Power anomaly processing method and system for self-organizing closed-loop feedback driving iterative optimization

Abstract

The invention provides a self-organizing closed-loop feedback driven iterative optimization power exception processing method and system, and relates to the technical field of power exception processing. Comprising the steps of inputting modal data into a hybrid expert model, and obtaining comprehensive features of the multi-modal data through a weighted fusion method; based on the comprehensive characteristics, detecting the abnormality of the power system; adopting a reinforcement learning method to generate a plurality of processing suggestions for the power system abnormity; obtaining feedback information of a user on various processing suggestions, inputting the feedback information into the self-organizing mapping model for clustering, and dynamically adjusting the weight of each modal data in the hybrid expert model based on a clustering result; the feedback information serves as a reward for generating multiple processing suggestions in reinforcement learning, and generation of the multiple processing suggestions in the next round is optimized; and iteration is carried out until an optimal processing suggestion is obtained. According to the method, the accuracy, the real-time performance and the individuation degree of anomaly detection and processing can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system abnormality processing, and in particular relates to a power abnormality processing method and system driven by self-organizing closed-loop feedback iterative optimization. Background Art

[0002] As the scale and complexity of power systems continue to increase, traditional power operation and maintenance methods can no longer meet the high requirements of modern power systems for real-time, accuracy and efficiency. Power system anomaly detection and processing is crucial to ensuring the stability and safety of the system. It can promptly detect equipment failures, operational anomalies and other problems, thereby avoiding or reducing power outages and ensuring the continuity and stability of power supply.

[0003] Most of the existing power system anomaly detection methods rely on manual inspections, rule-based detection methods, and monitoring of a single data source. These methods are not only inefficient, but also susceptible to human subjectivity and external environmental factors, making it difficult to achieve large-scale, high-frequency real-time monitoring and rapid response.

[0004] Under the limitations of traditional methods, anomaly detection in power systems usually faces the following problems: First, the data source is single, which makes it difficult to provide multi-dimensional and multi-angle system status information. Monitoring from a single data source (such as relying only on sensor data) is often difficult to accurately capture various abnormal conditions of the system.

[0005] Secondly, traditional anomaly detection methods are mostly based on fixed rule settings or historical thresholds and lack the ability to adapt and self-learn. In practical applications, the power system environment is complex and changeable, and static rules alone are difficult to cope with frequently updated system states, resulting in missed detections and false detections. At the same time, existing methods also lack a dynamic optimization mechanism, making it difficult to continuously improve detection performance based on feedback, and unable to achieve continuous optimization of anomaly handling solutions.

[0006] In addition, the abnormal detection and processing solutions in the power system usually fail to fully consider the changes in user feedback and actual operation requirements, and lack the ability to adapt to personalized needs. In actual operation, users often choose different processing solutions based on specific circumstances, and traditional systems find it difficult to use user choices and feedback to generate optimized solutions. Summary of the invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a self-organizing closed-loop feedback driven iterative optimization power anomaly processing method and system, which automatically generates and iteratively optimizes anomaly processing solutions based on the feedback of multi-dimensional information, allowing users to select processing suggestions that best meet actual needs, and uses user selections as input for further optimization, thereby forming an adaptive closed-loop feedback system, which effectively improves the accuracy, real-time and personalization of anomaly detection and processing.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: A first aspect of the present invention provides a power anomaly processing method driven by self-organizing closed-loop feedback iterative optimization.

[0009] The power anomaly processing method driven by self-organizing closed-loop feedback iterative optimization includes the following steps: Step 1: Obtain multimodal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multimodal data through the weighted fusion method; Step 2: Detect power system anomalies based on comprehensive features; Step 3: Use reinforcement learning methods to generate multiple treatment suggestions for power system anomalies; Step 4: Obtain user feedback on various processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering results; Step 5: Use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning to optimize the generation of multiple processing suggestions in the next round; Step 6: Iterate the above steps 1 to 5 until the optimal processing suggestion is obtained.

[0010] A second aspect of the present invention provides a power anomaly processing system driven by self-organizing closed-loop feedback and iterative optimization.

[0011] The self-organizing closed-loop feedback-driven iteratively optimized power anomaly processing system includes: The feature fusion module is configured to: obtain multi-modal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multi-modal data through the weighted fusion method; The anomaly detection module is configured to: detect anomalies of the power system based on comprehensive features; The suggestion generation module is configured to: generate a plurality of treatment suggestions for power system anomalies using a reinforcement learning method; The first closed-loop feedback module is configured to: obtain user feedback information on multiple processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering result; The second closed-loop feedback module is configured to: use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning, and optimize the generation of the next round of multiple processing suggestions; The iterative optimization module is configured to iterate the feature fusion module to the second closed-loop feedback module until the optimal processing suggestion is obtained. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization as described in the first aspect of the present invention.

[0012] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the self-organizing closed-loop feedback driven iterative optimization method for handling power anomalies as described in the first aspect of the present invention are implemented.

[0013] One or more of the above technical solutions have the following beneficial effects: The present invention creatively provides a self-organizing closed-loop feedback driven iterative optimization method and system for handling power anomalies. The proposed multi-information feedback self-organizing iterative technology breaks through the limitations of traditional anomaly detection and processing methods. Traditional methods usually rely on a single data source and fixed rules, which are difficult to cope with complex power system scenarios, resulting in insufficient accuracy and flexibility in monitoring and processing. In contrast, the present invention uses multi-source data such as text, images, and audio through multimodal data fusion, so that the system can fully perceive and understand the operating status of power equipment, thereby generating more comprehensive and accurate exception handling suggestions. This self-organizing iterative optimization method based on multi-source information enables the system to continuously fine-tune the exception handling suggestions according to the latest feedback in each round of iteration, and can automatically adapt to new abnormal situations in complex and changeable power scenarios, greatly improving the system's exception handling efficiency and accuracy.

[0014] The present invention introduces a closed-loop mechanism of user selection and feedback, and incorporates the user's selection and feedback during the exception handling process into the system's adaptive learning process. This closed-loop feedback system not only allows users to select the processing suggestions that best meet actual needs from multiple solutions, but also uses the user's selection as the input for the next round of iterations of the system, so that the system can gradually adapt to user preferences and personalized needs in the process of self-optimization. By continuously adjusting the generation algorithms and strategies, the system can respond to changes in user needs in real time and provide a personalized solution for power exception handling. The significance of this technology is that by combining user feedback with adaptive optimization, the intelligent and personalized handling of power system exceptions is realized, which greatly improves the accuracy and flexibility of exception detection and processing, and provides strong support for more efficient smart grid management and power system operation and maintenance.

[0015] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0017] Figure 1 This is a flow chart of the method of the first embodiment. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0019] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0020] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0021] Embodiment 1 This embodiment discloses a power anomaly processing method driven by self-organizing closed-loop feedback and iterative optimization.

[0022] like Figure 1 As shown, the power anomaly processing method driven by self-organizing closed-loop feedback iterative optimization includes the following steps: Step 1: Obtain multimodal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multimodal data through the weighted fusion method; Step 2: Detect power system anomalies based on comprehensive features; Step 3: Use reinforcement learning methods to generate multiple treatment suggestions for power system anomalies; Step 4: Obtain user feedback on various processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering results; Step 5: Use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning to optimize the generation of multiple processing suggestions in the next round; Step 6: Iterate the above steps 1 to 5 until the optimal processing suggestion is obtained.

