A multi-fault mode relay protection setting value intelligent checking system and method

Abstract

The application relates to the technical field of power system relay protection, and discloses a multi-fault mode relay protection setting value intelligent checking system and method, wherein the multi-fault mode relay protection setting value intelligent checking method comprises the following steps: constructing a neural network model; setting a curriculum strategy; performing an adversarial training; outputting an evolution path; and performing risk early warning; after a distributed photovoltaic power is accessed to a power distribution system, the fault characteristics of the power grid change in nature, and the change is manifested as the change of fault current size, direction and sequence component characteristics; under the multi-fault mode, the distributed photovoltaic power can produce an external pumping effect to cause the reduction of a protection range and the refusal of operation, an auxiliary effect can cause the expansion of the protection range and the overstep tripping, and a reverse fault current can cause the misoperation of a healthy feeder protection; through the adversarial learning framework, the multi-fault mode under different fault positions and different DG capacities is intelligently checked, and the protection configuration problem under high penetration rate distributed photovoltaic access is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection technology, and more specifically, to a multi-fault mode relay protection setting intelligent verification system and method. Background Technology

[0002] With the large-scale grid connection of new energy sources and the increasing complexity of power grid structures, the verification of relay protection settings faces severe challenges.

[0003] Traditional verification methods are ill-suited to the new characteristics of diversified failure modes, blurred boundaries, and risk evolution and accumulation, especially in identifying unknown or marginal failures and uncovering non-intuitive multi-step cumulative risks.

[0004] With the widespread integration of distributed photovoltaic (PV) systems into distribution networks, the fault characteristics of the power grid have undergone fundamental changes. Distributed PV integration alters the fault current distribution characteristics of the system, including changes in the magnitude and direction of short-circuit currents. Traditional fault characteristic analysis methods, primarily based on the principle of linear superposition, struggle to accurately reflect the transient fault characteristics of distribution networks containing a high proportion of power electronic devices. Research indicates that when distributed PV employs positive-sequence component control, it only provides positive-sequence current. Under different fault locations and varying connection capacities, this significantly impacts overcurrent and distance protection, leading to issues such as changes in protection range, cascading tripping, and failure to operate. In this complex multi-fault mode, existing technologies have limited capabilities in actively and efficiently searching for the "most dangerous" fault mode and verifying settings accordingly. This may result in erroneous or non-operational protection settings under complex disturbances, threatening the safe and stable operation of the power grid.

[0005] Therefore, there is an urgent need for an intelligent verification method that can proactively uncover potential risks and adapt to complex failure modes. Summary of the Invention

[0006] This invention provides an intelligent verification system and method for multi-fault mode relay protection settings, which solves the technical problems in related technologies such as poor adaptability to complex and unknown fault modes and difficulty in actively exploring the evolution path of setting vulnerability.

[0007] This invention provides an intelligent verification method for multi-fault mode relay protection settings, comprising:

[0008] Construct a neural network-based value checker and path explorer;

[0009] Based on the fixed value checker and the path explorer, a multi-stage course learning strategy from easy to difficult is set to guide the training process of the path explorer.

[0010] According to the stage settings of the course learning strategy, adversarial self-play training between the path explorer and the setpoint checker is performed in each stage. The path explorer generates candidate fault or event evolution sequences, the setpoint checker evaluates the effectiveness of the protection setpoint under the candidate sequences, and alternately updates the parameters of the two models.

[0011] Once the adversarial self-play training converges, the trained path explorer outputs a typical evolutionary path representing vulnerability, and the trained value checker verifies the protection value.

[0012] The trained value checker and typical evolutionary paths representing vulnerabilities are used to perform online or offline value check and risk warning.

[0013] Furthermore, the construction of the neural network-based constant value checker and path finder includes:

[0014] The setpoint checker is constructed using a recurrent neural network or a long short-term memory network to process serialized event inputs and combine them with a multilayer perceptron to determine the validity probability of the protection setpoint.

[0015] The path explorer is constructed using a policy network or generative adversarial network based on reinforcement learning to generate the candidate fault or event evolution sequence from random noise and current power grid state information.

[0016] Furthermore, the multi-stage learning strategy, which progresses from easy to difficult, includes:

[0017] Multiple course difficulty levels are defined based on the length of the evolution path, the number of event types, and the complexity of event combinations;

[0018] Based on the path explorer's performance metrics at the current difficulty level, plan the advancement conditions from the current difficulty level to the next difficulty level.

[0019] Furthermore, the adversarial self-play training between the execution path explorer and the fixed value checker includes:

[0020] The path explorer generates candidate fault or event evolution sequences with the goal of generating evolution sequences that enable the setpoint checker to output a low probability of validity, while minimizing the perturbation of the evolution sequences.

[0021] The setpoint checker evaluates the effectiveness of the candidate evolutionary sequences generated by the path explorer, the known safe evolutionary sequences, and the known risky evolutionary sequences, with the goal of accurately distinguishing between safe evolutionary sequences, risky evolutionary sequences, and the evolutionary sequences generated by the path explorer.

