Complex valued neural network dynamic modeling method for power grid fault identification and prediction

Abstract

The invention discloses a complex valued neural network dynamic modeling method for power grid fault identification and prediction, and relates to the field of smart power grids. According to the method, a separated complex valued neural network is constructed, voltage phasors and power vectors in a complex number form are input, and power grid characteristics are extracted through not less than 3 complex number convolution layers and not less than 2 full connection layers. And training the model by using historical loop closing operation data not less than 100,000 times, wherein the training period does not exceed 4 hours. The model can accurately predict loop closing current amplitude and output voltage deviation, and a continuous prediction value is converted into a switch state decision signal through an embedded binarization module. According to the invention, the precision and efficiency of power grid loop closing operation and fault prediction are improved, and core model support is provided for intelligent operation and maintenance of the power grid.

Description

[0001] This application is a divisional application of the following application: Application date 20250416, application number CN202510474595.2, invention title: A fault identification and prediction method based on low-voltage distribution network topology modeling and intelligent algorithm. Technical Field

[0002] This invention relates to the field of smart grids, specifically a dynamic modeling method for complex-valued neural networks used for power grid fault identification and prediction. Background Technology

[0003] As the final link in the power system, the low-voltage distribution network directly connects thousands of households, and its safety and reliability are crucial to economic development and people's livelihood. With rapid urbanization and increasing electricity demand, low-voltage distribution networks face challenges such as aging equipment, frequent faults, and difficulty in quickly determining the scope of power outages. These problems not only affect users' normal electricity use but also bring huge economic losses and social pressure to power companies. Researching fault prediction and outage area assessment technologies for low-voltage distribution networks has become a key issue in improving the safety and reliability of the power system.

[0004] As the "last mile" of the power system, the low-voltage distribution network bears the crucial responsibility of delivering electricity to end users, and its operational status directly impacts residents' lives, industrial production, and socio-economic stability. Due to its wide coverage, diverse equipment, and complex operating environment, the low-voltage distribution network has a failure rate exceeding 5%, resulting in millions to tens of millions of yuan in economic losses annually due to power outages caused by faults. Current operation and maintenance systems face three core problems: traditional manual inspection methods for fault location suffer from long response times and low accuracy, failing to meet the demands for rapid response; the lack of multi-source data integration capabilities in assessing the outage area leads to inaccurate results; and insufficient intelligence hinders the application of big data and artificial intelligence technologies.

[0005] Research on fault simulation and outage area assessment technologies is of significant necessity. Fault simulation technology allows power companies to predict and quickly locate faults. For example, the location method proposed by the China Electric Power Research Institute, based on current waveforms and flexible interconnection equipment, achieves accurate location even without data synchronization, significantly reducing processing time. Outage area assessment technology can quickly determine the outage area and restore power supply. Assessment technologies based on the four distribution network substations and PMS systems improve power system stability by integrating multi-source data. The application of these technologies also supports the intelligent transformation of the power industry, enabling intelligent management through the introduction of big data, artificial intelligence, and blockchain technologies.

[0006] The existing technological system has significant bottlenecks. In fault location technology, fault ranging methods are suitable for simple lines, while signal injection and outdoor detection methods are for complex lines and sections, respectively; however, both have limitations. Although power outage assessment technology can integrate multi-source data, insufficient intelligence limits its efficiency. Regarding the application of intelligent technologies, complex-valued neural networks and fuzzy networks improve the accuracy of prediction and diagnosis, and blockchain technology enhances data security; however, breakthroughs are still needed in key technologies such as dynamic modeling and multi-source fusion.

[0007] The future development of technology presents three major trends. First, the direction of intelligentization and digitalization will introduce more advanced artificial intelligence algorithms and big data analysis technologies to achieve real-time monitoring and intelligent management of distribution networks. Second, multi-source data fusion technology will become a research focus, improving the accuracy of analysis by integrating data from different devices and systems. Third, the application of blockchain technology will further ensure data security and trustworthiness, achieving tamper-proof recording of distribution network data. These development directions will drive the evolution of low-voltage distribution network operation and maintenance towards self-learning and self-optimization.