[0023] The embodiment of the present invention proposes a self-organizing iterative technology that integrates multi-source feedback data, and realizes dynamic optimization of power anomaly processing through multimodal data fusion, adaptive learning and closed-loop feedback mechanism. This technology can not only automatically generate and iteratively optimize the anomaly processing scheme based on the feedback of multivariate information, but also allow users to select the processing suggestions that best meet actual needs, and use the user's selection as input for further optimization, thereby forming an adaptive closed-loop feedback system. This technology has broad application prospects in ensuring the safety and stability of power systems, and can effectively improve the accuracy, real-time and personalization of anomaly detection and processing.

[0024] The implementation steps of this embodiment will be explained in detail below.

[0025] Step 1: In the power system, in order to accurately identify and handle abnormal situations, it is necessary to obtain relevant information from multiple data sources and pre-process the data. By bringing together multi-source data (such as equipment operation data, user input, historical operation records, and external environment data), the system can have a more comprehensive understanding of the operating status of power equipment, thereby enhancing the reliability and accuracy of anomaly detection. The collected data may come from IoT devices, monitoring systems, and power management systems, providing a rich source of information for subsequent analysis and modeling.

[0026] Furthermore, the data collection part is completed through multi-channel and multi-source data integration: Equipment operation data is mainly collected through IoT devices, including sensors, smart meters, circuit breaker monitors, etc., which can obtain equipment status data, operating parameters (such as voltage, current, temperature, etc.) and abnormal alarm records in real time.

[0027] User input data comes from the operator's logs, input records, and feedback information; this data can be collected through the human-computer interaction interface, mobile device application, or operation terminal.

[0028] Historical operation records include past maintenance and repair records and previous exception handling reports. These data can help the system understand the maintenance history and failure modes of the equipment, thereby providing a reference for anomaly detection.

[0029] External environmental data is obtained through integrated meteorological monitoring systems, geographic information systems (GIS), and other channels, covering factors such as temperature, humidity, wind speed, and rainfall. These data also have an important impact on the operation of the power system.

[0030] All collected data are sent to the data processing center through the real-time transmission module to ensure that the anomaly detection system has full awareness of the power system status.

[0031] Furthermore, in the data preprocessing process, special technical means are used to process different data types. For image data (such as equipment site photos, monitoring screenshots, etc.), it may include image denoising, brightness adjustment and other operations; for text data (such as logs, alarm information, etc.), it includes removing invalid fields, removing non-text content (emoticons, video links), removing noise (HTML tags, special characters), etc.; for audio data (such as alarm sounds emitted by equipment, environmental noise, etc.), preprocessing includes frequency domain analysis and removing environmental noise.

[0032] Step 2: After the preprocessing of multi-source data is completed, the system needs to fuse multiple data such as text, images and audio to more comprehensively understand the abnormal state of the power system. By using the hybrid expert cross-modal fusion method, the features of different data types can be vectorized and integrated into a unified multimodal feature representation to form initial abnormality handling suggestions. This step ensures that the system can fully utilize the advantages of multimodal data to achieve more efficient and accurate anomaly detection and processing.

[0033] Furthermore, the system first extracts features from different modal data. For text data, large-scale pre-trained models such as GPT-3 and GPT-4 can be used for processing. These models generate contextual representations of text through autoregression and can capture rich semantic information in the text. Assuming that the input text data is Ti, the text feature representation after processing by the GPT model is VT, and the formula is:

[0034] Based on its autoregressive structure, the GPT model can generate text representations with contextual relationships, enabling the system to better understand complex text data such as logs and equipment status descriptions in the power system.

[0035] For image data, Vision Transformer (ViT) can be used for processing. The ViT model divides the image into a series of fixed-size image patches, flattens these image patches into a linear sequence, and extracts features through a multi-layer self-attention mechanism. The ViT model can capture local and global features in the image and has a good understanding of the spatial relationship between different regions in the image. Assuming that the input image data is Ii, the image feature after processing by the ViT model is represented as VI, and the formula is:

[0036] The self-attention mechanism of the ViT model enables it to flexibly focus on key areas of the image, thereby generating image representations with spatial context information, enabling the system to perform well in tasks such as fault detection of power equipment images and equipment status monitoring.

[0037] For audio data, the Transformer for Audio (TFA) model can be used for processing. The TFA model converts audio signals into spectrograms or time-frequency features, and analyzes the relationship between the time and frequency domains of the audio through the self-attention mechanism. This model can effectively capture the timing characteristics and frequency changes in audio, and is suitable for identifying alarm sounds, abnormal equipment noise, etc. Assuming that the input audio data is Ai, the audio features processed by the TFA model are represented as VA, and the formula is:

[0038] The self-attention mechanism of the TFA model can focus on key time-frequency information when processing audio data, thereby generating an audio representation with time and frequency domain contextual relationships, enabling the system to more accurately analyze audio abnormal signals in the power system, such as identifying abnormal operating sounds or alarm signals of equipment.

[0039] The data of each modality is converted into feature vectors respectively, preparing for the next step of multimodal information fusion.

[0040] Furthermore, after the feature extraction is completed, the system performs multimodal feature fusion on these feature vectors. In this embodiment, efficient feature fusion of multimodal data is performed based on a hybrid expert model. The hybrid expert model (MOE) distributes multimodal input data to the most appropriate expert network through a gating network to specifically process the data features of each modality. The MOE model can intelligently select the best experts to process different modalities (such as text, image, and audio data), ensuring that the data of each modality is optimally feature extracted, thereby improving the efficiency and accuracy of the system.

[0041] For input data of different modalities (text, image, audio), the gating network first assigns each modality to the corresponding expert network (such as the Vision Transformer (ViT), GPT, Transformer for Audio (TFA) models mentioned above) for feature extraction and generates the corresponding modal feature representation.

[0042] Assume that the input modal data is text data: , image data: , audio data: , the data of each modality is extracted through its corresponding expert network to obtain independent feature representation:

[0043] in, , , Respectively represent expert networks for processing text, image, and audio data; the output feature representation is , and , corresponding to the characteristics of each mode respectively.

[0044] The gating network generates weights independently according to the content of each modality feature to determine the importance of each modality in the final fusion. Specifically, the gating network generates a weight value for each modality data, indicating the contribution of the modality feature in multimodal fusion. Assuming that the weights of text, image and audio data are gT, gI, gA, the weight of each modality is calculated as follows:

[0045] Among them, gi is the weight generated by the gating network for modality i, where i is one of T, I, and A, and T, I, and A represent text, image, and audio data respectively; σ is the activation function (such as Sigmoid), Wg and bg are the weight and bias terms of the gating network respectively, Represents the input data of a certain modality.

[0046] This independent weight distribution method avoids the direct splicing of modal features, making the features of each modality act independently, thereby effectively reducing the interference between modalities. After obtaining the weight and feature representation of each modal feature, the features of each modality are weighted and fused through the weight of the gating network to generate the final multimodal fusion feature. :

[0047] This weighted fusion method enables the system to flexibly adjust the importance of each modal feature in the fusion according to the modal characteristics of the current input data, thereby improving the effectiveness and information richness of multimodal feature fusion. The information is then fed into a subsequent deep learning model for further processing, such as a Transformer model. Through further deep feature extraction, the system can extract deeper semantic information from the fused features, improving the accuracy and robustness of anomaly detection:

[0048] That is the extracted comprehensive features.

