Power industry-oriented algorithm model operation monitoring method and system

Abstract

The invention provides a power industry-oriented algorithm model operation monitoring method and system, and relates to the field of model monitoring. Comprising 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 the causal chain with algorithm model parameters, and recording a sudden change event and an associated parameter update log; a four-dimensional space-time lock synchronization module is utilized to add geographic coordinates and timestamps, and data space-time consistency is ensured; the multi-target game strategy engine module constructs an optimization strategy based on historical misinformation events and resource consumption records in combination with SLA time constraints of real-time detection tasks; the DEED-Trigger anti-intuition triggering module triggers an emergency response according to the deviation between the real-time data and the optimization strategy; and the verification platform injects a composite fault through a digital twin technology to test the causal chain robustness, and dynamically adjusts the causal chain structure and strategy.

Description

Technical Field

[0001] The present invention relates to the field of model monitoring, and in particular to an algorithm model operation monitoring method and system thereof for the power industry. Background Art

[0002] With the accelerated construction of smart grids, the demand for digitalization and intelligence of power systems is becoming increasingly urgent; algorithm models are core decision-making support tools, and their operational stability directly affects grid security and economic dispatch; the multi-dimensional spatiotemporal data generated by massive heterogeneous devices place higher demands on the dynamic perception and collaborative processing capabilities of real-time monitoring systems, and there is an urgent need to break through the limitations of traditional monitoring frameworks; existing technologies are mostly based on single-dimensional static monitoring mechanisms, which make it difficult to cope with the challenges of causal reasoning and strategy optimization under complex working conditions.

[0003] Current power monitoring systems often use a combination of threshold alarms and offline analysis; for example, fixed-frequency filters are used to suppress power frequency noise, alarm thresholds are set based on expert experience, and time series databases are used to store historical data and conduct periodic inspections. Some solutions introduce machine learning models for anomaly detection, but most of them are single-objective optimizations and lack dynamic game mechanisms under multi-dimensional constraints. Other studies have attempted to improve data reliability through redundant verification, but spatiotemporal consistency guarantees still rely on manual calibration, making it difficult to achieve automatic synchronization at the hardware level.

[0004] The existing technology lacks dynamic adaptability in the data preprocessing link to suppress power frequency interference and pulse noise, which makes the causal chain construction susceptible to instantaneous interference; the spatiotemporal synchronization mechanism relies on simple timestamp alignment and does not solve the data packet consistency deviation caused by optical fiber transmission delay and geographic coordinate quantization error; the multi-objective optimization strategy uses static weight distribution, which makes it difficult to cope with the dynamic game requirements of false alarm rate, resource efficiency and response speed in real-time detection tasks. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the shortcomings of the existing technology, the present invention provides an algorithm model operation monitoring method and system for the power industry to solve the problem proposed in the above background technology that the power frequency interference and pulse noise suppression in the data preprocessing link lack dynamic adaptability, resulting in the causal chain construction being affected by instantaneous interference; the time and space synchronization mechanism relies on simple timestamp alignment and does not solve the problem of data packet consistency deviation caused by optical fiber transmission delay and geographic coordinate quantization error; the multi-objective optimization strategy adopts static weight distribution, which is difficult to meet the dynamic game requirements of false alarm rate, resource efficiency and response speed in real-time detection tasks.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method and system for monitoring the operation of an algorithm model for the power industry, including a dynamic causal chain construction module, a four-dimensional space-time lock synchronization module, a multi-objective game strategy engine module, a DEED-Trigger counter-intuitive triggering module, and a verification platform;

[0009] The dynamic causal chain construction module collects physical quantity data of power equipment through the sensor network, generates a causal chain after noise suppression and data preprocessing, binds the causal chain to the algorithm model parameters, and records physical quantity mutation events and associated parameter update logs;

[0010] The four-dimensional space-time lock synchronization module adds geographic coordinates and timestamps to the causal chain, ensuring data space-time consistency through hardware-level verification;

[0011] The multi-objective game strategy engine module generates an optimization strategy by building a balance between false alarm rate, resource efficiency and detection speed based on the historical false alarm event type distribution and resource consumption records stored in the system and the preset real-time detection task service level agreement (SLA) time constraints;

[0012] The DEED-Trigger counter-intuitive triggering module triggers an emergency response based on the deviation between the real-time data and the optimization strategy of the multi-objective game strategy engine module;

[0013] The verification platform injects composite faults through digital twin technology to test the robustness of the causal chain, and dynamically adjusts the causal chain structure and strategy based on the test results.

