Digital twin-driven transformer substation state artificial intelligence evaluation system

Abstract

The invention relates to the technical field of digital twinning, discloses a digital twinning-driven transformer substation state artificial intelligence evaluation system, and aims to solve the problems that the diagnosis precision is reduced and the model adaptability is insufficient when a traditional digital twinning model is inconsistent with a physical entity state. The system constructs a closed-loop architecture comprising a state perception and deviation quantification module, a systematic risk and cognitive focus distribution module, a focused causal diagnosis and introspection verification module and a cognitive twin adaptive evolution module. The method can actively sense the state deviation of a physical entity, intelligently distributes diagnosis resources based on the evaluation of systematic risks, and achieves the efficient diagnosis of a fault source. The system determines the real state of a physical entity through introspection verification, and performs closed-loop updating on a reference twin model by using the result, so that the reference twin model can be continuously evolved, and the accuracy, the automation level and the full life cycle adaptability of state monitoring and fault diagnosis of a complex industrial system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, specifically to a digital twin-driven artificial intelligence assessment system for substation status. Background Technology

[0002] Digital twin technology, serving as a core bridge connecting the physical world and the digital space, has been widely applied in key fields such as aerospace, power systems, and intelligent manufacturing. By constructing high-fidelity physical models, digital twins can achieve status monitoring, performance prediction, and virtual testing throughout the entire lifecycle of physical assets, providing strong support for optimizing operation and maintenance decisions.

[0003] However, existing digital twin technologies still face significant challenges in practical applications. Traditional digital twin models are typically built based on design parameters or initial states, and these parameters tend to remain static once set. During long-term operation, physical entities experience state drift due to normal aging, wear and tear, potential failures, or changes in the external environment, causing their actual operating characteristics to gradually deviate from the twin model's baseline description. This "state mismatch" between the model and the entity severely weakens the accuracy of the twin system in deviation detection and fault diagnosis, and may even lead to misleading conclusions due to an incorrect model baseline.

[0004] Furthermore, when a system detects a state deviation, existing diagnostic methods often lack a deep understanding of the concurrent occurrence of multiple faults and the propagation chain of fault effects in complex systems. They typically struggle to automatically and efficiently locate the root cause from massive amounts of data and a multitude of possible causes, relying heavily on expert experience for manual intervention. This makes the diagnostic process inefficient and unsuitable for large-scale, highly dynamic industrial scenarios. Existing technologies also fail to provide an effective closed-loop mechanism that enables digital twins to autonomously learn from diagnostic results and thereby complete self-correction and evolution of the model, thus fundamentally solving the aforementioned "state mismatch" problem. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing substation condition assessment methods typically respond passively after detecting anomalies, lacking a mechanism to dynamically adjust the focus of analysis based on the risks posed by equipment failures to the entire system. Furthermore, existing methods struggle to perform in-depth causal reasoning and quantitative identification when determining the root physical causes of condition deviations, resulting in inaccurate diagnostic conclusions and an inability to effectively track the gradual performance degradation process of equipment.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a digital twin-driven artificial intelligence assessment system for substation status. This system is specifically designed in terms of structure and function to achieve risk-oriented in-depth diagnosis and adaptive evolution of the model.

[0007] Specifically, the system includes:

[0008] The state perception and deviation quantification module is used to establish and output the theoretical state vector of the physical equipment based on the reference twin model of the substation equipment, obtain the actual measured state vector of the physical equipment, and calculate the state deviation residual vector between the theoretical state vector output by the reference twin model and the actual measured state vector.

[0009] The systemic risk and cognitive focus allocation module is used to infer the impact of each physical device on the system based on preset failure modes to determine vulnerability scores, and to generate a cognitive focus weight vector based on the vulnerability scores.

[0010] The focused causal diagnosis and introspection verification module is used to perform causal reasoning on the physical devices corresponding to the state deviation residual vector based on the cognitive focus weight vector to generate a set of fault hypotheses, and to conduct active experiments in the twin copies of the fault hypotheses to verify and identify the optimal fault hypothesis, thereby outputting the identified current state parameter vector.

[0011] The cognitive twin adaptive evolution module is used to receive the identified current state parameter vector and update the baseline twin model in the state perception and deviation quantization module based on the identified current state parameter vector.

[0012] The state perception and deviation quantification module includes:

[0013] The baseline twin model running unit is used to run the baseline twin model and generate the theoretical state vector based on the built-in initial parameter vector that characterizes the ideal health state of the equipment and by receiving real-time operating condition data.

[0014] The state data acquisition and deviation calculation unit is used to acquire the actual measured state vector of the physical device, and calculate the state deviation residual vector based on the state deviation residual vector calculation formula, the theoretical state vector and the actual measured state vector;

[0015] A diagnostic activation unit is used to activate the systemic risk and cognitive focus allocation module when the norm of the state deviation residual vector continuously exceeds a preset threshold.

[0016] The formula for calculating the state deviation residual vector is as follows:

[0017] ;

[0018] in, Let the state deviation residual vector be... This is the actual measured state vector of the physical device; The theoretical state vector output by the baseline twin model. , This provides real-time operating data for physical equipment. For the parameter vector adjusted within the fault hypothesis space, This is the theoretical simulation output for twin copies.

