Method and system for predicting intelligent state of power grid equipment based on digital twinning

Abstract

The invention discloses an intelligent state prediction method and system for power grid equipment based on digital twinning, and belongs to the technical field of intelligent operation and maintenance of power equipment. Comprising the following steps: collecting data through multiple sensors, performing edge preprocessing and hierarchical network transmission, and performing credible evidence storage by adopting an alliance chain technology; based on credible data, generating an equipment state vector by utilizing evidence theory fusion, constructing a three-layer physical model, and outputting an optimal state estimation sequence of the high-fidelity digital twin through model order reduction and data assimilation; carrying out anomaly detection and fault mode identification based on the sequence, constructing a dynamic causal network to realize fault root cause analysis, and fusing physical simulation and data driving prediction to output a state evolution result; and finally, formalizing the maintenance task into a multi-objective optimization problem, evaluating the risk by calculating the maintenance entropy, and screening and outputting a recommendation scheme based on an autophagy mechanism. According to the method, closed-loop operation and maintenance with data credibility, model high fidelity, diagnosis interpretability and decision intelligence are realized.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and intelligent operation and maintenance technology, specifically relating to a method and system for intelligent condition prediction of power grid equipment based on digital twins. Background Technology

[0002] With the rapid development of IoT, big data, and AI technologies, equipment condition monitoring based on digital twins has become an important research direction for intelligent operation and maintenance of power systems. Existing technologies typically involve deploying sensors to collect equipment operating data and using physical models or data-driven models to construct digital twins to achieve online analysis and prediction of equipment status.

[0003] However, existing digital twin application solutions still have significant shortcomings: First, at the data layer, the acquisition, transmission, and processing of multi-source heterogeneous monitoring data lack effective reliability assurance mechanisms, making the data susceptible to interference or tampering. Furthermore, simple data weighted averaging fusion methods struggle to effectively handle conflicts and uncertainties among evidence, resulting in insufficient reliability and accuracy of the observed states input into the digital twin. Second, at the model layer, complex multiphysics coupling models constructed to achieve high-fidelity simulation have high computational loads and cannot meet the demands of online real-time simulation. While simplified models improve speed, they sacrifice the ability to reflect key mechanisms such as microscopic aging within the equipment, making it difficult to balance the fidelity and real-time performance of the digital twin's predicted states. Finally, at the application layer, fault diagnosis based on digital twins often relies on data-driven black-box models, and diagnostic conclusions lack interpretable causal logic support. Simultaneously, maintenance decisions are often disconnected from state prediction results, failing to deeply integrate and quantitatively optimize fault root causes, lifespan predictions, resource constraints, and operational risks, resulting in limited decision-making intelligence.

[0004] Based on the shortcomings of the existing technology, the technical problem that this invention aims to solve is: how to provide a method and system for intelligent state prediction of power grid equipment that can ensure the reliability and deep integration of data throughout the process, construct a dynamic digital twin with both high fidelity and high efficiency, realize interpretable fault diagnosis and accurate state prediction, and automatically generate a comprehensive optimal maintenance plan. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a digital twin-based intelligent condition prediction and maintenance system for power grid equipment, so as to solve the technical problems existing in the prior art, such as low data reliability, contradiction between model fidelity and real-time performance, poor interpretability of fault diagnosis, and insufficient level of intelligence in maintenance decision-making.

[0006] To achieve the above objectives, the present invention adopts the following technical solution;

[0007] A method for intelligent state prediction of power grid equipment based on digital twins includes the following steps:

[0008] Raw data is collected by various types of sensors deployed on power grid equipment. After edge-side preprocessing and hierarchical network transmission, key feature data is reliably stored.

[0009] Based on the data with trusted evidence storage, the device state vector is generated by fusing evidence theory, a dynamic digital twin is constructed, and the optimal state estimation sequence of the digital twin is output.

[0010] Based on the optimal state estimation sequence, anomaly detection, fault diagnosis and state trend prediction of power grid equipment are performed to obtain state assessment information including diagnosis results and prediction results.

[0011] Based on the diagnostic results and state evolution prediction results, the problem of matching maintenance tasks with resources is formalized into a multi-objective optimization problem, and candidate maintenance schemes are generated. By solving the multi-objective optimization problem and evaluating the uncertainty of the schemes, recommended maintenance schemes for physical power grid equipment are generated.

[0012] As a further aspect of the present invention, the step of credibly storing key feature data includes:

[0013] Sensitive locations are determined based on monitoring requirements and physical model simulation. Multiple sensors are deployed to collect raw signals. The sensors include at least vibration sensors, UHF sensors, distributed fiber optic temperature sensors, and online dissolved gas monitoring sensors in oil.

[0014] Edge computing units are deployed at sensor nodes or regional data aggregation points to perform local real-time processing of raw signals;

[0015] The pre-processed data is transmitted through a three-layer network architecture consisting of a field device layer, a station control layer, and a wide area backbone layer.

[0016] At critical nodes in data transmission, consortium blockchain technology is used to store key feature data and its metadata.

[0017] As a further aspect of the present invention, the localized real-time processing includes:

[0018] The vibration signal is bandpass filtered to remove power frequency interference, and time-frequency domain analysis is performed to extract features such as amplitude, frequency, and phase; the UHF signal is amplitude discriminated, pulse classified, and its statistical features are calculated; the temperature and gas concentration data are subjected to moving average or median filtering to smooth noise, and trend indicators such as rate of change are calculated.

[0019] As a further embodiment of the present invention, the field device layer transmits preprocessed data to the data concentrator of the current bay via CAN or Modbus fieldbus protocol; the station control layer aggregates data from each bay to the substation edge server via an industrial Ethernet network conforming to the IEC61850 standard; and the wide area backbone layer transmits status datasets with unified timestamps to the cloud or regional data center via a power dispatch data network or a 5G power virtual private network.

[0020] As a further aspect of the present invention, the adoption of consortium blockchain technology includes:

[0021] Calculate the hash value of key feature data and its metadata, package the hash value, the hash value of the previous data block, and the digital signature of this node to generate a new block; verify the validity of the new block through a practical Byzantine fault-tolerant consensus mechanism deployed among power grid companies, equipment manufacturers, and maintenance units, and add it to the chain.

[0022] As a further aspect of the present invention, the step of constructing a dynamic digital twin includes:

[0023] The system receives preprocessed data from various types of sensors with trusted evidence storage and generates a unified device state vector description based on Dempster-Shafer evidence theory.

[0024] Multi-level physical model construction: Based on the equipment's design drawings, material parameters, and physical laws, a three-layer digital twin model is constructed to describe the intrinsic evolution law of the equipment's state.