[0008] Research on fault simulation and outage area assessment technologies for low-voltage distribution networks is a core component in improving the safety and reliability of power systems. By introducing advanced technologies such as artificial intelligence, big data, and blockchain, power companies can achieve intelligent management of distribution networks and significantly improve fault handling efficiency. With continuous technological development, this field will provide sustained support for the intelligent transformation of the power industry, becoming a crucial infrastructure for ensuring energy security and promoting green and low-carbon development. Summary of the Invention

[0009] The technical problem to be solved by this invention is a fault identification and prediction method based on low-voltage distribution network topology modeling and intelligent algorithms. This method is based on a multi-module system and solves a series of technical problems such as fault location and resolution in low-voltage distribution networks.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows;

[0011] A fault identification and prediction method based on low-voltage distribution network topology modeling and intelligent algorithms is characterized by the following steps: data acquisition and preprocessing, complex-valued neural network dynamic modeling, fuzzy network reasoning, fault propagation path tracing, intelligent decision generation and verification, and system coordination and optimization.

[0012] As a preferred technical solution of the present invention, the steps are implemented through the collaborative operation of a multi-module system.

[0013] As a preferred technical solution of the present invention, the data acquisition and preprocessing are as follows: Voltage, current, power, and equipment status data of the power distribution network are collected in real time through an IoT terminal; data is collected at intervals of no more than 100ms using the MQTT protocol, and edge computing nodes are configured to achieve localized preprocessing with a latency of ≤200ms; a novel technology is used to preprocess the data, specifically: this technology cleans and repairs the power data through four consecutive steps: first, high-frequency noise is filtered out, including but not limited to instantaneous spikes in voltage data; then, missing values, including but not limited to periods not recorded due to sensor failure, are filled in by learning the change patterns of the data before and after using an intelligent model; next, sliding window and polynomial fitting techniques are used to eliminate periodic interference and make the data curve smoother; finally, by calculating the middle 50% distribution range of the data, outliers exceeding the reasonable range are marked; these four steps are executed sequentially to keep the data complete, smooth, and anomaly-free; the data is normalized, converting voltage and current to... The system establishes a time series database to store three months of historical data for model training and validation; a novel algorithm is introduced to dynamically adjust the data acquisition accuracy, ensuring high-quality and real-time data.

[0014] Furthermore, the novel algorithm is specifically described as follows: This algorithm is a dynamically adjusted prediction method that continuously optimizes the results by combining predicted values ​​and actual measured values; when the measured values ​​become unreliable, the algorithm reduces its dependence on the measured values ​​and relies more on the predicted values; when the measured values ​​return to normal, the algorithm re-trusts the measured values; this algorithm, which automatically adjusts the weights of prediction and measurement according to the actual situation, makes the results both accurate and stable.

[0015] As a preferred technical solution of the present invention, a novel approach is to introduce dynamic modeling using a complex-valued neural network, specifically by constructing a discrete complex-valued neural network and inputting a voltage phasor in complex form: With power vector: The algorithm extracts network features through at least three complex convolutional layers and at least two fully connected layers; it optimizes and trains the algorithm model, using complex techniques to stabilize the training process and training on at least 100,000 historical loop closing operation data, with a training cycle of ≤4 hours; it outputs the loop closing current amplitude with an error of ≤5% and the output voltage deviation with an error of ≤2%; it embeds a binarization module to convert continuous value prediction results into switch state decision signals; and it introduces an attention mechanism to optimize the model's feature extraction capability for key nodes, thereby improving prediction accuracy and response speed.

[0016] As a preferred technical solution of the present invention, the fuzzy network inference specifically involves: constructing a hierarchical network structure of "substation feeder area users" containing no less than 100 nodes; defining a 5-level trapezoidal membership function including but not limited to "normal", "early warning", and "fault" to quantify equipment status and environmental factors; inputs including but not limited to fuzzy variables such as transformer oil temperature and ambient humidity; updating the posterior probability of nodes through a belief propagation algorithm; and an inference response time ≤300ms. The outputs are the three most likely fault causes and their confidence levels. An expert knowledge base is integrated to support dynamic updates of fuzzy rules to adapt to changes in the distribution network operating environment. An incremental learning mechanism is introduced, and new fault cases automatically update the conditional probability table to improve the model's adaptability.

