A method for risk simulation of power distribution rooms based on digital twins

By using digital twin technology to process multi-source heterogeneous data from the power distribution room, a consistent virtual-real state representation and dynamic calibration were achieved. This solved the data processing and decision-making challenges in risk management in existing technologies, and improved the stability of risk prediction and the pertinence and operability of operation and maintenance decisions.

CN123121545APending Publication Date: 2026-11-17BINZHOU WEIQIAO NATIONAL SCIENCE & TECHNOLOGY ADVANCED TECHNOLOGY RESEARCH INSTITUTE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611331629.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing risk management methods for power distribution rooms are unable to handle multi-source heterogeneous data for unified expression and alignment, cannot achieve consistent virtual and real state representation and dynamic calibration of state parameters, and lack unified modeling means to handle continuous state variables and discrete event variables at the same time. They are also unable to perform multi-scenario risk evolution extrapolation and probability characterization, and cannot effectively support operation and maintenance decisions.

Method used

A risk extrapolation method based on digital twins is adopted. Through multi-source heterogeneous operation data preprocessing, digital twin virtual-real synchronization and dynamic calibration, improved state space diffusion model and joint tree accident chain reasoning technology, multi-scenario risk evolution trajectory is generated and executable operation and maintenance decisions are output.

Benefits of technology

It achieves the fusion and representation of multi-source heterogeneous data and virtual-real synchronization, enhances the ability to characterize multivariate coupling, noise disturbance and slow degradation process, improves the stability and transferability of risk prediction, and quantifies the output of risk propagation path, impact range and disposal time window, providing actionable operation and maintenance decision suggestions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121545A_ABST
    Figure CN123121545A_ABST
Patent Text Reader

Abstract

This invention discloses a method for risk extrapolation of substations based on digital twins, comprising: collecting and preprocessing substation operation data to form multi-source heterogeneous time-series data; constructing a digital twin of the substation and generating a twin state vector; performing virtual-real synchronization and dynamic calibration to obtain a calibrated twin state; generating condition variables and constructing a basic sequence; constructing an improved state-space diffusion model to obtain a set of risk evolution trajectories; performing evidence extraction and event-based representation to generate an evidence sequence for reasoning; and performing extrapolation based on a joint tree algorithm to obtain operation and maintenance decisions. This invention, by combining an improved state-space diffusion model with the joint tree algorithm for accident chain extrapolation, achieves quantitative prediction of substation risk evolution paths and outputs executable operation and maintenance decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent power distribution technology, and in particular to a method for risk estimation of power distribution rooms based on digital twins. Background Technology

[0002] Current power distribution room safety management primarily employs threshold alarms, setpoint logic judgments, or experience-based static assessments for operational monitoring and defect handling. This typically relies on electrical quantities, equipment status, protection alarms, and environmental monitoring data, triggering alarms according to fixed thresholds or rules, which are then analyzed and handled by personnel. However, with the increasing number of online monitoring points and the expanding data access scope in power distribution rooms, operational data exhibits diverse sources, inconsistent sampling periods, inconsistent timestamps, and frequent missing and outlier values. Furthermore, management data such as maintenance work orders and historical records differ significantly in structure from real-time monitoring data, making it difficult for traditional methods to uniformly represent and align multi-source heterogeneous data. It also makes it difficult to consistently correlate equipment objects, primary topology, protection logic, and environmental conditions, thus hindering the achievement of consistent virtual-to-real state representation and dynamic calibration of state parameters. In practical applications, issues such as mismatched alarm thresholds for the same risk under different operating conditions, false alarms triggered by noise fluctuations, and missed alarms due to slow degradation processes going undetected are common. Risk assessment results also struggle to be continuously updated with changes in operating modes, environment, and equipment aging, affecting the accuracy and timeliness of power distribution room risk management.

[0003] The operational risks of power distribution rooms typically exhibit chain-like evolution and propagation characteristics, often manifesting as an accident chain process of accumulating hidden dangers, triggering disturbances, causing faults, implementing protection actions, and spreading consequences. This involves multi-variable coupling relationships, primary topology propagation relationships, and the sequence of protection linkage actions. Most existing methods focus on anomaly detection and health scoring of single monitoring quantities or single devices, lacking a unified modeling approach capable of simultaneously handling continuous state variables and discrete event variables. This makes it difficult to perform multi-scenario extrapolation and probabilistic characterization of risk evolution within future time windows. Even when trend prediction is conducted, it is often single-point or localized, failing to cover multi-branch uncertainties. Existing methods lack accident chain extrapolation technology that incorporates primary topology relationships and protection linkage relationships into a unified reasoning framework. This makes it difficult to quantify and output risk propagation paths, impact ranges, and response time windows at the platform level, and also makes it difficult to translate extrapolation results into actionable operation and maintenance decision recommendations. Consequently, they cannot effectively support the formulation and verification of inspection and retesting priorities, maintenance arrangements, load adjustments, operation mode switching, and isolation and transfer actions.

[0004] Therefore, how to provide a method for risk simulation of power distribution rooms based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a risk extrapolation method for power distribution rooms based on digital twins. This invention comprehensively utilizes multi-source heterogeneous operational data preprocessing, digital twin virtual-real synchronization and dynamic calibration, an improved state-space diffusion model, and joint tree accident chain reasoning technology. It describes the entire process from operational data acquisition and preprocessing, constructing a digital twin and generating twin state vectors, obtaining calibrated twin states through dynamic calibration, generating condition variables and basic sequences, generating multi-scenario risk evolution trajectories based on the improved state-space diffusion model, evidence extraction and event-based expression, and finally extrapolating and outputting operation and maintenance decisions based on the joint tree algorithm. The improved state-space diffusion model adopts multi-type condition variable embedding, discrete-continuous hybrid diffusion modeling, bidirectional state-space structure, and cross-variable interaction modules. The joint tree algorithm adopts a clique construction method based on functional module partitioning and a clique connection sorting method based on causal propagation path priority to realize the quantitative extrapolation of risk evolution paths, propagation ranges, and response time windows, and output executable operation and maintenance decisions.

[0006] A method for risk simulation of a power distribution room based on digital twins according to an embodiment of the present invention includes: Collect operational data from the power distribution room, preprocess the operational data, and form multi-source heterogeneous time-series data; A digital twin of the power distribution room is constructed based on multi-source heterogeneous time-series data, and a twin state vector is generated; Based on multi-source heterogeneous time series data, the twin state vector is synchronized with the real data and dynamically calibrated to update the state parameters and obtain the calibrated twin state. Conditional variables are generated based on the calibration twin state, and a basic sequence is constructed. The basic sequence consists of historical time window data of the calibration twin state. An improved state-space diffusion model is constructed, and conditional diffusion denoising is performed on the basic sequence by embedding multiple types of conditional variables. Joint generation processing is performed based on a discrete-continuous hybrid diffusion modeling structure. Bidirectional temporal representation processing is performed using a bidirectional state-space structure. Interactive modeling processing is performed through a cross-variable interaction module to obtain a set of multi-scenario risk evolution trajectories. Evidence is extracted and event-based representation is performed on a set of multi-scenario risk evolution trajectories to generate evidence sequences for reasoning. The joint tree algorithm is used to deduce the accident chain of evidence sequences. Based on the functional module division of the cluster construction method, the functional modules of the power distribution room are mapped to the corresponding clusters, and the cluster potential function of each cluster is initialized. The causal propagation path priority sorting method is introduced to form a joint tree structure to obtain the operation and maintenance decision.