[0025] Model reduction and real-time processing: To meet the computational efficiency requirements of online simulation, a projection method based on Krylov subspace is used to reduce the order of the three-layer digital twin model.

[0026] Data assimilation and dynamic calibration use the device state vector as the observation value and the low-dimensional surrogate model as the predictor to construct a data assimilation system, which outputs the optimal state estimation sequence of the digital twin. This sequence is the real-time output of the high-fidelity digital twin.

[0027] As a further aspect of the present invention, the step of generating a unified device state vector description based on Dempster-Shafer evidence theory includes:

[0028] For each type of sensor data, define a recognition framework that includes normal, attentive, and abnormal states;

[0029] Based on historical data or expert experience, assign a basic probability value (BPA) to the observations of each sensor.

[0030] Using the Dempster combination rule, BPAs from different sensors are fused to obtain a comprehensive confidence distribution of consistency regarding the overall health status of the device.

[0031] ;

[0032] This represents the basic probability assignment (BPA) from different sensors. These represent the proposition subsets in each sensor recognition framework; Describes a subset of propositions supported by two sources of evidence. and The intersection is ; The conflict coefficient measures the degree of conflict between two sources of evidence.

[0033] Based on the comprehensive confidence distribution, a multidimensional, unified device state vector is generated as the observation input for the digital twin.

[0034] As a further aspect of the present invention, the three-layer digital twin model includes:

[0035] The equipment was modeled using the finite element method to create a three-dimensional geometric system-level model; electromagnetic-thermal-mechanical multi-physics coupled component-level models were created for core components such as transformer windings, iron cores, and circuit breaker contacts; and microscopic aging kinetic material-level models based on the Arrhenius equation and fractional derivatives were created for key materials such as insulating paper, transformer oil, and epoxy resin.

[0036] As a further aspect of the present invention, the step of using the Krylov subspace-based projection method to reduce the order of the three-layer digital twin model is described.

[0037] The steps for reducing the order of a three-layer digital twin model using a Krylov subspace-based projection method include:

[0038] Linearize the three-layer digital twin model near its steady-state equilibrium point, or select a set of representative operating conditions that can cover the typical operating range of the equipment for simulation calculation, so as to obtain the dynamic response data of the model under different input excitations.

[0039] Based on the acquired dynamic response data, a low-dimensional subspace is constructed using the Krylov subspace method. This subspace is spanned by a series of basis vectors. The basis vectors essentially capture and represent the dominant vibrational modes or dynamic modes with the highest energy and the most significant impact on the dynamic behavior of the system in the full-order physical model.

[0040] By using mathematical projection, the high-dimensional state equation of the three-layer digital twin model is projected onto a low-dimensional subspace spanned by the dominant mode. The state equation of the model is then reconstructed in the low-dimensional subspace spanned by the dominant mode, resulting in a low-dimensional proxy model with a significantly reduced dimension of state variables, but retaining the key dynamic characteristics of the original model.

[0041] As the core prediction engine of the digital twin, the surrogate model enables rapid state extrapolation calculations.

[0042] As a further aspect of the present invention, the step of constructing the data assimilation system includes:

[0043] Construct a state-space model with a reduced-order model as the state equation and a fused state vector as the observation equation;

[0044] Using either an extended Kalman filter or an unscented Kalman filter algorithm, the following loop is executed at each sampling time step:

[0045] Prediction: Using the state estimate from the previous moment, the current state of the equipment is predicted through the state equation.

[0046] Update: Compare the predicted state with the actual observed device state vector at the current moment to calculate the deviation between the predicted and observed values; based on this deviation and the algorithm's internal estimation of model and observation uncertainties, calculate an optimal correction weight.

[0047] Calibration: Using the corrected weights, the predicted state is corrected to obtain a better state estimate for the current moment; at the same time, this process can also adaptively adjust the key adjustable parameters in the model so that the model can better match the current actual behavior of the device.

[0048] The update step is executed iteratively, continuously outputting the optimal state estimation sequence of the dynamically calibrated digital twin.

[0049] As a further aspect of the present invention, the steps of performing anomaly detection, fault diagnosis, and state trend prediction of power grid equipment include:

[0050] The residual sequence between the digital twin's predicted values ​​and the actual monitoring data is calculated, anomalies are determined based on dynamic thresholds, and a pre-trained classification model is used to identify fault modes of the anomalies.

[0051] By integrating equipment physical mechanisms, historical operation and maintenance reports, typical failure cases and domain expert experience, an initial causal knowledge graph is constructed, and a conditional independence test algorithm is used based on multivariate time series data to discover potential causal relationships, forming a dynamic causal network.

[0052] The identified fault categories are input into the dynamic causal network as observational evidence, and the posterior probability of each potential fault root cause is calculated using a Bayesian network-based inference algorithm.

[0053] All candidate root causes are ranked according to the calculated posterior probabilities, one or more root causes with the highest probabilities are identified, and the associated equipment components and severity levels are determined. Finally, a structured diagnostic report is output, which includes clear root cause conclusions, quantified confidence levels, and fault evolution and propagation path inferences obtained by causal network back-tracing.

[0054] Using the optimal state estimation sequence as initial conditions, the low-dimensional surrogate model is driven to perform multi-step forward physical simulation prediction. At the same time, an LSTM network is used for data-driven prediction. The two prediction results are then fused through adaptive weighting to output the comprehensive state evolution trajectory of the device in a specified future period, the predicted value of the remaining useful life (RUL), and the state evolution prediction results over time.

[0055] As a further aspect of the present invention, the posterior probability calculation formula is as follows:

[0056] ;

[0057] Indicates that given observational evidence Under these conditions, potential root causes of failure The posterior probability of occurrence;

[0058] Indicates potential root causes of failure The prior probability of occurrence; Indicates the root cause of the failure Observational evidence under conditions that actually occur The conditional probability; Indicates observational evidence Marginal probability of occurrence.

[0059] As a further aspect of the present invention, the mathematical model of the multi-objective optimization problem is as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] The total cost of all selected tasks to be executed; This indicates the overall risk index of the entire maintenance plan; This represents the total number of maintenance tasks. Number the task; Decision variable, indicating whether to perform the task. ; Execute the task Total estimated cost; To give decision variables Next, the task The planned completion time; This indicates the latest completion time among all tasks; For the task Risk weighting coefficient; For the task The task;

[0065] The constraints include: resource constraints, task dependency constraints, time window constraints, personnel skill matching constraints, and task integrity constraints.