[0017] As a preferred embodiment of the present invention, the fault propagation path tracing specifically involves: establishing a fault feature library containing no fewer than 5,000 cases, defining, but not limited to, voltage surges. ≥15%, zero-sequence current The algorithm employs a novel fault propagation path tracking method, based on ≥30 Class A criteria. Specifically, this algorithm is a shortest path algorithm used to find the shortest path between two points in a network composed of nodes and edges. In distribution network fault analysis, this algorithm abstracts equipment including, but not limited to, substations, lines, and transformers as network nodes, and the connecting lines between devices as edges, assigning corresponding weights including, but not limited to, fault propagation time and line resistance values. When a device fails, the algorithm calculates the shortest path for the fault to propagate to other nodes by traversing all possible paths, helping to quickly locate the fault's impact range. In practical applications, the algorithm dynamically adjusts path weights in real time to accurately identify critical propagation paths, guiding maintenance personnel to prioritize cutting off dangerous lines, thereby effectively reducing the power outage area and shortening the recovery time. It utilizes DS evidence theory to integrate multi-source fault features, achieving a decision confidence level of ≥95%, and develops a fault signal propagation path tracking algorithm, improving search efficiency by 40%. It outputs a preliminary judgment of the fault type, location, and impact range, and combines a fuzzy logic rule base to handle uncertain features.

[0018] As a preferred technical solution of the present invention, the intelligent decision generation and verification specifically involves: First, using an improved three-state coding algorithm to quickly locate the fault area and combining it with the firefly algorithm to optimize the search path and improve search efficiency; then, generating detailed switch operation steps to restore power supply; when the system detects a malfunction of the protection device, it automatically triggers a secondary verification mechanism, reconfirms the cause of the fault through a probability analysis model, and then simulates the operation plan in a virtual power grid to ensure the plan is feasible and the recovery time is controlled within 15 minutes, with an accuracy rate of over 95%; the system synchronizes the real power grid status in real time, intuitively displays the fault location and operation sequence through a heat map, and supports modular collaboration.

[0019] As a preferred technical solution of the present invention, the system collaboration and optimization specifically includes: data synchronization between modules through a Kafka message queue; data preprocessing followed by synchronization to a complex value neural network and a fuzzy network; fault feature analysis results being fed back to the intelligent decision-making module for secondary verification; recovery scheme execution results being fed back to the data acquisition module to update the historical database, supporting online learning and dynamic optimization, and optimizing the association matrix weight coefficients based on historical data; integrating an expert knowledge base to support dynamic updates of fuzzy rules, adapting to changes in the distribution network operating environment and quantifying key technical parameters; and introducing blockchain technology to ensure data security and immutability, thereby improving the system's credibility and stability.

[0020] The beneficial effects of adopting the above technical solution are as follows: First, through dynamic modeling using complex-valued neural networks, the system can accurately capture the complex dynamic characteristics of the power system and predict the errors in loop current and voltage deviations, controlling them within 5% and 2% respectively, significantly improving the accuracy of fault prediction. Simultaneously, fuzzy network inference combined with an expert knowledge base and incremental learning mechanism can quickly identify fault causes and output confidence levels, with an inference response time ≤300ms, improving the real-time performance and reliability of fault diagnosis. Second, fault propagation path tracing is optimized. The shortest path algorithm and DS evidence theory are used to quickly locate the fault propagation path, helping maintenance personnel to prioritize cutting off dangerous lines, effectively reducing the outage area and shortening the recovery time, with a decision confidence level ≥95%. Furthermore, intelligent decision generation and verification optimize the search path through improved algorithms and the Firefly algorithm, generating detailed switching operation steps, with a recovery time controlled within 15 minutes and an accuracy rate exceeding 95%. The feasibility of the solution is ensured through virtual power grid simulation. System collaboration and optimization achieve data synchronization through Kafka message queues, ensuring the system can adapt to complex and ever-changing operating environments. The introduction of blockchain technology further enhances data security and immutability. These innovations not only improve power supply reliability and reduce power outage time for users, but also reduce inspection and maintenance costs, optimize resource allocation, and bring significant economic and social benefits, providing strong technical support for the intelligent development of the power system. Attached Figure Description

[0021] Figure 1 : A diagram showing the structural relationships of the main research content of the project;

[0022] Figure 2 Schematic diagram of a neural network;

[0023] Figure 3 Technology roadmap;

[0024] Figure 4 : Roadmap for Fuzzy Bayesian Networks in Power Distribution;

[0025] Figure 5: Schematic diagram of a fuzzy Bayesian network;

[0026] Figure 6 Fault diagnosis technology roadmap. Detailed Implementation

[0027] The principles of this disclosure will now be described with reference to several exemplary embodiments illustrated in the accompanying drawings. Although exemplary embodiments of this disclosure are shown in the drawings, it should be understood that these embodiments are described merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way.