[0007] Optionally, the operational data includes electrical quantity data, equipment status data, protection and alarm data, environmental monitoring data, and operation and maintenance management data.

[0008] Optionally, the process of forming multi-source heterogeneous time-series data includes: The runtime data is timestamped and time aligned. Data from different sampling periods are resampled according to the alignment period. The resampling method is to take the mean, maximum, minimum or end value within the alignment period. Missing values ​​and outliers are handled in the running data. Missing values ​​are handled by forward filling, backward filling, linear interpolation or filling based on the historical window mean. Outliers are handled by truncation based on physical range, outlier removal based on sliding window statistics or abrupt correction based on the difference threshold between adjacent time points. The processed running data is uniformly encoded to form multi-source heterogeneous time series data. The uniform encoding includes standardization of continuous variables and categorical encoding of discrete variables. Standardization is to transform continuous variables by subtracting the preset mean and dividing by the preset standard deviation. Categorical encoding is to map discrete variables to one-hot codes.

[0009] Optionally, generating the twin state vector includes: Construct equipment twins, topology twins, protection twins, and environmental twins. Equipment twins include circuit breakers, switchgear, transformers, and cables. Topology twins include circuits, bus segments, interconnections, and power supply paths. Protection twins include action types, tripping logic, and interlocking logic. Environmental twins include temperature, humidity, and ventilation / dehumidification status. Establish relationships between digital twins, connect the device twin and the topology twin, map the protection twin and the device twin, and establish the interaction relationship between the environment twin and the device twin. Record these relationships in a graph structure, where nodes represent twin object instances and edges represent the relationships between twin object instances. Continuous state variables and discrete event variables are extracted from multi-source heterogeneous time-series data to generate twin state vectors. The twin state vectors are formed by splicing the continuous state variable sequence and the discrete event variable sequence along a unified time axis. The continuous state variable sequence includes electrical quantity variables and environmental variables, and the discrete event variable sequence includes equipment state variables and protection alarm event variables.

[0010] Optionally, obtaining the calibrated twin state includes: An observation vector is constructed based on multi-source heterogeneous time series data. The observation vector is formed by combining the observation values ​​of continuous state variables and discrete event variables along a unified time axis, and corresponds to the variable dimension of the twin state vector. Perform virtual-real synchronization, map the observation vector to the state variables of the corresponding twin object in the digital twin to obtain the predicted state, and calculate the residual vector, which is the difference vector between the observation vector and the predicted state in the same variable dimension; The state parameters are dynamically calibrated and updated based on the residual vector. The dynamic calibration update is to update the state parameters by adding the calibration increment to the state parameters at the previous time step. The calibration increment is calculated by the residual vector through a preset mapping function. The updated twin state vector is output as the calibration twin state.

[0011] Optionally, the step of generating conditional variables and constructing a basic sequence based on the calibrated twin state includes: Extract the operating mode, topology relationship identifier, protection configuration parameters, equipment health status identifier, and environmental condition identifier from the calibration twin state, and concatenate them in sequence to form condition variables; Set the length of the historical time window and the sampling interval, and extract historical time window data from the calibration twin state to form a basic sequence. The basic sequence includes the historical sequence of continuous state variables and the historical sequence of discrete event variables. The basic sequence is formatted for input, and the historical sequence of continuous state variables is arranged into a continuous input matrix by time step, and the historical sequence of discrete event variables is arranged into a discrete input matrix by time step. The condition variables are then associated with the continuous input matrix and the discrete input matrix to form the model input sequence.

[0012] Optionally, obtaining the multi-scenario risk evolution trajectory set includes: An improved state-space diffusion model is constructed, including multi-type condition variable embedding, discrete-continuous hybrid diffusion modeling, bidirectional state-space structure, and intervariate interaction module; A conditional diffusion denoising process is performed on the basic sequence using multi-type conditional variable embedding. The conditional variables are embedded to obtain a conditional embedding vector, and then embedded to obtain a diffusion step embedding vector. The basic sequence is encoded by continuous input to obtain a continuous feature sequence, and then encoded by discrete input to obtain a discrete feature sequence. The continuous noise estimation sequence and the discrete noise estimation sequence are obtained through a denoising unit. The joint generation process is based on discrete-continuous hybrid diffusion modeling. In the forward diffusion process, Gaussian noise is added to the continuous feature sequence to form a continuous noise sequence, and class perturbation is applied to the discrete feature sequence to form a discrete noise sequence. In the reverse diffusion process, the continuous noise estimation sequence and the discrete noise estimation sequence are used as the denoising results to update the continuous noise sequence and the discrete noise sequence to obtain the continuous denoised sequence and the discrete denoised sequence. Bidirectional temporal representation processing is performed using a bidirectional state space structure. In the denoising unit, forward state recursion is performed on the continuous denoised sequence and the discrete denoised sequence to obtain the forward hidden state sequence, and reverse state recursion is performed to obtain the reverse hidden state sequence. The forward hidden state sequence and the reverse hidden state sequence are then combined into a bidirectional temporal representation sequence by a fusion method. Interactive modeling is performed through a cross-variable interaction module. Interactive transformations are performed on the bidirectional time series representation sequence in the variable dimension to obtain an interactive feature sequence. The interactive feature sequence is then input into the decoding output unit to output a set of multi-scenario risk evolution trajectories within the prediction time window.

[0013] Optionally, generating the evidence sequence for reasoning includes: The set of multi-scenario risk evolution trajectories is traversed by scenario number to obtain the continuous state variable prediction sequence and discrete event variable prediction sequence for each scenario; State evidence is generated based on the prediction sequence of continuous state variables. Risk characteristics are calculated and categorized. Risk characteristics include rate of change, duration of continuous exceedance, peak amplitude, and cumulative amount. The rate of change is the difference between the values ​​of adjacent time steps divided by the time step interval. The duration of continuous exceedance is the number of time steps in continuous time steps that exceed the preset threshold multiplied by the time step interval. The peak amplitude is the maximum value minus the minimum value within the prediction time window. The cumulative amount is the sum of the values ​​of each time step within the prediction time window multiplied by the time step interval. Event evidence is generated by predicting discrete event variables, and state evidence and event evidence are combined along a unified timeline to form an evidence sequence. Each evidence item in the evidence sequence is assigned a scenario number and a twin object identifier to obtain an evidence sequence oriented towards reasoning.

[0014] Optionally, obtaining the operation and maintenance decision includes: Establish an accident chain reasoning graph, determine nodes and directed edges based on the primary topology and protection linkage relationship of the power distribution room, and map the evidence items in the evidence sequence to the corresponding nodes in the accident chain reasoning graph; Based on the functional module division of the cluster construction method, a cluster set is generated. The nodes in the accident chain reasoning graph are grouped according to the power distribution room functional module and mapped to the corresponding cluster. Evidence injection is performed on each cluster to initialize the cluster potential function. A joint tree structure is constructed based on the causal propagation path priority group connection sorting method. The connection order between groups is determined according to the accident chain propagation direction and the protection linkage order. Each group is connected to form a joint tree structure, and a separator set is determined for adjacent groups. The separator set is the set of common nodes of adjacent groups. Message passing is performed on the joint tree structure. Message passing includes evidence collection and passing from leaf cluster to root cluster and evidence distribution from root cluster to leaf cluster. The posterior probability distribution of each node is calculated. The propagation path, scope of impact, and response time window of the accident chain are determined by the posterior probability distribution, and operation and maintenance decisions are output. The response time window is determined by the time point when the posterior probability first reaches the preset probability threshold and the time interval during which the posterior probability continues to be higher than the preset probability threshold.