[0014] Preferably, the dynamic causal chain construction module collects the voltage, current and temperature physical quantity data of the power equipment in real time through the sensor network. The original signal output by the sensor first enters the power frequency noise suppression unit, which uses a notch filter with a center frequency of 50Hz to filter out the power frequency interference. The filtered signal is input to the pulse noise elimination unit, and the pulse noise elimination unit calculates the mean μ and variance σ of the current time series data with a sliding window of 10 sampling points in length. 2, dynamically set the noise threshold to μ±3σ, when the amplitude of a data point exceeds the threshold, it is marked as an impulse noise point and replaced with the sliding average of the adjacent time series, and the replacement window length is 10 sampling points; the preprocessed data stream is input into the causal chain generator, which extracts the change gradient ΔP / Δt of the physical quantity and matches it with the update direction Δθ of the algorithm model parameters. If the direction of ΔP is consistent with that of Δθ, a causal chain node is established and an initial weight of 0.5 is assigned. The connection strength between nodes is dynamically updated by the product N×ΔP / Δt of the historical activation times and the current data change rate ΔP / Δt, and the update cycle is 1 second; when a physical quantity mutation event is detected, the mutation event type, timestamp and associated model parameter version number are recorded at the same time. Event types include amplitude exceeding the limit and second-order derivative exceeding the threshold. A mutation event log is generated and stored in the system database; the log is bound to the causal chain node to support subsequent module calls. The causal chain generator traces back the associated algorithm model parameters along the causal chain path. If the mutation amplitude exceeds the preset stable interval [P_min, P_max], the corresponding parameter dimension is locked and its update is frozen. At the same time, the hash value H is generated using the SHA-256 algorithm. The input data is the physical quantity sequence P_seq, the algorithm model parameter version number θ_ver and the timestamp t. If the hash values ​​H_t and H_{t-1} of two consecutive time series fail to verify, the data rollback is triggered, the verified causal chain copy is loaded from the cache and the associated path is recalculated.

[0015] Preferably, the geographic coordinate encoding unit of the four-dimensional space-time lock synchronization module quantizes the longitude L, latitude B and altitude H of the device into an integer value of one thousandth, and the encoding format is L×1000||B×1000||H×1000, generating a 12-bit geographic code GEO; the timestamp generator calculates the chaotic time series through the optical fiber transmission delay compensation algorithm, specifically substituting the UTC time t, the optical fiber refractive index n, and the transmission distance d into the chaotic equation to generate an irreversible timestamp T; in the data encapsulation stage, the causal chain hash value H_{t-1} of GEO and the previous time series is written into the data packet header, and the hash value H_{t-1} is the latest valid hash generated by the dynamic causal chain construction module. The receiving end needs to verify that H_{t-1} is Whether it is consistent with the previous time series hash in the local cache, after the receiving end parses it, it first checks whether the GEO is within the preset power grid topology coordinate range [GEO_min, GEO_max]. If it is out of range, the data packet is discarded; secondly, it verifies whether the hash sequence satisfies the strict monotonic increasing property of H_t>{t-1}. If the verification fails, the abnormal source IP address is recorded and an alarm is triggered; when the time and space verification fails for three consecutive times, the system switches to the backup communication channel, which injects a redundant data packet containing the mean value of the physical quantity of the first three valid time series, the benchmark value of the model parameters stored in the system and the complete GEO. At the same time, the main channel pause strategy is generated. If the main channel passes the time series recovery verification for five consecutive times, the control is gradually returned and the missing causal chain change records are synchronized.

[0016] Preferably, the multi-objective game strategy engine module extracts the type distribution matrix F and resource consumption record matrix R of false alarm events in the past 24 hours from the historical false alarm event database stored in the system, and constructs the false alarm rate-resource efficiency two-dimensional game matrix M=F×R T ; According to the time constraints of the real-time detection task, which are defined by a preset service level agreement (SLA) profile, including parameters such as the maximum allowable detection delay and task priority, a strategy subset S in the matrix with a detection speed ≥ SLA requirements is screened; a genetic algorithm is used to iteratively optimize S, with an initial population size of 100 and a crossover probability decreasing by 0.02 per generation from 0.8. In each generation, 20% of individuals with the lowest false alarm rate are retained as strategies, and the remaining individuals are subjected to a random perturbation of ±5% on the resource efficiency dimension through mutation operations; the optimization termination condition is that the strategy false alarm rate fluctuation is less than 1% and the resource efficiency improvement stagnates for more than 10 generations, and the final optimized strategy is output; if a sudden increase in the false alarm rate exceeding the historical mean 2σ or the resource efficiency has not been improved for three consecutive generations and the detection speed has dropped by more than 5% during the iteration, the strategy backtracking mechanism is triggered, the current iteration pool is cleared, and the game matrix of the previous stable version is loaded, the strategy weight is reset to prioritize detection speed, and the mutation perturbation range is limited to no more than 75% of the stability boundary. After three rounds of conservative iteration verification without conflict, the optimization continues.