[0022] Furthermore, the loss function L of the path explorer GThis includes a term for evaluating the validity of the generated sequence based on the setpoint checker and a term based on the perturbation amount of the generated sequence:

[0023]

[0024] Where L G The loss function represents the pathfinder's loss function. This indicates that the random noise Z follows a distribution p. z The expectation of (z); f(D(G(z))) represents a function that increases as D(G(z)) decreases, where D(G(z)) represents the output of the setpoint checker's evaluation of the generated path's effectiveness; G(z) represents the evolutionary path generated by the path explorer based on the noise z; p z (z) represents the prior distribution of random noise; λ G The weight hyperparameters represent the terms of the balanced loss function. The perturbation amount of the generated path is represented; f(D(G(z))) is a function that increases as D(G(z)) decreases, and is the output of the constant value checker's evaluation of the validity of the path generated by the generator; p z (z) is the prior distribution of random noise, λ G It is an adjustable hyperparameter used to balance the goal of disabling the set value and the goal of keeping the path disturbance small.

[0025] Furthermore, the loss function L of the setpoint checker D This includes items for evaluating known safe evolutionary sequences, items for evaluating known risk evolutionary sequences, and items for evaluating evolutionary sequences generated by the path explorer:

[0026]

[0027] Where L D Represents the loss function of the fixed-value checker; Indicates the safe path S safe Follows distribution p data (S safe ) expectations; Indicates the risk path S risk Follows distribution p data (S risk ) expectations; This indicates that the noise z follows a distribution p. z The expectation of (z); D(S) safe D(S) represents the output of the fixed-value checker's evaluation of the effectiveness of a known safe path; risk) represents the output of the fixed-value checker's effectiveness evaluation of a known risk path; D(G(z)) represents the output of the fixed-value checker's effectiveness evaluation of a generated path; log(·) represents the natural logarithm operation; G(z) represents the path generated by the path explorer; p data (S safe ) and p data (S risk These represent the data distributions of the real safe path and the real risk path, respectively.

[0028] Furthermore, the step of outputting a typical evolutionary path representing vulnerability using the trained path explorer after the adversarial self-play training has converged includes:

[0029] The path explorer, after multiple training iterations, generates a candidate set of vulnerability evolution paths;

[0030] The evolutionary paths in the candidate set are sorted, filtered, and clustered to obtain the typical evolutionary paths representing the vulnerability.

[0031] Furthermore, the step of verifying the protection setting using the trained setting verifier includes:

[0032] The protection setting to be verified, the current power grid status, and the potential fault or event sequence are taken as inputs and provided to the set value verifier after training.

[0033] Based on the output of the setting verifier, the effectiveness of the protection setting under the current power grid state and potential fault or event sequences is evaluated.

[0034] Furthermore, the steps of performing online or offline value verification and risk warning on the trained model and the typical evolutionary path representing vulnerability output include:

[0035] The matching degree between the output typical evolution path representing vulnerability and the real-time acquired power grid operation data is calculated.

[0036] When the matching degree exceeds the preset warning threshold, a risk warning signal is generated. The risk warning signal includes the potential vulnerability evolution direction, the estimated risk level, and corresponding prevention measures suggestions.

[0037] This invention provides an intelligent verification system for multi-fault mode relay protection settings, used to execute the aforementioned intelligent verification method for multi-fault mode relay protection settings, comprising:

[0038] The model building module is used to build and initialize the neural network model, which includes a setpoint validator for verifying the validity of setpoints and a path explorer for exploring fault evolution paths.

[0039] The course strategy module is used to design progressive course learning strategies for path explorers, defining different difficulty levels based on the complexity of the evolutionary path;

[0040] The adversarial training module is used to realize the interactive learning process between the path explorer and the setpoint verifier at different difficulty levels, enabling the path explorer to generate challenging fault sequences and the setpoint verifier to accurately evaluate the validity of the setpoints.

[0041] The results output module is used to extract key evolutionary paths that characterize the vulnerability of the system from the path explorers that have converged during training, and to establish a mature value verification capability.

[0042] The risk application module is used to apply the output vulnerability evolution path and the trained model to the actual power grid to achieve intelligent verification of protection settings and early warning of security risks.

[0043] The beneficial effects of this invention are as follows: by constructing a path explorer and a setting verifier, and by adopting an adversarial self-play training strategy that combines course learning, it is able to proactively and intelligently discover the vulnerability evolution paths that lead to relay protection setting failure or miscoordination in complex power grid environments.

[0044] It can efficiently identify potential risks and provide precise guidance for preventive control and setpoint optimization of the power grid, thereby enhancing the overall safety and reliability of power grid operation. Attached Figure Description

[0045] Figure 1 This is a flowchart of an intelligent verification method for multi-fault mode relay protection settings in this invention;

[0046] Figure 2 This is a flowchart of step 1 in this invention;

[0047] Figure 3 This is a flowchart of step 2 in this invention;

[0048] Figure 4 This is a flowchart of step 3 in this invention;

[0049] Figure 5 This is a flowchart of step 4 in this invention;

[0050] Figure 6 This is a flowchart of step 5 in this invention. Detailed Implementation

[0051] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0052] At least one embodiment of the present invention discloses an intelligent verification method for multi-fault mode relay protection settings, such as... Figures 1 to 6 As shown, it includes:

[0053] Step 1: Construct a neural network-based value checker and path explorer;

[0054] This step mainly involves initializing a neural network-based constant verifier (discriminator D) and a pathfinder (generator G). These two models are the core innovative components of this application.

[0055] Step 1.1, construct a fixed value checker;

[0056] The input data types for the setting checker D include: a feature vector describing the power grid state (e.g., containing power flow, bus voltage, generator output, topology connections, etc. on critical lines), a set of protection setting parameters, and a given sequence of fault or event evolutions.