[0015] The beneficial effects of this invention are: This invention achieves the fusion and representation of multi-source heterogeneous data (electrical quantities, equipment status, protection alarms, environmental monitoring, and operation and maintenance management) within the same data system by performing time alignment, missing and anomaly handling, and unified encoding on power distribution room operation data. It also completes virtual-real synchronization and dynamic calibration within a digital twin, enabling continuous state variables and discrete event variables to form a consistent twin state vector and a calibrated twin state. Compared to existing threshold alarm or static evaluation methods, this invention utilizes an improved state-space diffusion model with multi-type condition variable embedding, discrete-continuous hybrid diffusion modeling, a bidirectional state-space structure, and intervariate interaction modules. This allows for the generation of multi-scenario risk evolution trajectory sets under given operating modes, topology identifiers, protection configurations, and operating conditions. This enhances the ability to characterize multivariate coupling, noise disturbances, weak signal vulnerabilities, and slow degradation processes, thereby reducing false alarms and false negatives and improving the stability and transferability of risk prediction.

[0016] This invention further extracts evidence and expresses events from a set of multi-scenario risk evolution trajectories, forming an evidence sequence oriented towards reasoning. Based on a joint tree algorithm, it performs accident chain deduction, constructing a joint tree reasoning structure consistent with the order of primary topology propagation and protection linkage. Through a cluster construction method based on functional module division, it achieves direct mapping of evidence injection, cluster potential initialization, and deduction results at the equipment module, bay, and circuit levels. By prioritizing causal propagation paths in cluster connection sorting, it ensures that the message transmission process aligns with the accident chain propagation direction and protection action chain, improving the interpretability and consistency of the deduction path. Compared to existing technologies that only output alarm signals or single-point scores, this invention can quantitatively output the posterior probability of risk events, risk propagation paths, impact ranges, and handling time windows. It also outputs executable operation and maintenance decision suggestions such as inspection and retesting, maintenance arrangements, load adjustments, operation mode switching, or isolated power supply, improving the pertinence, closed-loop nature, and engineering feasibility of risk handling in power distribution rooms. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a risk simulation method for power distribution rooms based on digital twins proposed in this invention; Figure 2 This is a structural block diagram of an improved state-space diffusion model for a digital twin-based risk extrapolation method for power distribution rooms proposed in this invention. Figure 3 This is a functional diagram of the joint tree algorithm for a digital twin-based risk estimation method for power distribution rooms proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for risk simulation of power distribution rooms based on digital twins, comprising: Collect operational data from the power distribution room, preprocess the operational data, and form multi-source heterogeneous time-series data; A digital twin of the power distribution room is constructed based on multi-source heterogeneous time-series data, and a twin state vector is generated; Based on multi-source heterogeneous time series data, the twin state vector is synchronized with the real data and dynamically calibrated to update the state parameters and obtain the calibrated twin state. Conditional variables are generated based on the calibration twin state, and a basic sequence is constructed. The basic sequence consists of historical time window data of the calibration twin state. An improved state-space diffusion model is constructed, and conditional diffusion denoising is performed on the basic sequence by embedding multiple types of conditional variables. Joint generation processing is performed based on a discrete-continuous hybrid diffusion modeling structure. Bidirectional temporal representation processing is performed using a bidirectional state-space structure. Interactive modeling processing is performed through a cross-variable interaction module to obtain a set of multi-scenario risk evolution trajectories. Evidence is extracted and event-based representation is performed on a set of multi-scenario risk evolution trajectories to generate evidence sequences for reasoning. The joint tree algorithm is used to deduce the accident chain of evidence sequences. Based on the functional module division of the cluster construction method, the functional modules of the power distribution room are mapped to the corresponding clusters, and the cluster potential function of each cluster is initialized. The causal propagation path priority sorting method is introduced to form a joint tree structure to obtain the operation and maintenance decision.

[0020] In this embodiment, the operational data includes electrical quantity data, equipment status data, protection and alarm data, environmental monitoring data, and operation and maintenance management data.

[0021] In this embodiment, the formation of multi-source heterogeneous time-series data includes: The runtime data is timestamped and time aligned. Data from different sampling periods are resampled according to the alignment period. The resampling method is to take the mean, maximum, minimum or end value within the alignment period. The running data undergoes missing value and outlier handling. Missing value handling employs forward imputation, backward imputation, linear interpolation, or imputation based on the historical window mean. Outlier handling employs physical range truncation, outlier removal based on sliding window statistics, or abrupt correction based on a difference threshold between adjacent time points. Specifically, the difference threshold is: The difference threshold is set using the three-standard-deviation method. For each monitored variable, based on the normal operating baseline data for seven consecutive days, time alignment is first completed, and then the mean and standard deviation of the difference between adjacent time points are calculated. The upper limit of the difference is set as the mean of the difference plus three times the standard deviation of the difference, and the lower limit of the difference is set as the mean of the difference minus three times the standard deviation of the difference. When the adjacent difference exceeds the upper limit or falls below the lower limit, it is determined as a sudden change point, and the value at the current time is corrected to the linear interpolation of the two adjacent points. The processed running data is uniformly encoded to form multi-source heterogeneous time series data. The uniform encoding includes standardization processing of continuous variables and categorical encoding processing of discrete variables. The standardization processing is to transform the continuous variables by subtracting the preset mean and dividing by the preset standard deviation. The categorical encoding processing is to map the discrete variables to one-hot codes. The preset mean and preset standard deviation are set to 0.1 and 1.5, respectively.