[0017] Preferably, the DEED-Trigger counter-intuitive triggering module receives the real-time causal chain data stream output by the dynamic causal chain building module, including the physical quantity sequence, the mutation event record and the hash value H_t of the current time series. The DEED-Trigger counter-intuitive triggering module calculates the KL divergence D_{KL}(P||Q) between the expected power distribution Q and the actual distribution P in real time. When D_{KL} exceeds the dynamic threshold θ=0.1×log1+ΔP / Δt, the negative entropy oscillation detection is started, and the entropy change rate ΔS / Δt of the actual power in 5 time series windows is continuously monitored. If ΔS / Δt <0 and the power fluctuation direction is opposite to the strategy expectation, it is determined to be a negative entropy oscillation; at this time, the hyperbolic space folding operation is triggered, the device geographic coordinates are mapped to the hyperbolic manifold space, and the shortest path between nodes is calculated through the Poincare disk model. The path selection criterion is the curvature change rate |ΔK / Δs| < 0.01. After the path is selected, its topology structure is reversely synchronized to the causal chain construction module, and the connection weight is forced to be updated to 1.2 times the original value. The updated causal chain topology structure needs to pass the hash check of the four-dimensional space-time lock synchronization module. If the check fails, it rolls back to the previous version and triggers an alarm, and is verified in real time through the verification platform.

[0018] Preferably, the verification platform injects a composite fault test every 30 minutes through digital twin technology, and the failure mode includes the transformer bushing vibration amplitude ≥ 5mm / s 2 Concurrent scenarios with abnormal deviations from the circuit breaker electric field strength of ≥10KV / m; during the test, the platform runs the real causal chain C_real and the virtual chain C_virtual with injected faults in parallel. If a node break is detected in C_virtual, the reconstruction engine calculates the priority based on the product N×w of the node's historical activation times N and the connection weight w, and prioritizes the reconstruction of nodes with N×w>100. At the same time, it refers to the mutation event log recorded by the dynamic causal chain construction module. If the node is associated with a high-frequency mutation event, such as the number of triggers in the past hour is ≥5, its priority is increased by 50%; the reconstructed causal chain needs to pass the full-node SHA-256 hash check and be compared for consistency with the current hash value H_t generated by the dynamic causal chain construction module, and the difference with the expected output of the digital twin platform is less than 5%. If the conditions are met, the original causal chain is replaced and takes effect, otherwise it rolls back to the previous version and marks the fault mode as unfixed.

[0019] Preferably, the impulse noise elimination unit of the dynamic causal chain building module adopts an adaptive threshold update mechanism to recalculate the mean μ and variance σ of the data distribution every 100ms. 2The noise judgment threshold is dynamically adjusted to μ±kσ, where μ is the sliding window mean and σ is the standard deviation. k is adaptively adjusted according to the recent noise density. The default value is k=3. If the noise point ratio in the past 10 seconds is greater than 5%, k=2.5. If the noise point ratio in the past 10 seconds exceeds 5%, k is reduced from 3 to 2.5 to enhance the filtering strength. The update formula of the causal chain node connection strength is S_{new}=S_{old}+N×ΔP / Δt, where N is the number of historical activations and ΔP / Δt is the current gradient. When S_{new}>100, the weight decay mechanism is triggered and the weight is reduced according to S_{new}=S_{new}×0.9. The hash value H is generated using the SHA-256 algorithm. The input data is a concatenated string of the physical quantity sequence P_seq, the algorithm model parameter version number θ_ver, and the timestamp t. When the verification fails, the system loads the verified causal chain copy in the last 5 minutes from the cache and recalculates the association path of the mutation event.

[0020] Preferably, when the backup communication channel of the four-dimensional space-time lock synchronization module is started, the multi-objective game strategy engine module suspends the generation of new strategies, maintains the output of the last valid strategy, and monitors the hash verification status of the main channel at the same time. If the main channel recovers H_t>H_{t-1} for 5 consecutive time sequences, the control right handover process is started, specifically gradually increasing the data flow from 20% of the backup channel to 100% of the main channel, 10% each time, and verifying the space-time consistency at each stage; after the handover is completed, the system synchronizes the causal chain change records that were missing during the redundancy period, and uses a difference merging algorithm to align the timestamps and resolve conflicts between the backup channel data and the main channel data. When resolving conflicts, the operation with the latest timestamp is retained first; if the timestamps are the same, the main channel data is selected, and the conflict event is recorded in the audit log.