[0019] The systemic risk and cognitive focus allocation module includes:

[0020] The fault impact simulation unit is used to embed a whole-station power grid topology model, inject a virtual fault of a single device into the whole-station power grid topology model, and positively simulate the impact of the virtual fault on the stability of the whole station, and quantify it into the systemic impact index of the physical device.

[0021] The focus weight generation unit is used to map the systemic impact index to the vulnerability score, and generate the cognitive focus weight vector based on the vulnerability scores of all devices and the cognitive focus weight vector calculation formula.

[0022] The formula for calculating the cognitive focus weight vector is as follows:

[0023] ;

[0024] in, For equipment Cognitive focus weight, For equipment Vulnerability score, The total number of devices. , For the focal concentration parameter, The natural number exponential function with base e is given by the expression . It is the index variable for the summation operation in the denominator. Used to iterate from 1 to All physical devices, and from 1 to The vulnerability scores of all physical devices are summed. For equipment Vulnerability score.

[0025] The focused causal diagnosis and introspection verification module includes:

[0026] The causal graph reasoning unit is used to build a knowledge graph corresponding to the physical device, and use the state deviation residual vector as evidence to perform source reasoning on the knowledge graph to generate the fault hypothesis set;

[0027] An active experimental unit is used to create an independent twin copy for each fault hypothesis in the fault hypothesis set, and actively adjust the physical parameters corresponding to the fault hypothesis in the twin copy with the goal of reproducing the actual measured state vector of the physical device.

[0028] The optimal hypothesis determination unit is used to determine the fault hypothesis that has the smallest difference from the actual measured state vector of the physical device as the optimal fault hypothesis, and to determine the identified current state parameter vector corresponding to the optimal fault hypothesis based on the formula for determining the actual measured state vector and the current state parameter vector of the physical device.

[0029] The formula for determining the current state parameter vector is as follows:

[0030] ;

[0031] in, This is the actual measurement state vector of the physical device. For the aforementioned baseline twin model, This provides real-time operating data for physical equipment. This is the identified current state parameter vector corresponding to the optimal fault hypothesis. The theoretical simulation output for twin copies, To minimize the parameter solution operator, Norm operator.

[0032] The cognitive twin adaptive evolution module includes:

[0033] The benchmark update unit is used to replace the original initial parameter vector of the benchmark twin model with the identified current state parameter vector, thereby completing the closed-loop update of the system cognitive benchmark.

[0034] The focused causal diagnosis and introspection verification module is also used to prioritize the diagnosis of multiple devices that have the state deviation residual vector based on the cognitive focus weight vector, and to prioritize the causal inference and active experiment of the device with the highest weight.

[0035] The reference twin model is a dynamic model that couples at least two of the physical fields: electromagnetic field, thermal flow field, and mechanical stress field.

[0036] This invention provides a digital twin-driven artificial intelligence assessment system for substation status. It offers the following advantages:

[0037] 1. By setting up a systemic risk and cognitive focus allocation module, this invention can quantify the vulnerability score of each device based on its impact on the stability of the entire station, and further generate a cognitive focus weight vector. This mechanism enables the system to prioritize subsequent in-depth diagnostic resources on the key devices that pose the greatest potential risk to the system, avoiding indiscriminate and time-consuming analysis of all devices, thereby significantly improving the overall efficiency and response speed of status assessment.

[0038] 2. This invention achieves in-depth diagnosis from phenomenon to essence by setting up a focused causal diagnosis and introspection verification module. This module not only generates a set of fault hypotheses through causal reasoning, but also conducts active experiments in twin copies to quantitatively solve the current state parameter vector that can best reproduce the measured state of the physical equipment. This process elevates the diagnostic conclusion from traditional qualitative judgment to quantitative identification, greatly improving the accuracy of the diagnostic results.

[0039] 3. This invention constructs a complete cognitive closed loop by setting up a cognitive twin adaptive evolution module. This module uses the parameter vectors identified by diagnosis, which reflect the current real state of the device, to update the initial parameter vectors in the benchmark twin model. In this way, the evaluation benchmark is no longer a static and unchanging model, but can keep pace with the performance degradation and state evolution of the physical device, realizing the adaptive evolution of the model and ensuring the long-term accuracy and effectiveness of the system in state evaluation throughout the entire life cycle of the device. Attached Figure Description

[0040] Figure 1 A structural block diagram of a digital twin-driven artificial intelligence assessment system for substation status, provided in one embodiment of the present invention;

[0041] Figure 2 This is an internal workflow diagram of the state perception and deviation quantification module of the present invention;

[0042] Figure 3 This is an internal workflow diagram of the systemic risk and cognitive focus allocation module of this invention;

[0043] Figure 4 This is an internal workflow diagram of the focused causal diagnosis and introspection verification module of the present invention;

[0044] Figure 5 This is a schematic diagram of the structure of the equipment fault knowledge graph of the present invention;

[0045] Figure 6 This is a schematic diagram of the whole-station power grid topology model and vulnerability assessment process of the present invention;

[0046] Figure 7 This is a schematic diagram of the cognitive twin adaptive evolution closed-loop update mechanism of the present invention. Detailed Implementation

[0047] The technical solutions in 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please refer to the appendix. Figure 1 , attached Figure 1 This is a structural block diagram of a digital twin-driven artificial intelligence assessment system for substation status according to an embodiment of the present invention. The digital twin-driven artificial intelligence assessment system for substation status provided by the present invention may include: a status perception and deviation quantification module 100, a systemic risk and cognitive focus allocation module 200, a focused causal diagnosis and introspection verification module 300, and a cognitive twin adaptive evolution module 400.