[0022] In this embodiment, generating the twin state vector includes: The construction of equipment twins, topology twins, protection twins, and environmental twins is carried out. Equipment twins include circuit breakers, switchgear, transformers, and cables; topology twins include circuits, bus segments, interconnections, and power supply paths; protection twins include action types, tripping logic, and interlocking logic; and environmental twins include temperature, humidity, and ventilation / dehumidification status. Specifically, the construction of equipment twins, topology twins, protection twins, and environmental twins involves: Equipment twins are achieved by creating object modules for circuit breakers, switchgear, transformers and cables respectively. Each object module contains a unique identifier, rated parameters, structural attributes, operating status interfaces and health status fields, and is bound to the corresponding measurement points in the operating data. Topology twins are achieved by constructing a connection diagram of loops, bus segments, interconnections and power supply paths. The connection diagram uses nodes to represent equipment or bays and edges to represent electrical connections, and records the on / off status to obtain the currently reachable power supply path. Protection twins are achieved by establishing protection configuration tables and logic mapping tables. The protection configuration table records the action type and setting group, and the logic mapping table records the input, criteria and output of tripping logic and interlocking logic, and binds them to the corresponding equipment objects. Environmental twins are achieved by establishing temperature and humidity and ventilation and dehumidification object modules. The environmental object modules record the monitoring points, sampling values, area affiliation and ventilation and dehumidification device status, and establish interaction relationships with area objects such as switch cabinets and busbar rooms to form an environmental impact mapping. Establish digital twin association relationships by connecting the device twin and the topology twin, mapping the protection twin and the device twin, and establishing the interaction relationship between the environmental twin and the device twin. Record these relationships using a graph structure, where nodes represent twin object instances and edges represent the association relationships between twin object instances. Specifically, establishing these digital twin association relationships involves: The digital twin association relationship is achieved by establishing a unified identifier for each twin object instance and configuring a relationship table. Based on the primary wiring and bay division, the object instances of circuit breakers, switchgear, transformers, and cable equipment are matched one by one with the topology object instances of circuits, bus sections, interconnections, and power supply paths to generate the connection relationship between equipment and topology. Based on the protection device access point, action object, and output circuit, the protection object instances of action type, trip logic, and blocking logic are matched with the corresponding equipment object instances to generate the mapping relationship between protection and equipment. Based on the area where the environmental monitoring points are located and the coverage of ventilation and dehumidification, the environmental object instances with temperature, humidity and ventilation and dehumidification status are matched with the affected equipment object instances to generate the interaction relationship between the environment and the equipment. All relationships are written into the graph structure, with twin object instances as nodes and connection relationships, mapping relationships and interaction relationships as directed edges. The relationship type, associated port and on / off status information are recorded in the edge attributes. Continuous state variables and discrete event variables are extracted from multi-source heterogeneous time-series data to generate twin state vectors. The twin state vectors are formed by splicing the continuous state variable sequence and the discrete event variable sequence along a unified time axis. The continuous state variable sequence includes electrical quantity variables and environmental variables, and the discrete event variable sequence includes equipment state variables and protection alarm event variables.

[0023] In this embodiment, obtaining the calibrated twin state includes: An observation vector is constructed based on multi-source heterogeneous time series data. The observation vector is formed by combining the observation values ​​of continuous state variables and discrete event variables along a unified time axis, and corresponds to the variable dimension of the twin state vector. To perform virtual-real synchronization, the observation vector is mapped to the state variables of the corresponding twin object in the digital twin to obtain the predicted state, and a residual vector is calculated. The residual vector is the difference vector between the observation vector and the predicted state in the same variable dimension. Specifically, the virtual-real synchronization is performed as follows: The virtual-real synchronization is performed through the mapping table between measurement points and twin variables and the state update process. Based on the mapping table, each dimension of data in the observation vector is located to the corresponding twin object instance and state variable field. The observation value at a given time is written to the real-time state cache of the twin object according to a unified timestamp, and the event state of the twin object is updated according to the timestamp trigger method for discrete events. Based on the topological connection relationship of the digital twin and the state parameters of the previous moment, a state advancement is performed on each twin object to obtain the predicted state at the same moment. The residual vector is formed by subtracting the observation vector and the predicted state item by item in the same variable dimension, and the residual vector and the corresponding twin object identifier are stored together. Dynamic calibration and updating of state parameters are performed based on residual vectors. This dynamic calibration update involves adding a calibration increment to the state parameters from the previous time step. The calibration increment is calculated from the residual vector using a preset mapping function. The updated twin state vector is output as the calibrated twin state. Specifically, the preset mapping function is: The preset mapping function uses piecewise linear mapping, grouping the residual vector according to the twin object and variable type. For each group of residuals, a calibration increment is calculated. When the absolute value of the residual does not exceed the preset threshold, the calibration increment is zero. When the absolute value of the residual exceeds the threshold but does not exceed the saturation threshold, the calibration increment is the proportional coefficient multiplied by the residual. When the absolute value of the residual exceeds the saturation threshold, the calibration increment is the proportional coefficient multiplied by the saturation threshold while retaining the residual sign. The proportional coefficient, preset threshold, and saturation threshold are given in the configuration table according to different twin objects and different state parameters, and are set to 0.2, 0.1, and 1.5 respectively. After the calibration increment is matched with the corresponding state parameter, it is used to perform the update of the state parameter plus the calibration increment at the previous time step.

[0024] In this embodiment, the step of generating condition variables and constructing a basic sequence based on the calibration twin state includes: Extract the operating mode, topology relationship identifier, protection configuration parameters, equipment health status identifier, and environmental condition identifier from the calibration twin state, and concatenate them in sequence to form condition variables; Set the length of the historical time window and the sampling interval, and extract historical time window data from the calibration twin state to form a basic sequence. The basic sequence includes the historical sequence of continuous state variables and the historical sequence of discrete event variables. The basic sequence is formatted for input, and the historical sequence of continuous state variables is arranged into a continuous input matrix by time step, and the historical sequence of discrete event variables is arranged into a discrete input matrix by time step. The condition variables are then associated with the continuous input matrix and the discrete input matrix to form the model input sequence.