[0021] Preferably, during the genetic algorithm iteration process of the multi-objective game strategy engine module, the crossover operation adopts a single-point crossover method, and the crossover point is randomly selected at the 1 / 3 or 2 / 3 position of the strategy code, and the offspring strategy is generated by splicing the crossover fragments of the parent strategy; the mutation operation imposes a random perturbation ΔR = ±5% × R_base on the resource efficiency dimension, where R_base is the benchmark efficiency value; when the strategy backtracking mechanism is triggered, the system loads the previous stable version of the game matrix M_prev from the cache, and resets the strategy weight distribution formula to W = 0.7 × Speed ​​+ 0.3 × Efficiency, limiting the mutation perturbation range ΔR ≤ 0.75 × R_stable, where R_stable is the benchmark value of the stability boundary; in the conservative iterative verification stage, only strategies are allowed to participate in crossover in each round of iteration, and the mutation probability is reduced to 0.1 until the standard optimization process is restored after three consecutive rounds without conflict.

[0022] Preferably, in the hyperbolic space folding operation of the DEED-Trigger module, the (x, y) formula for mapping geographic coordinates to hyperbolic manifold space is: The shortest path between nodes is calculated using the hyperbolic Dijkstra algorithm, and the path weight is the inverse of the curvature change rate 1 / |ΔK / Δs|. After the emergency strategy is generated, its topology is updated to the causal chain construction module through the reverse synchronization protocol. After the receiving end verifies the continuity of the timestamp, the connection weight is forcibly updated to NewWeight. If the difference with the original weight is greater than 20%, the verification platform's real-time testing process is triggered to ensure the robustness of the updated causal chain.

[0023] Preferably, the digital twin technology of the verification platform adopts a high-fidelity physical simulation engine, the transformer bushing vibration model is constructed based on finite element analysis, and the circuit breaker electric field model is solved based on Maxwell equations; during the composite fault test, the vibration surge event is realized by superimposing a sinusoidal wave interference signal, and the amplitude increases linearly from 0 to 5 mm / s 2 The abnormal electric field strength is simulated by modifying the dielectric constant ε; the parallel operation of the virtual causal chain C_virtual and the real chain C_real adopts time slice round-robin scheduling, switching the execution context every 10ms; in the node reconstruction priority calculation, if multiple nodes meet N×w>100, they are reconstructed in descending order according to the product value. The reconstructed causal chain needs to pass the full-node SHA-256 hash check, and the difference threshold with the expected output is set to 5%. If the difference exceeds the threshold, an alarm is triggered and a fault analysis report is generated.

[0024] Preferably, the causal chain generator of the dynamic causal chain building module monitors the second-order derivative d of the physical quantity in real time after the node connection is established. 2 P / dt 2 , if |d 2 P / dt 2 If the value exceeds the preset threshold α, it is determined to be a mutation event and the causal chain is traced back to the associated algorithm model parameter θ_i. After locking θ_i, its update is frozen and the parameter version θ_ver is recorded. If the hash value continuity check fails, the system compares the distance between the current hash H_t and the 10 most recent hash values ​​in the cache. If the minimum distance is greater than 3, it is determined to be data tampering and the security isolation mechanism is triggered, disconnecting the abnormal data source connection. During the data rollback process, the causal chain copy loaded from the cache must be verified by digital signature. The signature algorithm uses RSA-2048, and the private key is stored in the hardware.

[0025] Preferably, the redundant data packet injection of the backup communication channel adopts priority queue management with a queue length of 100. If the queue is full, the oldest data packet is discarded; in the control right handover process, each stage of traffic switching, such as 20% to 30%, needs to verify the time and space verification pass rate of the main channel. If the pass rate of a certain stage is less than 95%, it will fall back to the previous stage and extend the verification time by 10 seconds; wherein the difference merging algorithm adopts operation conversion (OT) technology to sort the causal chain change records of the backup channel by timestamp, and perform conflict detection with the main channel records. The conflict resolution rule is to retain the operation with the latest timestamp. If the timestamps are the same, the main channel operation is retained first.

[0026] Preferably, in the strategy retention mechanism of the multi-objective game strategy engine module, after each generation of iteration, the strategy's false alarm rate threshold is set to 80% of the contemporary average value, and the resource efficiency threshold is set to 90% of the contemporary maximum value; when determining strategy conflicts, if the false alarm rate suddenly increases by more than the historical average 2σ, the parameter snapshot function is automatically triggered to save the current algorithm model parameters θ and the causal chain state C to the read-only storage area; in the hyperbolic space folding operation, if the shortest path calculation exceeds 500ms, the alternative path selection algorithm is enabled, and the top three paths are selected in ascending order of curvature change rate, and the final emergency strategy is determined through a weighted voting mechanism.

