Power grid fault report decision verification method based on digital twin replay and interpretable neural network

Abstract

The invention discloses a power grid fault report decision verification method based on digital twin replay and an interpretable neural network, and relates to the technical field of intelligent operation and maintenance and artificial intelligence decision verification of a power system. According to the method, structured analysis of a fault report is generated through AI, power grid digital twinning simulation replay and neural network reasoning with characteristic interpretability are combined, report conclusions and simulation results are automatically compared, interpretability analysis such as SHAP and LIME is integrated, and transparent tracing and characteristic contribution degree quantitative verification of the fault report reasoning process are achieved. Compared with a traditional power grid fault report processing method depending on manual checking or black box AI model output, the method can achieve automatic consistency criterion verification and whole-process data traceability of a fault report reasoning conclusion, and the credibility, transparency and batch engineering application efficiency of a power grid operation and maintenance decision are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent operation and maintenance of power systems and artificial intelligence application technology, and in particular to a method for decision verification of power grid fault reports by combining digital twin simulation with interpretable neural networks. Background Technology

[0002] With the deep integration of next-generation information and communication technologies, artificial intelligence, and big data analytics, power systems are accelerating their upgrade towards intelligence, automation, and digitalization. Numerous online monitoring devices and SCADA systems provide a rich real-time data foundation for power grid operation. AI-assisted tools based on deep learning, such as intelligent diagnosis, fault analysis, and predictive maintenance, have been initially applied in various aspects of power grid operation and maintenance, equipment management, and dispatch decision-making. Currently, AI models such as Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Transformers can uncover complex temporal characteristics in power grid operation, enabling automated fault detection, location, and anomaly identification, thus improving the data processing capabilities and intelligence level of large-scale power grid systems.

[0003] However, existing AI-based power grid fault diagnosis and report generation technologies still face a series of prominent challenges: (1) AI models usually output conclusions in a "black box" manner, and the internal reasoning process, feature action mechanism and causal relationship chain of the model are difficult for engineers to intuitively grasp, resulting in a lack of interpretability and engineering transparency in the results; (2) Most AI reports are based on static training of historical data, which cannot be effectively mapped and verified with the dynamic evolution of real working conditions or real-time simulation results, and the reasoning content is prone to deviation from the actual situation, making it difficult to support engineering decisions in key links; (3) In actual batch operation and maintenance, AI-generated reports often require manual verification and cross-verification, which increases the human burden and is prone to omissions, affecting the overall intelligent operation and maintenance efficiency and engineering safety; (4) At present, there is a lack of integrated automatic verification tools that can realize AI reasoning results, digital twin physical simulation and feature interpretability analysis, and cannot provide closed-loop verification and visualization support for the entire process of report conclusions, simulation restoration and feature tracing.

[0004] With the increasing complexity of power grid structures and the rising requirements for operational safety, the aforementioned shortcomings have become a core bottleneck restricting the large-scale implementation of AI in the power industry. Therefore, there is an urgent need for a novel method that integrates digital twin simulation playback, interpretable neural network inference, and feature contribution measurement to achieve multi-source data verification of AI-generated fault reports, transparency of the inference process, and automated verification of results, thereby ensuring the reliable, safe, and efficient operation of intelligent power grid management. Summary of the Invention

[0005] This addresses the prominent issues in existing technologies, such as unverifiable AI power grid fault reports, unexplainable reasoning processes, low efficiency of manual verification, disconnect between reasoning content and actual operating conditions, and unclear feature contributions.

[0006] This invention proposes a power grid fault reporting decision verification method based on digital twin replay and interpretable neural networks. The method includes the following steps:

[0007] 1) Perform structured parsing on AI-generated power grid fault reports, automatically extract key information from the fault reports, including fault time, involved equipment, main cause inference, action process, etc., to form a verification target set;

[0008] 2) Based on the extracted verification objectives, drive the power grid digital twin simulation engine to dynamically restore the actual or predicted fault evolution process and generate time-series chemical condition data corresponding to the report;

[0009] 3) Utilize neural network models that integrate interpretable AI algorithms (such as SHAP and LIME) to perform inference analysis on twin simulation replay data, outputting judgment conclusions such as fault type, key equipment, and fault time. At the same time, quantify the contribution of each input feature to the inference result and form a feature contribution ranking.

[0010] 4) Based on the conclusions of the fault report, the results of twin simulation reasoning, and the analysis of feature contribution, a consistency criterion is set to automatically verify the engineering consistency and credibility of the AI ​​report;

[0011] 5) Visualize the verification results, feature contribution heatmaps, simulation process time-series curves, etc., and integrate them to generate a structured verification report, which is convenient for manual review and engineering archiving;

[0012] 6) Supports batch processing and automatic push, adapts to various application scenarios such as large-scale smart grids, distribution networks and dispatching, and realizes closed-loop verification and management of AI-assisted decision-making results.

[0013] Preferably, in the method, the structured parsing process includes:

[0014] The fault report text is automatically segmented and tagged using regular expressions and a named entity recognition (NER) model, extracting fields such as device, time, and cause to obtain the verification target set.