[0025] In this embodiment, obtaining the multi-scenario risk evolution trajectory set includes: An improved state-space diffusion model is constructed, including multi-type condition variable embedding, discrete-continuous hybrid diffusion modeling, bidirectional state-space structure, and intervariate interaction module, wherein: Multi-type condition variable embedding is set up to encode the condition variables of operating mode, topology identifier, protection configuration, health status and environmental conditions into condition embedding vectors, and inject them into the input layer and residual layer of the denoising unit. A discrete-continuous hybrid diffusion modeling structure is set up to split the basic sequence into continuous state variable channels and discrete event variable channels, which correspond to continuous Gaussian noise addition, denoising chain and discrete category perturbation and scrambling chain, respectively. The continuous noise estimation sequence and the discrete noise estimation sequence are jointly output in the same denoising unit. The backbone network of the denoising unit is configured as a bidirectional state space structure. The hidden state sequence is generated in parallel by the forward state recursion branch and the reverse state recursion branch and synthesized in a fusion manner. A cross-variable interaction module is set on the bidirectional hidden state sequence to perform interactive transformation on the variable dimension to generate an interactive feature sequence. The continuous state variable prediction sequence and the discrete event variable prediction sequence within the prediction time window are recovered by the decoding output unit, thereby forming a set of multi-scenario risk evolution trajectories. A conditional diffusion denoising process is performed on the base sequence using multi-type conditional variable embedding. The conditional variables are embedded to obtain a conditional embedding vector, which is then embedded through a diffusion step to obtain a diffusion step embedding vector. The base sequence is encoded by continuous input to obtain a continuous feature sequence, and by discrete input encoding to obtain a discrete feature sequence. A denoising unit then yields both a continuous noise estimation sequence and a discrete noise estimation sequence, where: The conditional variable embedding and diffusion step embedding are specifically as follows: Conditional variable embedding involves using one-hot encoding to map the discrete identifiers of each field in the conditional variable to a fixed-dimensional vector through linear transformation, and then using linear transformation to map the continuous quantities of each field to a fixed-dimensional vector before concatenating them to obtain a conditional embedding vector. Diffusion step embedding involves generating sine and cosine sequences of diffusion step numbers according to a preset length and concatenating them to obtain a diffusion step encoding vector, which is then mapped through linear transformation to a diffusion step embedding vector with the same dimension as the conditional embedding vector. The preset length is set to 128. The continuous input encoding and discrete input encoding are specifically as follows: continuous input encoding is to linearly transform the continuous input matrix in the basic sequence according to the variable dimension and add time position encoding to obtain a continuous feature sequence arranged in time steps; discrete input encoding is to linearly transform the one-hot encoding of the discrete input matrix in the basic sequence according to the category dimension and add time position encoding to obtain a discrete feature sequence arranged in time steps. A joint generation process based on discrete-continuous hybrid diffusion modeling is performed. During the forward diffusion process, Gaussian noise is added to the continuous feature sequence to form a continuous noise sequence, and class perturbation is applied to the discrete feature sequence to form a discrete noise sequence. During the reverse diffusion process, the estimated continuous and discrete noise sequences are used as denoising results to update the continuous and discrete noise sequences, resulting in a continuous denoised sequence and a discrete denoised sequence. The process of generating a continuous noise sequence by adding Gaussian noise is as follows: the continuous feature sequence is forward diffused according to a preset number of diffusion steps. In any diffusion step, standard Gaussian random noise of the same dimension is linearly combined with the continuous feature sequence according to a preset noise schedule to generate the noise addition result of the diffusion step. The noise addition result serves as the state of the continuous noise sequence in the current diffusion step. The noise schedule is used to determine the retention ratio of the original continuous features and the injection ratio of Gaussian noise in the diffusion step, so that the proportion of Gaussian noise gradually increases as the number of diffusion steps increases until the continuous noise sequence approaches pure Gaussian noise. The preset number of diffusion steps is set to one hundred steps, and the preset noise schedule adopts a linear schedule, which linearly increases the noise intensity of each step from one to the power of negative four to zero point zero two. The execution category perturbation forms a discrete noise sequence, specifically: the discrete feature sequence is positively perturbed by the same number of diffusion steps. In any diffusion step, the discrete category of each time step is randomly replaced with a perturbation probability of 0.5. The perturbation method is to keep the original category unchanged with a first probability and replace the original category with another category in the category set with a second probability. The replacement category is selected by uniform sampling in the category set. As the number of diffusion steps increases, the perturbation probability gradually increases, so that the category distribution of the discrete noise sequence gradually approaches a uniform distribution. The process of updating the continuous noise sequence and the discrete noise sequence to obtain the continuous denoised sequence and the discrete denoised sequence specifically involves: In the reverse diffusion process, the process is gradually pushed back from the maximum diffusion step to the zero diffusion step. In each diffusion step, the continuous noise estimation sequence output by the denoising unit is used as the estimate of the continuous noise in the current step. The corresponding proportion of noise contribution is deducted from the current continuous noise state and the corresponding proportion of signal component is restored in reverse order to obtain the continuous state of the previous diffusion step, thus forming a continuous denoising sequence. The discrete noise estimation sequence output by the denoising unit is converted into denoising scores or posterior probability distributions for each category. The current discrete noise state is then updated according to the back transition to obtain the discrete state of the previous diffusion step, thus forming a discrete denoising sequence. This process continues until the zero diffusion step is pushed back to output the continuous state variable prediction sequence and the discrete event variable prediction sequence. A bidirectional temporal representation processing method is used, employing a bidirectional state-space structure. Within the denoising unit, forward state recursion is performed on the continuous and discrete denoised sequences to obtain a forward hidden state sequence, and reverse state recursion is performed to obtain a reverse hidden state sequence. The forward and reverse hidden state sequences are then combined into a bidirectional temporal representation sequence using a fusion method. The forward and reverse state recursion are specifically as follows: Forward state recursion refers to processing continuous and discrete denoised sequences sequentially in the denoising unit. At each time step, the hidden state of the previous time step and the input features of the current time step are used to calculate the current hidden state by linear transformation and nonlinear activation. The current hidden state is then used as the input state of the next time step, forming a forward hidden state sequence that covers the entire time window. Reverse state recursion refers to processing continuous denoised sequences and discrete denoised sequences in the same way in chronological order. At each time step, the hidden state is calculated using the hidden state of the next time step and the input features of the current time step. The current hidden state is then used as the input state of the previous time step, forming a reverse hidden state sequence that covers the entire time window. Interactive modeling is performed through a cross-variable interaction module. An interactive transformation is applied to the bidirectional time-series representation sequence along the variable dimension to obtain an interactive feature sequence. This interactive feature sequence is then input into a decoding output unit, which outputs a set of multi-scenario risk evolution trajectories within the prediction time window. Specifically, the decoding output unit comprises: The decoding output unit is used to restore the interactive feature sequence into an output sequence consistent with the prediction target. It includes a continuous decoding head and a discrete decoding head. The continuous decoding head performs a linear transformation on the interactive feature sequence at each time step and outputs a continuous state variable prediction vector. The discrete decoding head performs a linear transformation on the interactive feature sequence at each time step and outputs a score vector for each discrete event category. The score vector is converted into a category probability vector through a normalization function, and the event category is selected according to the highest probability. Finally, the continuous state variable prediction sequence and the discrete event variable prediction sequence within the prediction time window are obtained. The sets of prediction sequences corresponding to different sampling scenarios are combined into a multi-scenario risk evolution trajectory set.

[0026] In this embodiment, generating a reasoning-oriented evidence sequence includes: The set of multi-scenario risk evolution trajectories is traversed by scenario number to obtain the continuous state variable prediction sequence and discrete event variable prediction sequence for each scenario; State evidence is generated based on the prediction sequence of continuous state variables. Risk characteristics are calculated and categorized. These risk characteristics include the rate of change, duration of exceeding limits, peak amplitude, and cumulative amount. The rate of change is the difference between values ​​at adjacent time steps divided by the time step interval. The duration of exceeding limits is the number of time steps exceeding a preset threshold multiplied by the time step interval. The peak amplitude is the maximum value minus the minimum value within the prediction time window. The cumulative amount is the sum of the values ​​at each time step within the prediction time window multiplied by the time step interval. Specifically, the generation of state evidence involves: For each twin object, the prediction sequence of continuous state variables is calculated within the prediction time window, including the rate of change sequence, duration of continuous exceedance, peak amplitude, and cumulative amount. The risk level of each risk feature is determined, and the risk level is coded into discrete level values ​​according to the level. The state evidence items are composed of twin object identifier, variable identifier, risk feature type, level value, and corresponding time window start and end time. The state evidence items generated for all twin objects and all continuous state variables are summarized on a unified time axis. Event evidence is generated by predicting discrete event variables, and state evidence and event evidence are combined along a unified timeline to form an evidence sequence. Each evidence item in the evidence sequence is assigned a scenario number and a twin object identifier, resulting in a reasoning-oriented evidence sequence. The generation of event evidence specifically involves: The event category code is read step by step for the discrete event variable prediction sequence. The time step when the event category switches from non-event coding to event coding is the event start time step, and the time step when it switches from event coding to non-event coding is the event end time step. The event start time step and end time step are converted into corresponding timestamp intervals, and the event category code within the interval is used as the event type code. The twin object identifier, event variable identifier, event type code, and event start and end timestamp interval are used to form an event evidence item. When the same twin object has the same event type in adjacent time windows and the interval does not exceed the merging interval, the corresponding event evidence items are merged into one event evidence item to obtain an event evidence set that is used to combine with the state evidence on a unified time axis.