[0028] In the following description of the embodiments, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0029] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described feature, integral, step, operation, or component, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, or components.

[0030] Example 1: Overall Steps Overview

[0031] 1. Method Overview

[0032] Reference Appendix Figure 1 This method aims to achieve accurate identification and prediction of faults in low-voltage distribution networks through the collaborative work of multiple modules. Specific steps include: data acquisition and preprocessing, dynamic modeling using complex-valued neural networks, fuzzy network inference, fault propagation path tracing, intelligent decision generation and verification, and system coordination and optimization. Through these steps, the system can quickly locate faults, predict fault propagation paths, and generate efficient recovery solutions, significantly improving the fault handling efficiency and power supply reliability of the distribution network.

[0033] 2. Data Acquisition and Preprocessing

[0034] Data acquisition and preprocessing are fundamental to fault identification and prediction. Voltage, current, power, and equipment status data of the distribution network are collected in real time via IoT terminals. Data is acquired at intervals of no more than 100ms using the MQTT protocol, and edge computing nodes are configured for localized preprocessing to ensure data processing latency is ≤200ms. Data preprocessing includes the following four consecutive steps: High-frequency noise filtering: removing high-frequency noise such as instantaneous spikes from voltage data. Missing value imputation: filling in missing values ​​caused by sensor malfunctions by learning the changes in data before and after using an intelligent model. Periodic interference elimination: eliminating periodic interference in the data using sliding window and polynomial fitting techniques to smooth the data curves. Outlier detection: marking and processing outliers that exceed the reasonable range by calculating the intermediate distribution range of the data.

[0035] Finally, the data is normalized to convert voltage and current to the [0,1] interval, and a time series database is established to store three months of historical data for model training and validation.

[0036] 3. Dynamic Modeling of Complex-Valued Neural Networks

[0037] Dynamic modeling using complex-valued neural networks is one of the core innovations of this method. By constructing a separable complex-valued neural network, inputting voltage phasors and power vectors in complex form, and extracting network features using at least three complex convolutional layers and two fully connected layers, the network is trained. During model training, complex techniques are employed to stabilize the training process, and at least 100,000 historical loop-closing operation data points are used for training, with a training cycle of ≤4 hours. The model outputs the loop-closing current amplitude and voltage deviation, with errors controlled within 5% and 2%, respectively. Furthermore, an embedded binarization module converts continuous value prediction results into switching state decision signals, and an attention mechanism is introduced to optimize the model's feature extraction capability for key nodes, further improving prediction accuracy and response speed.

[0038] 4. Fuzzy Network Inference

[0039] Fuzzy network inference is used to handle uncertainties and fuzzy information in distribution networks. A hierarchical network structure of "substation feeder area users" is constructed, containing no fewer than 100 nodes. A five-level trapezoidal membership function (e.g., "normal," "early warning," "fault") is defined to quantify equipment status and environmental factors. Fuzzy variables such as transformer oil temperature and ambient humidity are input, and the posterior probabilities of nodes are updated through a belief propagation algorithm, with an inference response time ≤300ms. The three most probable fault causes and their confidence levels are output, and an expert knowledge base is integrated to support dynamic updates of fuzzy rules to adapt to changes in the distribution network operating environment. An incremental learning mechanism is introduced, automatically updating the conditional probability table for new fault cases, improving the model's adaptability.

[0040] 5. Fault propagation path tracing

[0041] Fault propagation path tracing is achieved by establishing a fault feature database containing no fewer than 5000 cases, defining criteria such as voltage surge ΔV ≥ 15% and zero-sequence current I0 ≥ 30A, and combining this with a shortest path algorithm to track fault propagation paths. This algorithm abstracts equipment such as substations, lines, and transformers as network nodes, with connecting lines between devices as edges, and assigns corresponding weights to the edges (such as fault propagation time and line resistance). When a device fails, the algorithm traverses all possible paths to calculate the shortest path for the fault to propagate to other nodes, helping to quickly locate the scope of the fault's impact. In practical applications, the algorithm dynamically adjusts path weights in real time, accurately identifies critical propagation paths, and guides maintenance personnel to prioritize cutting off dangerous lines, thereby effectively reducing the power outage area and shortening the restoration time. Furthermore, it utilizes DS evidence theory to fuse multi-source fault features, achieving a decision confidence level of ≥ 95%, and incorporates a fuzzy logic rule base to handle uncertain features.

