Method for early warning of cascading failure risk

Abstract

The application discloses a cascading failure risk early warning method, comprising the following steps: collecting power grid operation data and arranging the operation data into an operation point state vector; constructing a device level graph model; calculating the distance from the operation point to a plurality of stability boundaries to form a stability boundary distance feature; generating a device level essential attribute risk index; encoding the device operation feature, the stability boundary distance, the essential attribute risk index and the device type into an initial hidden vector; performing graph propagation through a double-channel graph propagation model comprising a mechanism guided propagation channel and a data driven propagation channel based on the device level graph model and the initial hidden vector to obtain a mechanism channel hidden state and a data channel hidden state; obtaining a fused risk hidden state through a stability boundary distance adaptive fusion gate; and decoding the risk based on the fused risk hidden state. The cascading failure risk early warning method realizes early warning of cascading failure risk when the device is not out of limit through double-channel graph propagation and stability boundary distance quantization.

Description

Technical Field

[0001] This invention belongs to the field of power system security analysis and early warning technology, specifically relating to an early warning method for cascading failure risks. Background Technology

[0002] With the large-scale integration of new energy sources, the enhancement of inter-regional power transmission channels, the increase in the proportion of power electronic equipment, and the increase in load volatility, the power grid operation status exhibits greater uncertainty, nonlinearity, and time-varying characteristics. Cascading failures are usually not directly caused by a single device exceeding its limits, but are formed by the combined effects of factors such as line overload, power flow shift, insufficient node voltage support, transient power angle instability, frequency deviation, protection actions, safety control strategy triggering, and insufficient reserves.

[0003] Existing methods for identifying cascading fault risks in power grids mainly include the following categories: The first category is methods based on static rules or safety constraints. These methods assess risk based on rules such as whether line power flow exceeds limits, node voltage exceeds limits, cross-sectional power flow exceeds limits, and frequency exceeds limits. While simple to implement, these methods can only make judgments after equipment has exceeded limits, failing to identify the accumulation of risk before limits are exceeded, resulting in a significant lag in early warning.

[0004] The second category is based on transient stability simulation or continuous power flow methods. These methods analyze the system's distance from instability through time-domain simulation, transient energy functions, continuous power flow, voltage stability margin, and power angle stability margin. While these methods can assess system stability margin, they are computationally intensive and time-consuming, making it difficult to meet the needs of online real-time early warning. Furthermore, they typically output system-level stability margins, making it difficult to pinpoint specific high-risk devices and risk propagation paths.

[0005] The third category is cascading fault early warning methods based on monitoring or alarm information. This method collects data such as substation monitoring information, protection information, and equipment alarm information to identify the correlation between alarms and perform fault diagnosis, early warning, and assessment. This type of method relies on already occurring alarm information, and also suffers from early warning lag, and it is difficult to quantitatively explain the physical propagation mechanisms between alarms.

[0006] The fourth category is risk assessment methods based on machine learning or graph neural networks. This method represents the power grid as a graph structure and uses models such as graph convolutional networks, spatiotemporal graph networks, and recurrent neural networks to extract electrical and topological features to achieve risk prediction. While these methods can learn implicit patterns in historical data, they have the following shortcomings: First, ordinary graph propagation models lack constraints related to power grid mechanisms. Traditional graph neural networks rely heavily on physical topology propagation characteristics, but cascading fault propagation is not only related to topological adjacency but also to power flow direction, power flow transfer sensitivity, protection action relationships, and cross-sectional constraints. Relying solely on topological adjacency can easily lead to misjudgments. Second, a single model cannot simultaneously consider both physical mechanisms and historical data patterns. Mechanistic models are highly interpretable but struggle to cover the implicit fault evolution patterns under complex operating conditions. Data-driven models can learn historical patterns, but without machine learning mechanisms, they cannot effectively address these patterns. First, it lacks the ability to interpret the results. Second, it is prone to producing results that do not conform to the laws of the power grid. Third, it lacks the interpretability of the output results. Existing deep learning methods mostly output system-level risk scores, which are difficult to explain at the same time as high-risk equipment, risk propagation path, dominant stability boundary type, and whether the risk comes from mechanistic factors or historical similar patterns. Fourth, it lacks the ability to provide early warnings when equipment has not yet exceeded its limits. The risk of cascading failures often begins to accumulate before the equipment exceeds its limits. For example, the line thermal margin continues to decline, the cross-sectional power flow continues to approach the limit, and the node reactive power margin continues to decrease. Existing methods usually identify these issues only after an alarm or limit exceedance occurs, resulting in a lag in early warning.

[0007] In summary, existing technologies for early warning of cascading failure risks still suffer from problems such as lack of explicit quantification of the distance from the operating point to the stability boundary, lack of power grid mechanism constraints in the graph propagation model, difficulty in taking into account both physical mechanisms and historical data patterns, insufficient early warning capability when equipment does not exceed the limit, and insufficient interpretability of output results. There is an urgent need for a new early warning method for cascading failure risks that can comprehensively consider the above factors. Summary of the Invention

[0008] This invention provides an early warning method for cascading failure risks to solve the aforementioned technical problems, specifically employing the following technical solution: A method for early warning of cascading failure risks includes the following steps: Collect power grid operation data, including electrical operation characteristics, topology status, protection and control status, historical risk-related characteristics, and key section and operation mode characteristics, and organize the above data into an operation point status vector; The power grid is constructed using a device-level graph model, which includes a set of device nodes, a set of edges connecting the relationships between devices, a multi-relationship adjacency matrix, and device node features. Calculate the distance from the operating point to multiple types of stability boundaries to form stability boundary distance features. The multiple types of stability boundaries include thermal stability boundaries, voltage stability boundaries, frequency stability boundaries, protection action boundaries, and critical section limit boundaries. Based on the equipment type mask, the applicable stability boundary type of the equipment is determined. After the stability boundary distance is transformed into a risk-based value, the equipment-level essential attribute risk index is generated by combining power flow transfer sensitivity, historical failure chain participation and current load level. The equipment operating characteristics, stability boundary distance, essential attribute risk indicators and equipment type are encoded as initial latent vectors. Based on the equipment-level graph model and the initial latent vectors, graph propagation is performed through a dual-channel graph propagation model that includes a mechanism-guided propagation channel and a data-driven propagation channel to obtain the mechanism channel latent state and the data channel latent state. Based on the stable boundary distance, the change in stable boundary distance, the risk index of essential attributes, and the risk scores of the two channels, the hidden states of the mechanism channel and the hidden states of the data channel are dynamically fused through the stable boundary distance adaptive fusion gate to obtain the fused risk hidden state; Based on the fusion of risk hidden states, risk decoding is performed to output system-level risks, device-level risks, path-level risks, and dominant stability boundary types, thereby achieving early warning of cascading failure risks.