[0049] Through this fusion method, data from different modalities are integrated into a comprehensive feature vector with richer information content, which can more comprehensively reflect the current status of the equipment.

[0050] Furthermore, the system uses the fused multimodal feature vector to generate initial exception handling suggestions. The feature vector contains key information from multi-source data, and the system will generate a preliminary exception handling plan based on this information. The generated exception handling suggestions include the type of detected anomaly, the degree of impact, and the recommended preliminary handling measures. After multimodal information fusion, the system can more effectively integrate the advantages of multi-source data to form a multi-dimensional and comprehensive anomaly detection and handling plan for the power system, thus laying the foundation for subsequent feedback iteration and optimization.

[0051] Step 3: After completing the fusion of multimodal information, the system inputs the fused feature vector into the anomaly detection module, and uses the Trie-based detection agent (TDA) module for real-time analysis to identify potential abnormal events. The TDA module uses a lightweight Trie structure to parse the data level by level in an efficient manner to generate preliminary anomaly handling suggestions. This step converts the detected abnormal conditions into multiple processing solutions, providing diverse options for subsequent user selection and feedback.

[0052] Furthermore, firstly, the TDA module quickly detects anomalies through a lightweight Trie structure. The Trie structure is a prefix tree structure suitable for high-frequency, low-latency data processing, which can efficiently store and retrieve abnormal templates. In the power system, when each piece of data flows in, the TDA module will match it with the nodes in the Trie tree layer by layer, and detect potential abnormal events based on predefined abnormal patterns. The advantage of the Trie structure is that it can dynamically expand nodes to adapt to data updates and changes. In response to the evolving types of anomalies in the power system, the TDA module can add new nodes or adjust existing nodes in real time, thereby achieving flexible anomaly detection. At the same time, the TDA module combines statistical methods such as extreme value theory (EVT) to quantify the detected anomaly frequency and characteristics to ensure the accuracy and timeliness of anomaly detection.

[0053] Furthermore, the anomaly detection based on the Trie structure determines whether an anomaly occurs by counting the occurrence frequency of log templates. Assume that the log template set is ,in, Represents each log template respectively; the number of occurrences of the i-th log template is N(si), for each template si, count its number of occurrences N(si) in the time window. Normalized to get the occurrence probability:

[0054] It can be understood that before counting the occurrence frequency of the log template, the comprehensive features are matched layer by layer with the nodes in the pre-built Trie tree, and the occurrence frequency of the log template corresponding to the node after matching is counted, specifically: Extract the feature vector Vfinal from the current device state, and calculate the cosine similarity between the feature vector Vfinal and the feature vector VTrie of each path in the Trie tree. The formula is as follows:

[0055] Where Vfinal is the feature vector of the multi-modal comprehensive features after the hybrid expert model is fused, and VTrie is the feature vector of the Trie tree path; When the calculated similarity is greater than the set threshold, the log template is considered to be matched successfully; For each successfully matched log template si, update its occurrence frequency N(si).

[0056] Furthermore, the above contents are further elaborated as follows: (1) Definition of the above log template: A log template is a structured pattern extracted from the original log information, usually consisting of a fixed part and a variable part. For example, the template for the log message "Task 123 completed in 2ms" may be "Task <id>completed in <time>". In the Trie tree, the content stored in the node is the segment (or key field) of the log template. By inserting the log segment into the Trie tree, each log will correspond to a path in the tree layer by layer.

[0057] (2) Related description: The nodes of the Trie tree represent the components of the log template (such as field values, paragraph contents).

[0058] The log template as a whole corresponds to the path in the Trie tree, and the path from the root node to the leaf node forms the complete structure of the log template.

[0059] (3) Log template setting location: Log templates are usually automatically generated by log parsing tools (such as Drain or LogParser) at the beginning of system operation, or they can be defined by experts and stored in the knowledge base. The setting location and purpose are as follows: (4) Storage location: Log templates can be stored in the database or memory of the log management module as input for Trie tree construction.

[0060] (5) Relationship with Trie tree: Templates serve as the source of Trie tree node data, and the collection of all templates forms the structure of the Trie tree for efficient matching and statistics.

[0061] (6) Log template frequency statistics method: Statistical mechanism: The statistics of log frequency are implemented based on the layer-by-layer matching of the Trie tree. The specific steps are as follows: Log parsing: Each original log is decomposed into fields or segments according to template rules.

[0062] Trie tree matching: Match the decomposed log segments to the nodes of the Trie tree layer by layer.

[0063] Count update: After a successful match, the counter of the number of occurrences is increased at the leaf node (i.e. the complete template position).

[0064] (7) Algorithm flow: Initialize the Trie tree: insert the log template set into the Trie tree, where each node corresponds to a segment.

[0065] Real-time statistics: find the corresponding leaf nodes through layer-by-layer matching and count them incrementally.

[0066] Dynamic update: If the log template set changes (such as adding a new template), it can be achieved by inserting a new path in the Trie tree.

[0067] To identify anomalies, the Extreme Value Theory (EVT) is used to calculate the anomaly score for each template. is a random variable representing the number of times the template appears, and its extreme value distribution can be expressed by the generalized extreme value distribution (GEV):

[0068] Among them, ξ, μ, σ are shape parameters, location parameters and scale parameters.

[0069] The score calculated based on extreme value theory can help determine whether a log template is abnormal. Specifically, when the probability of a log template occurring is When a certain threshold is exceeded, the system marks it as abnormal.

[0070] Furthermore, the above contents are further elaborated as follows: (1) Extreme Value Theory (EVT): Extreme value theory studies the extreme behavior of data (such as maximum or minimum values). In log anomaly detection, extreme value theory can be used to evaluate whether the extreme frequency of occurrence of log templates deviates from the normal distribution.

[0071] The generalized extreme value distribution (GEV) is used to model the extreme behavior of random variables (such as the number of template occurrences), and the shape parameter ξ, location parameter μ, and scale parameter σ describe the characteristics of the distribution.

[0072] (2) Probability of occurrence of log template P(si): P(si) is the frequency ratio of the occurrence of log template si, which is a relatively simple statistic used to measure the prevalence of the template in log data.

[0073] (3) Key points of the relationship between extreme value theory and P(si): The connection between extreme value theory and P(si) is established through the number of template occurrences N(si). The details are as follows: Number of occurrences N(si): N(si) is the actual count value of log template si within a certain time window and is the basis of frequency statistics.

[0074] Calculation of probability P(si): is a normalized statistic Extreme Value Theory Modeling: Treat N(si) as a random variable and analyze its extreme value distribution (such as the maximum number of occurrences) to identify abnormalities. EVT uses the GEV distribution to fit the extreme cases of N(si) and determines whether it is abnormal by setting a threshold.

[0075] (4) The relationship between extreme value theory and P(si): Extreme value theory provides a description of the tail behavior of probability distribution, namely, the extreme value distribution model of N(si), focusing on those situations whose occurrence frequency is far higher (or lower) than the average level.