[0042] 6. Intelligent Decision Generation and Verification

[0043] Intelligent decision generation and verification are achieved through the following steps: Fault area location: An improved three-state coding algorithm is used to quickly locate the fault area, and the search path is optimized by combining the firefly algorithm to improve search efficiency. Recovery plan generation: Detailed switching operation steps are generated to restore power supply. Secondary verification mechanism: When the system detects a malfunction of the protection device, secondary verification is automatically triggered to reconfirm the cause of the fault through a probabilistic analysis model. Virtual power grid simulation: The operation plan is simulated in a virtual power grid to ensure that the plan is feasible and the recovery time is controlled within 15 minutes, with an accuracy rate of over 95%. Visualization and collaboration: The system synchronizes the real power grid status in real time, intuitively displays the fault location and operation sequence through a heat map, and supports modular collaboration.

[0044] 7. System Coordination and Optimization

[0045] System collaboration and optimization are achieved through the following methods: Data Synchronization: Data synchronization between modules is achieved through Kafka message queues. After data preprocessing, it is synchronized to the complex-valued neural network and fuzzy network. Fault feature analysis results are fed back to the intelligent decision-making module for secondary verification. Online Learning and Dynamic Optimization: The execution results of the recovery plan are fed back to the data acquisition module to update the historical database, supporting online learning and dynamic optimization, and optimizing the association matrix weight coefficients based on historical data. Expert Knowledge Base Integration: An expert knowledge base is integrated, supporting dynamic updates of fuzzy rules to adapt to changes in the distribution network operating environment and quantifying key technical parameters. Blockchain Technology Application: Blockchain technology is introduced to ensure data security and immutability, improving the system's credibility and stability.

[0046] 8. Summary

[0047] This method, through multi-module collaborative operation and combining advanced technologies such as complex-valued neural networks, fuzzy network inference, shortest path algorithms, and intelligent decision generation and verification, achieves accurate identification and prediction of faults in low-voltage distribution networks. The system can quickly locate faults, predict fault propagation paths, and generate efficient recovery plans, significantly improving the fault handling efficiency and power supply reliability of distribution networks. In the future, with continuous optimization and application of the technology, this method will provide strong support for the intelligent transformation of the power industry.

[0048] Example 2: Intelligent handling of single-phase grounding faults in low-voltage distribution networks in residential areas

[0049] A single-phase ground fault occurred in a residential area due to aging wiring. The system achieved rapid fault location and recovery through multi-module collaboration.

[0050] 1. Data Acquisition and Preprocessing:

[0051] Data such as voltage (220V±0.5%), current (0500A), transformer oil temperature (85℃), and humidity (75%) are collected at 100ms intervals via FTU / DTU terminals and transmitted to edge computing nodes using the MQTT protocol, with a preprocessing delay of 180ms.

[0052] Wavelet thresholding is used to filter out voltage spikes, LSTM is used to complete the missing 10 minutes of current data (error ≤3%), SavitzkyGolay filtering is used to eliminate 50Hz current noise, and IQR is used to detect and mark outliers with current >400A.

[0053] After normalization, three months of historical data are stored in a time-series database, supporting queries at the 50ms level.

[0054] 2. Complex-valued neural network modeling:

[0055] Reference Appendix Figure 2 A split complex value neural network (SCVNN) is constructed, which takes a complex voltage phasor (220V∠0°) and a power vector (100+j50kVA) as input, and extracts features through three complex convolutional layers and two fully connected layers.

[0056] Training based on 100,000+ historical closed-loop data, the complex mean square error is ≤0.01, the predicted closed-loop current error is 4.2%, and the voltage deviation is 1.8%.

[0057] 3. Fuzzy Bayesian inference:

[0058] Reference Appendix Figure 5 A hierarchical network (120 nodes) is constructed for "substation feeder A, transformer area 1, user 1", and oil temperature of 85℃ is defined as the "early warning" membership degree.