[0057]

[0058] Where S represents the set of fault or event evolution sequences, e1, e2, ... These represent the 1st, 2nd, and k1th faults or events, respectively, where k1 is the total number of events in the evolutionary sequence.

[0059] Constructing a fixed-value checker D: This checker is a deep neural network model. In some implementations, it can be a structure that includes an embedding layer, a sequence processing layer, and an output layer.

[0060] Optionally, the embedding layer: incorporates the input power grid state characteristics, protection setting parameters, and each event in the event sequence. Each vector is mapped to a low-dimensional dense vector using a learnable embedding matrix to facilitate subsequent neural network processing.

[0061] Optionally, the sequence processing layer employs a Long Short-Term Memory (LSTM) recurrent neural network to process the event embedding vector sequence, capturing the temporal dependencies and cumulative effects between events. The LSTM unit finely controls the flow of information in the time series and the maintenance of long-term memory through its internal input gate, forget gate, and output gate. Its state update can be summarized as follows:

[0062] f t =σ(W f ·(h t-1 x t )+b f );

[0063] i t =σ(W i ·(h t-1 x t )+b i );

[0064]

[0065] o t =σ(W o ·(h t-1 x t )+b o );

[0066] h t =ot⊙tanh(C t );

[0067] Where f t This represents the forget gate gating vector, which controls the cell state C at the previous time step. t-1 How much information should be retained up to the current time? t This represents the input gate gating vector, which controls the current input X. t The proportion of information written into the cell state; Represents the candidate cell state, the new information after transformation of the current input and the previous hidden state; C t This represents the current cell state, incorporating both forgetting and new input; o t h represents the output gate gating vector, which controls the proportion of the cell state output to the hidden state; t W represents the hidden state at the current moment, which is the output of the sequence processing layer. f W i W C W o b represents the weight matrices for the forget gate, input gate, candidate cell state, and output gate, respectively; f b i b C bo Let h represent the bias vectors for the forget gate, input gate, candidate cell state, and output gate, respectively; t-1 x t ) indicates the concatenation of the previous hidden state with the current input; h t-1 and C t-1 These represent the hidden state and cell state at the previous time step, respectively; σ represents the Sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents the Hadamard product, indicating element-wise multiplication.

[0068] With this structure, the model can effectively learn the dynamic characteristics of the event sequence. Optionally, the output layer concatenates the final hidden state (representing the encoded information of the entire event sequence) output by the sequence processing layer with the power grid state embedding vector and the fixed parameter embedding vector, or fuses them through an attention mechanism, and then inputs them into several fully connected layers (forming a multilayer perceptron) for nonlinear transformation and feature interaction, and finally outputs the final scalar value through a sigmoid activation function.

[0069] The specific output of the fixed value checker D is: a scalar value V. S ∈(0,1), representing the probability that the protection setting remains effective given the evolution sequence S, the current grid state, and the setpoint. For example, in a complex scenario simulating a high-resistance ground fault on a critical transmission line in a regional power grid with large-scale wind power integration, accompanied by a chain reaction of wind turbine disconnection leading to significant power fluctuations, this event sequence, the current grid operating parameters (such as power flow of each line and bus voltage), and the relevant protection setting are input into a trained setting verifier D. If the output V S If the value is lower than the preset safety threshold (e.g., 0.5), it indicates that the current protection setting may fail or not cooperate in this specific complex scenario.

[0070] Step 1.2, Construct the path explorer;

[0071] The input data types for the path explorer G include: a random noise vector Z sampled from a standard normal or uniform distribution, and optional, an embedding vector of current grid key state features (used to generate more targeted evolution paths relevant to the current state context).

[0072] Construct a path explorer G: This explorer is also a deep neural network model whose goal is to generate a potential fault or event evolution sequence S′ consisting of multiple events. In some implementations, a Transformer-based decoder structure or a recurrent neural network with an attention mechanism can be used as the core of the generated sequence.

[0073] For example, if an RNN-based sequence generation method is used, at each time step t′, the RNN unit receives the event e′ generated in the previous time step. t′-1 The embedding representation and the current hidden state h′ t′-1 As input, and combined with information from z, output the hidden state h′ at the current time step. t′ and the event e′ generated t′ The probability distribution of the event. The event type can be sampled from a predefined set of events, and the key parameters of the event (such as fault location, fault occurrence time, disturbance amplitude, etc.) can also be sampled from the parameter distribution output by the network or directly predicted.

[0074] For example, when generating a sequence of simulated power grid cascading faults, the path explorer G might first generate the first event e′1 (e.g., "A B-phase ground fault occurs on 220kV line L1 near bus A, with a fault resistance of 50 ohms") based on the input random noise z and the initial state information of the power grid; then, based on the information of z and e′1, it generates the second event e′2 (e.g., "0.1 seconds later, due to the malfunction of protection 1, the adjacent line L3 trips"); then, based on z, e′1, and e′2, it generates e′3 (e.g., "0.3 seconds later, due to the tripping of L3, power transfer occurs, and line L4 experiences an 80% overload"), and so on, until a sequence reaching a predetermined maximum length is generated or the model outputs a special termination marker.

[0075] The specific output of the pathfinder G is a potential sequence of failures or events:

[0076] S′=G(z);

[0077] Where S′ represents the fault or event evolution path generated by the path explorer G based on the input random noise vector z, and z represents the random noise vector.