[0066] As a further aspect of the present invention, the step of solving the multi-objective optimization problem and evaluating the uncertainty of the solution includes:

[0067] Parallel optimization algorithms are used to solve the established multi-objective optimization problem. The parallel optimization algorithm explores different regions of the solution space simultaneously on multiple computing cores and quickly generates a set of feasible candidate maintenance schemes that satisfy all hard constraints. Each candidate scheme specifies in detail the personnel to be executed, the spare parts required, the start and end times, the tools used, and the estimated costs for each task.

[0068] To quantify the uncertainty risks faced by each candidate maintenance plan in actual execution, maintenance entropy is introduced as an evaluation indicator. Maintenance entropy is calculated comprehensively based on the fluctuation probability of resources on which each task in the plan depends, the probability of risk in the working environment, and the technical complexity of the task itself.

[0069] Maintenance Entropy The calculation formula is:

[0070] ;

[0071] Indicates the first step in the implementation of the candidate maintenance plan. The probability of an uncertain event occurring; entropy. The higher the value, the greater the risk of the plan failing or deviating from the intended outcome.

[0072] Based on the autophagy mechanism, the solution screening and feedback are simulated to optimize and screen candidate maintenance solutions.

[0073] Calculate the maintenance entropy of all candidate solutions and set an acceptable entropy threshold based on historical data or management strategies;

[0074] For high-risk solutions with entropy values ​​exceeding the threshold, attempts should be made to automatically repair them, such as finding alternative resources, adding time buffers to the critical path, or decomposing complex tasks to reduce their uncertainty.

[0075] After repair, the maintenance entropy is recalculated. If the entropy value is still higher than the threshold, the risk of the solution is deemed uncontrollable, and it is degraded.

[0076] All schemes with entropy values ​​below the threshold are retained and ranked according to comprehensive performance indicators. The best one or more recommended maintenance schemes are then output for the decision-maker to make the final decision.

[0077] Another aspect of this application provides a digital twin-based intelligent condition prediction system for power grid equipment, the maintenance system comprising:

[0078] The trusted data acquisition module is configured to collect data through multiple types of sensors, perform edge-side preprocessing, hierarchical network transmission, and trusted storage based on consortium blockchain.

[0079] The digital twin construction and update module is configured to generate device state vectors based on multi-source fusion of trusted data, construct and reduce the order of a three-layer physical model, and dynamically calibrate it through a data assimilation algorithm to output the optimal state estimation sequence.

[0080] The state assessment and prediction module is configured to perform anomaly detection, fault mode identification, and root cause analysis based on the optimal state estimation sequence to output diagnostic results, and to integrate physical simulation and LSTM prediction to output state evolution prediction results.

[0081] The intelligent maintenance decision-making module is configured to formalize and solve multi-objective maintenance optimization problems based on diagnostic results and state evolution prediction results, calculate maintenance entropy to assess risks, and screen and output recommended maintenance solutions based on the autophagy mechanism.

[0082] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are as follows:

[0083] Compared with existing technologies, the technical solution provided by this invention achieves the following significant beneficial effects by constructing a complete technical chain of trusted perception, high-fidelity twin, explainable cognition, and intelligent decision-making:

[0084] First, addressing the issue of insufficient credibility and integration in existing data layers, this invention ensures the integrity, authenticity, and immutability of data throughout the entire process from field collection to cloud application through a combination of edge preprocessing, layered transmission, and consortium blockchain evidence storage. Furthermore, it innovatively employs a multi-source data fusion method based on Dempster-Shafer evidence theory, effectively handling conflicts between different sensor evidence and generating a unified, reliable, and quantified device state vector, providing a high-quality data foundation for subsequent processing.

[0085] Secondly, addressing the contradiction between model-level fidelity and real-time performance, this invention creatively proposes a collaborative technical path: evidence fusion state vector + a three-layer physical model of system-component-material + Krylov subspace order reduction + data assimilation and dynamic calibration. This path not only embeds the complete evolution mechanism of the device from macroscopic structure to microscopic materials through a multi-layered physical model, but also significantly improves computational efficiency through model order reduction technology. Furthermore, through a data assimilation system centered on extended / unscented Kalman filtering, it achieves optimal fusion and dynamic calibration of model prediction and real-time observation, thereby outputting a high-fidelity, real-time optimal state estimation sequence that conforms to physical laws and closely reflects the actual device, fundamentally solving the problem of digital twins being both realistic and fast.

[0086] Finally, addressing the issues of poor interpretability and low intelligence in decision-making at the application layer, this invention constructs a dynamic causal network, combining data-driven causal discovery with domain knowledge graphs and utilizing Bayesian networks for probabilistic reasoning. This makes the fault diagnosis and root cause analysis process transparent and the conclusions credible. Simultaneously, by integrating physical mechanism simulation with LSTM data-driven trend prediction, the accuracy of predictions for key indicators such as remaining service life is improved. Building upon this foundation, this invention formalizes maintenance decision-making as a multi-objective optimization problem considering cost, schedule, and risk. It innovatively introduces maintenance entropy to quantify execution uncertainty and combines it with an autophagy mechanism for dynamic scheme selection and repair. Ultimately, it outputs a recommended maintenance scheme that is comprehensively optimal under multiple real-world constraints and possesses strong risk resistance, achieving a closed loop from accurate state perception to scientific operation and maintenance decision-making.

[0087] In summary, through the organic synergy and progressive advancement of the above-mentioned technical aspects, this invention systematically solves the key defects in the prior art and realizes an integrated solution for reliable, accurate, explainable, and intelligent management of the status of power grid equipment. Attached Figure Description

[0088] Figure 1 This is a flowchart of a method for predicting the intelligent state of power grid equipment based on digital twins.

[0089] Figure 2A flowchart of S100, a smart state prediction method for power grid equipment based on digital twins;

[0090] Figure 3 A flowchart of S200, a smart state prediction method for power grid equipment based on digital twins;

[0091] Figure 4 S300 is a flowchart of a digital twin-based intelligent state prediction method for power grid equipment.

[0092] Figure 5 This is a flowchart of S400, a smart state prediction method for power grid equipment based on digital twins. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0094] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0095] The core objective of this invention is:

[0096] This provides a closed-loop intelligent operation and maintenance solution that enables the entire process of power grid equipment operation and maintenance, from high-fidelity real-time status mapping to interpretable diagnosis and accurate prediction of abnormal faults, and intelligent generation of maintenance plans. Specifically, it aims to build a comprehensive system and method based on digital twins, which can:

[0097] 1) Ensure the reliability and deep integration of monitoring data throughout the entire process;

[0098] 2) Dynamically generate and update a digital twin that combines physical depth with online computing performance to accurately reflect the device status;

[0099] 3) Based on the output of the digital twin, transparent and interpretable root cause analysis of faults and accurate state trend prediction are achieved;

[0100] 4) Finally, the prediction and diagnosis results are automatically transformed into the optimal maintenance action plan that comprehensively considers resources, costs, schedule and risks, thereby effectively improving the reliability, economy and intelligence of power grid equipment operation and maintenance.