[0076] P(si) is a general probability estimate for each template, which can quickly determine the universality of the template.

[0077] (5) Combined use: In this embodiment, P(si) is used to preliminarily filter abnormal templates (eg, select templates that may appear more frequently).

[0078] EVT applies more detailed tail distribution modeling to the filtered templates to confirm whether they are anomalous.

[0079] (6) The process of calculating anomaly scores based on EVT: Log template frequency statistics: For each template si, count its occurrence times N(si) in the time window. Normalize to get the occurrence probability:

[0080] Extreme value distribution fitting: Use the historical data of N(si) to fit the generalized extreme value distribution GEV(ξ,μ,σ).

[0081] The parameters ξ, μ, σ are calculated by maximum likelihood estimation (MLE) or other methods.

[0082] Calculate the anomaly score: For the number of occurrences N(si) of the new log template, calculate its cumulative distribution function value FGEV(N(si)) in the fitted distribution.

[0083] Determine whether it is abnormal based on the set threshold (such as 95% or 99%).

[0084] (7) Combine extreme value theory with P(si) to detect anomalies: If P(si) is lower than a certain probability threshold, it is directly marked as an anomaly (indicating that the template itself is rare).

[0085] If P(si) is high, but EVT determines that FGEV(N(si)) is at the tail, it can also be marked as abnormal (indicating that frequent occurrence may represent system abnormality).

[0086] The extreme value anomaly score Si = 1 − FGEV(N(si);ξ,μ,σ), where FGEV is the cumulative distribution function of the generalized extreme value distribution.

[0087] As an example: Assume that the historical data of the number of occurrences of log template s1 is [50, 55, 60, 58], and the number of occurrences in the current time window N(s1)=120: Very high, exceeding the preset threshold, requiring further analysis.

[0088] Using historical data to fit the GEV distribution to N(s1), we get ξ=0.1, μ=58, σ=5.

[0089] Calculate the cumulative distribution value FGEV(120)≈0.99 for the current N(s1)=120, and the anomaly score S1=1−FGEV(120)=0.01.

[0090] If S1 exceeds the threshold (such as S1>0.95), S1 is judged to be abnormal.

[0091] Furthermore, after detecting an anomaly, the system generates initial processing suggestions and proposes multiple solutions for users to choose from. Based on the detection results, the TDA module analyzes the severity and possible impact of each anomaly, and then generates multiple versions of processing suggestions. Each recommended solution contains different operating steps, risk assessments, and countermeasures to better adapt to diverse scenarios. For example, for the same abnormal event, the system may generate three processing solutions: one is a low-risk and low-cost solution that is suitable for rapid response; another is a high-security solution that includes more comprehensive inspections and detailed processing steps; and another is an emergency solution that focuses on rapid repairs to avoid greater losses. By providing multiple processing solutions, the system not only increases the flexibility of user selection, but also improves its adaptability to the needs of different industries.

[0092] In this embodiment, further: In the power system anomaly handling scenario, reinforcement learning gradually optimizes strategies from system status, user feedback and execution results through interactive learning. The key to generating multiple handling suggestions is: Mapping anomaly detection results to the state of reinforcement learning; For each exception, define an action space corresponding to different processing solutions; The reward mechanism is used to evaluate the effectiveness of suggestions and optimize the generation strategy.

[0093] (1) Define the reinforcement learning problem State: The state is defined as the current abnormal state of the power system, which may include the abnormality type, severity, impact scope, etc.

[0094] The design of the state vector S can combine the output of the anomaly detection module (such as the detection results of the TDA module) and historical data.

[0095] Action: Actions are different processing solutions generated by the system, for example: Quick fix solutions (low risk, low cost); Comprehensive inspection plan (high assurance, detailed steps); Emergency response plan (prioritize system recovery).

[0096] The action space can be extended from predefined templates to dynamically generated ones, combining the operating specifications and historical experience of the power industry.

[0097] Reward: Rewards measure the actual effectiveness of the proposed solution, including: user feedback (such as user satisfaction score), system performance indicators (such as processing completion time, residual risk assessment), and cost and resource consumption (such as operation cost and time).

[0098] The reward function can be designed as: ; in, is the weight parameter.

[0099] (2) Training reinforcement learning model: Modeling method: Use the Deep Q Network (DQN) or Policy Gradient method to train the model. Input state S and output the Q value or probability distribution corresponding to each solution.

[0100] Training process: Initialization: construct an initial strategy and generate solutions in a random or rule-based manner; Interactive learning: The system selects actions (processing suggestions) based on the current state S, observes user feedback or execution results, and records rewards R; Policy update: Optimize the policy function based on the reward so that the recommendations generated in the future are more in line with the goal.

[0101] Generates various treatment suggestions: Diversified processing solution generation: In the action space of reinforcement learning, multiple solution templates are predefined for each anomaly: Quick Fix: Actions are aimed at saving time and resources; Comprehensive inspection: Actions cover more complex diagnostic and repair steps; Emergency response: Actions prioritize control and mitigation of potential hazards.

[0102] The model dynamically adjusts the action generation rules by learning user preferences and actual execution results.

[0103] Introducing stochasticity or multi-objective optimization: To ensure diversity, an exploration mechanism can be introduced in action selection (e.g. Strategy). Multi-objective optimization (Pareto optimization) is used to generate recommended combinations that meet different needs.

[0104] (3) Closed-loop optimization of user feedback: After the user selects a solution, the system records the user's choice and execution effect and uses it as feedback for the next round of optimization. The reward function and action generation strategy of the reinforcement learning model are adjusted based on the feedback, so that future processing suggestions are more in line with actual needs.

[0105] (4) Example description: Assume that an overload exception occurs in a certain device, and the system status S includes: exception type: overload. Severity: high. Impact range: medium.

[0106] Three processing solutions generated by reinforcement learning: 1. Quick fix: Action: Reduce current load and redistribute power.

[0107] Rewards: Low cost, but some risk may remain.

[0108] 2. Comprehensive inspection plan: Action: Shut down the equipment, fully check the load and operating parameters, and perform equipment calibration.

[0109] Rewards: High cost, but risk is largely eliminated.

[0110] 3. Emergency treatment plan: Action: Start the backup equipment and switch the power supply immediately.

[0111] Reward: Power restored quickly, but further follow-up is required.

[0112] Example formula: In state S, the reinforcement learning model selects action Ai:

[0113] Select the optimal solution based on the Q value, or generate multiple optimal solutions at the same time.

[0114] Furthermore, finally, the system pushes the generated processing scheme to the user, waiting for the user's selection and feedback. The user can view the detailed information of each processing scheme through the interface, including potential risks, expected effects, implementation steps, etc. The user selects the appropriate scheme according to their own business needs and actual conditions, and can provide feedback. This feedback will be used as input for subsequent iterative optimization, helping the system to gradually adjust the generation strategy and further improve the quality of anomaly detection and processing suggestions. Through this flexible and diverse scheme generation and selection mechanism, the system can meet the needs of different scenarios in the power system with higher accuracy and personalization, laying a solid foundation for subsequent adaptive learning and closed-loop optimization.