[0009] Furthermore, the construction of the device-level graph model includes: Busbars, lines, transformers, generators, loads, protection devices, and key sections are all abstracted into equipment nodes; Construct the relationships between devices, including physical connection relationships, power flow direction relationships, power flow transfer relationships, protection linkage relationships, cross-section ownership relationships, and historical co-occurrence relationships; Correspondingly, a multi-relationship adjacency matrix is ​​constructed, including a physical topology connection relationship adjacency matrix, a power flow direction relationship adjacency matrix, a power flow transfer sensitivity relationship adjacency matrix, a protection association relationship adjacency matrix, a cross-section constraint relationship adjacency matrix, and a historical co-occurrence relationship adjacency matrix; Normalize each adjacency matrix.

[0010] Furthermore, the distance from the calculated running point to the multi-class stability boundary is calculated using the following formula:

[0011] in, Indicates the thermal stability boundary distance of device i. This represents the voltage stability boundary distance of device i. This represents the frequency stability boundary distance of device i in its current operating state. Indicates the protection action boundary distance of device i. This represents the stability boundary distance between device i and the critical section constraints.

[0012] Among them, S i (t) represents the current load of the device. For thermal stability limits, To prevent constants with a denominator of zero;

[0013] in, For node reactive power margin, Based on the baseline reactive power demand, For normalization,

[0014]

[0015] Where f(t) is the system frequency, f0 is the rated frequency, ROCOF(t) is the rate of frequency change, and R... reserve (t) represents the available reserve capacity, and γ1, γ2, and γ3 are weighting parameters. Indicates the variable Restricted to the interval [0,1] in, To set the protection action value, I i (t) represents the current current or equivalent action quantity.

[0016] Among them, P i (t) represents the current tidal current at the cross-section. Limits for cross-sectional tidal flow.

[0017] Furthermore, the generation device-level intrinsic attribute risk index adopts the following formula:

[0018] Where b is the stable boundary type, Mask i,b For device i, whether boundary b, r applies i,b (t)=1-clip(d i,b (t), 0, 1) represents the risk-based boundary distance, Sens i (t) represents the power flow transfer sensitivity, Hist i For historical failure chain participation, Load i (t) represents the current load level, β b β s β h β l These are the weight parameters.

[0019] Furthermore, the dual-channel graph propagation model includes an input encoder, a mechanism-guided propagation channel, a data-driven propagation channel, a stable boundary distance adaptive fusion gate, and a risk decoder; The input encoder encodes the device operating characteristics, stability boundary distance, essential attribute risk indicators, and device type into latent vectors. The mechanism-guided propagation channel propagates risks based on the physical mechanisms of the power grid. The data-driven propagation channel learns the propagation relationships of implicit risks in historical fault chains, alarm co-occurrences, limit-crossing co-occurrences, and scheduling and handling records. The stable boundary distance adaptive fusion gate dynamically adjusts the contributions of the two channels based on how close the current device is to the stable boundary; The risk decoder outputs risk results based on the fused device hidden state.

[0020] Furthermore, the mechanism-guided propagation channel includes: Construction mechanism propagation weights:

[0021] Among them, Aij topo Indicates physical connection relationship, Aij pf Indicates the direction of the trend, Aij sens Aij represents the power flow transfer sensitivity. protect Indicates a protected association, Aij section This represents the cross-sectional constraint relationship, where α1 to α5 are weighting coefficients. Introducing a stable boundary distance modulation factor:

[0022] in, This represents the risk-normalized distance of device i to the stability boundary. Sens represents the risk-based distance of device j to the stability boundary. ij Indicates the strength of the electrical influence of device i on device j, ProtectLink ij This indicates whether there is a protection or control linkage relationship between devices, where σ represents the Sigmoid activation function and η1 to η4 are weight parameters. The mechanism propagation weights modulated by the stable boundary distance are obtained as follows:

[0023] Message propagation is performed based on the modulated mechanism propagation weights, and the hidden state of the mechanism channel is updated.

[0024] Furthermore, the data-driven propagation channel includes: Constructing a data-driven adjacency matrix:

[0025] Among them, CoFault ij CoAlarm represents the frequency with which devices i and j co-occur in the historical failure chain. ij Indicates the frequency of alarm occurrences, CoOverload ij CoDispatch indicates the frequency of out-of-limit co-occurrence. ij Indicates the frequency of co-occurrence of dispatch and handling; Extracting features from historical time windows:

[0026] Among them, z i (t) represents the time-series feature representation of device i at time t, TemporalEncoder represents the time feature encoder, and h i (tk),...,h i (t) represents the historical hidden state sequence of device i from time tk to time t; Data-driven graph propagation and updating of hidden states of data channels are performed based on data-driven adjacency matrices and historical time window features.

[0027] Furthermore, the stable boundary distance adaptive fusion gate includes: Calculate the fusion gate coefficient:

[0028] in, This indicates the hidden state of the mechanism channel. D represents the hidden state of the data channel. i Denotes the stable boundary distance, ΔD i BRI represents the change in distance from the stability boundary. i Indicators representing inherent risk attributes This represents the risk score of the mechanism pathway. W represents the data channel risk score. f Let b represent the fusion gate weight matrix. f σ represents the fusion gate bias term, and σ represents the Sigmoid activation function. The hidden state of the merged device is:

[0029] g i This represents the fusion gate coefficient of device i. When the device is close to the stability boundary, has high power flow transfer sensitivity, or has a high risk of protection action, the weight of the mechanism channel is increased; when there are similar fault chains or alarm combinations in the historical samples, the weight of the data-driven channel is increased.

[0030] Furthermore, the risk decoder includes: System-level risk decoding:

[0031] Among them, Decoder sys This represents a system-level risk decoder, and Pool represents a graph-level pooling function. This represents the set of all device nodes' merged hidden states; Device-level risk decoding:

[0032] Among them, Decoder dev Indicates a device-level risk decoder; Path-level risk decoding:

[0033] Among them, Decoder path This represents a path-level risk decoder, where p represents a candidate risk propagation path consisting of several device nodes and relational edges. The mechanism propagation weights of the path edge (i,j) are represented. The risk levels are classified as R0 safe, R1 concern, R2 warning, R3 severe warning and R4 emergency risk.