[0049] In a specific workflow, the state perception and deviation quantization module 100 runs first. Internally, this module establishes and runs a baseline twin model of the substation equipment. This model receives real-time operating condition data from external input and outputs a theoretical state vector of the physical equipment. Simultaneously, the module acquires the actual measured state vector from the physical equipment. By performing vector subtraction between the theoretical and actual measured state vectors, the residual state deviation vector between the two is calculated.

[0050] When the norm of the state deviation residual vector continues to exceed the preset threshold, the state perception and deviation quantification module 100 outputs the residual vector and related information to the systemic risk and cognitive focus allocation module 200.

[0051] After receiving the data, the Systemic Risk and Cognitive Focus Allocation Module 200 begins operation. First, based on the built-in whole-station power grid topology model, the module 200 quantifies and determines the vulnerability score of each physical device by injecting virtual faults into each device and simulating their impact on the overall system stability. Then, using the vulnerability scores of all physical devices as input, the module 200 applies a Softmax function to generate a cognitive focus weight vector. This process is specifically implemented through the following formula:

[0052] ;

[0053] in, For physical equipment Cognitive focus weight; For physical equipment Vulnerability score; The total number of physical devices being evaluated; This is the preset focus concentration parameter; To find the sum index, iterate from 1 to... All physical devices.

[0054] After generating the cognitive focus weight vector, the systemic risk and cognitive focus allocation module 200 outputs this vector along with the original state deviation residual vector to the focused causal diagnosis and introspection verification module 300.

[0055] The focused causal diagnosis and introspection verification module 300 prioritizes the diagnosis of multiple physical devices exhibiting deviations based on the received cognitive focus weight vector. The module prioritizes processing the physical devices with the highest weights, generating a set of fault hypotheses through causal reasoning. To identify the optimal fault hypothesis from this set, the module creates an independent twin copy for each fault hypothesis and seeks to minimize the difference between the simulation output of the twin copy and the actual measured state vector of the physical device by adjusting the physical parameters associated with that fault hypothesis. This process of finding the optimal solution is achieved by solving the following optimization problem:

[0056] ;

[0057] in, This is the actual measured state vector. As a baseline twin model, For real-time operating condition data, For the parameter vector adjusted within the fault hypothesis space, The theoretical simulation output for twin copies, To minimize the parameter solution operator, Norm operator.

[0058] Once the above optimization problem is solved, the focused causal diagnosis and introspection verification module 300 determines the optimal fault hypothesis and obtains the identified current state parameter vector that can characterize the current true physical condition of the equipment. Subsequently, the focused causal diagnosis and introspection verification module 300 uses this parameter vector Output to the cognitive twin adaptive evolution module 400.

[0059] The cognitive twin adaptive evolutionary module 400 receives the identified current state parameter vector. Then, perform the update operation.

[0060] Specifically, the cognitive twin adaptive evolution module 400 will vectorize this parameter. The data is transmitted to the State Awareness and Deviation Quantization Module 100 to replace the original initial parameter vector in the baseline twin model. This operation completes a closed-loop process from state awareness, risk assessment, deep diagnostics to model update, enabling the baseline twin model to reflect the current state of the physical device, rather than just the initial ideal state.

[0061] Please refer to the appendix. Figure 2 The specific implementation of the state perception and deviation quantization module 100 will now be described. A core component of the state perception and deviation quantization module 100 is a reference twin model. This reference twin model is a high-fidelity digital model established for a specific substation device (e.g., a power transformer), which internally contains mathematical equations describing the physical behavior of the device.

[0062] In one embodiment, the baseline twin model is a dynamic model that couples at least two of the physical fields: electromagnetic, thermal, and mechanical stress fields. Based on finite element analysis or other numerical methods, this model can simulate the multiphysics response of a physical device under different operating conditions.

[0063] When the baseline twin model is initialized, it is internally configured with an initial parameter vector representing the physical device in an ideal health state. This vector It contains a set of key physical property parameters that represent the new or optimal state of a physical device. Taking a power transformer as an example, this initial parameter vector... This may include: the DC resistance of the winding at a reference temperature, the thermal conductivity of the insulating paperboard, the heat dissipation coefficient of the transformer oil, and the Young's modulus of the core material, etc.

[0064] During system operation, the baseline twin model is configured to receive real-time operating condition data. As its external input. Real-time operating condition data. Data from field acquisition systems, such as SCADA (Supervisory Control and Data Acquisition) systems, can specifically include: the load current borne by the physical equipment, the operating voltage of the power grid, and the ambient temperature of the environment in which the physical equipment is located.