[0115] Step 4: After generating a preliminary anomaly handling solution, the system will present multiple solutions to the user for selection. The user can view the specific details of each solution through the interface, make a choice based on actual needs, and provide feedback on the solution. User feedback will include satisfaction with the solution, desired improvement directions, or personalized needs. The system will record and store this feedback information to provide data support for subsequent iterative optimization and adaptive learning, thereby continuously improving the accuracy and personalization of anomaly detection and handling suggestions.

[0116] Furthermore, first of all, the system displays a variety of processing solutions for users to choose from. On the user interface, the system presents the details of each solution in a clear and actionable format, including: specific operation steps, expected results, risk assessment, required resources and time, etc. Users can compare the pros and cons of different solutions at a glance. For example, some solutions may focus on quick fixes to reduce system risks, while others may focus on thorough repairs of the root causes, but they take a long time. By providing a wealth of choices, the system can better meet the needs of users in different scenarios and give users the initiative in the exception handling process.

[0117] Furthermore, users can submit feedback after viewing the solutions. User feedback contains many aspects of information: for example, users can give a satisfaction evaluation of the implementation effect of each solution, indicate whether the solution meets actual needs, or make personalized suggestions for improvement. In addition, users can also suggest new solutions or adjust certain steps in existing solutions to better meet specific operational requirements. The types of feedback information may include: operational difficulty, resource consumption, expected effects and risks, etc. The user's actual experience and feedback information can play an important optimization role in this system with diverse choices, helping the system to better understand and adapt to user needs.

[0118] Furthermore, finally, the system uses the user's choice and feedback as input for iterative optimization. The system will record the feedback information of each solution, store it in the feedback database, and use it as the basic data for adaptive learning in the subsequent solution generation and optimization process. These feedback data can be analyzed by machine learning methods. For example, the system can use reinforcement learning algorithms to use user choices and satisfaction as reward signals to gradually optimize the generation strategy and solution recommendation logic. By continuously analyzing and learning user feedback, the system can gradually adjust and optimize the anomaly detection solution to make it more intelligent and personalized, thereby generating processing suggestions that are more in line with the actual needs of users in subsequent iterations. Through this closed-loop feedback mechanism, the system realizes adaptive improvement of the anomaly handling process, significantly improving the efficiency and accuracy of anomaly detection and processing in the power system.

[0119] Step 5: In this embodiment, based on user selection and feedback, the system builds a dynamic feedback loop through a self-organizing algorithm (such as a self-organizing map SOM) to continuously optimize the generation process of exception handling suggestions. This step involves further fusion adjustment of multimodal information, dynamic optimization of the feedback update mechanism, and iterative improvement of exception handling suggestions. Through this closed-loop optimization system, the system can continuously adjust the anomaly detection model parameters and generation strategy to better meet user needs and ensure the efficiency and accuracy of the processing suggestions.

[0120] Furthermore, after each round of operation, the system collects data from the user's operations, choices, and feedback, including the user's evaluation of the exception handling suggestions (such as satisfaction, accuracy) and the suggested adjustments. For example, for text feedback, models such as GPT are used to extract key opinions and semantic features. For rating and evaluation data, their numerical values ​​or label information are used directly.

[0121] Input the collected feedback data into the Self-Organizing Map (SOM) for clustering. SOM can automatically identify patterns in feedback data, find out user preferences for different scenarios and common exception handling requirements, and thus provide a basis for the system to adjust model parameters. Map the feedback data to a low-dimensional space and cluster the feedback samples through SOM. Identify different feedback modes (such as high satisfaction, low satisfaction, specific modal preferences, etc.) and form a distribution map of user preferences. Based on the feedback clustering results, the system can understand which specific scenarios or modalities are more likely to cause user dissatisfaction, thereby locating the optimization direction.

[0122] Furthermore, based on the feedback patterns identified by SOM, the system dynamically adjusts the weights and gating mechanisms of each expert network in the MOE model. For the modalities or preferences of users with high-frequency feedback, the system can appropriately increase the activation weights of the corresponding experts, thereby strengthening the feature extraction and fusion of related modalities.

[0123] Assume that the expert weights of text, image, and audio modalities are gT, gI, and gA respectively. Each weight value is dynamically adjusted according to the results of SOM clustering. Assume that the calculation formula of the adjusted weight is:

[0124] in: is the importance score of mode i in the feedback mode obtained through SOM cluster analysis; represents the importance score of the text, image or audio modality, such as the importance score of modality k (when k = text); represents the sum of the importance scores of text, image, and audio modalities; T, I, and A represent text, image, and audio data, respectively, which is equivalent to adding all modalities together (all modalities added together = 1). For example, if the text modality is more important in the feedback, the fT value is larger. β is a smoothing parameter used to control the difference between the weights of each expert. When the importance of a modality is A higher value will tend to give a higher weight to the mode.

[0125] Furthermore, an iterative optimization of dynamic feedback loop is constructed. In each round of feedback loop, the system performs weighted fusion of multimodal data according to the latest weight setting to generate new exception handling suggestions. After each round of iteration, the system re-collects user feedback and continues to update expert weights and fusion strategies to form a dynamic closed-loop feedback optimization process. Weighted fusion feature representation: In each round of iteration, multimodal fusion features are generated based on the current expert weights gT, gI, and gA. :

[0126] Feedback-driven iterative optimization: In the next round, the system adjusts the values ​​of fT, fI, and fA based on the user feedback from the previous round, and updates the weights gT, gI, and gA. This process is repeated until the fusion features meet the user's needs and achieve the best performance.

[0127] In addition, you can also set the convergence condition. When the changes in the expert weights gT, gI, and gA are less than the threshold ϵ in several consecutive rounds or the fusion features are stable, stop the iteration. Convergence condition formula:

[0128] Represents the expert weight of modality i in the next iteration; Represents the expert weight of modality i in the current iteration.

[0129] Furthermore, the system continuously optimizes the anomaly detection algorithm based on new anomaly types. Based on user feedback information, the system dynamically adjusts the parameters and thresholds of the anomaly detection algorithm, for example, using extreme value theory (EVT). Specifically, the system re-evaluates the anomaly thresholds and category boundaries in the detection algorithm in each iteration to more accurately identify new anomalies. For example, if the system detects that certain anomaly categories appear frequently in user feedback, the system will pay more attention to these categories in the updated iteration and appropriately lower their detection thresholds to enhance the sensitivity of the algorithm.

[0130] Furthermore, after completing the multimodal optimization and feedback update, the system regenerates and iterates the exception handling suggestions. The processing suggestions after each round of optimization are evaluated and fine-tuned based on the latest feedback data to ensure the applicability and rationality of the suggestions. When generating new processing solutions, the system considers previous user feedback as key input to adjust the generated recommendation content and recommendation logic. These suggestions are designed to adapt to dynamically changing power system scenarios, such as abnormal conditions that arise in new environments or under new conditions. After outputting the new processing solution, the system continues to collect user choices and feedback in order to enter the next round of feedback optimization. This feedback-driven iterative optimization process not only improves the accuracy of anomaly detection and processing, but also gives the system the ability to gradually improve itself, realizing the intelligent, adaptive and personalized power anomaly processing.