[0027] Preferably, when the hash value verification of the dynamic causal chain building module fails, the system starts the causal chain repair process, specifically extracting the connection weights of the same node ID from the last five valid causal chain copies, and calculating their weighted average as the repair value. The weight distribution formula is w_repair=∑(w_i×e {-|t-t_i|} ), where t_i is the replica timestamp; the repaired causal chain needs to pass the transient fault injection test of the digital twin platform. The test mode is single-node failure. If the number of node breakages in the repair chain during the test is ≤2, the repair is considered successful. Otherwise, the full chain reconstruction process is retriggered.

[0028] (3) Beneficial effects

[0029] The present invention provides an algorithm model operation monitoring method and system for the power industry. It has the following beneficial effects:

[0030] 1. The present invention significantly improves the accuracy and efficiency of algorithm model monitoring in the power industry through dynamic causal chain construction and multi-objective game strategy optimization. It adopts power frequency noise suppression and adaptive pulse filtering technology, combined with the dynamic update mechanism of causal chain node weights, to effectively reduce the impact of data interference. Through intelligent analysis of historical false alarm events and resource consumption, it generates an optimization strategy that takes into account false alarm rate, resource efficiency and detection speed, achieving a fault detection accuracy improvement of more than 30% and a reduction of ineffective resource consumption by 20%.

[0031] 2. The present invention uses four-dimensional space-time lock synchronization and digital twin composite fault injection technology to ensure data space-time consistency and system robustness. It uses hyperbolic space folding emergency response and redundant communication switching mechanism to quickly reconstruct the causal chain topology in complex fault scenarios, ensuring that the abnormal detection delay of power equipment is reduced to milliseconds, and the system self-repair success rate is increased to 98%, significantly enhancing the stability and safety of power system operation. DETAILED DESCRIPTION

[0032] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0033] An embodiment of the present invention provides an algorithm model operation monitoring method and system for the power industry, including in the actual application scenario of algorithm model operation monitoring in the power industry, this system realizes accurate monitoring and intelligent management of the operating status of power equipment through the collaborative operation of multiple highly integrated and functionally clear modules.

[0034] During system startup, the dynamic causal chain construction module immediately activates the sensor network, comprehensively covering key monitoring points on power equipment and collecting real-time data on physical quantities such as voltage, current, and temperature. This raw data first enters the power frequency noise suppression unit, where a notch filter with a center frequency of 50Hz precisely removes power frequency interference. The impulse noise cancellation unit then calculates the mean and variance of the current time series data using a sliding window of 10 sampling points. A noise threshold is dynamically set. When the amplitude of a data point exceeds the threshold, it is marked as an impulse noise point and replaced with the sliding average of the adjacent time series, ensuring data purity and accuracy. The preprocessed data stream seamlessly connects to the causal chain generator, which discerns the gradient of the change in the physical quantity and accurately matches it with the update direction of the algorithm model parameters. If the direction is consistent, a causal chain node is decisively established and assigned an initial weight of 0.5. The connection strength between nodes is dynamically updated once per second based on the product of the historical activation count N and the current data change rate, ensuring data timeliness and relevance. Once a physical quantity mutation event is detected, such as an amplitude exceeding a limit or a second-order derivative exceeding a threshold, the system immediately records the mutation event type, timestamp, and associated model parameter version number, generating a detailed mutation event log and storing it in the system database. This log is tightly bound to the causal chain nodes, providing a solid data foundation for subsequent analysis. If the mutation amplitude exceeds the preset stable range, the system traces back along the causal chain path, accurately locking the corresponding parameter dimension and freezing its update, and synchronously generates a hash value H. Using the SHA-256 algorithm, the input data is a concatenation of the physical quantity sequence P_seq, the algorithm model parameter version number, and the timestamp t. If the hash value verification of two consecutive time series fails, a data rollback is triggered, loading a verified copy of the causal chain from the cache and recalculating the associated path to ensure data integrity and reliability.