[0065] The function of the baseline twin model is to, based on its built-in initial parameter vector, and with real-time operating data As a boundary condition, a theoretical state vector is calculated and output by solving its internal physical equations. The theoretical state vector output by this method. This describes the physical state that a physical device should have under ideal health conditions and current operating conditions. Taking a power transformer as an example, the theoretical state vector... This can include: calculated winding hot spot temperature, theoretical dissolved gas concentration in oil, and theoretical mechanical stress distribution at key structural points. The calculation process can be expressed as: ,in The set of mathematical operations representing the baseline twin model.

[0066] In the state awareness and deviation quantization module 100, in addition to running the benchmark twin model to generate theoretical state vectors... In addition, this module is also configured to acquire the actual measured state vectors of physical equipment from the substation's monitoring system. .

[0067] Actual measured state vector The data comes from various sensors deployed on the physical device itself.

[0068] In one specific embodiment, these sensors may include fiber optic grating sensors for measuring temperature, chromatographic analyzers for online monitoring of dissolved gases in oil, ultra-high frequency sensors or acoustic sensors for monitoring partial discharge, and phasor measurement units (PMUs) for measuring electrical quantities. The data collected by these sensors is integrated by a data acquisition and processing unit to form a theoretical state vector. The actual measured state vector corresponding in dimensional and physical sense For example, if Including theoretical winding temperature and theoretical acetylene concentration, then This corresponds to the actual winding temperature and actual acetylene concentration, which are directly measured or indirectly calculated by the sensor.

[0069] Obtain the theoretical state vector With the actual measured state vector Subsequently, the state perception and deviation quantization module 100 performs a vector subtraction operation to generate a state deviation residual vector. The calculation process is defined by the following formula:

[0070] ;

[0071] in, This is the residual vector of state deviation; This is the actual measured state vector of the physical device; This is the theoretical state vector output by the baseline twin model. The subtraction operation is performed element-wise, meaning that each element in the residual vector represents the deviation between the measured and theoretical values ​​of a specific physical state quantity.

[0072] The generated state deviation residual vector Each component of the vector quantitatively characterizes the degree to which the physical device deviates from its ideal health state in a specific state dimension. When the health state of the physical device remains unchanged, the component values ​​of this vector will fluctuate around zero, with the amplitude depending on the measurement noise and model accuracy. When the physical device experiences performance degradation or early failure, one or more components will continuously deviate from zero, and this vector becomes the direct and objective basis for subsequent diagnostic analysis.

[0073] In the state awareness and deviation quantization module 100, to avoid frequent activation of subsequent complex diagnostic processes due to normal measurement noise or minor model errors, this module further executes a diagnostic activation mechanism. This mechanism is used to determine the state deviation residual vector. Whether the deviation being characterized is significant and persistent.

[0074] In one specific embodiment, the state perception and deviation quantization module 100 includes a diagnostic activation unit. This unit receives the state deviation residual vector generated by the aforementioned steps. And conduct a quantitative evaluation of it.

[0075] To perform this assessment, the diagnostic activation unit first calculates the state deviation residual vector. The norm is a scalar value used to characterize the overall magnitude of the bias vector. The norm can be calculated using the Euclidean norm (i.e., the L2 norm), but this invention is not limited to this. This calculation transforms a multidimensional bias vector into a single metric that quantifies its severity.

[0076] The diagnostic activation unit then compares the calculated norm value with a pre-set diagnostic threshold. This threshold is determined based on the specific type of physical equipment, statistical analysis of historical operating data, and an engineering definition of the acceptable error range.

[0077] The diagnostic activation is not based on a single, instantaneous comparison. The diagnostic activation unit determines that the norm value consistently exceeds a preset threshold—for example, if the calculated norm value is greater than the threshold in each of multiple consecutive preset sampling periods. This determination of deviation persistence effectively filters out transient interference, ensuring that only stable, physically meaningful state deviations trigger subsequent analysis.

[0078] When the diagnostic conditions are deemed met, the state perception and deviation quantization module 100 will generate the current state deviation residual vector. Along with relevant physical device identification information and timestamp information, the data is sent to the systemic risk and cognitive focus allocation module 200 to initiate subsequent risk assessment and in-depth diagnostic processes. If the diagnostic conditions are not met, the system remains in the state-aware loop without initiating subsequent modules.

[0079] Please refer to the appendix. Figure 3 and attached Figure 6 The specific implementation of the systemic risk and cognitive focus allocation module 200 will now be described. The systemic risk and cognitive focus allocation module 200 is configured to start operating after receiving the state deviation residual vector sent by the state awareness and deviation quantification module 100. Its core function is to determine a vulnerability score for each critical physical device in the substation.

[0080] Within the systemic risk and cognitive focus allocation module 200, a pre-configured whole-station power grid topology model is provided. This model digitally represents all key physical equipment and their electrical connections within the substation, where physical equipment (such as transformers and circuit breakers) are defined as nodes, and electrical connections (such as buses and lines) are defined as edges. Each node and edge is associated with its own electrical parameters, such as rated capacity, impedance, and admittance.