[0027] In this embodiment, obtaining the operation and maintenance decision includes: An accident chain reasoning graph is established, nodes and directed edges are determined based on the primary topology and protection linkage relationships of the power distribution room, and evidence items in the evidence sequence are mapped to corresponding nodes in the accident chain reasoning graph. Specifically, establishing the accident chain reasoning graph involves: Using the functional modules and key states and events of the power distribution room as nodes, the node identifier consists of module identifier and variable identifier. Value space and time attributes are set for each node. The topology propagation directed edge is determined according to the primary topology connection relationship. The edge direction is set according to the direction of power transmission or fault impact propagation. The protection action directed edge is determined according to the protection linkage relationship. The edge direction is set in the order of trigger condition node, protection action node, and disconnection and transfer node. The connection information in the bus segment, circuit, connection relationship and power supply path is loaded into the topology propagation directed edge. The association information in the action type, trip logic and blocking logic is loaded into the protection action directed edge. Each evidence item in the evidence sequence is located to the corresponding node according to the twin object identifier and variable identifier. The threshold value or event type code of the evidence item is written into the observation record of the node at the corresponding timestamp. A cluster set is generated based on a cluster construction method that divides the data into functional modules. Nodes in the accident chain reasoning graph are grouped according to the functional modules of the power distribution room and mapped to corresponding clusters. Evidence injection is performed on each cluster to initialize the cluster potential function, where: The generated clique set is specifically generated as follows: based on the functional module identifiers carried by the nodes in the accident chain reasoning graph, the nodes are grouped according to the functional modules of circuit breaker bays, switch cabinet units, busbar sections, outgoing circuits, transformer bays, and tie units. Each group of nodes forms a clique. For directed edges with topological propagation relationships or protection linkage relationships, the common variables that the nodes at both ends of the edge need to share in the reasoning are extracted, and the common variables are added to the cliques at both ends of the edge to make adjacent cliques have the intersection of common variables. A unique clique identifier is assigned to each clique and the set of nodes in the clique is recorded to obtain the clique set used for joint tree construction. The initialization of the cluster potential function specifically involves: establishing a cluster potential function table for each cluster, which covers the value combinations of nodes within the cluster and assigning a prior potential value to each value combination. The prior potential value is determined by a probability table. Evidence items in the evidence sequence are mapped to corresponding nodes within the cluster according to twin object identifiers and variable identifiers. The observed values ​​of the evidence items at the corresponding timestamps are written into the observation records within the cluster. For value combinations that are inconsistent with the observed values, the potential value is set to zero. For value combinations that are consistent with the observed values, the prior potential value is maintained or replaced with the evidence potential value. The potential values ​​of all value combinations in the cluster potential function table are normalized so that the sum of the potential values ​​is one, thus completing the initialization of the cluster potential function. A joint tree structure is constructed based on a causal propagation path-first cluster connection sorting method. The connection order between clusters is determined according to the accident chain propagation direction and the protection linkage sequence. Each cluster is connected to form a joint tree structure, and a separator set is determined for adjacent clusters. The separator set is the set of common nodes of adjacent clusters. The construction of the joint tree structure is as follows: The joint tree structure is constructed based on the causal propagation path in the clique set and the incident chain reasoning graph. The connection priority between cliques is determined by the incident chain propagation direction and the protection linkage order. The connection relationship at the clique level is reflected as the directed order of the inducing clique, intermediate state clique, protection action clique and consequence clique. Candidate connections are established for clique pairs in the clique set that have causal propagation relationship or protection linkage relationship. The connection set that meets the acyclic requirement of the tree structure is selected from the candidate connections so that there is a unique connection path between any two cliques. The separator set of adjacent cliques is determined by the common node set of the two cliques. The separator set is recorded as the attribute of the connection edge to obtain the joint tree structure used for message passing. Message passing is performed on the joint tree structure. Message passing includes evidence collection from leaf clusters to the root cluster and evidence distribution from the root cluster to leaf clusters. The posterior probability distribution of each node is calculated, where: The execution message passing is specifically as follows: In the joint tree structure, a root clique is selected and the directed traversal relationship of the tree is determined. Evidence collection and transmission are carried out in the direction from leaf clique to root clique. The leaf clique performs marginalization on the variables of its own clique potential function that are outside the delimiter set variables to obtain delimiter set messages. The delimiter set messages are sent to the adjacent parent clique. After receiving the delimiter set messages from each adjacent child clique, the parent clique multiplies each delimiter set message with its own clique potential function item by item and completes normalization to generate a delimiter set message facing the parent clique and continues to send it upward. Evidence distribution and transmission proceed from the root clique to the leaf clique. The parent clique calculates the partition set message to be sent to each child clique based on the fused clique potential function and the partition set message from the parent clique. After receiving the partition set message, the child clique updates the clique potential function and normalizes it by multiplying it by the received partition set message and eliminating the influence of the old partition set message, so that the marginal distribution of adjacent cliques on the partition set is maintained. The calculation of the posterior probability distribution of each node is as follows: After the message is transmitted, the clique potential function of each clique is in a calibration state consistent with all evidence. The posterior probability distribution of the node is calculated by the calibration clique potential function of any clique containing the current node. The calculation method is to perform marginalization on the variables other than the target node in the calibration clique potential function to obtain the marginal distribution of the target node, and normalize the marginal distribution so that the sum is one. When the target node appears in multiple cliques at the same time, the marginal distribution obtained by marginalizing any clique containing the node is taken as the posterior probability distribution of the node, and the consistency of the split set is used to ensure that the calculation results of different cliques are consistent. The incident propagation path, impact range, and response time window are determined through posterior probability distribution, and operational decisions are output. The response time window is determined by the time point when the posterior probability first reaches a preset probability threshold and the time interval during which the posterior probability remains above the preset probability threshold. Specifically, the determination of the incident propagation path, impact range, and response time window through posterior probability distribution, and the output of operational decisions, are as follows: The accident chain propagation path is determined by the directed edges in the accident chain reasoning graph and the posterior probability distribution of each node. Nodes with a posterior probability not less than a preset probability threshold of 0.7 are selected as valid nodes. The continuous directed path formed by connecting valid nodes in the reasoning graph with directed edges is defined as the accident chain propagation path. The scope of influence is obtained by mapping the functional modules involved in the accident chain propagation path with the loops, bus segments, connection relationships and power supply paths of the topological twin. The mapping result includes the affected bus segment identifier, loop identifier and power supply path identifier. The handling time window determines the posterior probability sequence of the target risk event node on the prediction time axis. The time point when the preset probability threshold is first reached is defined as the handling start time. The time interval during which the posterior probability is continuously higher than the preset probability threshold is defined as the handling time window. The operation and maintenance decision outputs the handling object according to the functional module identifier, and outputs the handling action set in combination with the accident chain propagation path and impact range. The handling action set includes inspection, retesting, maintenance, load adjustment, operation mode switching and isolation transfer.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a distribution room scenario in a 110kV substation. The distribution room includes typical equipment such as main incoming line cabinets, bus tie cabinets, feeder cabinets, and station service transformers. Monitoring points for temperature and humidity, dew point, cabinet and contact temperature, partial discharge (TEV), electrical quantities including voltage, current, active and reactive power, door magnets, and the status of fans and dehumidifiers are installed within the cabinets. These points are also connected to protection device operation and alarm records, maintenance work orders, and defect records. Common problems in this scenario include: frequent threshold alarm fluctuations and false alarms when humidity fluctuations and load changes overlap; weak and unstable early signals of partial discharge and insulation degradation, making timely capture difficult through static assessment and leading to missed alarms; and difficulty in mapping alarms to the accident chain of topology propagation and protection linkage, making it difficult for maintenance personnel to determine the scope of impact and the appropriate response time window.