[0034] Furthermore, the model training adopts a multi-task joint training objective:

[0035] Where Loss represents the total training loss of the model, Loss risk Loss is used to predict losses at the system level. boundary To stabilize the boundary distance regression loss, Loss device Loss is used to predict losses at the equipment level. path Loss for propagation path identification physics The loss is the physical consistency constraint loss, where λ1 to λ5 represent the weight coefficients of each loss term. The physical consistency constraints include: the risk should not decrease unnecessarily when the distance to the stability boundary decreases; the closer the equipment is to the stability boundary, the higher its risk should be; and the risk of related equipment with high power flow transfer sensitivity should increase with the increase of disturbance. The model is trained using offline simulation samples, reinforcement samples near the boundary, and historical fault chain samples. An online fine-tuning mechanism is set up to update the model with small steps based on the posterior label. At the same time, anti-forgetting constraints are added to prevent the loss of stable risk identification capabilities obtained in the offline training phase.

[0036] The advantage of this invention lies in the early warning method for cascading failure risks provided. By explicitly quantifying the distance from the operating point to multiple stability boundaries and combining it with equipment type masks to generate essential attribute risk indicators, the cascading failure risk has an interpretable physical margin basis. It can clearly indicate which category the risk comes from: thermal stability, voltage stability, frequency stability, protection action, or cross-sectional limit. This solves the problem that existing methods directly output the risk level and cannot explain how much margin there is between the current state and the stability boundary.

[0037] The advantages of this invention also lie in the early warning method for cascading fault risks provided. By constructing a dual-channel graph propagation model that includes a mechanism-guided propagation channel and a data-driven propagation channel, and setting a stable boundary distance adaptive fusion gate to dynamically adjust the contribution of the two channels, the model can not only perform interpretable risk propagation based on the power grid physical topology, power flow direction, power flow transfer sensitivity, protection association, and cross-sectional constraints, but also learn the implicit fault evolution laws in historical fault chains, alarm co-occurrence, and limit-crossing co-occurrence. This solves the problem that a single mechanism model is difficult to cover complex operating modes or that a single data model is prone to producing results that do not conform to the laws of the power grid.

[0038] The advantages of this invention also lie in the early warning method for cascading failure risks provided. By identifying risks when the distance to the stable boundary continues to decrease even before the equipment exceeds the limit, the method outputs system-level risk scores, equipment-level risk rankings, path-level propagation chains, and dominant stable boundary types. This enables dispatchers to obtain complete early warning information, including high-risk equipment, risk propagation paths, and explanations of risk causes, before a failure occurs. This solves the problem of delayed early warning caused by existing methods that typically identify risks only after an alarm or limit exceedance has occurred, as well as the problem that existing deep learning methods only output system-level risk scores with insufficient interpretability. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the early warning method for cascading failure risk proposed in this application. Detailed Implementation

[0041] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0042] In the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection" and "linkage" should be interpreted broadly, and can refer to mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0043] like Figure 1 The diagram illustrates an early warning method for cascading failure risks according to this application, comprising the following steps: S1: Collect power grid operation data, including electrical operation characteristics, topology status, protection and control status, historical risk-related characteristics, and key section and operation mode characteristics. Organize the above data into an operating point state vector. S2: Construct an equipment-level graph model of the power grid. The equipment-level graph model includes a set of equipment nodes, a set of relationships between equipment, a multi-relationship adjacency matrix, and equipment node characteristics. S3: Calculate the distance from the operating point to multiple types of stability boundaries, forming stability boundary distance characteristics. These multiple stability boundaries include thermal stability boundaries, voltage stability boundaries, frequency stability boundaries, protection action boundaries, and key section limit boundaries. S4: Determine the applicable stability boundary type for each equipment based on the equipment type mask. After risk-based transformation of the stability boundary distance, combine it with power flow transfer sensitivity, historical fault chain participation, and current load level to generate equipment-level essential attribute risk indicators. S5: Encode the equipment operation characteristics, stability boundary distances, essential attribute risk indicators, and equipment type into initial latent vectors. Based on the equipment-level graph model and the initial latent vectors, perform graph propagation through a dual-channel graph propagation model that includes a mechanism-guided propagation channel and a data-driven propagation channel to obtain the mechanism channel latent state and the data channel latent state. S6: Based on the stable boundary distance, the change in stable boundary distance, the essential attribute risk index, and the risk scores of the two channels, the hidden states of the mechanism channel and the data channel are dynamically fused using the stable boundary distance adaptive fusion gate to obtain the fusion risk hidden state. S7: Based on the fusion risk hidden state, risk decoding is performed, outputting system-level risk, device-level risk, path-level risk, and dominant stable boundary type, achieving early warning of cascading failure risks. This invention solves the problem of existing methods lacking an interpretable physical margin basis by explicitly quantifying the distance from the operating point to the stable boundary; it solves the problem that a single model cannot cover the hidden failure evolution laws under complex operating modes by using a dual-channel graph propagation model that takes into account both physical mechanisms and historical data patterns; it solves the problem of fusion rigidity caused by simply adding mechanisms and data by achieving dynamic fusion through the stable boundary distance adaptive fusion gate; it solves the problem of insufficient interpretability of output results in existing methods by outputting three levels of risk: system-level, device-level, and path-level; and it solves the problem of delayed early warning in existing methods by identifying risks before the equipment exceeds its limits. The following further explains the above steps.

[0044] For step S1: Collect power grid operation data, including electrical operation characteristics, topology status, protection and control status, historical risk-related characteristics, and key section and operation mode characteristics, and organize the above data into an operation point state vector.

[0045] Collect power grid operation data for the current moment and within historical time windows, including: bus voltage, phase angle, active power injection, reactive power injection; active power flow, reactive power flow, current, load factor and thermal stability limit of lines; generator output, reserve capacity, and predicted output of new energy sources; protection settings, protection action time limits, and protection activation / deactivation status; key transmission sections and their power flow limits; historical fault chains, alarm sequences, over-limit records and dispatch handling records.

[0046] The above data is organized into a runtime state vector:

[0047] Among them, H t Indicates electrical operating characteristics, T t P represents the topological state. t Indicates protection and control status, R t Indicating historical risk-related characteristics, C t This represents the key cross-sections and operational characteristics. The above data is organized into an operational point state vector, characterizing the comprehensive operational state of the power grid at time t. The collected data is used for subsequent construction of equipment-level graphical models, calculation of stability boundary distances, and generation of risk indicators.

