CN-121980400-A - Substation fixed value remote calibration method based on mixed enhanced intelligent and elastic flow

CN121980400ACN 121980400 ACN121980400 ACN 121980400ACN-121980400-A

Abstract

The invention discloses a transformer substation fixed value remote calibration method based on a hybrid enhanced intelligent and elastic flow, which comprises the following steps of S1, three-level screening and scene adaptation calling, S2, fixed value template intelligent generation and optimization, S3, on-line closed loop circulation and intelligent decision, S4, heterogeneous parameter semantic normalization, S5, working condition perception difference comparison and false difference filtering, S6, elastic downloading execution and real-time monitoring, S7, feedback optimization, wherein a target is precisely locked through a three-level screening structure, a calling strategy is classified and dynamically adjusted through a CNN-LSTM model, a fixed value list to be installed is generated by depending on the on-line closed loop circulation and hybrid enhanced intelligent framework, space limitation is broken, heterogeneous parameter semantic normalization is realized through a BERT model, the comparison ambiguity is eliminated by combining the working condition perception dynamic threshold and the false difference filtering, the abnormal condition is handled based on a BPMN2.0 elastic engine, the complete closed loop is verified and the feedback optimization is performed after the system is pushed, and the iterative upgrade of a system is promoted.

Inventors

ZHANG JINLONG
HU FAN
ZHOU JIANWU
HE WEIHUA
Ye Huoping
JI LI
ZHANG XIAOYONG
LIU RUI
YAN ZIPEI
ZHANG ZHI

Assignees

国网湖北省电力有限公司黄冈供电公司

Dates

Publication Date: 20260505
Application Date: 20251226

Claims (10)