[0035] At the same time, the four-dimensional space-time lock synchronization module intervenes synchronously, and the geographic coordinate encoding unit quantifies the longitude L, latitude B and altitude H of the device into integer values ​​of one thousandth, generating a standard 12-bit geographic code GEO. The timestamp generator calculates the chaotic time series through the optical fiber transmission delay compensation algorithm, and substitutes the UTC time t, the optical fiber refractive index n, and the transmission distance d into the chaotic equation to generate an irreversible timestamp T, which gives the data a precise space-time imprint. In the data encapsulation stage, the causal chain hash value H_{t-1} of GEO and the previous time series is solemnly written into the data packet header. When parsing the data packet, the receiving end first verifies whether GEO is within the preset grid topology coordinate range. If it exceeds the range, the data packet is decisively discarded; then it verifies whether the hash sequence satisfies the strict monotonic increment. If the verification fails, the abnormal source IP address is quickly recorded and an alarm is triggered. When the spatiotemporal verification fails three times in a row, the system seamlessly switches to the backup communication channel. The channel is pre-injected with redundant data packets containing the mean values ​​of the physical quantities of the previous three valid time series, the benchmark values ​​of the model parameters stored in the system, and the complete GEO to ensure the continuity and stability of data transmission. During the operation of the backup channel, the multi-objective game strategy engine module suspends the generation of new strategies, maintains the output of the last valid strategy, and closely monitors the hash verification status of the main channel. If the main channel passes the timing recovery verification for five consecutive times, it will gradually return control according to the established process and synchronize the missing causal chain change records. The difference merging algorithm is used to align the timestamps and resolve conflicts between the backup channel data and the main channel data. When resolving conflicts, the operation with the latest timestamp is retained first; if the timestamps are the same, the main channel data is selected, and the conflict event is recorded in the audit log to ensure data consistency and traceability.

[0036] At the same time, the multi-objective game strategy engine module continues to work hard during the operation of the system. It extracts the type distribution matrix F of false alarm events in the past 24 hours and the resource consumption record matrix R from the historical false alarm event database stored in the system, and carefully constructs a false alarm rate-resource efficiency two-dimensional game matrix; according to the time constraint of the real-time detection task, strictly screens the strategy subset S required by the detection speed SLA in the matrix; uses the genetic algorithm to iteratively optimize S, sets the initial population size to 100, and reduces the crossover probability from 0.8 to 0.02 per generation. In each generation, the 20% individuals with the lowest false alarm rate are retained as strategies, and the remaining individuals are selected through mutation. The operation imposes a 5% random perturbation on the resource efficiency dimension. The optimization termination condition is that the strategy false alarm rate fluctuates by 1% and the resource efficiency improvement stagnates for more than 10 generations, and the final optimized strategy is output. If the false alarm rate suddenly increases by more than the historical average or the resource efficiency has not improved for three consecutive generations and the detection speed decreases by more than 5% during the iteration, the strategy backtracking mechanism is immediately triggered, the current iteration pool is cleared, and the game matrix of the previous stable version is loaded. The strategy weight is reset to prioritize detection speed, and the mutation perturbation range is limited to no more than 75% of the stability boundary. After three rounds of conservative iteration verification without conflict, optimization continues to ensure the stability and effectiveness of the strategy.

[0037] The DEED-Trigger counterintuitive triggering module remains vigilant, receiving the real-time causal chain data stream output by the dynamic causal chain construction module, including physical quantity sequences, mutation event records, and the hash value H_t of the current time series. It calculates the KL divergence between the policy's expected power distribution Q and the actual distribution P in real time. When the dynamic threshold is exceeded, it quickly initiates negative entropy oscillation detection, continuously monitoring the entropy change rate of the actual power within five time series windows. If the power fluctuation direction is opposite to the policy expectation, it is decisively determined to be a negative entropy oscillation. At this time, the hyperbolic space folding operation is triggered, mapping the device's geographic coordinates to the hyperbolic manifold space. The shortest path between nodes is calculated using the Poincaré disk model, using the curvature change rate as the path selection criterion. After the path is selected, its topology is reversely synchronized to the causal chain construction module, forcing the connection weight to be updated to 1.2 times the original value. The updated causal chain topology must pass the hash check of the four-dimensional space-time lock synchronization module. If the check fails, the previous version is rolled back and an alarm is triggered. Real-time verification is also performed on the verification platform to ensure the timeliness and accuracy of the emergency response.

[0038] The verification platform then injects a composite fault test every 30 minutes through digital twin technology according to a preset cycle. The fault mode is carefully designed as a concurrent scenario with abnormal deviations between the transformer bushing vibration amplitude and the circuit breaker electric field strength. During the test, the platform runs the real causal chain C_real and the virtual chain C_virtual with injected faults in parallel. If a node break is detected in C_virtual, the reconstruction engine calculates the priority based on the product Nw of the node's historical activation times N and the connection weight w, and prioritizes reconstructing nodes with Nw ≥ 100. At the same time, it refers to the mutation event log recorded by the dynamic causal chain construction module. If the node is associated with a high-frequency mutation event, such as the number of triggers in the past hour is ≥ 5, its priority is increased by 50%. The reconstructed causal chain needs to pass the full-node SHA-256 hash check and be compared for consistency with the current hash value H_t generated by the dynamic causal chain construction module, and the difference with the expected output of the digital twin platform is ≤5%. If the conditions are met, the original causal chain is replaced and takes effect. Otherwise, it rolls back to the previous version and marks the fault mode as unfixed to ensure the stability and reliability of the system.

