CN-120611219-B - Algorithm model operation monitoring method and system for power industry

CN120611219BCN 120611219 BCN120611219 BCN 120611219BCN-120611219-B

Abstract

The invention provides an algorithm model operation monitoring method and system for the power industry, and relates to the field of model monitoring. The method comprises the steps of collecting physical quantity data of power equipment through a dynamic causal chain construction module, generating a causal chain after noise suppression and preprocessing, binding with algorithm model parameters, recording mutation events and associated parameter update logs, utilizing a four-dimensional space-time lock synchronization module to attach geographic coordinates and time stamps to ensure data space-time consistency, constructing an optimization strategy by a multi-target game strategy engine module based on historical misinformation events and resource consumption records and combining SLA time constraint of a real-time detection task, triggering emergency response by a DEED-Trigger anti-intuitionism triggering module according to deviation between real-time data and the optimization strategy, injecting compound fault test causal chain robustness by a digital twin technology through a verification platform, and dynamically adjusting causal chain structure and strategy.

Inventors

Liang Hantao
ZHAO YANYANG
CHEN JIAN
ZHENG HUA
CHEN ZHIHONG
YANG JIAYING
LIU JUNJIAN
ZHI FANGLONG
ZENG FANQIANG

Assignees

南方电网数字电网研究院股份有限公司

Dates

Publication Date: 20260508
Application Date: 20250529

Claims (8)