[0048] For step S2: Construct a device-level graph model of the power grid. The device-level graph model includes a set of device nodes, a set of edges between devices, a multi-relationship adjacency matrix, and device node features.

[0049] In the embodiments of this application, constructing a device-level graph model includes: Busbars, lines, transformers, generators, loads, protection devices, and critical sections are uniformly abstracted as equipment nodes. Relationships between these equipment are constructed, including physical connection relationships, power flow direction relationships, power flow transfer relationships, protection linkage relationships, section affiliation relationships, and historical co-occurrence relationships. Correspondingly, multi-relationship adjacency matrices are constructed, including adjacency matrices for physical topology connection relationships, power flow direction relationships, power flow transfer sensitivity relationships, protection association relationships, section constraint relationships, and historical co-occurrence relationships. Each adjacency matrix is ​​then normalized.

[0050] Specifically, to facilitate a unified representation of the risks associated with busbars, lines, transformers, generators, protection devices, and critical sections, this invention constructs the power grid as an equipment-level diagram:

[0051] Where V represents the set of device nodes, E represents the set of edges connecting the devices, and A t H represents a multi-relation adjacency matrix. t This indicates the characteristics of the device node, and t represents the sampling time or the number of the operating section.

[0052] The following adjacency matrix is ​​constructed accordingly: A t =A topo A pf A sens A protect A section A data Among them, A topo Indicates physical topology connections; A pf Indicates the relationship between current direction; A sens Indicates the power flow transfer sensitivity relationship; A protect This indicates the relationship between protection linkage, backup protection, and safety control actions; A section Indicates the constraint relationship between the equipment and the critical section; A data This represents the implicit relationships formed by historical fault chains, alarm co-occurrences, and limit violation co-occurrences. To avoid inconsistencies in the dimensions of different matrices, each adjacency matrix is ​​normalized before being input into the model.

[0053] This application unifies the abstraction of heterogeneous power grid equipment into equipment nodes, realizing a unified risk representation of buses, lines, transformers, generators, protection devices, and key sections within the same graph framework. By constructing six types of relational edges, it breaks through the limitation of traditional graph neural networks that rely solely on physical topological adjacency, making risk propagation not only related to topological adjacency but also to power flow direction, power flow transfer sensitivity, protection action relationships, and section constraints, thereby improving the consistency between the graph model and the physical mechanism of the power grid.

[0054] For step S3: Calculate the distance from the operating point to multiple stability boundaries to form stability boundary distance characteristics. The multiple stability boundaries include thermal stability boundary, voltage stability boundary, frequency stability boundary, protection action boundary, and critical section limit boundary.

[0055] For the current running point X t The distances from the calculation device to various stability boundaries are used to form stability boundary distance features:

[0056] Among them, D i (t) represents the stable boundary distance vector of device i.

[0057] This represents the thermal stability boundary distance of device i, primarily applicable to branch-type equipment such as lines and transformers, used to describe the current load's margin from the thermal stability limit. A smaller value indicates that the load is closer to the thermal stability threshold.

[0058] This value represents the voltage stability boundary distance of device i. It is mainly applicable to voltage-related nodes such as busbars, load centers, and new energy grid connection points, and is used to describe the current voltage state or reactive power margin's distance from the voltage stability boundary. The smaller the value, the closer it is to the risk of voltage instability.

[0059] This represents the frequency stability boundary distance of device i under its current operating state, and is mainly related to system frequency deviation, frequency change rate, and reserve capacity. For non-frequency dominant devices, the system-level frequency stability distance can be mapped to the relevant device nodes. The smaller this value, the higher the frequency stability risk.

[0060] This represents the protection action boundary distance of device i, used to describe the margin between the current electrical quantities such as current, voltage, and frequency and the protection action setting or control triggering condition. The smaller this value, the closer it is to the protection action or safety control action triggering condition.

[0061] This represents the stability boundary distance associated with device i and the critical section constraints. It is primarily used to describe the margin of the current power flow distance section limit for the section to which the device belongs or is associated. The smaller this value, the closer the relevant section is to the limit.

[0062] The unified convention of this invention is as follows: A smaller distance to the stability boundary indicates that the distance is closer to the risk boundary; a distance less than 0 indicates that the limit has been exceeded; the distance is normalized before entering the model; within the model, 1-d can be used to represent the risk-normalized distance.

[0063] For branch equipment such as lines and transformers, calculate:

[0064] Among them, S i (t) represents the current load of the device. For thermal stability limits, To prevent constants with a denominator of zero.

[0065] For busbars or load center nodes, calculations are performed based on reactive power margin or voltage stability critical point:

[0066] in, For node reactive power margin, Based on the baseline reactive power demand, For normalization, Frequency stability boundary distance:

[0067] Among them, frequency risk quantity:

[0068] Where f(t) is the system frequency, f0 is the rated frequency, ROCOF(t) is the rate of frequency change, and R... reserve (t) represents the available reserve capacity, and γ1, γ2, and γ3 are weighting parameters. Indicates the variable It is restricted to the interval [0,1].

[0069] Calculated based on the difference between the protection setting and the current electrical quantity:

[0070] in, To set the protection action value, I i (t) represents the current current or equivalent action quantity, and the protection action boundary distance is used to indicate the degree to which the equipment state approaches the protection action triggering condition.

[0071] Calculations for key transmission sections:

[0072] Among them, P i (t) represents the current tidal current at the cross-section. Limits for cross-sectional tidal flow.

[0073] This application achieves explicit quantification of the operating point to thermal stability, voltage stability, frequency stability, protection action, and cross-sectional limit boundaries by calculating five types of stability boundary distances separately, thus providing an interpretable physical margin basis for cascading failure risks. By unifying the meaning of distances and normalizing the process, boundary distances of different dimensions can be compared and fused within the same framework, providing standardized input features for subsequent adaptive fusion gates.

[0074] For step S4: Determine the applicable stability boundary type for the device based on the device type mask, perform risk-based transformation on the stability boundary distance, and generate device-level essential attribute risk indicators by combining power flow transfer sensitivity, historical fault chain participation, and current load level.

[0075] Since different devices have different applicable stability boundaries, this invention uses a device type mask to generate device-level intrinsic attribute risk indicators. For device i, the definition is:

[0076] Where b is the stable boundary type, Mask i,b For device i, whether boundary b and r apply i,b (t)=1-clip(d i,b (t), 0, 1) represents the risk-based boundary distance, Sens i (t) represents the power flow transfer sensitivity, Histi For historical failure chain participation, Load i (t) represents the current load level, β b β s β h β l These are weighted parameters. In this approach, the main indicators used for the line are thermal stability, protection, and cross-section related indicators; for the bus, voltage stability and reactive power margin indicators are mainly used; for the generator, frequency, reserve, and power angle related indicators can be used; and for cross-section nodes, the cross-section limit distance is mainly used.

