 main transformer oil temperature rise fault again, the weight of the "shutdown and maintenance path after load transfer" is significantly reduced, and the system prioritizes other more reliable maintenance paths (such as directly operating the cooler). This achieves a rapid closed-loop effect of "preventing the recurrence of the previous error mode."

[0060] This invention addresses four key issues in intelligent substation operation and maintenance: data quality risks, insufficient model confidence, conflicting safety rules and constraints, and delayed execution feedback. It constructs a closed-loop decision-making process consisting of four levels of judgment, diversion, and feedback nodes.

[0061] The three-level data verification mechanism in step 2 effectively recovers abnormal data that was originally unusable due to missing data or physical inconsistencies, avoiding the risk of misjudgment caused by directly inputting abnormal data into the model. The confidence level dual-threshold diversion and rule-model conflict feedback mechanism in step 4 adds secondary confirmation and correction methods to the judgment in the medium confidence interval, effectively reducing the proportion of low confidence outputs being switched to manual processing. The security constraint verification and violation scheme location-backtracking-replacement mechanism in step 6 achieves minimal correction of violation schemes instead of full rework. The high-bias real-time knowledge graph weighting mechanism in step 8 achieves rapid closed-loop suppression of error patterns, making up for the response lag defect of traditional schemes that only perform periodic batch retraining.

[0062] In summary, this invention, while maintaining the deep learning model, can effectively improve the automation rate, security, and closed-loop response speed of the substation intelligent operation and maintenance system through optimized process architecture design, providing a reliable technical support framework for the engineering implementation of substation intelligent agents.

[0063] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A smart operation and maintenance decision-making method for substation equipment based on multi-level confidence judgment and security constraint backoff, characterized in that: include: Step 1: Collect multi-source operating data of power equipment and preprocess it to form structured time-series data; Step 2: The preprocessed data is comprehensively evaluated from three dimensions: completeness, consistency, and timeliness. Based on the evaluation results, the data is divided into three levels: Level I valid, Level II partially abnormal, and Level III severely abnormal. Level II data is imputed and re-verified, and Level III data is physically consistent and reversed before being sent to the subsequent process. If the correction fails, it is transferred to manual processing. Step three, using a deep learning model to extract features and classify faults of the effective data output in step two, output the probability distribution of each fault type and its maximum probability value P max and the preliminary diagnosis result; Step 4: Set the high confidence threshold T H and low confidence threshold T L According to P max The interval is divided into three parts: if P max ≥T H Then directly use the model's diagnostic results; if T L ≤P max <T H Then, the rule and model jointly make the decision, and in case of conflict, the decision is fed back to guide the model to recalculate the attention; if P max <T L Then it will be transferred to manual analysis; Step 5: Based on the diagnostic results confirmed in Step 4, generate no fewer than two functionally equivalent candidate maintenance solutions to form a candidate solution set. Step 6: Perform hard safety rule verification on each solution in the candidate solution set. For non-compliant solutions, perform location, backtracking and equivalent replacement correction. Remove solutions that still violate the rules after correction. Compare and score the candidate solution set that passes the verification from three dimensions: number of operation steps, power outage impact and resource matching degree, and sum the scores by weight. Select the solution with the highest comprehensive score as the final recommended solution output. Step 7: Push the solution output in Step 6 to the operations and maintenance personnel for review and execution, and collect feedback data during the execution process simultaneously; Step 8: Compare the actual execution data with the recommended solution, calculate the operation sequence deviation, resource deviation, and effect deviation, and take the maximum value of the three as the comprehensive deviation. Perform graded processing according to the comprehensive deviation level: when the deviation is low, store it in the case library for regular training; when the deviation is medium, perform differentiated attribution analysis; when the deviation is high, immediately trigger the weighted update of the associated paths in the knowledge graph.

2. The method according to claim 1, characterized in that, In step two, the completeness assessment uses the missing rate as an indicator: R miss =(N miss / N total )×100% Where N tota l represents the total number of fields for the preset measurement points, N miss This represents the number of missing fields. The conformity assessment uses the violation rate as an indicator: R violate =(N violate / N rule )×100% Where N rule N is the preset total number of physical constraint rules. violate The number of rules violated includes three-phase current balance constraints, voltage-power consistency constraints, and switching state and electrical quantity consistency constraints. Timeliness assessment uses data delay metrics: D delay =t now -t latest Where t now t represents the current system time. latest It provides the latest valid data timestamp, combined with sampling breakpoint detection and out-of-order detection.