[0131] Step 6: Based on anomaly detection and user feedback, the system further adopts adaptive learning and closed-loop optimization mechanisms, converting user feedback into reward signals through reinforcement learning (such as deep Q-learning algorithm), and continuously adjusting and optimizing the generation strategy of the model. With the continuous input of user feedback, the adaptive learning module dynamically adjusts the generation parameters so that the anomaly handling solution gradually meets the actual operation requirements and user preferences. Through this closed-loop system, the system can identify and adapt to changes in user needs, realize the personalization and accuracy of anomaly handling suggestions, and gradually accumulate experience to form a self-learning cycle.

[0132] Furthermore, first, the system incorporates user feedback into the reward mechanism of reinforcement learning. After each generation and processing of suggestions and obtaining user feedback, the system uses the user's choice and feedback as a reward or penalty signal to update the strategy of the generation model. Specifically, the Q-learning algorithm uses user feedback to adjust the Q value, and uses the learned rewards (such as high satisfaction feedback) or penalties (such as low satisfaction feedback) to guide the strategy adjustment when generating solutions next time. The system continuously accumulates user feedback data during the generation of multiple versions of suggestions, thereby adjusting the strategy and maximizing the long-term accumulated rewards. This method can not only optimize solution generation in the current scenario, but also accumulate experience points to provide support for future solution generation.

[0133] Furthermore, the system optimizes strategies through the deep Q-learning algorithm to ensure adaptability. Deep Q-learning combines the advantages of Q-learning and deep neural networks and can handle complex features and state spaces. In the deep Q-learning architecture, the system inputs the feature vectors of multimodal data into the deep neural network model together with user feedback, and uses the neural network to predict the reward values ​​of different schemes. As the number of iterations increases, the system continuously updates the network weights so that the recommendation generation strategy in a given power anomaly scenario is more in line with actual needs. Since deep Q-learning can handle complex feature relationships in high-dimensional state space, the system can learn user preferences and scenario changes more comprehensively through this method.

[0134] In order to continuously optimize the generated exception handling suggestions through user feedback, the system introduces a reinforcement learning model. Suppose at time step t, the suggestions generated by the system The reward is , the status is , the goal is to learn a policy , so that the cumulative reward maximize:

[0135] Where γ is the discount factor; k represents the delay time of the reward rt+k relative to the current time step t. If k=0, it means that the reward rt+k=rt is the reward obtained immediately at the current time step t. If k>0, it means that the reward rt+kr is the delayed reward obtained at the future time step t+k. It represents the immediate reward obtained by the system after executing the action according to the current strategy π at time step t+k.

[0136] When reinforcement learning algorithms such as Q-learning update their strategies, they use the following update formula:

[0137] Among them, α is the learning rate. It represents the expected cumulative reward that can be obtained by selecting action at in the current state st. Q(st,at) is the core function in reinforcement learning, which is used to measure the value of selecting action at in state st. Q(st+1,a) represents the expected cumulative reward that can be obtained by selecting action a in the next state st+1. a is an action in reinforcement learning, that is, the action that the agent can take in a certain state.

[0138] Using user feedback as a reward signal, the system continuously adjusts its strategy so that the generated exception handling suggestions gradually approach the optimal one.

[0139] More specifically, in this embodiment, the following detailed explanations and examples are given in combination with the application scenarios of the power system: (1) Status The concretization state Indicates the description and environmental information of the current power system abnormality. It may include the following aspects: 1) Anomaly detection information Exception type: Such as equipment overload, line interruption, overtemperature, communication failure, etc.

[0140] Example value: Device overload.

[0141] Abnormal severity: The impact level of the anomaly, which can be expressed through confidence (such as probability value) or classification label generated by the anomaly detection module.

[0142] Example value: Severe (confidence 0.9).

[0143] Abnormal impact range: The equipment or areas involved, such as the number of substations and line nodes affected.

[0144] Example value: 5 substations are affected.

[0145] 2) Environmental status information Equipment operating parameters: Real-time status information of the equipment, such as voltage, current, temperature, load and other key parameters.

[0146] Example value: The current load rate is 80%.

[0147] Historical processing records: The historical processing data of the current exception in the system, including the solutions used and the effect feedback.

[0148] Example value: Quick fix solutions were used 3 times in the past with a success rate of 80%.

[0149] Power dispatch information: Real-time grid operation information, such as the availability of backup equipment, changes in power demand, etc.

[0150] Example values: A spare unit is available, and the load demand increases by 1.

[0151] (2) Action The concretization action Indicates the processing suggestions generated by the system in the current state. It can be multiple predefined solutions or dynamically generated solution strategies.

[0152] 1) Low-risk solution: Quick fixes, suitable for lower risk scenarios, are simple to perform but may have limited effects.

[0153] Example value: Reduce load distribution to alternate lines.

[0154] 2) Comprehensive solution: In-depth inspection and repair of all potential factors that may cause abnormalities, suitable for high assurance requirements.

[0155] Example value: Full shutdown to check equipment operating parameters and calibrate loads.

[0156] 3) Emergency plan: Prioritize quickly restoring power supply or controlling risks, suitable for scenarios where major accidents need to be avoided.

[0157] Example value: Switch to backup power, shut down affected equipment.

[0158] (3) Rewards The concretization award Used to measure the effectiveness of the previous round of suggestions, including user feedback and execution results. The reward can be designed as a weighted combination of the following parts: 1) User feedback: The user's satisfaction rating for the generated suggestions, for example, on a scale of 1 to 5.

[0159] Example value: User rating is 4.

[0160] 2) Treatment effect: The degree of reduction in residual risk, for example, the percentage reduction in failure rate.

[0161] Example value: 30% reduction in risk.

[0162] The time it takes to complete a repair, such as how quickly the proposed solution is executed.

[0163] Example value: Completion time is 15 minutes.

[0164] 3) Resource consumption: The cost of implementing the plan, such as manpower, time, power resources, etc.

[0165] Example value: The processing cost is 500 yuan.

[0166] 4) The reward function can be defined as: ; in, The weight is adjustable.

[0167] (4) Application examples of reinforcement learning framework In time : state : The current anomaly is equipment overload, the severity is medium (confidence level 0.85), the impact range is 2 substations, the current load rate is 85%, and backup equipment is available.

[0168] action : The system generates 3 suggestions: Quick fix: Reduce load distribution; Comprehensive inspection: Stop the machine and check all equipment parameters; Emergency treatment: Switch to backup equipment.

[0169] award : The user chooses the quick fix solution, which takes 10 minutes to complete, reduces the risk by 20%, gives a user rating of 5, and costs 300 yuan.

[0170] Reward calculation: .

[0171] Furthermore, based on reinforcement learning, the system generates multiple versions of solutions for users to choose from, in order to gradually optimize the generation strategy. After each round of adaptive learning, the system generates multiple versions of exception handling solutions for users to choose and compare. Each version of the solution covers different risks, costs, and expected effects, and the user's choice will be further fed back to the system to provide data for the next round of learning. Through the selection feedback of multiple versions of the solution, the system can more accurately identify the user's needs in actual operations and gradually adjust the generation strategy. For example, if the user chooses a high-security solution multiple times between a low-cost solution and a high-security solution, the system will be more inclined to recommend similar high-security solutions in future generation, ultimately making the solution generation more in line with user needs.