[0081] To determine the vulnerability score of each physical device The systemic risk and cognitive focus allocation module 200 sequentially performs a virtual fault injection and impact simulation calculation process for each physical device in the topology model to assess the physical devices. Taking a main transformer as an example, virtual fault injection refers to temporarily modifying the data structure of the topology model to reflect the physical equipment. The relevant parameters are used to simulate a state where the transformer experiences a severe fault and is taken out of service. One specific implementation is to set the branch impedance representing the main transformer to an extremely large value.

[0082] In terms of physical equipment After the virtual fault injection is completed, the Systemic Risk and Cognitive Focus Allocation Module 200 immediately performs a full-station power flow calculation based on the modified topology model. This power flow calculation, based on the basic principles of power system analysis, solves for the new steady-state operating parameters of the entire power grid under the virtual fault condition, including the voltage amplitude and phase angle of each bus, as well as the power distribution on each line.

[0083] After completing a power flow calculation, the systemic risk and cognitive focus allocation module 200 quantifies the impact of the virtual fault on the overall stability of the station by comparing the power grid state after the fault with that before the fault, in order to generate a systemic impact index.

[0084] In one embodiment, this indicator can be determined by calculating the sum of the voltage offsets of all busbars in the entire station, or by counting the number of lines that are overloaded and the degree of overload.

[0085] The calculated systemic impact index was used to determine the vulnerability score of the physical facility. .

[0086] In one embodiment, this indicator value can be directly used as a vulnerability score. Used, or normalized, as a vulnerability score. The Systemic Risk and Cognitive Focus Assignment Module 200 covers all aspects within the scope of system assessment. Each physical device repeats the above process, ultimately generating a corresponding vulnerability score for each physical device, forming a set of vulnerability scores. It is then transmitted to the focus weight generation unit within the module for further processing.

[0087] In the systemic risk and cognitive focus allocation module 200, when the previous process is all A set of vulnerability scores was generated for each physical device. Subsequently, a focus weight generation unit within the module will process the score set. The purpose of this processing is to transform each individual vulnerability score into a standardized cognitive focus weight vector that represents the priority of diagnostic resource allocation.

[0088] To achieve this, the focus weight generation unit applies a Softmax function. This function takes the vulnerability scores of all physical devices as input and outputs a corresponding cognitive focus weight for each physical device. This calculation process is defined by the following formula:

[0089] ;

[0090] In this formula, the symbols are defined as follows:

[0091] It is a physical device The calculated cognitive focus weight is a value between 0 and 1, and the sum of the cognitive focus weights of all physical devices is 1.

[0092] Physical equipment The vulnerability score is determined by the aforementioned virtual fault injection and impact simulation process;

[0093] It is a natural exponential function, whose function is to ensure that all calculated weight values ​​are positive and to amplify the differences in vulnerability scores between different physical devices;

[0094] This is a preset, positive-zero focus concentration parameter. This parameter is used to adjust the degree of concentration in the weight allocation. When... When the value of is large, the calculation results will cause the weights to be more significantly concentrated on the physical devices with the highest vulnerability scores; when When the value is small, the weight will be distributed to multiple physical devices in a relatively more even manner;

[0095] It is the total number of physical devices being evaluated;

[0096] It is the index variable for the summation operation in the denominator, which iterates from 1 to... All physical devices are used to calculate a summation term for normalization. Specifically used to iterate from 1 to All physical devices, and from 1 to The vulnerability scores of all physical devices are summed to obtain ,in, For equipment Vulnerability score.

[0097] By performing the above calculations, the focus weight generation unit generates the focus weight for each physical device in the system. A cognitive focus weight has been determined. Ultimately, this unit generates a complete cognitive focus weight vector. This vector, along with the original state deviation residual vector, is then output to the focused causal diagnosis and introspection verification module 300 to guide the subsequent diagnostic analysis process.

[0098] Please refer to the appendix. Figure 4 The specific implementation of the focused causal diagnosis and introspection verification module 300 will now be described. The focused causal diagnosis and introspection verification module 300 is configured to begin operation upon receiving information output by the systemic risk and cognitive focus allocation module 200. The input information received by this module mainly includes: state deviation residual vector. And the corresponding cognitive focus weight vector .

[0099] The first step performed by the focused causal diagnostics and introspection verification module 300 is to determine the priority order of diagnostic analyses. The goal of this step is to allocate limited computational and analytical resources first to the physical devices that pose the greatest potential risk to the overall system and have already exhibited state deviations.

[0100] In one specific embodiment, the focused causal diagnosis and introspection verification module 300 first identifies the residual vector of state deviation. The set of physical devices with significant non-zero components. Subsequently, the focused causal diagnosis and introspection verification module 300 extracts the cognitive focus weight values ​​corresponding to these physical devices that have shown bias from the received cognitive focus weight vector. .

[0101] After obtaining these weight values, the focused causal diagnosis and introspection verification module 300 performs a ranking operation on these physical devices. This ranking operation is based on the cognitive focus weight values ​​of each physical device. Based on this, the devices are arranged in descending order, with the physical devices having the highest weight values ​​listed first. This generates a diagnostic queue with a clear priority order.