[0015]

[0016] in, For the first One target device, For the moment of the event, For device status, The cause of the failure is inferred from the AI ​​report.

[0017] Preferably, the digital twin simulation model adopts a power grid physical model based on state transition:

[0018]

[0019] in, For system time State variables, For operation input, For system parameter set, This is the state evolution function.

[0020] Preferably, the neural network inference output is:

[0021]

[0022] in, Fault type The main cause is the equipment. The time when the fault occurred.

[0023] Preferably, the SHAP decomposition of the feature contribution is expressed as:

[0024]

[0025] in, For the first Each input feature The contribution of this feature to the reasoning result. This is the baseline term.

[0026] Preferably, the consistency verification criterion is defined as follows:

[0027]

[0028] in, To report the main cause reasoning, For simulation and neural network principal inference, Score the feature contribution. For the threshold, This is a logical judgment function.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention enables automated verification of the entire process of AI-generated power grid fault reports, engineering consistency comparison of inference conclusions, and quantitative analysis of the contribution of key features. This effectively improves the transparency, verifiability, and traceability of reports, reducing the burden of manual verification and the risk of missed diagnoses. The method supports deep fusion of multi-source data and traceability of the interpretability of AI inference features, significantly enhancing the scientific rigor and intelligence of power grid operation and maintenance decisions. It is particularly suitable for the rapid batch verification and archiving management of AI-assisted reports in large-scale smart grids and distribution networks, providing a solid technical foundation for achieving safe, efficient, and intelligent power grid operation and management. Attached Figure Description

[0031] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the specific embodiments will be briefly introduced below. This drawing outlines the complete process from fault report parsing to verification result output, including the logical connections and data flow paths of key steps.

[0032] Figure 1 This is an overall flowchart of the power grid fault reporting decision verification method based on digital twin replay and interpretable neural network of the present invention. Detailed Implementation

[0033] This invention proposes a power grid fault report decision verification method based on digital twin replay and interpretable neural network technology. This method is used for automatic consistency verification and full-process traceability of AI-generated fault reports. It improves the transparency, credibility, and engineering application efficiency of reports, exhibiting high robustness and adaptability, especially demonstrating significant advantages in handling complex power grid scenarios, multi-source data fusion, and batch report verification. The specific implementation process of this method will be described in detail below.

[0034] S1. Fault Report Structured Parsing and Validation Target Extraction

[0035] This step performs structured text parsing on AI-generated power grid fault reports. Using regular expressions and a Named Entity Recognition (NER) model, it automatically extracts key information from the report, such as fault time, critical equipment, root cause inference, and action process, forming an actionable set of verification targets. The parsing process goes beyond basic extraction; it also involves semantic association analysis to ensure the completeness and logical consistency of the extracted information. For example, extracted entities are cross-validated with the report context to avoid bias caused by isolated information. Specifically, the structured extraction result can be represented as:

[0036]

[0037] in, For the first One target device, For the moment of the event, For device status, The AI ​​report infers the cause of the failure. Furthermore, to quantify the accuracy of the extraction, a confidence assessment formula for entity extraction can be introduced:

[0038]

[0039] in, The confidence score function for each entity, ranging from [0,1], is the output probability based on the NER model.

[0040] S2. Construction of Digital Twin Simulation Replay Model and Fault Process Reconstruction

[0041] This step, based on the verification objectives and original measurement and control data extracted in S1, drives the power grid digital twin simulation engine to dynamically recreate the actual or predicted fault evolution process and output corresponding time-series operational condition data. The simulation model construction needs to consider the dynamic adjustment of the power grid topology and the fusion of multi-source data to simulate real fault propagation paths, such as integrating SCADA system data and PMU synchronization phasor measurements. Specifically, the digital twin simulation process is described using a state transition-based power grid physical model:

[0042]

[0043] in, For the system at time State variables (such as node voltage, current, switch state, etc.). For a moment Operational inputs (such as switch operation, relay protection action). This is a set of system physical parameters. The twin simulation engine can output the following timing data:

[0044]

[0045] in, For the first Node voltage timing, For the first Current timing of the branch circuit For the first The operating states of each switch. To evaluate the accuracy of the simulation, a state deviation index can be calculated:

[0046]

[0047] in, The observed actual state data.

[0048] S3. Explainable Neural Network Inference and Feature Contribution Analysis

[0049] A neural network model integrating interpretable AI algorithms (such as SHAP and LIME) is used to infer the time-series data obtained from twin simulation replay, outputting conclusions such as fault type, cause device, and fault time. The contribution of each input feature to the inference result is quantitatively analyzed. The neural network architecture can employ LSTM or Transformer variants to handle temporal dependencies, while embedding an attention mechanism to enhance interpretability. Specifically, the neural network model input is:

[0050]

[0051] The inference output is:

[0052]

[0053] in, Fault type The main cause is the equipment. This indicates the moment of failure.