[0078] For example, a sequence description might be generated: "When the system is in heavy load operation mode, an inter-turn short circuit first occurs on phase A of the high-voltage side of transformer T1. The protection fails to isolate it correctly. 0.8 seconds later, due to the continuation of the fault, the regional frequency drops to 49.2Hz, triggering the first stage of low-frequency load shedding of generator G2."

[0079] This sequence is intended to be a potentially high-risk evolutionary path that can expose weaknesses in existing protection setting configurations.

[0080] Step 2: Based on the fixed value verifier and the path explorer, set up a multi-stage course learning strategy from easy to difficult to guide the training process of the path explorer;

[0081] This step uses the path explorer G constructed in step 1 as input to further enhance its generation capabilities.

[0082] The purpose of this step is to progressively improve the pathfinder's ability to generate complex and practically meaningful vulnerability evolutionary paths through training tasks ranging from easy to difficult. The learning process here is a typical machine learning training strategy, and its specific implementation includes:

[0083] Step 2.1, define the course difficulty level;

[0084] Based on the maximum allowed length of the evolutionary path Set of allowed event types type (For example, from single fault types to combinations of multiple fault types, and to broader events including control system malfunctions or human error) and the allowable range of event parameters and the complexity C of their combinations. event Factors such as these predefine multiple difficulty levels:

[0085] Cur1, Cur2, ..., Cur M ;

[0086] Among them, Cur1, Cur2, Cur M These represent the 1st, 2nd, and Mth difficulty levels of the course, respectively, where M is the total number of difficulty levels.

[0087] For example, the initial level Cur1 might be set Only simple fault events such as metallic grounding are allowed, and the fault location, occurrence time, and other parameters are fixed or within a very small range. Subsequent levels gradually increase in severity. The value of the event parameter is gradually increased by introducing more complex event types (such as high-resistance grounding, system oscillation, and cascading failures caused by information system attacks), and the possible range and combination of event parameters are broadened or increased.

[0088] Step 2.2, plan the conditions for advancing the course;

[0089] Set at a difficulty level Jump to the next difficulty level The conditions. For example, when pathfinder G is at the current difficulty level. The generated values ​​can be judged as high-risk by the fixed-value checker D (e.g., D(G(z)) < θ). risk , where θ risk The proportion of paths that consistently achieve a predetermined success rate η in the most recent training batches, based on a pre-set risk assessment threshold (e.g., 0.3). succ(e.g., 70%), or when the total number of iterations of adversarial training at this difficulty level reaches a predetermined number, the system will automatically switch to the next more difficult course level.

[0090] At this point, the input data used for the judgment is: the path explorer G at the current difficulty level. A series of performance evaluation metrics during the training process, which may include, but are not limited to: the frequency of generating effective high-risk paths, the average length of generated paths, the event diversity of generated paths, and the average minimum perturbation of generated paths (measuring the "economics" of paths).

[0091] The specific output is: update and determine the course difficulty level parameter to be used in the next training phase, for example, increase it. The value in Set type Adding new, complex event types or adjusting the distribution range of event parameters can encourage exploration of more peripheral scenarios.

[0092] Step 3: According to the stage settings of the course learning strategy, perform adversarial self-play training between the path explorer and the setpoint checker in each stage. The path explorer generates candidate fault or event evolution sequences, the setpoint checker evaluates the effectiveness of the protection setpoint under the candidate sequences, and alternately updates the parameters of the two models.

[0093] This step is based on the model initialized in step 1 and the course difficulty level set in step 2, with the two being coordinated and optimized alternately.

[0094] This step is the core innovation of this application, which involves iteratively optimizing two models through an adversarial process.

[0095] Step 3.1: The path explorer generates candidate paths;

[0096] The input data type of the path explorer G is: random noise z (z is a value from the prior distribution p) z (z) Sampled noise vector and current course level Defined constraints (such as maximum path length) Allowed event type Set type (and its parameter range, event occurrence probability model, etc.).

[0097] Application path explorer G (whose internal neural network structure and parameters are continuously optimized during training).

[0098] The specific output is: a batch (N) of candidate vulnerability evolution paths:

[0099] S′1,S′2,...,S′ N ;

[0100] Where S′1, S′2, S′ N These represent the 1st, 2nd, and Nth candidate vulnerability evolution paths, respectively, where N is the total number of paths generated in this batch.

[0101] The training objective of the path finder G is to generate paths that minimize the validity probability output by the constant value checker D (i.e., D(S′) approaches 0, indicating that the constant value is highly likely to fail on that path). Simultaneously, to ensure the generated paths have a realistic probability and avoid generating meaningless extreme perturbations, a regularization term or objective can be introduced, namely, finding a path perturbation that minimizes the path perturbation. The path should be as small as possible. The path perturbation here can be defined, for example, as the sum of some measure of the deviation of each event in the path from the normal state, or as the sum of the costs of each human intervention event included in the path.