[0102] The focused causal diagnosis and introspection verification module 300 will strictly follow the order of this diagnostic queue and perform subsequent causal reasoning and active experimental verification on the physical devices in the queue in turn.

[0103] Specifically, the focused causal diagnosis and introspection verification module 300 will first initiate a detailed diagnostic process for the physical device ranked first in the queue (i.e., the one with the highest cognitive focus weight). After completing the diagnosis of that physical device, it will continue to process the next physical device in the queue if necessary. This mechanism ensures the targeted and efficient nature of the diagnostic process.

[0104] Please refer to the appendix. Figure 5 Within the focused causal diagnosis and introspection verification module 300, to perform reasoning from deviation phenomena to the root cause of faults, this module pre-builds and stores a device fault knowledge graph. The knowledge graph is a directed graph structure containing various types of nodes and directional edges that define the relationships between nodes.

[0105] In one specific embodiment, the knowledge graph contains four types of nodes:

[0106] Component nodes, functional nodes, failure mode nodes, and observable residual characteristic nodes. Taking a power transformer as an example, a component node can be either a "winding" or a "core".

[0107] Functional nodes can be either "carrying current" or "forming the main magnetic circuit";

[0108] The failure mode node can be "inter-turn short circuit in winding" or "local overheating of iron core";

[0109] The observable residual feature nodes directly correspond to the state deviation residual vector. The components are represented by various parameters, such as the "winding temperature deviation component" or the "acetylene concentration deviation component in the oil". The directed edges in the diagram represent the physical causal relationships between nodes. For example, a directed edge from the "winding inter-turn short circuit" failure mode node to the two observable residual feature nodes of "winding temperature deviation component" and "acetylene concentration deviation component in the oil" indicates that the failure mode will cause deviations in these two state quantities.

[0110] When the focused causal diagnosis and introspection verification module 300 begins performing a diagnosis on a physical device with the highest diagnostic priority, it will use the state deviation residual vector corresponding to that physical device. The non-zero component with significant amplitude serves as initial evidence for diagnosis.

[0111] Subsequently, the focused causal diagnosis and introspective verification module 300 executes a reverse causal reasoning algorithm based on the knowledge graph. Specifically, the algorithm uses the "observable residual feature node" corresponding to the initial evidence as the starting point for the search and traverses the graph in the opposite direction of the directed edges. The goal of this search is to identify all "failure mode nodes" that can act as causes and ultimately lead to the appearance of one or more observed residual features.

[0112] The final output of this reasoning process is a set of failure hypotheses consisting of one or more identified failure modes. Each element in this set, i.e., each failure hypothesis, provides a plausible explanation for the observed state deviation based on a physical causal chain. This set of failure hypotheses is then passed to the next processing unit within the module for subsequent active experimentation and optimal hypothesis identification.

[0113] In the focused causal diagnosis and introspection verification module 300, after the previous process generates a set of fault hypotheses, the module continues to execute an active experiment and optimal hypothesis identification process. The purpose of this process is to identify, from multiple possible fault hypotheses, the unique optimal fault hypothesis that can most accurately reproduce the currently observed deviation in the physical equipment state through quantitative simulation verification.

[0114] For each failure hypothesis in the failure hypothesis set, the focused causal diagnosis and introspection verification module 300 first creates an independent, dedicated twin copy for verifying that hypothesis. This twin copy is a complete copy of the baseline twin model, ensuring that the verification processes for different hypotheses are independent and do not interfere with each other.

[0115] For each twin copy, the focused causal diagnosis and introspection verification module 300 determines a subset of parameter vectors to be optimized based on the specific fault hypothesis it corresponds to. For example, if a fault hypothesis is "local overheating of the winding," the subset of parameter vectors to be optimized may include one or more relevant physical parameters such as the equivalent thermal resistance of a specific region of the winding or the thermal conductivity of the material in that region. These parameters to be optimized are referred to as parameter vectors below. .

[0116] Subsequently, the focused causal diagnosis and introspection verification module 300 performs a parameter optimization process on the twin copy. In one embodiment, this process is implemented using an optimization algorithm (e.g., particle swarm optimization or genetic algorithm). This algorithm iteratively adjusts the parameter vector. The value is determined and adjusted based on the current real-time operating data after each adjustment. and the adjusted parameter vector As input, the twin copy is run to generate a simulated state vector. The goal of the algorithm is to find a set of simulated state vectors that corresponds to the actual measured state vectors obtained from the physical devices. The optimal parameter values ​​that minimize the differences between them.

[0117] The mathematical objective of this optimization process is to solve the following optimization problem:

[0118] ;

[0119] In this formula, the symbols are defined as follows: It is the solution to the optimization problem, that is, the set of optimal parameter values ​​found under the current fault assumption. It is a mathematical operator that represents the search for a parameter vector that minimizes the value of a subsequent expression. . It is a norm operator used to calculate the distance or difference between two vectors, such as the Euclidean norm. It is the actual measured state vector of the physical device. Indicates real-time operating condition data and the parameter vector of the current iteration The input is the simulation output vector obtained after running the twin copy.