[0077] Understandably, by using equipment type masks to distinguish the applicable boundary types of different equipment, the problem of unified risk representation for heterogeneous equipment is solved, enabling different types of equipment such as lines, buses, generators, and cross-sections to use their respective relevant stability boundary indicators for risk assessment. By introducing power flow transfer sensitivity, historical fault chain participation, and current load level, the essential attribute risk indicators not only reflect the current operating status but also the criticality of the equipment in fault propagation and the degree of historical fault participation, thus improving the comprehensiveness and accuracy of risk indicators.

[0078] For step S5: Encode the equipment operating characteristics, stable boundary distance, essential attribute risk indicators and equipment type into initial latent vectors. Based on the equipment-level graph model and the initial latent vectors, perform graph propagation through a dual-channel graph propagation model that includes mechanism-guided propagation channels and data-driven propagation channels to obtain the mechanism channel latent state and the data channel latent state.

[0079] For step S6: Based on the stable boundary distance, the change in stable boundary distance, the risk index of essential attributes, and the risk scores of the two channels, the hidden state of fusion risk is obtained through the hidden state of the channel and the hidden state of the data channel in the dynamic fusion mechanism of the stable boundary distance adaptive fusion gate.

[0080] For step S7: Based on the fusion of risk hidden states, perform risk decoding and output system-level risks, device-level risks, path-level risks and dominant stability boundary types to achieve early warning of cascading failure risks.

[0081] In the embodiments of this application, the dual-channel graph propagation model includes an input encoder, a mechanism-guided propagation channel, a data-driven propagation channel, a stable boundary distance adaptive fusion gate, and a risk decoder.

[0082] The input encoder encodes device operating characteristics, stability boundary distance, inherent attribute risk indicators, and device type into latent vectors. The mechanism-guided propagation channel propagates risk based on the physical mechanisms of the power grid. The data-driven propagation channel learns from the implicit risk propagation relationships in historical fault chains, alarm co-occurrences, limit-crossing co-occurrences, and dispatch handling records. The stability boundary distance adaptive fusion gate dynamically adjusts the contributions of the two channels based on the degree to which the device approaches the stability boundary. The risk decoder outputs the risk result based on the fused device latent states.

[0083] By using equipment type masks to differentiate the applicable boundary types for different equipment, the challenge of unified risk characterization for heterogeneous equipment is solved. This allows different types of equipment, such as lines, buses, generators, and cross-sections, to use their respective relevant stability boundary indicators for risk assessment. By introducing power flow transfer sensitivity, historical fault chain participation, and current load levels, the intrinsic attribute risk indicators not only reflect the current operating status but also the equipment's criticality in fault propagation and its historical fault participation, thus improving the comprehensiveness and accuracy of the risk indicators.

[0084] The input encoder encodes device operating characteristics, stability boundary distance, and essential attribute risk indicators into latent vectors:

[0085] Where, x i (t) represents the original operating characteristics of the equipment, D i (t) is the stable boundary distance, BRI i (t) is the essential attribute risk indicator, Type i Encodes the device type.

[0086] In the embodiments of this application, the mechanism-guided propagation channel is used to express risk propagation under the physical mechanism of the power grid.

[0087] Mechanisms guiding the propagation channels include: Construction mechanism propagation weights:

[0088] Among them, Aij topo Indicates physical connection relationship, Aij pf Indicates the direction of the trend, Aij sens Aij represents the power flow transfer sensitivity. protect Indicates a protected association, Aij section This represents the cross-sectional constraint relationship, where α1 to α5 are weighting coefficients.

[0089] Then, a stable boundary distance modulation factor is introduced:

[0090] in, This represents the risk-normalized distance of device i to the stability boundary. Sens represents the risk-based distance of device j to the stability boundary. ij Indicates the strength of the electrical influence of device i on device j, ProtectLink ij This indicates whether there is a protection or control linkage between devices. σ represents the Sigmoid activation function, and η1 to η4 are weight parameters.

[0091] The mechanism propagation weights modulated by the stable boundary distance are obtained as follows:

[0092] Message propagation and latent state updates of the mechanism channel are performed based on the modulated mechanism propagation weights. By introducing a stability boundary distance modulation factor, the mechanism propagation weights can be dynamically adjusted according to the degree to which the equipment is approaching the stability boundary. The closer the equipment is to the risk boundary, the stronger its risk impact on its neighbors, which conforms to the actual physical laws of power grid fault propagation. By incorporating power flow transfer sensitivity and protection linkage into the modulation factor, the mechanism channel considers not only static topology connections but also dynamic electrical effects and protection strategy correlations, improving the accuracy of mechanism-guided propagation and the adaptability of the power grid.

[0093] Specifically, the mechanism channel message propagation is as follows:

[0094] in, The mechanism channel aggregation message received by device i represents the set of neighboring device nodes of device i, and j represents the neighboring device node that propagates the message to device i; Φphy represents the mechanism channel message transformation function; h j D represents the hidden state vector of device j; j BRI represents the stable boundary distance vector of device j. j This represents the inherent risk index of device j. The underlying state of the mechanism channel is then updated.

[0095] Among them, Ψ phy The mechanism channel state update function can be implemented using GRU, residual connections, feedforward neural networks, or graph neural network update units. Mechanism channel outputs include: device-level mechanism risk latent vector, device-level mechanism risk score, mechanism propagation path contribution, and dominant stability boundary type.

[0096] The data-driven propagation channel is used to learn the propagation relationships of historical fault chains, alarm co-occurrences, limit violation co-occurrences, and implicit risks in runtime samples. The data-driven adjacency matrix is ​​constructed as follows:

[0097] Among them, CoFault ij CoAlarm represents the frequency with which devices i and j co-occur in the historical failure chain. ij Indicates the frequency of alarm occurrences, CoOverload ij CoDispatch indicates the frequency of out-of-limit co-occurrence. ij This indicates the frequency of co-occurrence of scheduling and handling.

[0098] The data channel first extracts features from historical time windows:

[0099] Among them, z i (t) represents the time-series feature representation of device i at time t, TemporalEncoder represents the time feature encoder, and h i (tk),...,h i (t) represents the historical hidden state sequence of device i from time tk to time t.

[0100] Data-driven graph propagation and updating of hidden states of data channels are performed based on data-driven adjacency matrices and historical time window features.