1. The algorithm model operation monitoring method for the power industry is characterized by comprising the following steps of: Step 1, acquiring physical quantity data of electric equipment through a sensor network, inputting the data into a causal chain generator after noise suppression and data preprocessing, extracting a change gradient of the physical quantity P & lt & gt from the generator, matching the change gradient of the physical quantity P & lt & gt with an updating direction of algorithm model parameters theta, if P & lt & gt is consistent with the updating direction theta, establishing causal chain nodes and giving an initial weight of 0.5, connecting strength among the nodes through a historical activation frequency N & lt/EN & gt and the current data change rate P & lt/EN & gt, dynamically updating the product N & lt & gt by P & lt/EN & gt, and recording a mutation event type, a time stamp and an associated model parameter version number when a mutation event of the physical quantity is detected, wherein the mutation event type comprises an amplitude overrun and a second derivative overrun threshold, generating a mutation event log and storing the mutation event log into a system database, supporting subsequent module calling, and reversely tracing the associated algorithm parameters along the causal chain if the mutation amplitude exceeds a preset stable interval P & lt/EN & gt, locking the corresponding to the current data change rate P & lt- & gt, and triggering a time sequence of the hash algorithm, and a time sequence of the hash algorithm is generated by a hash value H-buffer sequence, and a sequence of the hash value is generated when the mutation value H & lt- & gt is detected, and the time sequence is replaced by a hash value of a buffer sequence; Step 2, adding geographic coordinates and time stamps to the causal chain through a geographic coordinate coding unit and a time stamp generator, and ensuring data space-time consistency through hardware-level verification; Step 3, based on the historical false alarm event type distribution and resource consumption record stored by the system, constructing a balance relation of false alarm rate, resource efficiency and detection speed by combining with a preset real-time detection task Service Level Agreement (SLA) time constraint to generate an optimization strategy; Step 4, triggering emergency response according to the deviation of the real-time causal chain data in the step 1 and the optimization strategy in the step 3, after the real-time causal chain data in the step 1 are input, firstly calculating KL divergence of strategy expected power distribution and actual power distribution, starting negative entropy concussion detection when the divergence value exceeds a dynamic threshold value, continuously monitoring entropy variable rate of actual power in5 time sequence windows, judging negative entropy concussion if the entropy value continuously drops and the power fluctuation direction is opposite to the strategy expectation, triggering hyperbolic space folding operation at the moment, mapping geographic coordinates of equipment to hyperbolic manifold space, recalculating shortest paths among nodes in the space, preferentially selecting paths with gentle curvature change as emergency strategies, reversely synchronizing the topological structure of a new path to the step 1, and forcedly updating the connection weight of the current causal chain; and 5, injecting the causal link robustness of the compound fault test by a digital twin technology, and dynamically adjusting the causal link structure according to the test result.
2. The method for monitoring the operation of the algorithm model for the power industry according to claim 1, wherein in the step 1, raw data collected by a sensor are firstly input into a power frequency noise suppression unit, 50Hz power frequency interference is filtered through a notch filter with fixed frequency, filtered data are input into a pulse noise elimination unit, the unit calculates the mean value and the variance of data distribution in real time, dynamically adjusts a noise judgment threshold value, marks as pulse noise when the amplitude of a data point exceeds the threshold value and replaces the pulse noise with a sliding average value of adjacent time sequences, inputs a data stream after preprocessing into a causal chain generator, extracts the change gradient of a physical quantity and matches with the updating direction of algorithm model parameters, if the physical quantity gradient is consistent with the adjusting direction of algorithm model parameters, a causal chain node is established and an initial weight is given, the connection strength between nodes is dynamically updated according to the product of the historical activation times and the current data change rate, and after the causal chain node is established, the type, the time stamp and the associated model parameter update log of the physical quantity mutation event are recorded in real time, and the causal chain node is bound and stored.
3. The algorithm model operation monitoring method for the electric power industry according to claim 1 is characterized in that in the step 2, a geocoordinate coding unit respectively quantizes equipment longitude, latitude and altitude into integer values of one thousandth, the integer values are spliced into 12-bit geocodes, a timestamp generator calculates chaotic time sequence through optical fiber transmission delay compensation, UTC time, optical fiber refractive index and transmission distance are associated to generate an irreversible timestamp, when data are packaged, a causal chain hash value of the geocodes and the previous time sequence is written into a data packet header, the causal chain hash value is the latest hash value of the current time sequence generated by a dynamic causal chain construction module, a receiving end analyzes and then performs space-time verification, firstly verifies whether the geocodes are in a preset power grid topological range, then verifies whether the hash value strictly monotonically increases, if any condition is not met, discards a data packet and records an abnormal source, then judges that the space-time verification fails, discards the data packet and records the abnormal source, and when the number of the space-time verification fails is accumulated for 3 times, the system automatically switches to a standby communication channel.
4. The method for monitoring the operation of the algorithm model for the power industry according to claim 3, wherein when the standby communication channel is started, a redundant data packet comprising a physical quantity average value of the first three effective time sequences, an algorithm model parameter reference value stored by a system and complete geocoding is injected into a link, and during the operation of the standby communication channel, new strategy generation is suspended, the output of the last effective strategy is maintained, the verification state of a main channel is monitored, and if the continuous 5 time sequence recovery verification of the main channel is passed, control rights are gradually returned to the main channel, and a causal link change record missing during redundancy is synchronized.
5. The method for monitoring the operation of the algorithm model for the power industry according to claim 1 is characterized in that in the step 3, the distribution of false alarm event types in the past 24 hours is extracted from historical false alarm event records stored in a system, resource consumption records are obtained from a system resource monitoring unit, a two-dimensional game matrix of the false alarm rate and the resource efficiency is constructed, a strategy subset with the detection speed reaching the SLA requirement is screened from the game matrix according to the time constraint of a real-time detection task, then iterative optimization is carried out on the subset by adopting a genetic algorithm, the cross probability is dynamically reduced in each iteration, 20% of individuals with the lowest false alarm rate are reserved as strategies, the other individuals randomly perturb in the dimension of the resource efficiency through mutation operation until the fluctuation of the false alarm rate of the strategies is less than 1% and the resource efficiency is improved by more than 10 iterations, and the final optimization strategy is output.
6. The method for monitoring the operation of the algorithm model for the power industry according to claim 5, wherein in the step 3, if a policy conflict is detected in the iteration process, the conflict judging condition is that the false alarm rate of the current policy is 2 times standard deviation of the historical mean value compared with the previous generation of sudden increase, or the resource efficiency is continuously 3 generations and is not improved, and the detection speed is reduced by more than 5%, a policy backtracking mechanism is triggered, namely, a current iteration pool is emptied, a game matrix of a previous stable version is loaded from a cache, the weight distribution of the policy is reset, the policy with the standard detection speed reaching the standard is preferentially reserved, the disturbance range of the variation operation is limited to be not more than 75% of the stability boundary, and the game matrix after reset can be continuously optimized after no conflict is verified through 3 rounds of conservative iteration.
7. The method for monitoring the operation of the algorithm model for the power industry according to claim 1, wherein in the step 5, a composite fault test is automatically injected at intervals by a digital twin technology, wherein the fault mode comprises a scene of concurrence of vibration sudden increase of a transformer bushing and electric field intensity abnormality of a circuit breaker, during the test, the digital twin technology runs a real causal chain and a virtual chain injected with faults in parallel, if a causal chain point fracture is detected, a reconstruction engine calculates a priority according to the product of the historical activation times of the nodes and the connection weight, meanwhile, a physical quantity mutation event log recorded by a dynamic causal chain construction module is referred to, nodes with product values greater than 100 are preferably reconstructed, the reconstructed causal chain needs to pass through full-node hash verification, the expected output difference of the digital twin platform is less than 5%, and the original causal chain can be replaced and validated.
8. The power industry oriented algorithm model operation monitoring system of the power industry oriented algorithm model operation monitoring method of claim 1 is characterized by comprising a dynamic causal chain construction module, a four-dimensional space-time lock synchronization module, a multi-target game strategy engine module, a DEED-Trigger anti-intuition triggering module and a verification platform; The dynamic causal chain construction module acquires physical quantity data of the power equipment through a sensor network, generates a causal chain after noise suppression and data preprocessing, binds the causal chain with algorithm model parameters, and records a physical quantity mutation event and an associated parameter update log; The four-dimensional space-time lock synchronization module attaches geographic coordinates and time stamps to the causal chain, and ensures data space-time consistency through hardware level verification; the multi-target game policy engine module constructs a balance relation of false alarm rate, resource efficiency and detection speed to generate an optimization policy by combining preset real-time detection task Service Level Agreement (SLA) time constraint based on historical false alarm event type distribution and resource consumption records stored by the system; The DEED-Trigger anti-intuition triggering module triggers emergency response according to deviation of real-time data and an optimization strategy of the multi-target game strategy engine module; The verification platform injects the causal link robustness of the compound fault test through a digital twin technology, and dynamically adjusts the causal link structure and strategy according to the test result.