[0120] The focused causal diagnosis and introspection verification module 300 independently executes the above optimization process once for each fault hypothesis, recording the minimum difference value reached by each hypothesis. By comparing the minimum difference values ​​corresponding to all hypotheses, the fault hypothesis that achieves the global minimum difference value is determined as the optimal explanation for this diagnosis. The set of optimal parameter values ​​corresponding to this optimal explanation is then used. Then, it is determined as the identified current state parameter vector and output to the cognitive twin adaptive evolution module 400.

[0121] In the focused causal diagnosis and introspection verification module 300, after the aforementioned active experiment and parameter optimization solution process has been completed for all fault hypotheses, the focused causal diagnosis and introspection verification module 300 will make a final comparison and determination of the results of each optimization process.

[0122] Specifically, the focused causal diagnosis and introspective verification module 300 compares the minimum fit difference calculated for each fault hypothesis under optimal parameter conditions, i.e. The minimum value.

[0123] The focused causal diagnosis and introspective verification module 300 determines the fault hypothesis that minimizes the global minimum fit difference value through comparison. This fault hypothesis is judged as the optimal physical explanation for the deviation of the current state because it most accurately reproduces the current measured state of the physical device.

[0124] The optimal set of parameter solutions corresponding to this optimal physical interpretation is the specific set of parameter vectors obtained when solving the following optimization problem:

[0125] ;

[0126] It is then formally identified as the current state parameter vector.

[0127] Identified current state parameter vector Instead of parameters representing an ideal state of health, these are quantitative representations of the actual physical properties of a device at the current moment. For example, if the optimal fault assumption is "aging of the insulating medium," then... The data will include specific values ​​for the equivalent conductivity and dielectric constant of the identified, altered insulating medium.

[0128] This identified current state parameter vector This constitutes the final output of the focused causal diagnosis and introspection verification module 300. The focused causal diagnosis and introspection verification module 300 uses this vector... Along with the identifier of its corresponding optimal fault hypothesis, it is output to the cognitive twin adaptive evolution module 400 for updating the baseline twin model.

[0129] Please refer to the appendix. Figure 7The specific implementation of the cognitive twin adaptive evolution module 400 will now be described. The cognitive twin adaptive evolution module 400 is configured to, upon receiving the identified current state parameter vector output by the focused causal diagnosis and introspection verification module 300... Then, a closed-loop update operation is performed.

[0130] In one specific embodiment, the steps for performing the closed-loop update operation on the cognitive twin adaptive evolution module 400 are as follows: First, the cognitive twin adaptive evolution module 400 receives the identified current state parameter vector. This vector It quantitatively characterizes the actual physical properties of the device at the current moment, which are determined by the aforementioned diagnostic and identification process.

[0131] Received the parameter vector Subsequently, the cognitive twin adaptive evolution module 400 transmits it to the state perception and bias quantization module 100. The purpose of this transmission operation is to update the internal parameters of the baseline twin model running within the state perception and bias quantization module 100.

[0132] Within the state awareness and deviation quantization module 100, this module performs a parameter replacement operation. Specifically, the initial parameter vector previously stored in the baseline twin model, which characterizes the ideal health state of the device, is replaced. The received and identified current state parameter vector will be used to... Covered and replaced.

[0133] By performing this replacement operation, the core parameters of the baseline twin model are updated from initial, static ideal state parameters to dynamic current state parameters that reflect the actual physical condition of the current physical device. This operation completes a closed loop from state deviation perception, risk assessment, causal diagnosis to model parameter updates, and is the core mechanism for the adaptive evolution of the system in this invention.

[0134] After the cognitive twin adaptive evolution module 400 completes the closed-loop update operation of the baseline twin model, the overall operation of the system will produce specific technical effects.

[0135] The direct consequence of the update operation is that the state awareness and deviation quantization module 100 will use the updated current state parameter vector in subsequent operating cycles. To calculate the theoretical state vector. That is, the new method for calculating the theoretical state vector becomes:

[0136] ;

[0137] in, It is the theoretical state vector calculated after the update. It still represents the set of mathematical operations for that benchmark twin model. It is real-time operating condition data, and It is the identified current state parameter vector.

[0138] Due to the parameter vector The solution itself is obtained by minimizing the difference between the output of the twin model and the physical measured value. Therefore, in the next calculation cycle after the update operation is completed, the newly calculated state deviation residual vector Its norm The value will decrease significantly and approach zero. This effect is equivalent to deducting the systematic bias caused by the identified fault or degradation from subsequent bias calculations.

[0139] In this way, the baseline twin model is no longer just a static model representing the initial ideal state of the device. Its physical parameters are continuously calibrated through closed-loop updates, enabling it to accurately reflect the real physical condition of the physical device at the current moment, including the effects of degradation or failure, thereby improving the fidelity of the digital twin model.