[0054] The SHAP decomposition of feature contribution is expressed as:

[0055]

[0056] in, For the first Each input feature The contribution of this feature to SHAP. This serves as the baseline. Interpretable methods such as SHAP and LIME are used to output feature contribution rankings and heatmaps to aid in fault tracing. To further quantify the overall explanatory power, a global explanatory score can be introduced.

[0057]

[0058] in, It represents the maximum absolute value of the contribution.

[0059] S4. Consistency Criterion Setting and Automatic Verification

[0060] This step, based on the AI ​​report conclusions, twin simulation inference results, and feature contribution analysis, establishes consistency verification criteria for automation engineering to assess the credibility of the report's inference results. The criterion design needs to consider multi-dimensional matching, including principal cause consistency, time window alignment, and contribution significance, to adapt to complex power grid scenarios. Specifically, the consistency criterion formula is as follows:

[0061]

[0062] in, The main reasoning conclusions of the AI ​​report. The main results output by twin simulation and neural network model, The feature contribution score, As the contribution threshold, This is a logical judgment function (1 if the condition is true, 0 otherwise). When When, it indicates that the reasoning conclusion has passed the consistency verification; when In such cases, manual review is required. To enhance the robustness of the criterion, a weighted consistency score can be introduced:

[0063]

[0064] in, For the weighting coefficients, satisfying .

[0065] S5. Visualization and Reporting of Validation Results

[0066] All verification results, feature contribution analysis, and simulation process time-series data are visualized using curves, heatmaps, etc., and integrated into a structured fault verification report for easy review and archiving by engineers. The visualization design emphasizes interactivity, such as supporting time-series data scaling and dynamic feature contribution filtering to enhance user experience. Output includes: fault evolution curves, equipment action sequences, SHAP feature contribution heatmaps, consistency verification conclusions, and review recommendations. Reports can be generated in multiple formats such as PDF and HTML and support automatic push to the intelligent operation and maintenance platform. To quantify the visualization effect, an information density index can be defined.

[0067]

[0068] in, For the first The effective information content of each visual element To display the area size, The total number of elements.

[0069] S6. Batch Processing and Platform Integration

[0070] This invention supports automated batch verification of multiple AI-generated power grid fault reports, facilitating engineering implementation and continuous operation and maintenance in large-scale smart grids, distribution networks, and dispatching scenarios. The batch processing mechanism involves parallel computing optimization, such as utilizing a distributed framework to allocate tasks, thereby shortening the overall response time. The method can be flexibly embedded into existing intelligent operation and maintenance, alarm, and decision management platforms of power companies to achieve efficient automated closed-loop management. To evaluate batch efficiency, a processing throughput formula can be introduced:

[0071]

[0072] in, The number of reports processed This represents the total processing time.

[0073] This invention integrates digital twin simulation playback, interpretable neural network inference, and feature contribution quantification methods to achieve automatic consistency verification and full-process traceability of power grid fault report inference results. This not only significantly improves the transparency, credibility, and engineering application efficiency of reports, but also further enhances decision support capabilities by introducing quantitative indicators and visualization tools. It effectively solves the problems of traditional AI power grid fault reports lacking verifiability, traceability, and batch processing capabilities. At the same time, it provides a reliable technical foundation for the intelligent transformation of power systems and promotes the transformation of fault diagnosis from experience-based to data-driven.

Claims

1. A power grid fault reporting decision verification method based on digital twin replay and interpretable neural network, characterized in that, Includes the following steps: Step 1: Perform structured parsing on the power grid fault reports generated by artificial intelligence, and automatically extract verification targets such as fault time, key equipment, and fault cause; Step 2: Based on the verification objective, drive the power grid digital twin simulation engine to dynamically restore the actual or predicted fault evolution process and output operating condition data related to the report; Step 3: Use an interpretable neural network model to infer the fault time series data obtained from the twin simulation replay, output the fault cause, key equipment, and key moment, and analyze and rank the contribution of each input feature; Step 4: Set consistency criteria between the report's reasoning results and the simulation model's reasoning results, and automatically verify the engineering consistency and credibility of the AI ​​report's reasoning conclusions; Step 5: Visualize the verification results, feature contribution analysis, and simulation replay process, and integrate them to output a structured verification report for easy manual review and archiving management.

2. The power grid fault reporting decision verification method according to claim 1, characterized in that, The interpretable neural network model is a neural network model that integrates interpretable algorithms such as SHAP and LIME, and is used to analyze and visualize the contribution of input features to the reasoning conclusion.

3. The power grid fault reporting decision verification method according to claim 1, characterized in that, The consistency criteria include automatic comparison of multiple indicators such as the main cause of the failure, the time of failure, and the ranking of key features, which are used to determine the engineering consistency between the report's reasoning conclusions and the simulation results.

4. The power grid fault reporting decision verification method according to claim 1, characterized in that, The output verification report includes simulation time-series curves, feature contribution heatmaps, consistency verification conclusions, and manual review suggestions.

5. The power grid fault reporting decision verification method according to claim 1, characterized in that, This method is applicable to large-scale smart grids, distribution networks, and dispatching scenarios, and supports batch automated verification and management of grid fault reports generated with the assistance of artificial intelligence.

Images