Description

Algorithm model operation monitoring method and system for power industry Technical Field The invention relates to the field of model monitoring, in particular to an algorithm model operation monitoring method and system for the power industry. Background Along with the accelerating promotion of smart grid construction, the digital and intelligent demands of an electric power system are increasingly urgent, an algorithm model is used as a core decision support tool, the running stability of the algorithm model directly influences the safety and economic dispatch of the electric power grid, the multi-dimensional space-time data generated by massive heterogeneous equipment provides higher requirements on the dynamic sensing and collaborative processing capacity of a real-time monitoring system, the limitation of a traditional monitoring framework is urgently broken, and the prior art is mainly based on a single-dimensional static monitoring mechanism and is difficult to deal with causal reasoning and strategy optimization challenges under complex working conditions. The current power monitoring system usually adopts a mode of combining threshold alarming and offline analysis, for example, a fixed frequency filter is used for inhibiting power frequency noise, an alarming threshold is set depending on expert experience, historical data is stored by utilizing a time sequence database and periodic inspection is carried out, a part of schemes are introduced into a machine learning model for anomaly detection, but single-objective optimization is carried out, a dynamic game mechanism under multi-dimensional constraint is lacked, and other researches try to improve data reliability through redundancy check, but space-time consistency guarantee still depends on manual calibration, and hardware-level automatic synchronization is difficult to realize. The prior art has the defects that the suppression of power frequency interference and impulse noise by a data preprocessing link lacks dynamic adaptability, so that the causal link construction is easily affected by instantaneous interference, a space-time synchronization mechanism depends on simple time stamp alignment, the data packet consistency deviation caused by optical fiber transmission delay and geographic coordinate quantization error is not solved, a multi-objective optimization strategy adopts static weight distribution, and the dynamic game requirements of false alarm rate, resource efficiency and response speed in a real-time detection task are difficult to deal with. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides an algorithm model operation monitoring method and system for the power industry, which are used for solving the problems that the construction of a causal chain is influenced by instantaneous interference due to the lack of dynamic adaptability of power frequency interference and impulse noise suppression in a data preprocessing link in the background art, the problem that a space-time synchronization mechanism is aligned by relying on simple time stamps, the problem of data packet consistency deviation caused by optical fiber transmission delay and geographic coordinate quantization errors is not solved, and the problem that a multi-objective optimization strategy adopts static weight distribution is difficult to meet the dynamic game requirements of false alarm rate, resource efficiency and response speed in a real-time detection task. (II) technical scheme The algorithm model operation monitoring method and the system thereof for the power industry comprise a dynamic causal chain construction module, a four-dimensional space-time lock synchronization module, a multi-target game strategy engine module, a DEED-Trigger anti-intuition triggering module and a verification platform; The dynamic causal chain construction module acquires physical quantity data of the power equipment through a sensor network, generates a causal chain after noise suppression and data preprocessing, binds the causal chain with algorithm model parameters, and records a physical quantity mutation event and an associated parameter update log; The four-dimensional space-time lock synchronization module attaches geographic coordinates and time stamps to the causal chain, and ensures data space-time consistency through hardware level verification; the multi-target game policy engine module constructs a balance relation of false alarm rate, resource efficiency and detection speed to generate an optimization policy by combining preset real-time detection task Service Level Agreement (SLA) time constraint based on historical false alarm event type distribution and resource consumption records stored by the system; The DEED-Trigger anti-intuition triggering module triggers emergency response according to deviation of real-time data and an optimization strategy of the multi-target game strategy engine modu