[0101] Specifically, data-driven graph propagation is performed:

[0102] in, This represents the data channel aggregation message received by device i, N. data (i) represents the set of data neighbors formed based on historical co-occurrence, fault chains, and alarm chains. Φ represents the strength of the data-driven association between device i and device j; data z represents the data channel message transformation function. j h represents the historical time window characteristics of device j. j This indicates the current hidden state of device j. The hidden state of the data channel is then updated accordingly.

[0103] Data channel outputs: device-level data risk latent vector, device-level data risk score, historical similar fault chains, and data-driven high-risk propagation paths.

[0104] By learning the implicit risk propagation relationships from historical fault chains, alarm co-occurrences, limit-crossing co-occurrences, and scheduling and handling records, the data-driven channel can capture complex fault evolution patterns in historical data, compensating for the insufficient coverage of non-physically related faults by the mechanism channel. By introducing historical time window features, the model can use time series information to identify risk accumulation trends, rather than making judgments based solely on a single moment. By strictly distinguishing the data usage boundaries between online inference and subsequent fine-tuning, future information leakage is avoided, ensuring the causal correctness of online inference.

[0105] It should be noted that the online inference phase only uses data observed at the current and historical moments; future actual events, outcomes, and posterior labels do not participate in the risk inference at the current moment, but are only used for subsequent online fine-tuning.

[0106] To avoid simply adding the mechanism channel and the data channel together, this invention sets up a stable boundary distance adaptive fusion gate, which dynamically adjusts the contributions of the two channels according to how close the current device is to the stable boundary.

[0107] In embodiments of this application, the stable boundary distance adaptive fusion gate includes: Calculate the fusion gate coefficient:

[0108] in, This indicates the hidden state of the mechanism channel. D represents the hidden state of the data channel. i Denotes the stable boundary distance, ΔD i BRI represents the change in distance from the stability boundary. i Indicators representing inherent risk attributes This represents the risk score of the mechanism pathway. W represents the data channel risk score. f Let b represent the fusion gate weight matrix. f Let represent the fusion gate bias term, and σ represent the Sigmoid activation function. This is achieved by varying the stability boundary distance ΔD. i By incorporating a fusion gate input, the model can perceive risk development trends and increase the weight of the mechanism channel as the boundary distance continues to decrease, thus enhancing its sensitivity to the accumulation of physical risks. By using the risk scores of the two channels as the fusion basis, the fusion decision can refer to the confidence level of each channel, avoiding forced fusion even when the output of a certain channel is obviously unreasonable. Through a dynamic adjustment mechanism, the rigidity of fusion caused by simply adding the mechanism and data channels is solved, enabling the model to automatically select the more reliable channel basis in different operating scenarios, thereby improving the accuracy and adaptability of the fusion results.

[0109] The hidden state of the merged device is:

[0110] g i This represents the fusion gate coefficient for device i. When the device is nearing a stability boundary, has high power flow transfer sensitivity, or faces a significant risk of protection action, the weight of the mechanism channel is increased. When similar fault chains or alarm combinations exist in historical samples, the weight of the data-driven channel is increased.

[0111] The risk decoder outputs a risk result based on the fused device hidden state. In embodiments of this application, the risk decoder includes: System-level risk decoding:

[0112] Among them, Decoder sys This represents a system-level risk decoder, and Pool represents a graph-level pooling function. This represents the set of all hidden states of the merged device nodes.

[0113] Device-level risk decoding:

[0114] Among them, Decoder dev This indicates a device-level risk decoder.

[0115] Path-level risk decoding:

[0116] Among them, Decoder path This represents a path-level risk decoder, where p represents a candidate risk propagation path consisting of several device nodes and relational edges. The mechanism propagation weight represents the path edge (i,j).

[0117] The output includes: system-level risk score, device-level risk score, ranking of high-risk devices, high-risk propagation path, dominant stability boundary type, and contribution ratio of mechanism and data channel.

[0118] The final risk levels are classified as R0 safe, R1 concern, R2 warning, R3 severe warning, and R4 emergency risk.

[0119] This application achieves a complete risk characterization from the macroscopic system state to the microscopic device state and then to the propagation path through three-level risk decoding: system-level, device-level, and path-level. This solves the problem that existing deep learning methods only output system-level risk scores and struggle to locate high-risk devices and propagation paths. By introducing path-level risk decoding, the model can identify the specific propagation chain of risk in the power grid, providing dispatchers with targeted handling suggestions.

[0120] The training method for the model built in this application is as follows: A training sample set is constructed, which includes three categories: offline simulation samples, boundary reinforcement samples, and historical failure chain samples.

[0121] For offline simulation samples, the following disturbance scenarios are constructed based on typical operating conditions: load growth, renewable energy fluctuations, line disconnection, unit shutdown, cross-sectional power flow approaching the limit, low voltage, protection actions, and multiple disturbances. Tags are generated through power flow calculation, stability analysis, or cascading fault evolution simulation.

[0122] Where: Y represents the set of training sample labels. This indicates the dominant stable boundary type label.

[0123] To enhance early warning capabilities, this invention focuses on constructing stable samples near the boundary for enhanced samples near the boundary.

[0124] The construction methods include: gradually approaching the stability boundary along the load growth direction; approaching the stability boundary along the direction of the decrease or fluctuation of new energy output; approaching the stability boundary along the power flow transfer direction of the critical line; approaching the stability boundary along the direction of the reduction of reactive power reserve; generating samples along the direction of the cross-sectional power flow approaching the limit; and generating samples along the direction of the protection action threshold.

[0125] Label samples near the boundary: stable boundary distance; whether they will enter a risk state within a future time window; high-risk equipment; high-risk propagation path.

[0126] For historical fault chain samples, extract the following from historical fault records: initial faulty device, subsequent tripping device sequence, power flow transfer chain, voltage drop chain, protection action chain, load loss scale, and dispatching and handling actions. Generate labels for real or approximate fault propagation paths.

[0127] In the embodiments of this application, model training adopts a multi-task joint training objective:

[0128] Where Loss represents the total training loss of the model, Loss risk Loss is used to predict losses at the system level. boundary To stabilize the boundary distance regression loss, Loss device Loss is used to predict losses at the equipment level. path Loss for propagation path identification physics The loss is the physical consistency constraint loss, and λ1 to λ5 represent the weight coefficients of each loss term.

[0129] Physical consistency constraints include: risk should not decrease unnecessarily as the distance to the stability boundary decreases; the closer a device is to the stability boundary, the higher its risk should be; and the risk of related devices with high sensitivity to power flow transfer should increase with the increase of disturbance.