[0059] Input oil temperature and humidity data, update node probabilities through belief propagation algorithm, and output the top 3 fault causes: line insulation aging (92%), insulator damage (88%), and short circuit caused by small animals (85%). The inference response time is 280ms.

[0060] 4. Fault path tracing:

[0061] The single-phase grounding mode with "zero-sequence current 35A, ΔV=18%" in the matching feature library is used to calculate the shortest path algorithm to determine the propagation path of the fault from transformer substation 1 to user 1 (weight: resistance + fault time).

[0062] By integrating the DS evidence theory, the decision confidence level is 96% and the search efficiency is improved by 40%.

[0063] 5. Intelligent decision-making and verification:

[0064] The three-state coding matrix algorithm locates the fault section (98% accuracy), and the firefly algorithm generates the switch operation sequence (disconnecting the feeder A circuit breaker → closing the tie switch).

[0065] When a false protection action is detected, a Bayesian secondary verification is triggered, and a digital twin simulation is used to verify the recovery scheme. The recovery time is 12 minutes, and the verification accuracy is 97%.

[0066] 6. System collaborative optimization:

[0067] Kafka synchronously preprocesses data to complex value neural networks and fuzzy networks, and the recovered results are fed back to update the historical database.

[0068] Blockchain records operation logs to ensure data immutability, optimizes the weight coefficients of the association matrix online, and improves positioning efficiency by 15%.

[0069] Example 3: Efficient Handling of Phase-to-Phase Short Circuit Faults in Industrial Zone Distribution Networks

[0070] Reference Appendix Figure 3 In an industrial park, a phase-to-phase short circuit occurred due to equipment overload. The system achieved precise fault handling through the integration of multiple technologies:

[0071] 1. Data Acquisition and Preprocessing:

[0072] The system collects three-phase voltage (380V±0.5%), current (01000A) and line load rate (85%). The edge computing preprocessing delay is 190ms, and the data integrity is 99.8%.

[0073] Wavelet noise reduction filters out current spikes, LSTM completes the missing 5 minutes of data (error ≤ 2.5%), Savitzky Golay filtering eliminates periodic load fluctuations, and IQR detects outliers with marked voltage < 300V.

[0074] 2. Complex-valued neural network modeling:

[0075] The input voltage phasor (380V∠30°) and power vector (200+j100kVA) are trained through three complex convolutional layers, and the prediction error of the closed loop current is 4.8% and the voltage deviation is 1.5%.

[0076] Training with 120,000 combined loop cases, training cycle of 3.8 hours, error ≤0.01.

[0077] 3. Fuzzy Bayesian inference:

[0078] Construct a network of "substation feeder B, area 2, user 2" (115 nodes), and define the load rate as "high" corresponding to the 75%-90% range.

[0079] Inference output: Cable insulation breakdown (94%), equipment overload (91%), lightning strike (87%), response time 290ms.

[0080] 4. Fault path tracing:

[0081] The shortest path algorithm determines the fault path (weight: fault propagation time) by matching the phase-to-phase short circuit characteristics of "ΔV=22% and I0=40A".

[0082] By integrating multi-source features, the decision confidence level reaches 95%, and the search time is reduced to 0.3 seconds.

[0083] 5. Intelligent decision-making and verification:

[0084] Reference Appendix Figure 6 The three-state matrix algorithm locates the fault section (accuracy of 98.5%), and the firefly algorithm generates the operation sequence (disconnect the feeder B section switch → close the backup line).

[0085] The digital twin verification solution has a recovery time of 14 minutes and a verification accuracy rate of 96%.

[0086] 6. System collaborative optimization:

[0087] The fault characteristic analysis results are fed back to the decision-making module for secondary verification, and the expert knowledge base dynamically updates the load rate threshold.

[0088] Online learning optimizes the correlation matrix, improving positioning efficiency by 15% and supporting the rapid recovery of industrial load.

[0089] Example 4: Parallel processing of multiple faults under extreme weather conditions

[0090] Reference Appendix Figure 4 Typhoon weather caused simultaneous phase-to-phase short circuits and single-phase grounding faults in a certain area. The system achieved parallel handling through multi-module collaboration:

[0091] 1. Data Acquisition and Preprocessing:

[0092] Data collected included voltage (220V±0.5%), current (0500A), wind speed (12m / s), and rainfall (50mm / h). Edge computing preprocessing had a delay of 170ms, and the data integrity was 99.7%.