[0102] Therefore, the loss function L of the pathfinder G G Represented as:

[0103]

[0104] Where L G The loss function represents the pathfinder's loss function. This indicates that the random noise z follows a distribution p. z The expectation of (z); f(D(G(z))) represents a function that increases as D(G(z)) decreases, commonly represented as -log(D(G(z))) and (1-D(G(z))). 2 D(G(z)) represents the output of the setpoint checker's evaluation of the generated path's validity; G(z) represents the evolutionary path generated by the path explorer based on the noise z; p z (z) represents the prior distribution of random noise; λ G The weight hyperparameters represent the terms of the balanced loss function. The perturbation amount of the generated path is represented; f(D(G(z))) is a function that increases as D(G(z)) decreases, and is the output of the constant value checker's evaluation of the validity of the path generated by the generator; p z (z) is the prior distribution of random noise, λ G It is an adjustable hyperparameter used to balance the goal of disabling the set value and the goal of keeping the path disturbance small.

[0105] Minimize L using optimization algorithms such as gradient descent G To update the parameters of G.

[0106] Step 3.2: The setpoint checker evaluates the path and updates it;

[0107] The input data type of the setpoint checker D is: the candidate path set S′ generated by the path explorer G in the current iteration, and a pre-prepared set S containing known safe paths obtained through high-precision power grid simulation (e.g., using tools such as PSCAD and RTDS) or by combining historical accident data and expert experience. safe (its corresponding expected output D(S) safe The path approaches 1) and the set of known risk paths S. risk (its corresponding expected output D(S) risk (approaching 0).

[0108] The fixed value checker D is applied (its internal neural network structure and parameters are continuously optimized during training).

[0109] The specific output results are: effectiveness evaluation scores D(S′) and D(S′) for the above-mentioned paths. safe ), D(S risk ). And update the parameters of D accordingly.

[0110] The goal of the fixed-value checker D is to accurately distinguish between genuine safe paths (from the dataset), genuine risky paths, and risky paths generated by the path explorer G with the intention of "deception." Its loss function L... D In some implementations, standard discriminator loss forms can be used, such as binary cross-entropy loss:

[0111]

[0112] Where L D Represents the loss function of the fixed-value checker; Indicates the safe path S safe Follows distribution p data (S safe ) expectations; Indicates the risk path S risk Follows distribution p data (S risk ) expectations; This indicates that the noise Z follows a distribution p. z The expectation of (z); D(S) safe D(S) represents the output of the fixed-value checker's evaluation of the effectiveness of a known safe path; risk ) represents the output of the fixed-value checker's effectiveness evaluation of a known risk path; D(G(z)) represents the output of the fixed-value checker's effectiveness evaluation of a generated path; log(·) represents the natural logarithm operation; G(z) represents the path generated by the path explorer; p data (S safe ) and p data (S riskThese represent the data distributions of the real safe path and the real risk path, respectively.

[0113] The parameters of D are updated by minimizing LD using optimization algorithms such as gradient descent, enabling a more accurate assessment of the risk level of various evolutionary paths.

[0114] Step 3.3, alternating training;

[0115] At the current course difficulty level, repeat steps 3.1 (training the pathfinder G) and 3.2 (training the constant value checker D) several times, or until the local convergence condition for this level is met. For example, within a training cycle, the parameters of the constant value checker D can be fixed first, and then the parameters of the pathfinder G can be k-valued. G This is the first iteration update; afterwards, the parameters of the fixed path explorer G are adjusted, and then the parameters of the fixed-value checker D are adjusted using k-step methods. D This process continues until the overall convergence condition is met at the current course difficulty level (e.g., the loss functions of both models change less than a preset small threshold within an evaluation window, or the maximum number of training iterations preset for this course level has been reached).

[0116] Step 4: After the adversarial self-play training converges, the trained path explorer outputs the typical evolutionary path representing vulnerability, and the trained value checker verifies the protection value.

[0117] This step is performed after the adversarial training in step 3 has fully converged, and depends on the optimal state of the aforementioned model.

[0118] Once all course learning phases have been completed, or the overall adversarial training process has reached the preset global convergence criterion (for example, after continuous training at the highest difficulty level for a period of time, the quality indicators of the generated vulnerability evolution path set (such as average risk score, path diversity index) tend to stabilize, and the discrimination accuracy of the setpoint checker D on the independent test set no longer shows a significant improvement), it can be considered that both models have been sufficiently trained, and the entire training process is over.

[0119] Step 4.1, extract vulnerability evolution paths;

[0120] The trained and best-performing pathfinder G* (i.e., the generator model whose best state is saved during training) is used as input.

[0121] Application Path Explorer G*.

[0122] The specific output is as follows: By driving G* multiple times (e.g., inputting a large number of random noise vectors Z following a prior distribution into G*), a large-scale, rich, and diverse set of vulnerability evolution path candidates is generated. Subsequently, optionally, these generated paths can be post-processed, for example, by sorting their risk assessment values ​​D*(S′) according to the final setter D*, and combining this with the path perturbation. By screening or using clustering algorithms to merge similar paths, a set of representative, typical evolutionary paths that pose the most serious potential threat to the currently assessed protection setting is obtained. At the same time, record the specific event sequences, event parameters, and possible triggering conditions or related initial power grid disturbance characteristics corresponding to these paths.

[0123] Step 4.2: Obtain the final value verifier;

[0124] The optimal value checker D* (i.e., the discriminator model with the best state saved during training) that has been trained and has the best performance is used as input.

[0125] Apply the fixed value checker D*.

[0126] The specific output is a high-performance verification model D*. Because it has undergone long-term adversarial training with G*, this model has the ability to accurately identify and quantify the risk assessment of the effectiveness of protection settings under various complex power grid boundary conditions and multiple fault evolution modes.