[0130] The model is trained using offline simulation samples, reinforcement samples near the boundary, and historical fault chain samples. An online fine-tuning mechanism is set up to update the model with small steps based on the posterior label. At the same time, anti-forgetting constraints are added to prevent the loss of stable risk identification capabilities obtained in the offline training phase.

[0131] When running online, the system executes according to a fixed cycle or event-triggered method: acquiring real-time running cross-sections, updating equipment-level graphical models, calculating stable boundary distances, generating essential attribute risk indicators, performing mechanism channel and data channel propagation separately, fusing the outputs of the two channels through an adaptive fusion gate, outputting system-level / equipment-level / path-level risks, determining whether to trigger warnings, and outputting risk explanation information.

[0132] The conditions for triggering early warning include: the system-level risk score exceeds the threshold, the stability boundary distance of a certain device is lower than the threshold, the boundary distance of a certain section decreases continuously, the risk of a certain propagation path increases continuously, multiple types of stability boundaries approach simultaneously, and the mechanism channel and data channel output high risk simultaneously.

[0133] The following is an example of an early warning output: Risk level: R2 warning System risk score: 0.76 Main risks: Thermal stability boundary + Cross-sectional limit boundary High-risk equipment lines: L1 / L3, busbar B2 High-risk transmission path: L1->L3->Section-S1->Bus-B2 Minimum boundary distance: 0.08 Mechanistic channel contribution: 74% Data channel contribution: 26% We recommend paying attention to: power flow at section S1, load rates at lines L1 / L3, and reactive power support status at bus B2. To adapt to changes in power grid topology, load structure, renewable energy output, protection settings, and operating modes, an online fine-tuning mechanism is implemented. The system caches online operating samples (including timestamps, graph status, boundary distances, predicted risks, actual events, scheduling actions, and posterior status). When a future time window expires, posterior labels are generated based on the actual operating results. If a high-risk prediction is made and a limit violation or failure subsequently occurs, it is marked as a positive sample. Samples that are predicted to be low-risk but subsequently exhibit anomalies are marked as underreported samples. Samples predicted to be high-risk but which did not experience any abnormalities due to scheduling and handling were marked as samples where handling was suppressed. Samples predicted to be high-risk but without any abnormalities or intervention are marked as suspected false alarms.

[0134] Online fine-tuning uses small-step updates:

[0135] in, This represents the online model parameters before the k-th online update; Represents the online model parameters after the (k+1)th online update; μ represents the online fine-tuning learning rate; Loss online This represents the online learning loss calculated based on online samples and posterior labels. This represents the gradient of the online sample loss function with respect to the model parameters. And add anti-forgetting constraints:

[0136] Among them, Loss total ρ represents the total loss of online updates after adding the anti-forgetting constraint, θ represents the anti-forgetting constraint coefficient, and θ represents the total loss after adding the anti-forgetting constraint. online θ represents the current parameters of the online model. offline Indicates the parameters of the offline pre-trained model. This represents the L2-squared distance between the parameters of the online model and the offline model. The posterior label is only used for future fine-tuning and does not participate in current inference, thus avoiding future information leakage.

[0137] This application employs multi-task joint training, enabling the model to simultaneously learn five tasks: system-level risk prediction, stability boundary distance regression, equipment-level risk prediction, propagation path identification, and physical consistency constraints. This improves the comprehensiveness of the model output and the consistency among the outputs. Through physical consistency constraint loss, the physical laws of the power grid are explicitly embedded into the training objective, ensuring that the model output conforms to fundamental physical principles such as increased risk with decreasing stability boundary distance and increased risk of equipment with high power flow transfer sensitivity due to disturbances. This avoids the results that might arise from purely data-driven models that do not conform to power grid laws. Furthermore, by combining online fine-tuning with anti-forgetting constraints, the model can adapt to dynamic changes in power grid topology, load structure, renewable energy output, protection settings, and operating modes, while maintaining the stability risk identification capability acquired during offline training. This addresses the issue of performance degradation over time after the model goes live.

[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A method for early warning of cascading failure risk, characterized in that, Includes the following steps: Collect power grid operation data, including electrical operation characteristics, topology status, protection and control status, historical risk-related characteristics, and key section and operation mode characteristics, and organize the above data into an operation point status vector; The power grid is constructed using a device-level graph model, which includes a set of device nodes, a set of edges connecting the relationships between devices, a multi-relationship adjacency matrix, and device node features. Calculate the distance from the operating point to multiple types of stability boundaries to form stability boundary distance features. The multiple types of stability boundaries include thermal stability boundaries, voltage stability boundaries, frequency stability boundaries, protection action boundaries, and critical section limit boundaries. Based on the equipment type mask, the applicable stability boundary type of the equipment is determined. After the stability boundary distance is transformed into a risk-based value, the equipment-level essential attribute risk index is generated by combining power flow transfer sensitivity, historical failure chain participation and current load level. The equipment operating characteristics, stability boundary distance, essential attribute risk indicators and equipment type are encoded as initial latent vectors. Based on the equipment-level graph model and the initial latent vectors, graph propagation is performed through a dual-channel graph propagation model that includes a mechanism-guided propagation channel and a data-driven propagation channel to obtain the mechanism channel latent state and the data channel latent state. Based on the stable boundary distance, the change in stable boundary distance, the risk index of essential attributes, and the risk scores of the two channels, the hidden states of the mechanism channel and the hidden states of the data channel are dynamically fused through the stable boundary distance adaptive fusion gate to obtain the fused risk hidden state; Based on the fusion of risk hidden states, risk decoding is performed to output system-level risks, device-level risks, path-level risks, and dominant stability boundary types, thereby achieving early warning of cascading failure risks. The generation device-level essential attribute risk index is adopted using the following formula: Where b is the stable boundary type, Mask i,b For device i, whether boundary b and r apply i,b (t)=1-clip(d i,b (t),0,1) represents the risk-based boundary distance. Indicates the variable Restricted to the interval [0,1], Sens i (t) represents the power flow transfer sensitivity, Hist i For historical failure chain participation, Load i (t) represents the current load level, β b β s β h β l These are weight parameters; The mechanism-guided propagation channel propagates risks based on the physical mechanisms of the power grid. The data-driven propagation channel learns the propagation relationships of implicit risks in historical fault chains, alarm co-occurrences, limit-crossing co-occurrences, and scheduling and handling records. The stable boundary distance adaptive fusion gate includes: Calculate the fusion gate coefficient: in, This indicates the hidden state of the mechanism channel. D represents the hidden state of the data channel. i Denotes the stable boundary distance, ΔD i BRI represents the change in distance from the stability boundary. i Indicators representing inherent risk attributes This represents the risk score of the mechanism pathway. W represents the data channel risk score. f Let b represent the fusion gate weight matrix. f σ represents the fusion gate bias term, and σ represents the Sigmoid activation function. The hidden state of the merged device is: g i This represents the fusion gate coefficient of device i. When the device is close to the stability boundary, has high power flow transfer sensitivity, or has a high risk of protection action, the weight of the mechanism channel is increased; when there are similar fault chains or alarm combinations in the historical samples, the weight of the data-driven channel is increased.