1. The substation fixed value remote calibration method based on the mixed enhanced intelligent and elastic flow is characterized by comprising the following steps of: s1, establishing a three-level screening structure in a dispatching end remote operation and maintenance system, namely collecting real-time measurement data of a screening object by a transformer substation name-primary equipment-protection device, inputting a CNN-LSTM hybrid neural network model to finish scene classification, calling a preset scene-calling priority mapping table to dynamically adjust a calling strategy, and outputting a scene tag and calling data; S2, constructing a scene-association rule mapping library, calling a corresponding association rule by taking an S1 scene tag as an index, generating an initial fixed value template by combining S1 calling data and commonality quantity, dynamically optimizing association coefficients by a DQN reinforcement learning module, and outputting an accurate fixed value template; s3, constructing a closed loop circulation channel on a browser end line, integrating a mixed enhanced intelligent framework containing expert rule base, case reasoning and machine learning by taking the accurate fixed value template of S2 as input, providing decision support, and generating a fixed value list to be downloaded after checking and confirming completion of circulation; S4, inputting heterogeneous parameters into a BERT pre-training language model aiming at the heterogeneous naming problem of parameters of different manufacturers in the fixed value list to be downloaded in S3, carrying out semantic normalization on heterogeneous expressions, constructing a cross-manufacturer semantic mapping library, and outputting a unified semantic interface; S5, constructing a dynamic difference comparison model for sensing working conditions, comparing the to-be-installed fixed value list subjected to semantic normalization processing in S4 with the current operation fixed value of the device, synchronously introducing real-time working condition parameters of equipment, establishing a dynamic threshold model correction standard, filtering unreal differences through a pseudo-difference recognition algorithm, and outputting a real difference list; and S6, deploying an elastic operation flow engine, presetting an abnormal scene treatment plan, calling a unified semantic interface of S4, reversely converting the unified semantic tag in the S5 real difference list into a heterogeneous parameter name which can be identified by a field device, inputting corresponding fixed value data, monitoring real-time feedback in the fixed value downloading process, dynamically adjusting an execution path, and outputting a safe execution result.
2. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the step S1, a three-level screening structure is built by adopting a tree index architecture, the substation name, the primary equipment type and the protection device number in the power grid topology ledger are respectively assigned with unique identifiers, and then the unique locking of the screening object is realized by the following formula: Wherein, the For the only call object to be called, For the three-level screening mapping function, A set is identified for the substation name, For a set of primary device types, the values are { line, transformer, breaker }, Numbering a set for a protection device; after locking the object, the real-time measurement data of the screened object is acquired through an IEC61850 special power communication protocol, namely current Voltage of Fault recording data New energy output data And cleaning the collected original data by adopting a 3 sigma criterion: Removing abnormal values to obtain a cleaned effective data set 。
3. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in S1, a CNN-LSTM hybrid neural network model adopts a secondary processing structure of feature extraction and time sequence analysis, and a CNN network is used for collecting cleaned effective data firstly Extracting local features, inputting the extracted features into an LSTM network to capture time sequence dependency relationship, and finally outputting a scene label, wherein a scene classification output formula is as follows: Wherein, the The value of the scene label is { steady state, fault, new energy grid connection }, For the CNN feature extraction function, A capture function for LSTM timing dependency; the CNN feature extraction process is executed according to the sequence of convolution-activation-pooling, and the method meets the following conditions: Wherein, the In order to extract the local feature matrix, For the convolution kernel weight matrix, In the case of a convolution operation, In order to convolve the offset term, To activate the function, for introducing the nonlinear feature, Is a maximum pooling function; After feature extraction is completed, inputting the feature matrix into an LSTM network according to time sequence, wherein the LSTM time sequence capturing process meets the following conditions: Wherein, the Is the first The state is output by the time LSTM, Is the first The time of day input characteristics are provided, Is the first The state is output at the moment, Is the first Cell state at time; Aggregating valid data After the scene classification is completed by inputting the CNN-LSTM hybrid neural network model, the system automatically invokes a preset scene-recall priority mapping table, and dynamically adjusts recall strategies according to in-table rules, wherein the specific strategies are that all fixed value items are recalled according to the conventional sequence when the scene label is steady, the recall priority of the zero sequence and overcurrent key values is set to be the highest when the scene label is a fault scene, and the recall response time is less than or equal to 1s, and when the scene label is a new energy grid-connected scene, the recall data subset of the protection of the inverter and the voltage related value of the grid-connected point are recalled preferentially, and finally the scene label and the adaptation scene are output 。
4. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the step S2, the construction logic of the scene-association rule mapping library is that matching relations between different power grid operation scenes and protection device types are firstly carded, then corresponding fixed value association coefficients are set for each matching relation, and the mapping relation is: Wherein, the For the constant value of the associated coefficient, As a function of the scene-to-coefficient mapping, Is a protection device type parameter; The specific mapping rule is that when In the steady state the device is in a state, 1.0 For line protection and 1.2 for transformer protection, when In the event of a failure of the device, 1.5 For ground fault and 1.3 for interphase fault, when When new energy is connected, the special association coefficient of the protection fixed value of the newly added inverter and the voltage fixed value of the grid-connected point is increased ; The commonality set Wherein Is CT transformation ratio, Is PT transformation ratio, Is the length of the line; Generating an initial fixed value template: , wherein, For the initial set-value template, Is a function generated based on a freemaker template engine.
5. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the S2, the optimization flow of the DQN reinforcement learning module is that the association error of the initial fixed value template and the historical optimal fixed value is calculated, then the association coefficient is adjusted based on the error, and the optimal coefficient is obtained through multiple iterations, specifically: (1) The state space definition, namely taking a scene tag, a current association coefficient and an association error as the state of reinforcement learning, wherein the state space expression is as follows: Wherein, the In order to strengthen the state of learning, For the current associated error, the error calculation mode is as follows: Wherein, the Setting a value for history optimal; (2) And defining an action space, namely setting the adjustment step length of the association coefficient as an action, wherein the expression of the action space is as follows: Wherein, the As a set of actions, Step length is adjusted for the association coefficient; (3) The rewarding function is defined by setting a rewarding value according to the ratio of the association error to the allowable maximum error, and the expression of the rewarding function is as follows: Wherein, the In order to be a prize value, To allow maximum correlation error; (4) Solving the optimal association coefficient, namely obtaining the optimal association coefficient by iteratively calculating the cumulative rewards maximum value, wherein the solving formula is as follows: Wherein, the In order to optimize the optimal correlation coefficient after the optimization, As a discount factor, the number of times the discount is calculated, For the number of iterations, Is the first A prize value for the second iteration; inputting the optimal correlation coefficient into the template generating function again to obtain a precise constant value template 。
6. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the step S3, an online closed loop circulation channel comprises 5 core nodes of setting, recalculating, auditing, approving and archiving, each node is configured with corresponding operation authority and verification rules, wherein the setting node is input with basic parameters by operation and maintenance personnel, the recalculating node automatically verifies the fixed value logical consistency, the auditing and approving node supports multistage approval and electronic signature, and the archiving node automatically stores fixed value single full flow records; The mixed enhancement intelligent framework adopts a three-level decision fusion logic, an independent decision result is obtained through expert rule base, case reasoning and machine learning, then a final decision is output according to weight fusion, and the decision is output: Wherein, the For decision results, the value is { pass, return modification, supplemental verification }, A weighted fusion decision function; the verification logic of the expert rule base is that the accurate fixed value template is compared with the standard rule set: Wherein, the For a standard rule set, 1 indicates compliance rules, and 0 indicates non-compliance; the logic of case reasoning is that the cosine similarity between the current scene state and the historical case state is calculated, and the maximum value of the similarity is selected as a matching basis: Wherein, the As a function of the cosine similarity, For the current scene state of the scene, As a set of historical case states, when Time multiplexing history association logic; Aiming at a new scene without matching history cases, the machine learning layer calls the DQN module of S2 to re-optimize the association coefficient, and outputs a decision result And finally, after three-level decision fusion, generating a fixed value sheet to be installed.
7. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the step S4, the BERT pre-training language model judges whether parameters to be normalized and standard semantic tags are the same parameters by calculating the semantic similarity of the parameters to be normalized and the standard semantic tags: Wherein, the For the purpose of semantic similarity, For the semantic vector of the parameter to be normalized, Semantic vectors that are standard semantic tags; When (when) When the heterogeneous parameter text is normalized to a standard semantic label Then building a cross-manufacturer semantic mapping library: Wherein, the For heterogeneous parameter text, the mapping library contains a forward mapping of heterogeneous names to standard labels and a reverse mapping of standard labels to heterogeneous names, and a unified semantic interface is output based on the mapping library.
8. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic process of claim 1, wherein in S5, the equipment real-time working condition parameters comprise ambient temperature Humidity of Accumulated running time of equipment ; The dynamic threshold model is based on the reference allowable deviation, and the correction terms are calculated respectively by combining the difference values of the temperature, the humidity and the operation time length, and the dynamic allowable deviation is obtained by superposition, wherein the correction formula is as follows: Wherein, the In order to allow for the dynamic range of the deviation, As a reference to the allowable deviation of the reference, For the temperature correction coefficient(s), As the difference between the actual temperature and the 25C standard temperature, For the humidity correction coefficient, the temperature of the air is adjusted, As the difference between the actual humidity and the 50% rh standard humidity, The coefficient is modified for the duration of the run, The difference value between the accumulated running time length and the standard time length; The pseudo-difference recognition algorithm adopts an isolated forest model, and difference data is input into the model to calculate an abnormal score, wherein the abnormal score formula is as follows: Wherein, the For the purpose of scoring the anomaly, For the sample The path length in the isolated tree is such that, As a result of the desired path length, As the average path length correction factor, For training sample number; Setting the abnormal score threshold to 0.8 when When the difference is determined to be a pseudo difference and filtered, when And further judging whether the difference value is larger than the dynamic allowable deviation, and finally generating a real difference list: Wherein, the Setting a value to be downloaded after semantic normalization, The value is set for the current operation of the semantically normalized device, Is a constant difference value.
9. The substation fixed value remote calibration method based on the hybrid enhanced intelligent and elastic flow according to claim 1, wherein in the step S6, the elastic operation flow engine builds a modular architecture based on the BPMN2.0 standard, and a normal execution path and a treatment plan of 5 types of core abnormal scenes are preset first, and then the execution path is dynamically switched through real-time monitoring data: (1) Defining a flow node and a circulation rule based on a BPMN2.0 standard, wherein a preset abnormal scene comprises a high risk item with a fixed value difference exceeding a dynamic allowable deviation by more than 20%, a communication timeout of more than or equal to 5S, an operation timeout of more than or equal to 30min, personnel permission deficiency and data verification failure, and corresponding treatment plans respectively comprise the steps of triggering safety isolation and three-level alarming by the high risk item, automatically returning the communication timeout to a previous stable state and retry connection, recording a log by the operation timeout and transferring to-do queue, refusing the operation by the permission deficiency and prompting application permission, returning to S5 for re-comparison by the data verification failure; (2) Semantic reverse conversion and data input, namely calling a reverse interface of a cross-manufacturer semantic mapping library constructed by S4, and inputting the S5 output real difference list The uniform semantic tag data in the format is converted into a format of heterogeneous parameter names which can be identified by a corresponding device on site Ensuring that the downloading instruction is matched with the parameter naming rule of the device; (3) Real-time monitoring and path adjustment, namely, through a substation secondary equipment remote operation and maintenance system, real-time monitoring 3 types of core indexes in the fixed value downloading process, namely, differential risk level, device communication state and operation execution time length, wherein monitoring data judges whether to trigger an abnormal plan or not through the following steps: Wherein, the For an abnormal trigger flag, 1 indicates a trigger, 0 indicates no trigger, In order for the risk level to be different, In order to be in a communication state, In order to operate the execution time period, Is an operation duration threshold; When (when) When the engine is automatically switched to the corresponding abnormal branch path to execute the treatment plan And when the fixed value downloading is completed along the normal path, finally outputting a safe execution result comprising a downloading result, a monitoring log and an exception handling record.
10. The method for remote calibration of a fixed value of a transformer substation based on a mixed enhanced intelligent and elastic process according to claim 1, further comprising S7, setting a value of an automatic system readback device within 10 seconds after the fixed value is downloaded Verifying fixed value rationality by combining current power grid scene labels and dynamic allowable deviation, and verifying a formula: Wherein, the For the verification result, 1 indicates reasonable, 0 indicates unreasonable, For the current power grid scene label, multiplexing the CNN-LSTM model of S1 to finish classification; After the verification is passed, tracking the running state of the equipment for one week, if abnormal conditions such as protection misoperation, refusal operation and the like do not occur in the period, judging that the equipment is a successful case, and obtaining case data Logging a set of historical cases The analysis of root cause is immediately triggered if abnormal conditions occur, the analysis result is converted into new constraint rules by combining fault wave recording, fixed value log and working condition parameter positioning problem root cause, and the expert rule base is updated Simultaneously feeding back the adjusted association coefficient to the DQN reinforcement learning module, and iteratively optimizing the association coefficient 。