[0127] Step 5: Apply the trained value checker and typical evolutionary paths representing vulnerabilities to perform online or offline value check and risk warning.

[0128] This step takes the optimal model and set of vulnerable paths obtained in step 4 as input to realize practical applications.

[0129] Step 5.1, application of fixed value verification;

[0130] The input data types are: a specific set of protection setting parameters that need to be verified in actual power grid operation, and real-time operating status data of the current power grid obtained through power grid online monitoring systems (such as WAMS, PMU data) (e.g., voltage amplitude and phase angle of each major bus, power flow and current values ​​of key transmission lines, etc.), or power grid status and possible potential fault / event sequences under specific preset operating modes or typical fault scenarios simulated through offline simulation platforms (such as BPA, PSASP, etc.).

[0131] The final value validator D* is applied. After formatting the input data according to the input requirements of the D* model, it is input into the D* model for forward calculation.

[0132] The specific output is the probability value of the effectiveness assessment of the protection settings for this set of inputs in a specific scenario, output by model D*. Based on this probability value, it can be determined whether the settings have potential operational risks. For example, if the probability value output by the setting verifier D* is lower than a pre-set safety alarm threshold (such as 0.1), the system determines that the current setting configuration has a high operational risk in the current scenario.

[0133] Step 5.2, Risk Warning Analysis;

[0134] The input data type is: the set of typical high-risk vulnerability evolution paths output in step 4.1. In addition, real-time operating condition data streams of the current power grid and ultra-short-term load forecasts and renewable energy output forecasts for a future period (such as 15 minutes or 1 hour) are obtained from the Energy Management System (EMS) or Distribution Automation System (DAS).

[0135] The application utilizes intelligent analysis algorithms based on advanced rule matching, Complex Event Processing (CEP) engines, or pattern recognition and similarity calculation. It combines current real-time operational data and predicted data with vulnerability evolution paths. The triggering conditions and early preceding event characteristics of each path are continuously compared and dynamically evaluated.

[0136] The specific output is as follows: Real-time identification of whether the current power grid operating conditions and short-term predicted operating conditions are highly similar to and match the initial stage and key triggering conditions of one or more known high-risk vulnerability evolution paths. If the calculated matching degree exceeds a preset early warning activation threshold, the system automatically generates and issues a risk warning signal. This signal may optionally include the potential vulnerability evolution direction, the estimated risk level, and the relevant path number, providing timely decision support information for power grid dispatching and operation personnel. For example, it may suggest that they take targeted preventative control measures (such as adjusting generator output or changing the power grid topology) or re-verify and adjust relevant protection settings.

[0137] A multi-fault mode relay protection setting intelligent verification system is used to execute the above-mentioned multi-fault mode relay protection setting intelligent verification method, including:

[0138] The model building module is used to build and initialize the neural network model, which includes a setpoint validator for verifying the validity of setpoints and a path explorer for exploring fault evolution paths.

[0139] The course strategy module is used to design progressive course learning strategies for path explorers, defining different difficulty levels based on the complexity of the evolutionary path;

[0140] The adversarial training module is used to realize the interactive learning process between the path explorer and the setpoint verifier at different difficulty levels, enabling the path explorer to generate challenging fault sequences and the setpoint verifier to accurately evaluate the validity of the setpoints.

[0141] The results output module is used to extract key evolutionary paths that characterize the vulnerability of the system from the path explorers that have converged during training, and to establish a mature value verification capability.

[0142] The risk application module is used to apply the output vulnerability evolution path and the trained model to the actual power grid to achieve intelligent verification of protection settings and early warning of security risks.

[0143] Here, the present invention provides an implementation example:

[0144] Consider a regional power grid with a large-scale photovoltaic (PV) power plant connected to the main grid via multiple 220kV lines. PV power output is fluctuating and intermittent, making the grid operation complex and potentially subject to various fault modes, such as PV inverter failure, single-phase or two-phase short circuits, and system oscillations. Intelligent verification of the distance protection settings for critical lines in this regional power grid is necessary to ensure correct operation of the protection systems under various complex fault modes.

[0145] Constructing the setting verification device D: An LSTM network is used to process event sequences, combined with an MLP to output the setting validity probability. Inputs include: the real-time topology of the regional power grid (obtained via a SCADA system), line parameters, and current protection settings (e.g., the setting for distance protection segment I of line L23 is set to impedance Z). set1 =1.5+j4.0Ω, action time is 0.05s and an event evolution sequence to be evaluated.

[0146] Constructing the path explorer G: Employing a Transformer-based sequence generation model.

[0147] Data Preparation: Historical power grid operation data, fault waveform data, and various preset fault simulation data generated using PSCAD are collected for model training and validation. For example, in scenario 1: A metallic three-phase short circuit occurs at the end of line L23. The current setpoint can operate correctly, marking it as a safe path, with a desired D output close to 1. In scenario 2: A two-phase short circuit occurs in the middle section of line L23 through a 50Ω resistor. Simultaneously, photovoltaic power station A experiences frequency control anomalies, resulting in a significant 20% reduction in output. The current setpoint may fail to operate, marking it as a risk path, with a desired D output close to 0.

[0148] Initial Course (Level 1): Path Explorer G generates only a single, simple fault event of length 1 (such as a metallic single-phase ground fault), with the fault location fixed at the beginning of the line.