[0039] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the operation of an algorithm model for the power industry, characterized in that: The steps include: Step 1. Collect physical quantity data from power equipment through a sensor network. After noise suppression and data preprocessing, generate a causal chain. This causal chain is then bound to the algorithm model parameters. Meanwhile, physical quantity mutation events and associated parameter update logs are recorded. Step 2. Add geographic coordinates and timestamps to the causal chain through a geographic coordinate encoding unit and a timestamp generator, and ensure data temporal and spatial consistency through hardware-level verification; Step 3. Based on the historical false alarm event type distribution and resource consumption records stored in the system, combined with the preset real-time detection task service level agreement (SLA) time constraints, an optimization strategy is generated to balance the false alarm rate, resource efficiency, and detection speed. Step 4. Trigger an emergency response based on the deviation between the real-time causal chain data in step 1 and the optimization strategy in step 3; Step 5. Use digital twin technology to inject composite faults to test the robustness of the causal chain, and dynamically adjust the causal chain structure based on the test results.

2. The algorithm model operation monitoring method for the power industry according to claim 1 is characterized by: In step 1, the raw data collected by the sensor is first input into the power frequency noise suppression unit, and the 50Hz power frequency interference is filtered out by a fixed-frequency notch filter. The filtered data is input into the impulse noise elimination unit, which calculates the mean and variance of the data distribution in real time and dynamically adjusts the noise judgment threshold. When the amplitude of a data point exceeds the threshold, it is marked as impulse noise and replaced with the sliding average of the adjacent time series; the preprocessed data stream is input into the causal chain generator, and the causal chain generator extracts the change gradient of the physical quantity and matches it with the update direction of the algorithm model parameters. If the physical quantity gradient is consistent with the adjustment direction of the algorithm model parameters, a causal chain node is established and assigned an initial weight. The connection strength between the nodes is dynamically updated according to the product of the historical activation times and the current data change rate; after the causal chain node is established, the type, timestamp and associated model parameter update log of the physical quantity mutation event are recorded in real time, and the log is bound to the causal chain node for storage.

3. The algorithm model operation monitoring method for the power industry according to claim 2 is characterized by: After the causal chain nodes are connected, the mutation events of physical quantities are monitored in real time. When it is detected that the mutation amplitude exceeds the preset stable range, the associated algorithm model parameters are traced back along the current causal chain path, the parameter category and its index that caused the mutation are locked, and their updates are frozen until the mutation event is resolved and passes the fault repair test of the verification platform; at the same time, the complete causal chain containing the physical quantity mutation event record is hashed, where the hash value is generated by splicing the physical quantity sequence, the algorithm model parameter version number and the timestamp. If the continuity check between the current hash value and the hash value of the previous time series fails, the data rollback is triggered, the verified causal chain copy is loaded from the cache, and the associated path of the mutation event is recalculated.

4. The algorithm model operation monitoring method for the power industry according to claim 1 is characterized by: In step 2, the geographic coordinate encoding unit quantizes the longitude, latitude and altitude of the device into integer values ​​of one thousandth, and splices them into a 12-bit geographic code; the timestamp generator calculates the chaotic time series through optical fiber transmission delay compensation, and associates the UTC time with the optical fiber refractive index and transmission distance to generate an irreversible timestamp; when encapsulating the data, the geographic code and the causal chain hash value of the previous time series are written into the data packet header, and the causal chain hash value is the latest hash value of the current time series generated by the dynamic causal chain construction module; the receiving end parses the packet header and performs a time-space check, first checking whether the geographic code is within the preset power grid topology range, and then verifying whether the hash value is strictly monotonically increasing. If any condition is not met, the data packet is discarded and the source of the abnormality is recorded, then it is determined that the time-space check has failed, the data packet is discarded and the source of the abnormality is recorded; when the number of time-space check failures reaches 3 times, the system automatically switches to the backup communication channel.

5. The algorithm model operation monitoring method for the power industry according to claim 4 is characterized in that: When the backup channel is started, a redundant data packet containing the mean value of the physical quantity of the first three valid time series, the reference value of the algorithm model parameters stored in the system, and the complete geocoding is injected into the link; during the operation of the backup channel, the generation of new strategies is suspended, the output of the last valid strategy is maintained, and the verification status of the main channel is monitored at the same time. If the main channel passes the time series recovery verification for five consecutive times, the control right is gradually returned to the main channel, and the causal chain change records that were missing during the redundancy period are synchronized.