[0101] The core idea of ​​this method is:

[0102] We will build a hierarchical and collaborative technical framework with the closed-loop drive of data-model-knowledge-decision as the main thread.

[0103] Accurate perception driven by trusted data: Through an edge-cloud collaborative architecture and blockchain evidence storage technology, the authenticity of the data source and the reliability of the transmission process are guaranteed. Evidence theory is used to intelligently fuse multi-source heterogeneous data, providing high-quality observation input for the entire system.

[0104] High-fidelity digital twins are driven by the fusion of mechanisms and data: a multi-level model integrating macroscopic, mesoscopic, and microscopic physical laws is constructed as the skeleton of the digital twin, and computational efficiency bottlenecks are addressed through model order reduction. Most importantly, data assimilation technology is introduced, using reliable observational data as a calibration benchmark to continuously and dynamically correct model predictions, making the digital twin not only similar in form but also in essence, becoming a living model capable of accurately mapping the internal health state of physical devices in real time.

[0105] Explainable cognition driven by causal knowledge: Moving beyond the traditional black-box model, a dynamic causal network driven by both data and knowledge is constructed. Using this network, anomalies detected by digital twins are transformed into root cause diagnoses of faults based on probabilistic reasoning and physical logic. This makes the condition assessment process transparent, the conclusions reliable, and provides clear targets for subsequent maintenance.

[0106] Intelligent decision-making driven by quantitative optimization: The diagnostic and prediction results are combined with real-world resource, time, and risk constraints, formalized into a multi-objective optimization problem. By introducing maintenance entropy to quantify uncertainty risks and simulating an autophagy mechanism to dynamically screen and repair solutions, an automatic transformation and closed-loop feedback from state awareness to optimal action decision-making is achieved.

[0107] like Figure 1 As shown, it illustrates an exemplary method for predicting the intelligent state of power grid equipment based on digital twins, specifically including the following steps:

[0108] The specific details of the above technical solution are as follows:

[0109] S100. Raw data is collected by various types of sensors deployed on power grid equipment, and after edge-side preprocessing and hierarchical network transmission, key feature data is reliably stored.

[0110] S200. Based on the data stored in a reliable manner, the device state vector is generated by fusing evidence theory, a dynamic digital twin is constructed, and the optimal state estimation sequence of the digital twin is output.

[0111] S300. Based on the optimal state estimation sequence, perform anomaly detection, fault diagnosis and state trend prediction of power grid equipment to obtain state assessment information including diagnosis results and prediction results;

[0112] S400. Based on the diagnostic results and state evolution prediction results, the maintenance task and resource matching problem is formalized into a multi-objective optimization problem, and candidate maintenance schemes are generated. By solving the multi-objective optimization problem and evaluating the uncertainty of the schemes, recommended maintenance schemes for physical power grid equipment are generated.

[0113] Please refer to Figure 2 The diagram illustrates a flowchart of an exemplary method S100 for predicting the intelligent state of power grid equipment based on digital twins, as described in this application.

[0114] In a digital twin-based intelligent state prediction method for power grid equipment, the S100 aims to establish a trusted data supply chain spanning the entire process from physical sensing to cloud-based data storage. Its logic follows a closed loop of precise data acquisition, on-site processing, reliable transmission, and trusted data storage: First, sensor deployment is optimized based on model simulation to ensure data validity from the source; then, real-time preprocessing is performed at the edge to extract structured features and reduce transmission burden; next, efficient and reliable data aggregation is achieved through a three-layer network adapted to the power system; finally, consortium blockchain technology is used to store key data, providing a complete, authentic, and tamper-proof data foundation for all subsequent analyses.

[0115] The specific steps include:

[0116] S110. Sensor deployment and data acquisition: Based on the preset monitoring requirements and the sensitive locations determined by physical model simulation, deploy various types of sensors to acquire raw signals;

[0117] In one possible implementation, the sensor includes at least: a piezoelectric vibration sensor for acquiring the mechanical vibration spectrum of the equipment; an ultra-high frequency (UHF) sensor for capturing electromagnetic wave signals generated by partial discharge; a distributed fiber optic temperature sensor for continuous temperature measurement of the equipment surface and key internal points; and an online monitoring sensor for dissolved gases in transformer oil based on infrared spectroscopy or electrochemical principles for real-time analysis of dissolved gases in transformer oil. , , , , Concentration of characteristic gases;

[0118] S120. Edge-side data preprocessing: Deploy edge computing units at each sensor node or regional data aggregation point. The edge computing units perform local real-time processing on the raw acquired signals.

[0119] In one possible implementation, the localized real-time processing includes:

[0120] The vibration signal is bandpass filtered to remove power frequency interference, and time-frequency domain analysis is performed to extract features such as amplitude, frequency, and phase; the UHF signal is amplitude discriminated, pulse classified, and its statistical features are calculated; the temperature and gas concentration data are subjected to moving average or median filtering to smooth noise, and trend indicators such as rate of change are calculated.

[0121] S130. Layered network data transmission, a three-layer data transmission network architecture consisting of field equipment layer, station control layer, and wide area backbone layer;

[0122] The field device layer transmits the preprocessed data to the local data concentrator via CAN or Modbus fieldbus protocols.

[0123] The station control layer aggregates data from each bay to the substation edge server via an internal industrial Ethernet network conforming to the IEC61850 standard.

[0124] The wide-area backbone layer transmits status datasets with unified timestamps to the cloud or regional data centers via the power dispatch data network or 5G power virtual private network.

[0125] S140. Data Trustworthy Storage: At key nodes before data transmission is uploaded to the edge server and after it is received in the data center, consortium blockchain technology is used to store key feature data and its metadata.

[0126] In one possible implementation, the adoption of consortium blockchain technology includes:

[0127] Calculate the hash value of key feature data and its metadata, package the hash value, the hash value of the previous data block, and the digital signature of this node to generate a new block; verify the validity of the new block through a practical Byzantine fault-tolerant consensus mechanism deployed among power grid companies, equipment manufacturers, and maintenance units, and add it to the chain.

[0128] Please refer to Figure 3 The diagram illustrates a flowchart of an exemplary method S200 for predicting the intelligent state of power grid equipment based on digital twins, as described in this application.