Description

Substation fixed value remote calibration method based on mixed enhanced intelligent and elastic flow Technical Field The invention relates to the technical field of power system automation, in particular to a substation fixed value remote calibration method based on a hybrid enhanced intelligent and elastic flow. Background The method comprises the following steps of determining the protection action behavior of a power grid relay protection device when the power grid fails, wherein the protection action behavior is directly determined when the power grid fails, and the setting accuracy, the execution timeliness and the correction reliability are the keys for guaranteeing the safe and stable operation of the power grid: Firstly, a traditional substation fixed value calibration mode is highly dependent on manual field operation, operation and maintenance personnel need to carry paper fixed value sheets to run from station to station, a protection device is connected through a special debugging tool to download a fixed value list, template manufacture and parameter input are manually completed, an OMS system lacks remote calling and automatic template generation functions, so that the template manufacture time consumption is extremely high; Secondly, the relay protection devices produced by different manufacturers have the problem of heterogeneous naming of parameters, particularly the names of constant value items of standardized disaster reduction strategies such as low-cycle schemes and the like are inconsistent with the naming rules in the devices, the traditional constant value comparison needs to be matched manually one by one, so that semantic ambiguity is easy to generate; in addition, the automatic readback verification link is absent after the fixed value downloading is finished, a closed loop feedback optimization mechanism is absent, the fixed value parameters and the scene information of a successful case cannot be effectively multiplexed to a historical case library, the fixed value calculation and decision link is difficult to be fed back due to the analysis result of the failed case, the system capacity cannot be continuously iterated, and the development requirements of power distribution network scale expansion and protection configuration refinement are difficult to be adapted; Therefore, it is necessary to design a substation constant value remote calibration method based on hybrid enhanced intelligence and elastic flow. Disclosure of Invention The invention aims to provide a substation fixed value remote calibration method based on a hybrid enhanced intelligent and elastic flow, which solves the problems of high dependence of manual field operation, missing generation of remote recall and automatic template and low efficiency of FTP circulation in the traditional mode proposed in the background technology, overcomes the defects of comparison ambiguity caused by heterogeneous naming of different manufacturer parameters, difficulty in distinguishing true and false differences due to unbonded working conditions of fixed thresholds, and the problems of lack of an elastic execution mechanism, abnormal handling plans, automatic readback verification and closed loop feedback optimization, and incapability of continuous iteration of a system. In order to achieve the purpose, the invention provides the technical scheme that the substation fixed value remote calibration method based on the mixed enhanced intelligent and elastic flow comprises the following steps: s1, establishing a three-level screening structure in a dispatching end remote operation and maintenance system, namely collecting real-time measurement data of a screening object by a transformer substation name-primary equipment-protection device, inputting a CNN-LSTM hybrid neural network model to finish scene classification, calling a preset scene-calling priority mapping table to dynamically adjust a calling strategy, and outputting a scene tag and calling data; S2, constructing a scene-association rule mapping library, calling a corresponding association rule by taking an S1 scene tag as an index, generating an initial fixed value template by combining S1 calling data and commonality quantity, dynamically optimizing association coefficients by a DQN reinforcement learning module, and outputting an accurate fixed value template; s3, constructing a closed loop circulation channel on a browser end line, integrating a mixed enhanced intelligent framework containing expert rule base, case reasoning and machine learning by taking the accurate fixed value template of S2 as input, providing decision support, and generating a fixed value list to be downloaded after checking and confirming completion of circulation; S4, inputting heterogeneous parameters into a BERT pre-training language model aiming at the heterogeneous naming problem of parameters of different manufacturers in the fixed value list to be downloaded in S3, carrying out semant