2. The method for early warning of cascading failure risk according to claim 1, characterized in that, The construction of the device-level graph model includes: Busbars, lines, transformers, generators, loads, protection devices, and key sections are all abstracted into equipment nodes; Construct the relationships between devices, including physical connection relationships, power flow direction relationships, power flow transfer relationships, protection linkage relationships, cross-section ownership relationships, and historical co-occurrence relationships; Correspondingly, a multi-relationship adjacency matrix is ​​constructed, including a physical topology connection relationship adjacency matrix, a power flow direction relationship adjacency matrix, a power flow transfer sensitivity relationship adjacency matrix, a protection association relationship adjacency matrix, a cross-section constraint relationship adjacency matrix, and a historical co-occurrence relationship adjacency matrix; Normalize each adjacency matrix.

3. The method for early warning of cascading failure risk according to claim 1, characterized in that, The distance from the calculation running point to the multi-class stability boundary is expressed by the following formula: in, Indicates the thermal stability boundary distance of device i. This represents the voltage stability boundary distance of device i. This represents the frequency stability boundary distance of device i in its current operating state. Indicates the protection action boundary distance of device i. This represents the stability boundary distance between device i and the critical section constraints. Among them, S i (t) represents the current load of the device. For thermal stability limits, To prevent constants with a denominator of zero; in, For node reactive power margin, Based on the baseline reactive power demand, For normalization, Where f(t) is the system frequency, f0 is the rated frequency, ROCOF(t) is the rate of frequency change, and R... reserve (t) represents the available reserve capacity, and γ1, γ2, and γ3 are weighting parameters. Indicates the variable Restricted to the interval [0,1] in, To set the protection action value, I i (t) represents the current current or equivalent action quantity. Among them, P i (t) represents the current tidal current at the cross-section. Limits for cross-sectional tidal flow.

4. The method for early warning of cascading failure risk according to claim 1, characterized in that, The dual-channel graph propagation model includes an input encoder, a mechanism-guided propagation channel, a data-driven propagation channel, a stable boundary distance adaptive fusion gate, and a risk decoder. The input encoder encodes the device operating characteristics, stability boundary distance, essential attribute risk indicators, and device type into latent vectors. The stable boundary distance adaptive fusion gate dynamically adjusts the contributions of the two channels based on how close the current device is to the stable boundary; The risk decoder outputs risk results based on the fused device hidden state.

5. The method for early warning of cascading failure risk according to claim 4, characterized in that, The mechanism guiding the propagation channel includes: Construction mechanism propagation weights: Among them, Aij topo Indicates physical connection relationship, Aij pf Indicates the direction of the trend, Aij sens Aij represents the power flow transfer sensitivity. protect Indicates a protected association, Aij section This represents the cross-sectional constraint relationship, where α1 to α5 are weighting coefficients. Introducing a stable boundary distance modulation factor: in, This represents the risk-normalized distance of device i to the stability boundary. Sens represents the risk-based distance of device j to the stability boundary. ij Indicates the strength of the electrical influence of device i on device j, ProtectLink ij This indicates whether there is a protection or control linkage relationship between devices, where σ represents the Sigmoid activation function and η1 to η4 are weight parameters. The mechanism propagation weights modulated by the stable boundary distance are obtained as follows: Message propagation is performed based on the modulated mechanism propagation weights, and the hidden state of the mechanism channel is updated.

6. The method for early warning of cascading failure risk according to claim 4, characterized in that, The data-driven propagation channel includes: Constructing a data-driven adjacency matrix: Among them, CoFault ij CoAlarm represents the frequency with which devices i and j co-occur in the historical failure chain. ij Indicates the frequency of alarm occurrences, CoOverload ij CoDispatch indicates the frequency of out-of-limit co-occurrence. ij Indicates the frequency of co-occurrence of dispatch and handling; Extracting features from historical time windows: Among them, z i (t) represents the time-series feature representation of device i at time t, TemporalEncoder represents the time feature encoder, and h i (tk),...,h i (t) represents the historical hidden state sequence of device i from time tk to time t; Data-driven graph propagation and updating of hidden states of data channels are performed based on data-driven adjacency matrices and historical time window features.

7. The method for early warning of cascading failure risk according to claim 4, characterized in that, The risk decoder includes: System-level risk decoding: Among them, Decoder sys This represents a system-level risk decoder, and Pool represents a graph-level pooling function. This represents the set of all device node fused hidden states; Device-level risk decoding: Among them, Decoder dev Indicates a device-level risk decoder; Path-level risk decoding: Among them, Decoder path This represents a path-level risk decoder, where p represents a candidate risk propagation path consisting of several device nodes and relational edges. The mechanism propagation weights of the path edge (i,j) are represented. The risk levels are classified as R0 safe, R1 concern, R2 warning, R3 severe warning and R4 emergency risk.

8. The method for early warning of cascading failure risk according to claim 1, characterized in that, The model training adopts a multi-task joint training objective: Where Loss represents the total training loss of the model, Loss risk Loss is used to predict losses at the system level. boundary To stabilize the boundary distance regression loss, Loss device Loss is used to predict losses at the equipment level. path Loss for propagation path identification physics The loss is the physical consistency constraint loss, where λ1 to λ5 represent the weight coefficients of each loss term. The physical consistency constraints include: the risk should not decrease unnecessarily when the distance to the stability boundary decreases; the closer the equipment is to the stability boundary, the higher its risk should be; and the risk of related equipment with high power flow transfer sensitivity should increase with the increase of disturbance. The model is trained using offline simulation samples, reinforcement samples near the boundary, and historical fault chain samples. An online fine-tuning mechanism is set up to update the model with small steps based on the posterior label. At the same time, anti-forgetting constraints are added to prevent the loss of stable risk identification capabilities obtained in the offline training phase.

Images