[0129] In a digital twin-based intelligent state prediction method for power grid equipment, S200 aims to create a dynamic digital twin capable of high-fidelity, real-time mapping of the physical equipment state. Its core logic is multi-source fusion, mechanism modeling, efficient simulation, and dynamic synchronization: First, it employs evidence theory to fuse reliable multi-source observation data, forming a unified description of the equipment state; then, it constructs a twin model embedding multi-level physical mechanisms of the system, components, and materials, and addresses its real-time computational bottleneck through model order reduction; finally, it uses a data assimilation algorithm to use the fused observations as a calibration benchmark, continuously and optimally correcting the model predictions to ensure that the output sequence of the digital twin remains synchronized with the actual state of the physical equipment.

[0130] The specific steps include:

[0131] S210. Multi-source data fusion: Receive data from various types of sensors preprocessed by S130 and stored in S140, and generate a unified device state vector description using evidence theory methods.

[0132] The generation of a unified device state vector description based on Dempster-Shafer evidence theory includes:

[0133] For each type of sensor data, define a recognition framework that includes normal, attentive, and abnormal states;

[0134] Based on historical data or expert experience, assign a basic probability assignment (BPA) to the observations of each sensor.

[0135] Using the Dempster combination rule, BPAs from different sensors are fused to obtain a comprehensive confidence distribution of consistency regarding the overall health status of the device.

[0136] ;

[0137] This represents the basic probability assignment (BPA) from different sensors. These represent the proposition subsets in each sensor recognition framework; Describes a subset of propositions supported by two sources of evidence. and The intersection is ; The conflict coefficient measures the degree of conflict between two sources of evidence.

[0138] Based on the comprehensive confidence distribution, a multi-dimensional, unified device state vector is generated as the observation input for the digital twin;

[0139] S220. Multi-level physical model construction: Based on the equipment's design drawings, material parameters, and physical laws, a three-level digital twin model is constructed to describe the intrinsic evolution law of the equipment's state.

[0140] The three-layer digital twin model includes:

[0141] The equipment was modeled using the finite element method to create a three-dimensional geometric system-level model; electromagnetic-thermal-mechanical multi-physics coupled component-level models were created for core components such as transformer windings, iron cores, and circuit breaker contacts; and microscopic aging kinetic material-level models based on the Arrhenius equation and fractional derivatives were created for key materials such as insulating paper, transformer oil, and epoxy resin.

[0142] S230. Model Reduction and Real-Time Implementation: To meet the computational efficiency requirements of online simulation, a projection method based on Krylov subspace is used to reduce the order of the three-layer physical model.

[0143] The steps for reducing the order of a three-layer physical model using a Krylov subspace-based projection method include:

[0144] Linearize the three-layer digital twin model near its steady-state equilibrium point, or select a set of representative operating conditions that can cover the typical operating range of the equipment for simulation calculation, so as to obtain the dynamic response data of the model under different input excitations.

[0145] Based on the acquired dynamic response data, a low-dimensional subspace is constructed using the Krylov subspace method. This subspace is spanned by a series of basis vectors. The basis vectors essentially capture and represent the dominant vibrational modes or dynamic modes with the highest energy and the most significant impact on the dynamic behavior of the system in the full-order physical model.

[0146] By using mathematical projection, the high-dimensional state equation of the three-layer digital twin model is projected onto a low-dimensional subspace spanned by the dominant mode. The state equation of the model is then reconstructed in the low-dimensional subspace spanned by the dominant mode, resulting in a low-dimensional proxy model with a significantly reduced dimension of state variables, but retaining the key dynamic characteristics of the original model.

[0147] As the core prediction engine of the digital twin, the surrogate model enables rapid state extrapolation calculations.

[0148] S240. Data assimilation and dynamic calibration: Using the device state vector as the observation value and the low-dimensional surrogate model as the predictor, a data assimilation system is constructed to output the optimal state estimation sequence of the digital twin. This sequence is the real-time output of the high-fidelity digital twin.

[0149] In one possible implementation, the steps for constructing a data assimilation system are as follows:

[0150] Construct a state-space model with a reduced-order model as the state equation and a fused state vector as the observation equation;

[0151] Using either an extended Kalman filter or an unscented Kalman filter algorithm, the following loop is executed at each sampling time step:

[0152] Prediction: Using the state estimate from the previous moment, the current state of the equipment is predicted through the state equation.

[0153] Update: Compare the predicted state with the actual observed device state vector at the current moment to calculate the deviation between the predicted and observed values; based on this deviation and the algorithm's internal estimation of model and observation uncertainties, calculate an optimal correction weight.

[0154] Calibration: Using the corrected weights, the predicted state is corrected to obtain a better state estimate for the current moment; at the same time, this process can also adaptively adjust the key adjustable parameters in the model so that the model can better match the current actual behavior of the device.

[0155] The update step is executed iteratively, continuously outputting the optimal state estimation sequence of the dynamically calibrated digital twin.

[0156] Please refer to Figure 4 The diagram illustrates a flowchart of an exemplary method S300 for predicting the intelligent state of power grid equipment based on digital twins, as described in this application.

[0157] In a digital twin-based intelligent state prediction method for power grid equipment, the S300 aims to perform in-depth analysis of the state sequence output by the digital twin, achieving a complete understanding from anomaly perception to root cause tracing and future prediction. Its logic follows a path of anomaly detection, knowledge fusion, probabilistic diagnosis, and trend prediction: accurate anomaly perception and pattern recognition are achieved by comparing the residuals of the twin prediction with actual observations; a dynamic causal network is constructed and continuously evolves, integrating data-discovered patterns and domain prior knowledge; based on this, probabilistic reasoning is performed using Bayesian networks to achieve interpretable and quantitative fault root cause localization; simultaneously, mechanism simulation and data-driven prediction are integrated to output the future state evolution trajectory and remaining lifespan of the equipment, providing a forward-looking basis for decision-making. Specific steps include:

[0158] S310. Anomaly detection and pattern recognition: continuously compare the real-time state estimation sequence of the high-fidelity digital twin with the pre-processed real-time monitoring data to calculate the residual sequence of key state variables; set a dynamic detection threshold for the residual sequence based on the statistical distribution of historical normal data or the prediction uncertainty interval of the model itself; when the statistical characteristics of the residual sequence continuously deviate from the normal range and exceed the dynamic threshold, it is determined that the equipment has a valid anomaly.