3. The method according to claim 1, characterized in that, In step two, the physical consistency reverse correction of Level III data is as follows: cross-validate the severely abnormal data using redundant measurement data from associated devices, attempt to correct the contradictory values ​​based on the power system topology, and re-execute the consistency check after correction. If the check fails three times in a row, an alarm is triggered and the data in that time window is removed from the automatic diagnosis queue.

4. The method according to claim 1, characterized in that, In step four, the specific method of joint decision-making by rules and model is as follows: the current feature vector is simultaneously input into the preset expert rule base for independent decision-making. If the rule decision result is consistent with the model decision result, the result is adopted. If they are inconsistent, the conflict information is encoded and sent back to the feature extraction layer in step three, triggering the model to recalculate the attention weight for the device dimension pointed to by the rule, generate a new confidence score and re-compare it with the threshold. The recalculation is performed at most twice. If they still cannot reach a consensus, the process is downgraded to manual processing.

5. The method according to claim 1, characterized in that, In step six, the hard safety rule verification is as follows: check whether the maintenance plan violates the preset safety rule library for each action. The preset safety rule library includes at least the five-prevention interlocking logic rule, the live safety distance rule, and the dispatching operation procedure rule. The location, backtracking, and equivalent replacement correction is as follows: mark the violating action and its position in the operation sequence, trace back the operation purpose of the action, and select a substitute action that meets the same operation purpose and does not violate the rules from the equivalent operation mapping table based on the topological connection relationship and functional redundancy of the power grid equipment. Each plan is only allowed one correction opportunity.

6. The method according to claim 1, characterized in that, In step six, each scheme S j The formula for calculating the number of operation steps is: Where N j For scheme S j Total number of operation steps The minimum number of steps in the candidate solution set; The power outage impact score is calculated according to the following rules: when At that time, Score (j) outage =100; when And U j When Score > 0 (j) outage =0; when And U j When >0, ; Among them U j For scheme S j The number of users affected by the power outage is calculated in real time through power supply path analysis of the current power grid topology. The minimum number of users affected in the candidate solution set; The formula for calculating resource matching score is: Where M represents scheme S j Total number of resource items required, s i Let s be the satisfaction level of the i-th resource. i The value is the ratio of the number of resources currently available to the number required by the plan, and s i ≤1.

7. The method according to claim 1, characterized in that, In step six, each scheme S j The formula for calculating the overall score is: Where α+β+γ=1, the weighting coefficients are preset according to business priority; if there are ties for the highest comprehensive score, the one with the higher power outage impact score will be selected first.

8. The method according to claim 1, characterized in that, In step eight, the operation sequence deviation is calculated using the edit distance ratio: Where Seq actual Seq is the actual operation sequence. rec Recommended operation sequence; Resource deviation is calculated using the material variation rate: Circuit=(N rec / N diff )×100% Where N rec N represents the total number of spare parts and tools in the recommended solution. diff This represents the number of differences between the actual materials used and the recommended materials. The effect deviation is calculated using the average relative deviation: Where K is the number of evaluation state variables, and V actual,i V is the actual recovered measurement value of the i-th state variable. rec,i R is the expected recovery value of the i-th state variable. i This represents the normal fluctuation range of the i-th state variable.

9. The method according to claim 1, characterized in that, In step eight, the comprehensive deviation Dev all =max(Dev seq Dev res Dev eff When Dev all ≤10% is considered low deviation, when Dev all A deviation of >40% or the occurrence of a safety alarm during execution is considered high deviation; a deviation in between is considered medium deviation.

10. The method according to claim 1, characterized in that, In step eight, the specific method for triggering the emergency update process of the knowledge graph when there is a high deviation is as follows: extract the equipment nodes, fault modes and recommended maintenance paths involved in the current case, multiply the recommended weight of the path associated with the triple in the knowledge graph by the attenuation coefficient, and generate an early warning notification to push to the management personnel. The differential attribution analysis for the mean deviation is as follows: examine Dev separately. seq Dev res Dev eff Three deviation indicators, if Dev seq >max(Dev res Dev eff And Dev seq If the percentage is ≤40%, it is attributed to changes in on-site conditions and only recorded for record-keeping; if Dev res >Dev seq or Dev eff >Dev seq If the model recommendation is inaccurate, the sample will be labeled as a difficult sample and stored in the optimization pool.

Images