[0029] In the current scenario, operational data undergoes time alignment, missing and anomaly handling, and unified encoding to form multi-source heterogeneous time-series data. Equipment twins, topology twins, protection twins, and environmental twins are constructed, generating twin state vectors. Dynamic calibration is performed using the residual between the observed vectors and the twin predicted states to obtain calibrated twin states. The calibrated twin states are extracted as a historical time window base sequence, and condition variables are formed by concatenating operating mode, topology identifier, protection configuration, health status, and environmental conditions. An improved state-space diffusion model, supported by condition variable embedding, discrete-continuous hybrid diffusion, bidirectional state-space structure, and intervariate interaction modules, performs conditional diffusion denoising on the base sequence and generates a multi-scenario risk evolution trajectory set. The trajectory set outputs a continuous state variable prediction sequence and a discrete event prediction sequence. Subsequently, the rate of change, duration of continuous exceedance, peak amplitude, and cumulative amount are calculated for the trajectories and categorized to form state evidence. Simultaneously, the start and end intervals of the discrete event sequences are identified to form event evidence, resulting in a unified evidence sequence. The joint tree algorithm generates a set of clusters by dividing them into functional modules and initializes the cluster potential. It connects the clusters according to the accident chain propagation direction and protection linkage sequence to form a joint tree structure and executes message passing. It outputs the posterior probability distribution of risk events, propagation path, scope of impact and handling time window, thereby forming operation and maintenance decisions such as inspection and retesting, ventilation and dehumidification adjustment, load adjustment, operation mode switching and isolation transfer.

[0030] During a typical high humidity and load fluctuation risk event, the relative humidity inside the cabinet continuously increased from about 60% to nearly 90% at one point. The difference between the dew point temperature and the temperature inside the cabinet gradually converged from a positive value and then showed short-term negative fluctuations. The temperature rise slope of some feeder cabinet contact points increased significantly, and the peak temperature rise exceeded the average level of cabinets in the same row. The TEV partial discharge amplitude shifted upward from the background level and showed intermittent enhancement, accompanied by intermittent alarms from the protection devices. The changes in door magnets, fan and dehumidification status showed a clear temporal correlation with the enhancement of partial discharge. Traditional threshold methods experience multiple short-term triggers during periods of rapid humidity fluctuation and are unstable in identifying weak partial discharge signals. This invention generates multi-scenario evolution trajectories after dynamic calibration and provides evidence sequences for condensation-related risk events, enhanced partial discharge events, and protection alarm events. The joint tree deduction outputs a propagation path with environmental modules, switchgear units, circuit-protection actions, and consequential events as the main chain, clearly mapping the impact range to specific bus segments and adjacent feeder circuits, and providing a response time window. Within the response time window, maintenance personnel adjust dehumidification and ventilation strategies and retest the target cabinet's infrared sensor. The retest results show that the contact temperature rise slope and partial discharge amplitude have returned to a stable range, and the alarm has not escalated further, forming a closed-loop response record and writing it back to the maintenance management data.

[0031] Table 1 Comparison of Risk Simulation Methods for Power Distribution Rooms

[0032] As shown in Table 1, the method of this invention performs best in both key indicators: false positive rate (2.8%) and false negative rate (3.1%). Compared with the false positive rate of 11.6% and false negative rate of 14.2% for threshold alarms, and the false positive rate of 9.4% and false negative rate of 12.8% for static evaluation, this invention can more stably distinguish between real risks and short-term anomalies even under conditions of noise disturbance, operating condition fluctuations, and weak signal risks. It significantly reduces false positives caused by short-term triggers and also reduces false negatives during the slow degradation phase. Compared with the false positive rate of 6.7% and false negative rate of 7.9% for LSTM prediction, 5.9% and 7.1% for Transformer prediction, and 7.5% and 8.6% for Bayesian network inference, this invention still maintains a lower level of false positives and false negatives, demonstrating stronger robustness and comprehensive recognition capabilities.

[0033] In terms of early warning capability and the ability to characterize the response time window, the early warning lead time of this invention is 19.4 minutes, the highest among the compared methods, significantly higher than the 4.8 minutes of threshold alarm and the 6.2 minutes of static evaluation, and also higher than the 12.1 minutes of LSTM prediction, the 13.5 minutes of Transformer prediction, and the 10.7 minutes of Bayesian inference. The time window error of this invention is 3.6 minutes, significantly less than the 14.9 minutes of threshold alarm and the 12.7 minutes of static evaluation, and better than the 7.8 minutes of LSTM prediction, the 6.9 minutes of Transformer prediction, and the 9.1 minutes of Bayesian inference, enabling a more accurate definition of the response start time and duration interval.

[0034] From the perspectives of incident chain simulation effectiveness and engineering feasibility, this invention leads in terms of propagation range hit rate (91.2%), path consistency rate (88.5%), and decision executability rate (93.7%), significantly higher than threshold alarms (52.4%, 41.6%, 60.5%), static evaluation (58.7%, 47.2%, 62.1%), and superior to LSTM prediction (71.3%, 66.8%, 76.9%), Transformer prediction (74.6%, 69.1%, 79.3%), and Bayesian inference (78.9%, 75.4%, 81.8%). The computation latency is 1.9 seconds, within the second range and lower than Bayesian inference's 2.6 seconds, meeting the requirements for online simulation and decision output.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for risk estimation of power distribution rooms based on digital twins, characterized in that, Includes the following steps: Collect operational data from the power distribution room, preprocess the operational data, and form multi-source heterogeneous time-series data; A digital twin of the power distribution room is constructed based on multi-source heterogeneous time-series data, and a twin state vector is generated; Based on multi-source heterogeneous time-series data, the twin state vector is synchronized with the real data and dynamically calibrated to update the state parameters and obtain the calibrated twin state. Conditional variables are generated based on the calibration twin state, and a basic sequence is constructed. The basic sequence consists of historical time window data of the calibration twin state. An improved state-space diffusion model is constructed, and conditional diffusion denoising is performed on the basic sequence by embedding multiple types of conditional variables. Joint generation processing is performed based on a discrete-continuous hybrid diffusion modeling structure. Bidirectional temporal representation processing is performed using a bidirectional state-space structure. Interactive modeling processing is performed through a cross-variable interaction module to obtain a set of multi-scenario risk evolution trajectories. Evidence is extracted and event-based representation is performed on a set of multi-scenario risk evolution trajectories to generate evidence sequences for reasoning. The joint tree algorithm is used to deduce the accident chain of evidence sequences. Based on the functional module division of the cluster construction method, the functional modules of the power distribution room are mapped to the corresponding clusters, and the cluster potential function of each cluster is initialized. The causal propagation path priority sorting method is introduced to form a joint tree structure to obtain the operation and maintenance decision.

2. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The operational data includes electrical quantity data, equipment status data, protection and alarm data, environmental monitoring data, and operation and maintenance management data.

3. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The formation of multi-source heterogeneous time-series data includes: The runtime data is timestamped and time aligned. Data from different sampling periods are resampled according to the alignment period. The resampling method is to take the mean, maximum, minimum or end value within the alignment period. Missing values ​​and outliers are handled in the running data. Missing values ​​are handled by forward filling, backward filling, linear interpolation or filling based on the historical window mean. Outliers are handled by truncation based on physical range, outlier removal based on sliding window statistics or abrupt correction based on the difference threshold between adjacent time points. The processed running data is uniformly encoded to form multi-source heterogeneous time series data. The uniform encoding includes standardization of continuous variables and categorical encoding of discrete variables. Standardization is to transform continuous variables by subtracting the preset mean and dividing by the preset standard deviation. Categorical encoding is to map discrete variables to one-hot codes.

4. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The generation of the twin state vector includes: Construct equipment twins, topology twins, protection twins, and environmental twins. Equipment twins include circuit breakers, switchgear, transformers, and cables. Topology twins include circuits, bus segments, interconnections, and power supply paths. Protection twins include action types, tripping logic, and interlocking logic. Environmental twins include temperature, humidity, and ventilation / dehumidification status. Establish relationships between digital twins, connect the device twin and the topology twin, map the protection twin and the device twin, and establish the interaction relationship between the environment twin and the device twin. Record these relationships in a graph structure, where nodes represent twin object instances and edges represent the relationships between twin object instances. Continuous state variables and discrete event variables are extracted from multi-source heterogeneous time-series data to generate twin state vectors. The twin state vectors are formed by splicing the continuous state variable sequence and the discrete event variable sequence along a unified time axis. The continuous state variable sequence includes electrical quantity variables and environmental variables, and the discrete event variable sequence includes equipment state variables and protection alarm event variables.

5. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The process of obtaining the calibrated twin state includes: An observation vector is constructed based on multi-source heterogeneous time series data. The observation vector is formed by combining the observation values ​​of continuous state variables and discrete event variables along a unified time axis, and corresponds to the variable dimension of the twin state vector. Perform virtual-real synchronization, map the observation vector to the state variables of the corresponding twin object in the digital twin to obtain the predicted state, and calculate the residual vector, which is the difference vector between the observation vector and the predicted state in the same variable dimension; The state parameters are dynamically calibrated and updated based on the residual vector. The dynamic calibration update is to update the state parameters by adding the calibration increment to the state parameters at the previous time step. The calibration increment is calculated by the residual vector through a preset mapping function. The updated twin state vector is output as the calibration twin state.

6. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The step of generating conditional variables and constructing a basic sequence based on the calibrated twin state includes: Extract the operating mode, topology relationship identifier, protection configuration parameters, equipment health status identifier, and environmental condition identifier from the calibration twin state, and concatenate them in sequence to form condition variables; Set the length of the historical time window and the sampling interval, and extract historical time window data from the calibration twin state to form a basic sequence. The basic sequence includes the historical sequence of continuous state variables and the historical sequence of discrete event variables. The basic sequence is formatted for input, and the historical sequence of continuous state variables is arranged into a continuous input matrix by time step, and the historical sequence of discrete event variables is arranged into a discrete input matrix by time step. The condition variables are then associated with the continuous input matrix and the discrete input matrix to form the model input sequence.

7. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The obtained set of multi-scenario risk evolution trajectories includes: An improved state-space diffusion model is constructed, including multi-type condition variable embedding, discrete-continuous hybrid diffusion modeling, bidirectional state-space structure, and intervariate interaction module; A conditional diffusion denoising process is performed on the basic sequence using multi-type conditional variable embedding. The conditional variables are embedded to obtain a conditional embedding vector, and then embedded to obtain a diffusion step embedding vector. The basic sequence is encoded by continuous input to obtain a continuous feature sequence, and then encoded by discrete input to obtain a discrete feature sequence. The continuous noise estimation sequence and the discrete noise estimation sequence are obtained through a denoising unit. The joint generation process is based on discrete-continuous hybrid diffusion modeling. In the forward diffusion process, Gaussian noise is added to the continuous feature sequence to form a continuous noise sequence, and class perturbation is applied to the discrete feature sequence to form a discrete noise sequence. In the reverse diffusion process, the continuous noise estimation sequence and the discrete noise estimation sequence are used as the denoising results to update the continuous noise sequence and the discrete noise sequence to obtain the continuous denoised sequence and the discrete denoised sequence. Bidirectional temporal representation processing is performed using a bidirectional state space structure. In the denoising unit, forward state recursion is performed on the continuous denoised sequence and the discrete denoised sequence to obtain the forward hidden state sequence, and reverse state recursion is performed to obtain the reverse hidden state sequence. The forward hidden state sequence and the reverse hidden state sequence are then combined into a bidirectional temporal representation sequence by a fusion method. Interactive modeling is performed through a cross-variable interaction module. Interactive transformations are performed on the bidirectional time series representation sequence in the variable dimension to obtain an interactive feature sequence. The interactive feature sequence is then input into the decoding output unit to output a set of multi-scenario risk evolution trajectories within the prediction time window.

8. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The generation of reasoning-oriented evidence sequences includes: The set of multi-scenario risk evolution trajectories is traversed by scenario number to obtain the continuous state variable prediction sequence and discrete event variable prediction sequence for each scenario; State evidence is generated based on the prediction sequence of continuous state variables. Risk characteristics are calculated and categorized. Risk characteristics include rate of change, duration of continuous exceedance, peak amplitude, and cumulative amount. The rate of change is the difference between the values ​​of adjacent time steps divided by the time step interval. The duration of continuous exceedance is the number of time steps in continuous time steps that exceed the preset threshold multiplied by the time step interval. The peak amplitude is the maximum value minus the minimum value within the prediction time window. The cumulative amount is the sum of the values ​​of each time step within the prediction time window multiplied by the time step interval. Event evidence is generated by predicting discrete event variables, and state evidence and event evidence are combined along a unified timeline to form an evidence sequence. Each evidence item in the evidence sequence is assigned a scenario number and a twin object identifier to obtain an evidence sequence oriented towards reasoning.

9. The method for risk estimation of a power distribution room based on digital twins according to claim 1, characterized in that, The obtained operation and maintenance decision includes: Establish an accident chain reasoning graph, determine nodes and directed edges based on the primary topology and protection linkage relationship of the power distribution room, and map the evidence items in the evidence sequence to the corresponding nodes in the accident chain reasoning graph; Based on the functional module division of the cluster construction method, a cluster set is generated. The nodes in the accident chain reasoning graph are grouped according to the power distribution room functional module and mapped to the corresponding cluster. Evidence injection is performed on each cluster to initialize the cluster potential function. A joint tree structure is constructed based on the causal propagation path priority group connection sorting method. The connection order between groups is determined according to the accident chain propagation direction and the protection linkage order. Each group is connected to form a joint tree structure, and a separator set is determined for adjacent groups. The separator set is the set of common nodes of adjacent groups. Message passing is performed on the joint tree structure. Message passing includes evidence collection and passing from leaf cluster to root cluster and evidence distribution from root cluster to leaf cluster. The posterior probability distribution of each node is calculated. The propagation path, scope of impact, and response time window of the accident chain are determined by the posterior probability distribution, and operation and maintenance decisions are output. The response time window is determined by the time point when the posterior probability first reaches the preset probability threshold and the time interval during which the posterior probability continues to be higher than the preset probability threshold.