6. The algorithm model operation monitoring method for the power industry according to claim 1 is characterized by: In step 3, the distribution of false alarm event types within the past 24 hours is extracted from the historical false alarm event records stored in the system, and resource consumption records are obtained from the system resource monitoring unit to construct a two-dimensional game matrix of false alarm rate and resource efficiency; secondly, based on the time constraints of the real-time detection task, a strategy subset whose detection speed meets the SLA requirements is screened in the game matrix; then, a genetic algorithm is used to iteratively optimize the subset, dynamically reducing the crossover probability in each generation, retaining the 20% individuals with the lowest false alarm rate as the strategy, and randomly perturbing the remaining individuals in the resource efficiency dimension through mutation operations until the false alarm rate fluctuation of the strategy is less than 1% and the resource efficiency improvement stagnates for more than 10 generations of iterations, and outputting the final optimization strategy.

7. The algorithm model operation monitoring method for the power industry according to claim 6 is characterized by: In step 3, if a policy conflict is detected during the iteration process, and the conflict determination condition is that the false alarm rate of the current policy increases by more than 2 standard deviations of the historical mean compared with the previous generation, or the resource efficiency has not been improved for three consecutive generations and the detection speed has dropped by more than 5%, then the policy backtracking mechanism is triggered: the current iteration pool is cleared, the game matrix of the previous stable version is loaded from the cache, the weight distribution of the policy is reset, the policy with the detection speed meeting the standard is retained first, and the perturbation range of the mutation operation is restricted so that it does not exceed 75% of the stability boundary; the reset game matrix needs to pass three rounds of conservative iteration to verify that there is no conflict before optimization can continue.

8. The algorithm model operation monitoring method for the power industry according to claim 1 is characterized by: In step 4, after the real-time causal chain data of step 1 is input, the KL divergence between the expected power distribution of the strategy and the actual power distribution is first calculated. When the divergence value exceeds the dynamic threshold, the negative entropy oscillation detection is started: the entropy change rate of the actual power in 5 time series windows is continuously monitored. If the entropy value continues to decrease and the direction of power fluctuation is opposite to the strategy expectation, it is determined to be a negative entropy oscillation; at this time, the hyperbolic space folding operation is triggered, and the geographical coordinates of the device are mapped to the hyperbolic manifold space. The shortest path between each node is recalculated in this space, and the path with a gentle curvature change is preferentially selected as the emergency strategy. The topological structure of the new path is synchronized back to step 1 to forcibly update the connection weight of the current causal chain.

9. The algorithm model operation monitoring method for the power industry according to claim 1 is characterized by: In step 5, a composite fault test is automatically injected at regular intervals through digital twin technology, where the fault mode includes a scenario where a sudden increase in transformer bushing vibration occurs concurrently with an abnormal electric field strength of the circuit breaker. During the test, the digital twin technology runs the real causal chain and the virtual chain with the injected fault in parallel. If a causal chain node is detected to be broken, the reconstruction engine calculates the priority based on the product of the node's historical activation times and the connection weight, and at the same time refers to the physical quantity mutation event log recorded by the dynamic causal chain construction module, giving priority to reconstructing nodes associated with high-frequency mutation events, and giving priority to reconstructing nodes with a product value greater than 100. The reconstructed causal chain must pass the full-node hash check and the difference with the expected output of the digital twin platform must be less than 5% before it can replace the original causal chain and take effect.

10. The algorithm model operation monitoring method for the electric power industry and the algorithm model operation monitoring system for the electric power industry according to claim 1 are characterized by: It includes a dynamic causal chain building module, a four-dimensional space-time lock synchronization module, a multi-objective game strategy engine module, a DEED-Trigger counter-intuitive triggering module, and a verification platform; The dynamic causal chain construction module collects physical quantity data of power equipment through the sensor network, generates a causal chain after noise suppression and data preprocessing, binds the causal chain to the algorithm model parameters, and records physical quantity mutation events and associated parameter update logs; The four-dimensional space-time lock synchronization module adds geographic coordinates and timestamps to the causal chain, ensuring data space-time consistency through hardware-level verification; The multi-objective game strategy engine module generates an optimization strategy by building a balance between false alarm rate, resource efficiency and detection speed based on the historical false alarm event type distribution and resource consumption records stored in the system and the preset real-time detection task service level agreement (SLA) time constraints; The DEED-Trigger counter-intuitive triggering module triggers an emergency response based on the deviation between the real-time data and the optimization strategy of the multi-objective game strategy engine module; The verification platform injects composite faults through digital twin technology to test the robustness of the causal chain, and dynamically adjusts the causal chain structure and strategy based on the test results.