[0172] Furthermore, the system forms an adaptive learning loop through continuous closed-loop feedback. In the process of generating suggestions, collecting feedback, and learning adjustment strategies, the system gradually accumulates an understanding of user preferences, while adapting to different abnormal situations in the power system. As the adaptive learning cycle continues, the system gradually establishes a self-optimizing closed-loop feedback mechanism, so that the abnormal handling plan can be dynamically adjusted as user needs and operating scenarios change. Through this continuous learning process, the system gradually forms a self-learning ability, which can not only generate the best solution in the current scenario, but also show good adaptability and flexibility in future unknown scenarios, providing strong support for the intelligent and personalized handling of power abnormalities.

[0173] Step 7: After completing adaptive learning and closed-loop optimization, the system outputs the optimized exception handling suggestions to actual operations, and conducts continuous monitoring and dynamic adjustments. By collecting operational feedback and monitoring the processing effects in real time, the system can continuously update model parameters to ensure that the processing scheme meets the actual scenarios and operating standards of the power system. While continuously collecting and analyzing user feedback, the system maintains the operation of the iterative optimization mechanism, enabling it to adaptively respond to new demands and changing power anomalies. In this embodiment, the criteria for the optimal processing suggestion include: (1) Convergence of cumulative rewards: In reinforcement learning, cumulative rewards are usually an important indicator for evaluating the quality of a strategy. If the cumulative rewards gradually converge after multiple rounds of iterations, it means that the strategy has become optimal, and it can be considered that the optimal processing suggestion has been obtained.

[0174] The change in the cumulative reward is less than a certain threshold (e.g., less than 1% in N consecutive iterations). The average reward of reinforcement learning reaches a predefined value.

[0175] (2) Stability of multimodal data weights: Whether the weight adjustment of the hybrid expert model reaches a stable state can be used as another criterion. When the clustering results and weight distribution in the self-organizing map model change little after multiple rounds of feedback, it means that the system's processing of multimodal data has become stable.

[0176] (3) User satisfaction feedback: If the processing suggestions generated by the system receive high satisfaction feedback from users (such as selecting the same type of suggestions multiple times), it can also be considered that the optimal suggestion has been obtained.

[0177] (4) Iteration limit: In order to control the computational cost, an upper limit on the number of iterations can be set (such as 100 rounds). When the upper limit is reached, the round with the best cumulative reward or user feedback is selected as the final output.

[0178] Therefore, "optimal processing suggestion" means that after multiple rounds of iterations, the cumulative reward of reinforcement learning tends to converge, the change range is less than the set threshold (such as 0.01), or the satisfaction score of user feedback reaches the preset target (such as more than 95%). In addition, when the preset maximum number of iterations (such as 100 rounds) is reached, the processing suggestion generated in the round with the largest cumulative reward is used as the final output.

[0179] Furthermore, first, the system directly applies the optimized processing scheme to actual operations. After each abnormal processing scheme is generated, the system pushes the scheme to the operation end, and the user or operator implements it in the actual power system. The system monitors the execution effect of the processing scheme in real time during the operation process, and collects data related to the execution for subsequent analysis. For example, the system can monitor factors such as equipment status, operating parameters, and external environmental changes during the processing process to understand the actual impact of the scheme. These real-time data can help the system evaluate the effectiveness of the processing suggestions, and promptly remind users or automatically generate new suggestions when undesirable situations are found.

[0180] Furthermore, the system continuously adjusts models and strategies based on real-time monitoring and user feedback. During the application of the treatment plan, the system will continuously collect user feedback and adjust model parameters based on the feedback results. This process can be achieved by analyzing information such as user preferences in operation, key indicators in execution results, and satisfaction in feedback. For example, if a treatment plan shows a high failure rate or fails to achieve the expected results in actual operation, the system will identify the shortcomings in the plan and fine-tune the parameters through the adaptive learning module to optimize the generation of future plans. The system adjusts the anomaly detection model based on the latest data feedback to ensure that it is more in line with on-site needs in future operations.

[0181] Furthermore, the system maintains an iterative optimization mechanism that enables it to adapt to new demands and changes in the power system. Over time, the environment and equipment conditions in the power system may change, and the system needs to be continuously updated and adapted. Through the continuous iterative optimization mechanism, the system can dynamically respond to new types of anomalies and user needs, forming a "learning by doing" self-learning mode. This continuous iterative mechanism ensures that the system is always kept up to date and continuously improves the accuracy and effectiveness of anomaly handling. In this way, the system can not only solve the current power anomaly problem, but also quickly adjust and adapt when the power system changes, ensuring that it always maintains efficient operation in a complex and changing power environment.

[0182] Embodiment 2 This embodiment discloses a power anomaly processing system driven by self-organizing closed-loop feedback and iterative optimization.

[0183] The self-organizing closed-loop feedback-driven iteratively optimized power anomaly processing system includes: The feature fusion module is configured to: obtain multi-modal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multi-modal data through the weighted fusion method; The anomaly detection module is configured to: detect anomalies of the power system based on comprehensive features; The suggestion generation module is configured to: generate a plurality of treatment suggestions for power system anomalies using a reinforcement learning method; The first closed-loop feedback module is configured to: obtain user feedback information on multiple processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering result; The second closed-loop feedback module is configured to: use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning, and optimize the generation of the next round of multiple processing suggestions; The iterative optimization module is configured to iterate the feature fusion module to the second closed-loop feedback module until the optimal processing suggestion is obtained.

[0184] It can be understood that this embodiment also includes: The data collection module is configured to: be responsible for data collection and preliminary cleaning; The multimodal information processing module is configured to: analyze the data and generate initial anomaly detection results; Self-organizing feedback module (first closed-loop feedback module): a dynamic optimization mechanism based on multiple rounds of feedback; The user interaction module is configured to: support the display of multiple versions of suggestions and the collection of feedback; Adaptive learning module (second closed-loop feedback module): optimizes the generation strategy through closed-loop learning of user feedback and adaptive algorithms; The output and monitoring module is configured to: provide final recommendations and implement them, and monitor feedback from actual operations.

[0185] Embodiment 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0186] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization as described in Example 1 of the present disclosure.

[0187] Embodiment 4 The purpose of this embodiment is to provide an electronic device.

[0188] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the self-organizing closed-loop feedback driven iterative optimization method for handling power anomalies as described in Example 1 of the present disclosure are implemented.

[0189] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0190] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0191] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.< / time> < / id>

Claims

1. A power anomaly processing method driven by self-organizing closed-loop feedback and iterative optimization, characterized in that: The following steps are involved: Step 1: Obtain multimodal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multimodal data through the weighted fusion method; Step 2: Detect power system anomalies based on comprehensive features; Step 3: Use reinforcement learning methods to generate multiple treatment suggestions for power system anomalies; Step 4: Obtain user feedback on various processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering results; Step 5: Use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning to optimize the generation of multiple processing suggestions for the next round; Step 6: Iterate the above steps 1 to 5 until the optimal processing suggestion is obtained.