[0149] Intermediate course (Level 3): Allows path lengths of 2-3, containing two different types of fault or disturbance combinations (such as line fault + photovoltaic power output disturbance), with fault location and disturbance amplitude varying within a certain range.

[0150] Advanced Course (Level 5): Allows path lengths of 4-5, includes a variety of complex event combinations (such as cascading failures, protection malfunctions, and system oscillations), and further expands the range of event parameters.

[0151] Course advancement conditions: When the proportion of paths generated by Path Explorer G at the current level that elicit a validity assessment of less than 0.2 from the Fixed Value Verifier D reaches more than 60%, the path explorer will advance to the next level.

[0152] According to the current strategy, the path explorer G generates a candidate event evolution sequence S′: {Event 1: A phase high resistance grounding occurs at the midpoint of line L23, with a fault resistance of 30Ω; Event 2: 0.2 seconds later, the output of photovoltaic power station B drops rapidly from 80MW to 30MW due to cloud cover}.

[0153] The fixed value checker D receives S′ and the security path S extracted from the dataset. safe and risk path S risk For example, the output of D(S′) is 0.3, and D(S safe The output is 0.9, D(S) risk The output is 0.1.

[0154] Based on the loss function L G and L D Update the parameters of G and D. For example, since the value of D(S) is 0.3, it indicates that the path generated by G is somewhat threatening but does not completely "fool" D. G will adjust its parameters to generate a more threatening path (e.g., increase the resistance value of the high-resistance grounding or accelerate the rate of decline of photovoltaic output).

[0155] The parameters will then be adjusted based on the evaluation results of all input paths to better differentiate the risks of different paths. This process is performed iteratively within level 3.

[0156] After training at all course levels, the final path explorer G* is used to generate a large number of vulnerability evolution paths. For example, a typical high-risk path output is: "Scenario: System operating under heavy load, power flow on line L23 is approaching stability limits. Path: {Event 1: Line L21 (adjacent to L23) experiences a transient single-phase ground fault, successfully reclosing after 0.1s; Event 2: At the moment of reclosing of L21, due to system disturbance, the protection of line L23 misjudges, and the output signal is abnormally blocked; Event 3: 0.05 seconds later, a permanent two-phase short-circuit fault actually occurs on line L23}". This path indicates a complex scenario where protection logic anomalies caused by disturbances on adjacent lines are superimposed with setting failure (failure to operate) due to subsequent real faults.

[0157] The final value checker D* has the ability to accurately evaluate the validity of the set value L23 in the above complex scenarios.

[0158] Online verification: The dispatcher intends to modify the distance I segment setting of line L23 to Z′. set1 = 1.2 + j3.5Ω, with an action time of 0.03s. Input the new setpoint, current real-time grid operation data (e.g., line L23 currently has a load factor of 70%, and total photovoltaic output accounts for 50% of the regional load), and an event sequence generated through simulation of "high impedance grounding near the end of line L23" into D*. If the validity probability of the D* output is 0.85, the new setpoint is considered acceptable in this scenario.

[0159] Risk Warning: The system detects that the current power grid operation mode is highly similar to the initial conditions of a typical high-risk path (such as the aforementioned cascading failure scenario) output in step 4 (similarity exceeding 90%). The system automatically issues a warning, prompting the dispatcher to pay attention to the protection status of lines L21 and L23, and suggests reviewing the protection settings of L23 or taking temporary safety measures.

[0160] This implementation method can improve the accuracy of relay protection setting verification and risk warning capabilities under multiple fault modes.

[0161] Improving the accuracy of setpoint verification under complex and unknown fault modes: Compared with traditional fixed-rule-based verification methods, in a test set containing 100 complex and edge faults (50 of which are known to cause setpoint failure and 50 of which should operate correctly), the traditional method correctly identified 35 failure scenarios and 40 correct operation scenarios, achieving an accuracy of 75%. Our method, utilizing a pre-trained setpoint verifier D*, correctly identified 48 failure scenarios and 47 correct operation scenarios, achieving an accuracy of 95%.

[0162] Proactively identifying potential vulnerability evolution paths: During a week-long simulation of a regional power grid, traditional methods failed to effectively predict any potential setpoint failure risks caused by multi-event evolution. This application's method proactively generates and analyzes over 1000 potential vulnerability evolution paths using a path explorer G*, and, combined with a setpoint verifier D*, successfully identifies three high-risk, non-intuitive vulnerability evolution paths that were previously unnoticed by Pahee, providing crucial evidence for risk prevention.

[0163] The results of the accuracy comparison of the fixed value verification are shown in Table 1:

[0164] Table 1: Comparison of the accuracy of fixed value verification

[0165]

[0166] The results of vulnerability evolution path mining capabilities are shown in Table 2:

[0167] Table 2: Vulnerability Evolution Path Discovery Capability

[0168] Evaluation indicators Traditional methods This application method Number of high-risk evolutionary paths proactively identified within a week 0 items 3 Average event complexity (number of events) for path identification N / A 3.2

[0169] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent verification of settings in multi-fault mode relay protection, characterized in that, include: Construct a neural network-based value checker and path explorer; Based on the fixed value checker and the path explorer, a multi-stage course learning strategy from easy to difficult is set to guide the training process of the path explorer. According to the stage settings of the course learning strategy, adversarial self-play training between the path explorer and the setpoint checker is performed in each stage. The path explorer generates candidate fault or event evolution sequences, the setpoint checker evaluates the effectiveness of the protection setpoint under the candidate sequences, and alternately updates the parameters of the two models. Once the adversarial self-play training converges, the trained path explorer outputs a typical evolutionary path representing vulnerability, and the trained value checker verifies the protection value. The trained value checker and typical evolutionary paths representing vulnerabilities are used to perform online or offline value check and risk warning.

2. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The construction of the neural network-based constant value checker and path finder includes: A long short-term memory network is used to construct a setpoint checker to process serialized event inputs and combine it with a multilayer perceptron to determine the validity probability of the protection setpoint. A path explorer is constructed using a generator structure based on generative adversarial networks to generate candidate fault or event evolution sequences from random noise and current power grid state information.

3. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The multi-stage learning strategy, which progresses from easy to difficult, includes: Multiple course difficulty levels are defined based on the length of the evolution path, the number of event types, and the complexity of event combinations; Based on the path explorer's performance metrics at the current difficulty level, plan the advancement conditions from the current difficulty level to the next difficulty level.

4. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The adversarial self-play training between the execution path explorer and the fixed value checker includes: The path explorer generates candidate fault or event evolution sequences with the goal of generating evolution sequences that result in a low probability of validity for the setpoint checker, while minimizing the perturbation of the evolution sequences. The setpoint checker evaluates the effectiveness of candidate evolutionary sequences generated by the path explorer, known safe evolutionary sequences, and known risky evolutionary sequences. Its goal is to accurately distinguish between safe evolutionary sequences, risky evolutionary sequences, and evolutionary sequences generated by the path explorer.

5. The intelligent verification method for multi-fault mode relay protection settings according to claim 4, characterized in that, The loss function of the path explorer This includes terms that evaluate the validity of the generated sequence based on a fixed-value checker and terms that evaluate the perturbation of the generated sequence: ； in The loss function represents the pathfinder's loss function. Indicates random noise Follows distribution Expectations; Indicates following A function that increases as it decreases. This represents the output of the fixed-value checker's evaluation of the validity of the generated path; The pathfinder is based on noise The generated evolutionary path; Represents the prior distribution of random noise; The weight hyperparameters represent the terms of the balanced loss function. This represents the amount of perturbation in the generated path; It is a follow The function that decreases and increases is the output of the constant value checker's evaluation of the validity of the paths generated by the generator. It is the prior distribution of random noise. It is an adjustable hyperparameter used to balance the goal of disabling the set value and the goal of keeping the path disturbance small.

6. The intelligent verification method for multi-fault mode relay protection settings according to claim 4, characterized in that, The loss function of the fixed value checker This includes items that evaluate known safety evolution sequences, items that evaluate known risk evolution sequences, and items that evaluate evolution sequences generated by path explorers: ； in Represents the loss function of the fixed-value checker; Indicates safe path Follows distribution Expectations; Indicates the risk path Follows distribution Expectations; Indicates the noise Follows distribution Expectations; This represents the output of the fixed-value checker's effectiveness assessment of a known safe path; This represents the output of the fixed-value checker's effectiveness assessment of known risk paths; This represents the output of the fixed-value checker's evaluation of the validity of the generated path. Represents the natural logarithm operation; This represents the path generated by the path explorer; and These are the data distributions for the real safe path and the real risk path, respectively.

7. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The step of outputting a typical evolutionary path representing vulnerability using the trained path explorer after the adversarial self-play training has converged includes: The path explorer, which has been trained multiple times, generates a candidate set of vulnerability evolution paths. The evolutionary paths in the candidate set are sorted, filtered, and clustered to obtain typical evolutionary paths that characterize vulnerability.

8. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The step of verifying the protection setting using the trained setting verifier includes: The protection setting to be verified, the current power grid status, and the potential fault or event sequence are taken as inputs and provided to the training-completed setting verification unit. Based on the output of the setting verifier, assess the effectiveness of the protection settings under the current power grid condition and potential fault or event sequences.

9. The intelligent verification method for multi-fault mode relay protection settings according to claim 1, characterized in that, The steps for online or offline value verification and risk warning based on the applied trained value verifier and typical evolutionary path representing vulnerability include: The matching degree between the output typical evolution path representing vulnerability and the real-time acquired power grid operation data is calculated. When the matching degree exceeds the preset warning threshold, a risk warning signal is generated. The risk warning signal includes the potential vulnerability evolution direction, the estimated risk level, and corresponding prevention measures suggestions.

10. A multi-fault mode relay protection setting intelligent verification system, characterized in that, A method for intelligent verification of multi-fault mode relay protection settings as described in any one of claims 1-9 includes: The model building module is used to build and initialize the neural network model, which includes a setpoint validator for verifying the validity of setpoints and a path explorer for exploring fault evolution paths. The course strategy module is used to design progressive course learning strategies for path explorers, defining different difficulty levels based on the complexity of the evolutionary path; The adversarial training module is used to realize the interactive learning process between the path explorer and the setpoint verifier at different difficulty levels, enabling the path explorer to generate challenging fault sequences and the setpoint verifier to accurately evaluate the validity of the setpoints. The results output module is used to extract key evolutionary paths that characterize the vulnerability of the system from the path explorers that have converged during training, and to establish a mature value verification capability. The risk application module is used to apply the output vulnerability evolution path and the trained model to the actual power grid to achieve intelligent verification of protection settings and early warning of security risks.

Images