[0159] Formula for calculating the residual sequence of key state variables:

[0160] ;

[0161] For residuals; These are actual observed values; Predicted values ​​for digital twins;

[0162] For each data segment identified as abnormal, quantitative indicators reflecting fault characteristics are extracted. These indicators are obtained from the time domain, frequency domain, or time-frequency domain transformation of the signal. Subsequently, a pre-trained classification model is used to classify the abnormality into one or more predefined fault categories based on the extracted feature indicator set. By learning the relationship between historical abnormal data samples and their corresponding fault labels, the classification model can automatically identify and classify unknown abnormal patterns, and output the identified fault category and its corresponding confidence level.

[0163] S320. Causal knowledge construction and discovery: Construct a dynamic causal network driven by both data and knowledge to understand fault mechanisms;

[0164] By integrating equipment physical mechanisms, historical operation and maintenance reports, typical failure cases and domain expert experience, a structured initial causal knowledge graph is constructed. This graph uses nodes to represent various equipment status parameters, failure modes, external environmental factors and maintenance actions, and directed edges to represent the causal relationships between nodes. Weights describing the probability or degree of influence of the relationship can also be attached.

[0165] Using the device state vector as input, the PC algorithm based on conditional independence test analyzes multivariate time series data, automatically identifies and infers potential causal relationships between state variables, automatically identifies potential causal links and their directions, thereby discovering the causal laws hidden behind the data.

[0166] The causal laws are compared, conflict-resolved, and logically integrated with the constructed initial causal knowledge graph. In the process, the initial graph is supplemented, revised, and even overturned, parts that are seriously inconsistent with the data evidence, ultimately forming a continuously evolving dynamic causal network. This network becomes the core knowledge foundation for subsequent interpretable and probabilistic diagnosis.

[0167] S330. Fault diagnosis and root cause analysis: Based on dynamic causal networks, fault location and confidence assessment are performed.

[0168] The identified fault categories are used as observational evidence and input into the dynamic causal network;

[0169] A Bayesian network-based inference algorithm is used to calculate the posterior probability of each potential root cause node in the network, given the observational evidence.

[0170] Formula for calculating posterior probability:

[0171] ;

[0172] Indicates that given observational evidence Under these conditions, potential root causes of failure The posterior probability of occurrence;

[0173] Indicates potential root causes of failure The prior probability of occurrence; Indicates the root cause of the failure Observational evidence under conditions that actually occur The conditional probability; Indicates observational evidence The marginal probability of occurrence;

[0174] All candidate root causes are ranked according to the calculated posterior probabilities, and one or more root causes with the highest probabilities are identified. The associated equipment components and severity levels are determined. Finally, a structured diagnostic report is output, which includes a clear conclusion on the root cause, a quantified confidence level, and a deduction of the failure evolution and propagation path based on causal network back tracing.

[0175] S340. Predict the future health status trend of the equipment;

[0176] Using the optimal state estimation sequence of the digital twin as the initial condition, input the planned or predicted operating conditions for the future; drive the digital twin to perform multi-step forward simulation, and based on its embedded physical, chemical and other degradation mechanism models, deduce the future evolution trajectory of key performance indicators under the constraints of physical laws.

[0177] Meanwhile, using the device state vector as input, the Long Short-Term Memory (LSTM) network is used to learn the temporal patterns of state changes and directly predict the future values ​​of key indicators.

[0178] An adaptive weighted or uncertainty quantification method is used to fuse physical simulation prediction results with data-driven prediction results, so as to balance the accuracy of mechanism and the adaptability of data. The final output is the comprehensive state evolution trajectory of the device in a future specified period, the predicted value of the remaining useful life (RUL), and the state evolution prediction results over time.

[0179] Please refer to Figure 5 The diagram illustrates a flowchart of an exemplary method S400 for predicting the intelligent state of power grid equipment based on digital twins, as described in this application.

[0180] In a digital twin-based intelligent state prediction method for power grid equipment, S400 aims to transform the diagnostic and prediction results of previous steps into executable optimal action plans. Its logic is problem modeling – solution generation – risk assessment – ​​dynamic selection: First, maintenance tasks are defined based on diagnostic reports and prediction results, and combined with resource and time constraints, the maintenance decision is formalized into a multi-objective optimization problem; then, a parallel optimization algorithm is used to quickly generate numerous feasible candidate solutions; innovatively, maintenance entropy is introduced to quantify the uncertainty of each solution caused by resource fluctuations, environmental risks, etc.; finally, simulating a biological autophagy mechanism, high-entropy solutions are automatically repaired or eliminated, outputting a recommended maintenance solution with optimal comprehensive cost, schedule, and risk, and robustness.

[0181] The specific steps include:

[0182] S410. Maintenance problem modeling: Based on the structured diagnostic report output by S330 and the state evolution prediction results output by S340, define the set of maintenance tasks to be executed; at the same time, obtain the currently available personnel, spare parts, special tools, time windows and other resource information from the resource management system to form a set of available resources; formalize the matching problem of maintenance tasks and resources into a multi-objective combinatorial optimization problem.

[0183] The mathematical model for the multi-objective optimization problem is as follows:

[0184] ;

[0185] ;

[0186] ;

[0187] ;

[0188] The goal is to minimize the total cost of all selected tasks. This indicates the total duration of the entire maintenance project. By optimizing task scheduling and resource allocation, the goal is to shorten the completion time of the last task as much as possible, thereby minimizing the total downtime or power outage of equipment. It represents the overall risk index of the entire maintenance plan. It quantifies the technical, safety and operational risks that the plan may face during the execution process through a weighted sum, with the goal of minimizing the total risk. This represents the total number of maintenance tasks. Number the task; Decision variable, indicating whether to perform the task. ; Execute the task Total estimated cost; To give decision variables Next, the task The planned completion time; This indicates the latest completion time among all tasks; For the task Risk weighting coefficient; For the task The task;

[0189] The constraints are:

[0190] Resource constraints: At any given time, the total amount of all types of resources used must not exceed the total amount available.

[0191] Task dependency constraints: Tasks with a sequential relationship must be executed in sequence, and subsequent tasks can only begin after the preceding task is completed, with a necessary safety or technical interval between them. This sequential relationship is determined based on the fault propagation paths and maintenance process requirements identified in the S330 structured diagnostic report.

[0192] Time window constraint: Each maintenance task must begin execution within the allowed time window. The lower limit of the time window is usually the currently schedulable time, while the upper limit is determined by the S340 state evolution prediction results, especially the predicted time node when the failure risk exceeds an acceptable level.

[0193] Personnel skill matching constraint: Personnel performing a task must possess at least all the skills required for that task.

[0194] Task integrity constraint: Each task should be completed continuously within a reasonable duration to avoid unnecessary splitting or interruption, so as to ensure maintenance quality and efficiency.