2. The method for handling power anomalies by self-organizing closed-loop feedback driven iterative optimization according to claim 1, characterized in that: The comprehensive features of the multimodal data are specifically acquired through the following process: Acquire multimodal data of the power system, including equipment operation data, user input data, historical operation records and external environment data, wherein the types of the multimodal data include text data, image data and audio data; The multimodal data is input into the hybrid expert model, and the multimodal data is assigned to the corresponding expert network according to the type through the gating network. The corresponding multimodal data features are extracted by the expert network to obtain independent feature representations; The gating network generates corresponding weights according to the content of each modality data feature; According to the weights corresponding to the features of each modal data, the feature representations of each modal data feature are weighted fused to obtain the fused features; The fused features are input into the deep learning model to further extract features and obtain comprehensive features.

3. The power anomaly processing method driven by self-organizing closed-loop feedback iterative optimization according to claim 1, characterized in that: Based on comprehensive features, power system anomalies are detected, including: Input the comprehensive features into the Trie-based network model, match the comprehensive features with the nodes in the pre-built Trie tree layer by layer, count the occurrence frequency of the log templates corresponding to the matched nodes, and determine whether there is an abnormality: When the occurrence probability of the log template is lower than the set detection threshold, it is directly determined that the power system is abnormal; When the probability of occurrence of the log template is higher than the set detection threshold, the pre-built generalized extreme value distribution GEV model is further combined to determine the abnormality; or, Combine the pre-built generalized extreme value distribution GEV model to determine abnormalities, including: Get the type si of the log template corresponding to the node after matching; Count the actual count value N(si) of the log template si within a certain time window, and use the historical data of N(si) to fit the parameters of the generalized extreme value distribution GEV model; Based on the generalized extreme value distribution GEV model, calculate the cumulative distribution function value of N(si) in the fitting distribution; Based on the cumulative distribution function value of N(si) in the fitting distribution, the extreme value anomaly score is calculated to determine whether the extreme value anomaly score exceeds the set threshold. If so, the power system is determined to be abnormal; or, Pre-build the nodes in the Trie tree, including: Get the log template and decompose the log template into fields or segments; Match the fields or segments of the decomposed log template to the nodes of the Trie tree layer by layer. Each node corresponds to a field or segment, and the pre-construction of the Trie tree node is completed; or, The comprehensive features are matched layer by layer with the nodes in the pre-built Trie tree, and the occurrence frequency of the log template corresponding to the matched nodes is counted, specifically: Extract the feature vector Vfinal from the current device state, and calculate the cosine similarity between the feature vector Vfinal and the feature vector VTrie of each path in the Trie tree. The formula is as follows: Where Vfinal is the feature vector of the multi-modal comprehensive features after the hybrid expert model is fused, and VTrie is the feature vector of the Trie tree path; When the calculated similarity is greater than the set threshold, the log template is considered to be matched successfully; For each successfully matched log template si, update its occurrence frequency N(si).

4. The power anomaly processing method driven by self-organizing closed-loop feedback iterative optimization according to claim 1, characterized in that: A reinforcement learning method is used to generate a variety of treatment suggestions for power system anomalies. The specific generation process includes: Defining states in reinforcement learning The abnormality detection information and environmental status information of the current power system, wherein the abnormality detection information includes the abnormality type, abnormality severity and abnormal impact range, and the environmental status information includes equipment operation parameters, historical processing records and power dispatch information; Defining actions in reinforcement learning A plurality of different treatment suggestions, wherein the plurality of treatment suggestions include different operation steps, risk assessments and countermeasures adapted to various scenarios, and the plurality of treatment suggestions are obtained through a plurality of predefined solutions or dynamically generated solution strategies; Defining Rewards in Reinforcement Learning It is the weighted sum of user feedback, processing effect and resource consumption; Select actions based on the current state, obtain user feedback information and calculate rewards, update the strategy function based on the reward value, and generate multiple processing suggestions.

5. The power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization according to claim 4, characterized in that: The goal of reinforcement learning is to learn a strategy that accumulates rewards maximize: Among them, γ is the discount factor; k represents the reward The delay time relative to the current time step t; It represents the immediate reward obtained by the system after executing the action according to the current strategy π at time step t+k; Update the strategy function based on the reward, specifically: Among them, α is the learning rate; represents the expected cumulative reward that can be obtained by selecting action at in the current state st; Q(st+1,a) represents the expected cumulative reward that can be obtained by selecting action a in the next state st+1; a is the action in reinforcement learning.

6. The power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization according to claim 1, characterized in that: Obtain user feedback on various processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weights of each modality data in the hybrid expert model based on the clustering results, including: Obtaining user feedback on various processing suggestions, including satisfaction with the solution, desired improvement direction, personalized needs, suggestions for new solutions and certain steps to adjust existing solutions. The types of feedback information include operational difficulty, resource consumption, expected effect and risk; A dynamic feedback loop is constructed through a self-organizing map model, feedback information is input into the self-organizing map model, feedback information is clustered, different feedback modes are identified and a user preference distribution map is formed, wherein the feedback modes include high satisfaction, low satisfaction and preference for a specific mode; Calculate the importance scores of each modality under different feedback modes, and adjust the weights of each modality data in the hybrid expert model based on the importance scores : ； Where: fi is the importance score of mode i under the set feedback mode; β is the smoothing parameter; Indicates the importance score of the text, image, or audio modality; represents the sum of the importance scores of text, image, and audio modalities; T, I, and A represent text, image, and audio data, respectively; When the weight change of each modality data after adjustment is less than the threshold in several consecutive rounds or the fusion feature performance is stable, the iteration is stopped.

7. The power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization according to claim 3, characterized in that: It also includes re-evaluating the detection threshold of abnormal categories in the Trie-based network model in each iteration based on user feedback: If certain abnormal categories appear frequently in user feedback information, the detection threshold of the corresponding abnormal category will be lowered in the updated iteration; or, In step 6, the criteria for the optimal processing suggestion include: After multiple rounds of iterations, the cumulative rewards of reinforcement learning tend to converge, and the change range is less than the set threshold; or the weight adjustment of the hybrid expert model reaches a stable state; Or the satisfaction score of user feedback reaches the preset target; Or when the preset maximum number of iterations is reached, the processing suggestion generated in the round with the largest cumulative reward is used as the final output.

8. A self-organizing closed-loop feedback-driven iteratively optimized power anomaly processing system, characterized in that: include: The feature fusion module is configured to: obtain multi-modal data of the power system, input each modal data into the hybrid expert model, extract features and generate weights based on the gating network, and finally obtain the comprehensive features of the multi-modal data through the weighted fusion method; The anomaly detection module is configured to: detect anomalies of the power system based on comprehensive features; The suggestion generation module is configured to: generate a plurality of treatment suggestions for power system anomalies using a reinforcement learning method; The first closed-loop feedback module is configured to: obtain user feedback information on multiple processing suggestions, input the feedback information into the self-organizing map model for clustering, and dynamically adjust the weight of each modality data in the hybrid expert model based on the clustering result; The second closed-loop feedback module is configured to: use the feedback information as a reward for generating multiple processing suggestions in reinforcement learning, and optimize the generation of the next round of multiple processing suggestions; The iterative optimization module is configured to iterate the feature fusion module to the second closed-loop feedback module until the optimal processing suggestion is obtained.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the power anomaly processing method of self-organizing closed-loop feedback driven iterative optimization as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the power anomaly processing method driven by iterative optimization and self-organizing closed-loop feedback are implemented as described in any one of claims 1 to 7.

Images