[0093] Wavelet noise reduction was used to process voltage fluctuations, LSTM was used to complete three missing data points (error ≤ 3%), Savitzky Golay filtering was used to eliminate wind and rain interference, and IQR was used to detect and mark outliers with current > 450A.

[0094] 2. Complex-valued neural network modeling:

[0095] The input voltage phasor (220V∠15°) and power vector (80+j40kVA) were trained with 150,000 cases, and the predicted loop current error was 4.5% and the voltage deviation was 1.7%.

[0096] The training cycle is 3.2 hours, and the error is ≤0.01.

[0097] 3. Fuzzy Bayesian inference:

[0098] Construct a network of "substation feeder C area 3 users 3" (130 nodes), and define the wind speed "extremely high" as the interval of 1015m / s.

[0099] Inference output: Line short circuit (93%), insulator flashover (90%), tree crushing the line (88%), response time 270ms.

[0100] 4. Fault path tracing:

[0101] Matching the multi-fault characteristics of "ΔV=20% and I0=35A", the shortest path algorithm determines two fault paths (area 3→user 3 and area 4→user 4).

[0102] The DS evidence theory fusion resulted in a decision confidence level of 97% and a 42% improvement in search efficiency.

[0103] 5. Intelligent decision-making and verification:

[0104] The three-state matrix algorithm can simultaneously locate two fault sections (accuracy of 98.2%), while the firefly algorithm generates a dual-path operation sequence with a recovery time of 13 minutes.

[0105] When a false protection action is detected, a secondary verification is triggered, with a digital twin verification accuracy rate of 98%.

[0106] 6. System collaborative optimization:

[0107] Kafka enables real-time data synchronization across multiple modules, while the blockchain records data throughout the entire process.

[0108] The expert knowledge base updates humidity and wind speed thresholds based on typhoon data, and optimizes the correlation matrix through online learning, improving positioning efficiency by 18%.

[0109] A comparison of the effects of the above embodiments

[0110] Technical indicators Example 2 Example 3 Example 4 Fault location accuracy 98.0% 98.5% 98.2% Average recovery time 12 minutes 14 minutes 13 minutes Data preprocessing delay 180ms 190ms 170ms Complex value network training error 0.009 0.008 0.007 Bayesian inference response time 280ms 290ms 270ms Digital twin verification accuracy 97% 96% 98% Online learning weight optimization cycle 2 hours 2.5 hours 1.8 hours

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic modeling method for complex-valued neural networks for power grid fault identification and prediction, characterized in that: Construct a discrete complex-valued neural network and input voltage phasors in complex form: With power vector: Network features are extracted through at least three complex convolutional layers and at least two fully connected layers; The algorithm model is optimized and trained. During the training process, complex number techniques are used to stabilize the training process and training is performed on no less than 100,000 historical loop closing operation data. The training cycle is ≤4 hours. The output loop closing current amplitude has an error of ≤5%, and the output voltage deviation has an error of ≤2%. The embedded binarization module converts the continuous value prediction results into switch state decision signals.

2. The complex-valued neural network dynamic modeling method for power grid fault identification and prediction according to claim 1, characterized in that, An attention mechanism is also introduced to optimize the model's ability to extract features from key nodes. Specifically, this mechanism is integrated after the complex convolutional layers and / or fully connected layers to calculate the correlation weights between features of different nodes in the network.

3. The complex-valued neural network dynamic modeling method for power grid fault identification and prediction according to claim 1, characterized in that, The amplitude of the voltage phasor is 220V or 380V.

4. The complex-valued neural network dynamic modeling method for power grid fault identification and prediction according to claim 1, characterized in that, The training times for the historical loop merging operation data are 100,000, 120,000, or 150,000 times.

5. The method for dynamic modeling of complex-valued neural networks for power grid fault identification and prediction according to claim 1, characterized in that, The training cycle is 3.2 hours, 3.8 hours, or 4 hours.

6. The complex-valued neural network dynamic modeling method for power grid fault identification and prediction according to claim 1, characterized in that, The loop current amplitude error of the complex-valued neural network output is 4.2% to 4.8%, and the voltage deviation error is 1.5% to 1.8%.

Images