[0195] S420. Parallel generation of candidate solutions: Parallel optimization algorithms are used to solve the established multi-objective optimization problem. The parallel optimization algorithm explores different regions of the solution space simultaneously on multiple computing cores, and quickly generates a set of feasible candidate maintenance solutions that satisfy all hard constraints. Each candidate solution specifies in detail the personnel to be executed for each task, the spare parts required, the start and end times, the tools used, and the estimated cost.

[0196] S430. Maintenance Entropy Calculation and Uncertainty Assessment: To quantify the uncertainty risks faced by each candidate maintenance plan in actual execution, maintenance entropy is introduced as an assessment indicator. Maintenance entropy is calculated comprehensively based on the fluctuation probability of resources on which each task in the plan depends, the risk probability of the working environment, and the technical complexity of the task itself.

[0197] Maintenance Entropy The calculation formula is:

[0198] ;

[0199] Indicates the first step in the implementation of the candidate maintenance plan. The probability of an uncertain event occurring; entropy. The higher the value, the greater the risk of the plan failing or deviating from the intended outcome.

[0200] S440. Solution screening and feedback based on autophagy mechanism: simulate the autophagy process of biological cells to optimize and screen candidate maintenance solutions;

[0201] Calculate the maintenance entropy of all candidate solutions and set an acceptable entropy threshold based on historical data or management strategies;

[0202] For high-risk solutions with entropy values ​​exceeding the threshold, attempts should be made to automatically repair them, such as finding alternative resources, adding time buffers to the critical path, or decomposing complex tasks to reduce their uncertainty.

[0203] After repair, the maintenance entropy is recalculated. If the entropy value is still higher than the threshold, the risk of the solution is deemed uncontrollable, and it is degraded.

[0204] All schemes with entropy values ​​below the threshold are retained and ranked according to comprehensive performance indicators. The best one or more recommended maintenance schemes are then output for the decision-maker to make the final decision.

[0205] Example 2

[0206] A digital twin-based intelligent condition prediction system for power grid equipment includes:

[0207] The trusted data acquisition module is configured to collect data through multiple types of sensors, perform edge-side preprocessing, hierarchical network transmission, and trusted storage based on consortium blockchain.

[0208] The digital twin construction and update module is configured to generate device state vectors based on multi-source fusion of trusted data, construct and reduce the order of a three-layer physical model, and dynamically calibrate it through a data assimilation algorithm to output the optimal state estimation sequence.

[0209] The state assessment and prediction module is configured to perform anomaly detection, fault mode identification, and root cause analysis based on the optimal state estimation sequence to output diagnostic results, and to integrate physical simulation and LSTM prediction to output state evolution prediction results.

[0210] The intelligent maintenance decision-making module is configured to formalize and solve multi-objective maintenance optimization problems based on diagnostic results and state evolution prediction results, calculate maintenance entropy to assess risks, and screen and output recommended maintenance solutions based on the autophagy mechanism.

[0211] The above modules interact with the message middleware through standard API interfaces and call services, together forming an intelligent operation and maintenance platform that integrates hardware and software and coordinates cloud and edge.

[0212] The sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0213] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for intelligent state prediction of power grid equipment based on digital twins, characterized in that, Includes the following steps: Raw data is collected by various types of sensors deployed on power grid equipment. After edge-side preprocessing and hierarchical network transmission, key feature data is reliably stored. Based on the data with trusted evidence storage, the device state vector is generated by fusing evidence theory, a dynamic digital twin is constructed, and the optimal state estimation sequence of the digital twin is output. Based on the optimal state estimation sequence, anomaly detection, fault diagnosis and state trend prediction of power grid equipment are performed to obtain state assessment information including diagnosis results and prediction results. Based on the diagnostic results and state evolution prediction results, the problem of matching maintenance tasks with resources is formalized into a multi-objective optimization problem, and candidate maintenance schemes are generated. By solving the multi-objective optimization problem and evaluating the uncertainty of the schemes, recommended maintenance schemes for physical power grid equipment are generated.

2. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 1, characterized in that, The steps for credibly storing key feature data include: Sensitive locations are determined based on monitoring requirements and physical model simulation. Multiple sensors are deployed to collect raw signals. The sensors include at least vibration sensors, UHF sensors, distributed fiber optic temperature sensors, and online dissolved gas monitoring sensors in oil. Edge computing units are deployed at sensor nodes or regional data aggregation points to perform local real-time processing of raw signals; Localized real-time processing: Bandpass filtering is applied to vibration signals to remove power frequency interference, and time-frequency domain analysis is performed to extract amplitude, frequency, and phase characteristics; amplitude discrimination, pulse classification, and statistical characteristic calculation are performed on UHF signals; moving average or median filtering is applied to temperature and gas concentration data to smooth noise, and the rate of change trend index is calculated. The pre-processed data is transmitted through a three-layer network architecture consisting of a field device layer, a station control layer, and a wide area backbone layer. At critical nodes in data transmission, consortium blockchain technology is used to store key feature data and its metadata.

3. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 1, characterized in that, The steps for constructing a dynamic digital twin include: The system receives preprocessed data from various types of sensors with trusted evidence storage and generates a unified device state vector description based on Dempster-Shafer evidence theory. Multi-level physical model construction: Based on the equipment's design drawings, material parameters, and physical laws, a three-layer digital twin model is constructed to describe the intrinsic evolution law of the equipment's state. The three-layer digital twin model includes: a three-dimensional geometric system-level model of the equipment established using the finite element method; an electromagnetic-thermal-mechanical multi-physics coupled component-level model established for core components such as transformer windings, iron cores, and circuit breaker contacts; and a microscopic aging dynamics material-level model established for key materials such as insulating paper, transformer oil, and epoxy resin based on the Arrhenius equation and fractional derivatives. Model reduction and real-time processing: To meet the computational efficiency requirements of online simulation, a projection method based on Krylov subspace is used to reduce the order of the three-layer digital twin model. Data assimilation and dynamic calibration involve using the device state vector as an observation and a low-dimensional surrogate model as a predictor to construct a data assimilation system. Dynamic calibration is then performed using an extended Kalman filter or an unscented Kalman filter algorithm to output the optimal state estimation sequence.

4. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 3, characterized in that, The steps for generating a unified device state vector description based on Dempster-Shafer evidence theory include: For each type of sensor data, define a recognition framework that includes normal, attentive, and abnormal states; Based on historical data or expert experience, assign a basic probability assignment (BPA) to the observations of each sensor. Using the Dempster combination rule, BPAs from different sensors are fused to obtain a comprehensive confidence distribution of consistency regarding the overall health status of the device. ； This represents the basic probability assignment (BPA) from different sensors. These represent the proposition subsets in each sensor recognition framework; Describes a subset of propositions supported by two sources of evidence. and The intersection is ; The conflict coefficient measures the degree of conflict between two sources of evidence. Based on the comprehensive confidence distribution, a multidimensional, unified device state vector is generated as the observation input for the digital twin.

5. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 3, characterized in that, The step of using the Krylov subspace-based projection method to reduce the order of the three-layer digital twin model; The steps for reducing the order of a three-layer digital twin model using a Krylov subspace-based projection method include: Linearize the three-layer digital twin model near its steady-state equilibrium point, or select a set of representative operating conditions that can cover the typical operating range of the equipment for simulation calculation, so as to obtain the dynamic response data of the model under different input excitations. Based on the acquired dynamic response data, a low-dimensional subspace is constructed using the Krylov subspace method. The low-dimensional subspace is spanned by a series of basis vectors. The basis vectors essentially capture and represent the dominant vibrational modes or dynamic modes with the highest energy and the most significant impact on the dynamic behavior of the system in the full-order physical model. By using mathematical projection, the high-dimensional state equation of the three-layer digital twin model is projected onto a low-dimensional subspace spanned by the dominant mode. The state equation of the model is then reconstructed in the low-dimensional subspace spanned by the dominant mode, resulting in a low-dimensional proxy model with a significantly reduced dimension of state variables, but retaining the key dynamic characteristics of the original model. As the core prediction engine of the digital twin, the surrogate model enables rapid state extrapolation calculations.

6. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 1, characterized in that, The steps for anomaly detection, fault diagnosis, and status trend prediction of power grid equipment include: The residual sequence between the digital twin's predicted values ​​and the actual monitoring data is calculated, anomalies are determined based on dynamic thresholds, and a pre-trained classification model is used to identify fault modes of the anomalies. By integrating equipment physical mechanisms, historical operation and maintenance reports, typical failure cases and domain expert experience, an initial causal knowledge graph is constructed, and a conditional independence test algorithm is used based on multivariate time series data to discover potential causal relationships, forming a dynamic causal network. The identified fault categories are used as observational evidence and input into the dynamic causal network. A Bayesian network-based inference algorithm is used to calculate the posterior probability of each potential fault root cause. All candidate root causes are ranked according to the calculated posterior probabilities, one or more root causes with the highest probabilities are identified, and the associated equipment components and severity levels are determined. Finally, a structured diagnostic report is output, which includes clear root cause conclusions, quantified confidence levels, and fault evolution and propagation path inferences obtained by causal network back-tracing. Using the optimal state estimation sequence as initial conditions, a low-dimensional surrogate model is driven to perform multi-step forward physical simulation prediction. At the same time, an LSTM network is used for data-driven prediction. The two prediction results are then fused through adaptive weighting to output the comprehensive state evolution trajectory of the device in a specified future period, the predicted value of the remaining useful life (RUL), and the state evolution prediction results over time.

7. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 6, characterized in that, The formula for calculating the posterior probability is as follows: ； Indicates that given observational evidence Under these conditions, potential root causes of failure The posterior probability of occurrence; Indicates potential root causes of failures The prior probability of occurrence; Indicates the root cause of the failure Observational evidence under conditions that actually occur The conditional probability; Indicates observational evidence The marginal probability of occurrence.

8. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 1, characterized in that, The mathematical model for the multi-objective optimization problem is as follows: ； ； ； ； The total cost of all selected tasks to be executed; This indicates the overall risk index of the entire maintenance plan; This represents the total number of maintenance tasks. Number the task; Decision variable, indicating whether to perform the task. ; Execute the task Total estimated cost; To give decision variables Next, the task The planned completion time; This indicates the latest completion time among all tasks; For the task Risk weighting coefficient; For the task The task; The constraints include: resource constraints, task dependency constraints, time window constraints, personnel skill matching constraints, and task integrity constraints.

9. The method for intelligent state prediction of power grid equipment based on digital twins according to claim 1, characterized in that, The steps for solving the multi-objective optimization problem and evaluating the uncertainty of the solution are as follows: Parallel optimization algorithms are used to solve the established multi-objective optimization problem. The parallel optimization algorithm explores different regions of the solution space simultaneously on multiple computing cores and quickly generates a set of feasible candidate maintenance schemes that satisfy all hard constraints. Each candidate scheme specifies in detail the personnel to be executed, the spare parts required, the start and end times, the tools used, and the estimated costs for each task. To quantify the uncertainty risks faced by each candidate maintenance plan in actual execution, maintenance entropy is introduced as an evaluation indicator. Maintenance entropy is calculated comprehensively based on the fluctuation probability of resources on which each task in the plan depends, the probability of risk in the working environment, and the technical complexity of the task itself. Maintenance Entropy The calculation formula is: ； Indicates the first step in the implementation of the candidate maintenance plan. The probability of an uncertain event occurring; entropy. The higher the value, the greater the risk of the plan failing or deviating from the intended outcome. Based on the autophagy mechanism, the solution screening and feedback are simulated to optimize and screen candidate maintenance solutions. Calculate the maintenance entropy of all candidate solutions and set an acceptable entropy threshold based on historical data or management strategies; For high-risk solutions with entropy values ​​exceeding the threshold, attempts are made to automatically repair them by finding alternative resources, adding time buffers to the critical path, or decomposing complex tasks to reduce their uncertainty. After repair, the maintenance entropy is recalculated. If the entropy value is still higher than the threshold, the risk of the solution is deemed uncontrollable, and it is degraded. All schemes with entropy values ​​below the threshold are retained and ranked according to comprehensive performance indicators. The best one or more recommended maintenance schemes are then output for the decision-maker to make the final decision.

10. A system for predicting the intelligent state of power grid equipment based on digital twins according to any one of claims 1-9, characterized in that, The system includes: The trusted data acquisition module is configured to collect data through multiple types of sensors, perform edge-side preprocessing, hierarchical network transmission, and trusted storage based on consortium blockchain. The digital twin construction and update module is configured to generate device state vectors based on multi-source fusion of trusted data, construct and reduce the order of a three-layer physical model, and dynamically calibrate it through a data assimilation algorithm to output the optimal state estimation sequence. The state assessment and prediction module is configured to perform anomaly detection, fault mode identification, and root cause analysis based on the optimal state estimation sequence to output diagnostic results, and to integrate physical simulation and LSTM prediction to output state evolution prediction results. The intelligent maintenance decision-making module is configured to formalize and solve multi-objective maintenance optimization problems based on diagnostic results and state evolution prediction results, calculate maintenance entropy to assess risks, and screen and output recommended maintenance solutions based on the autophagy mechanism.

Images