[0140] This improved fidelity enables the system to have greater sensitivity in detecting state deviations caused by new or evolving failure modes that are based on the current state. Furthermore, any future performance predictions or remaining lifetime assessments based on this updated twin model will start from a baseline that more accurately characterizes the current health of the equipment, thereby improving the accuracy of such analyses.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital twin-driven artificial intelligence assessment system for substation status, characterized in that, include: The state perception and deviation quantification module is used to establish and output the theoretical state vector of the physical equipment based on the reference twin model of the substation equipment, obtain the actual measured state vector of the physical equipment, and calculate the state deviation residual vector between the theoretical state vector output by the reference twin model and the actual measured state vector. The systemic risk and cognitive focus allocation module is used to infer the impact of each physical device on the system based on preset failure modes to determine vulnerability scores, and to generate a cognitive focus weight vector based on the vulnerability scores. The focused causal diagnosis and introspection verification module is used to perform causal reasoning on the physical devices corresponding to the state deviation residual vector based on the cognitive focus weight vector to generate a set of fault hypotheses, and to conduct active experiments in the twin copies of the fault hypotheses to verify and identify the optimal fault hypothesis, thereby outputting the identified current state parameter vector. The cognitive twin adaptive evolution module is used to receive the identified current state parameter vector and update the baseline twin model in the state perception and deviation quantization module based on the identified current state parameter vector.

2. The digital twin-driven substation status artificial intelligence assessment system according to claim 1, characterized in that, The state perception and deviation quantification module includes: The baseline twin model running unit is used to run the baseline twin model and generate the theoretical state vector based on the built-in initial parameter vector that characterizes the ideal health state of the equipment and by receiving real-time operating condition data. The state data acquisition and deviation calculation unit is used to acquire the actual measured state vector of the physical device, and calculate the state deviation residual vector based on the state deviation residual vector calculation formula, the theoretical state vector and the actual measured state vector; A diagnostic activation unit is used to activate the systemic risk and cognitive focus allocation module when the norm of the state deviation residual vector continuously exceeds a preset threshold. The formula for calculating the state deviation residual vector is as follows: ； in, Let the state deviation residual vector be... This is the actual measured state vector of the physical device; The theoretical state vector output by the baseline twin model. , This provides real-time operating data for physical equipment. For the parameter vector adjusted within the fault hypothesis space, This is the theoretical simulation output for twin copies.

3. The digital twin-driven substation status artificial intelligence assessment system according to claim 1, characterized in that, The systemic risk and cognitive focus allocation module includes: The fault impact simulation unit is used to embed a whole-station power grid topology model, inject a virtual fault of a single device into the whole-station power grid topology model, and positively simulate the impact of the virtual fault on the stability of the whole station, and quantify it into the systemic impact index of the physical device. The focus weight generation unit is used to map the systemic impact index to the vulnerability score, and generate the cognitive focus weight vector based on the vulnerability scores of all devices and the cognitive focus weight vector calculation formula.

4. The digital twin-driven artificial intelligence assessment system for substation status according to claim 3, characterized in that, The formula for calculating the cognitive focus weight vector is as follows: ； in, For equipment Cognitive focus weight, For equipment Vulnerability score, The total number of devices. , For the focal concentration parameter, The natural number exponential function with base e is used to characterize the natural number exponential function. It is the index variable for the summation operation in the denominator. Used to iterate from 1 to All physical devices, and from 1 to The vulnerability scores of all physical devices are summed. For equipment Vulnerability score.

5. The digital twin-driven artificial intelligence assessment system for substation status according to claim 1, characterized in that, The focused causal diagnosis and introspection verification module includes: The causal graph reasoning unit is used to build a knowledge graph corresponding to the physical device, and use the state deviation residual vector as evidence to perform source reasoning on the knowledge graph to generate the fault hypothesis set; An active experimental unit is used to create an independent twin copy for each fault hypothesis in the fault hypothesis set, and actively adjust the physical parameters corresponding to the fault hypothesis in the twin copy with the goal of reproducing the actual measured state vector of the physical device. The optimal hypothesis determination unit is used to determine the fault hypothesis that has the smallest difference from the actual measured state vector of the physical device as the optimal fault hypothesis, and to determine the identified current state parameter vector corresponding to the optimal fault hypothesis based on the formula for determining the actual measured state vector and the current state parameter vector of the physical device.

6. The digital twin-driven artificial intelligence assessment system for substation status according to claim 5, characterized in that, The formula for determining the current state parameter vector is: ； in, This is the actual measurement state vector of the physical device. For the aforementioned baseline twin model, This provides real-time operating data for physical equipment. This is the identified current state parameter vector corresponding to the optimal fault hypothesis. The theoretical simulation output for twin copies, To minimize the parameter solution operator, Norm operator.

7. The digital twin-driven artificial intelligence assessment system for substation status according to claim 1, characterized in that, The cognitive twin adaptive evolution module includes: The benchmark update unit is used to replace the original initial parameter vector of the benchmark twin model with the identified current state parameter vector, thereby completing the closed-loop update of the system cognitive benchmark.

8. The digital twin-driven artificial intelligence assessment system for substation status according to claim 1, characterized in that, The focused causal diagnosis and introspection verification module is also used to prioritize the diagnosis of multiple devices that have the state deviation residual vector based on the cognitive focus weight vector, and to prioritize the causal inference and active experiment of the device with the highest weight.

9. A digital twin-driven artificial intelligence assessment system for substation status according to claim 1, characterized in that, The reference twin model is a dynamic model that couples at least two of the physical fields: electromagnetic field, thermal flow field, and mechanical stress field.

Images