A power distribution network fault diagnosis method based on an optimized double-sided neural fuzzy Petri net

Abstract

This invention relates to a distribution network fault diagnosis method based on an optimized bilateral fuzzy Petri net, comprising: S1, screening suspected faulty components and adjusting the initial fault confidence based on the type of the suspected component; S2, constructing a bilateral fuzzy Petri net; S3, optimizing the MLP using the reinforcement learning gray wolf method, and obtaining the weight matrix of the bilateral fuzzy Petri net based on the optimized MLP; S4, configuring the parameters of the associated links according to the weight matrix, and calculating the fault probability of the suspected component through forward inference; S5, if the calculated fault probability is greater than a preset threshold, the suspected component is determined to be faulty; otherwise, return to S1 to re-screen suspected components until the fault diagnosis is completed and the process ends. This invention achieves efficient, robust, and engineering-oriented optimization of the bilateral NFPN parameters, meeting the needs of power system dispatching and reliability assessment.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of power system reliability analysis, intelligent optimization, and fuzzy reasoning, and particularly to a method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets. Specifically, it is applied to the optimization of fuzzy Petri net parameters in scenarios where the power receiving end (e.g., user-side microgrids, park distribution systems, distributed energy access points) and the transmission end (e.g., regional main grids, high-voltage transmission lines, grid-connected tie lines) operate in coordination. By adaptively optimizing the weight parameters of the fuzzy Petri net in this scenario, the uncertainty problem caused by the state coupling between the power receiving end and the transmission end is solved, thereby achieving accurate identification and highly robust fault diagnosis of faulty components in the distribution network. Background Technology

[0002] Currently, the global new energy industry is entering a phase of rapid development, profoundly transforming the energy structure and operation mode of power systems. According to data from the International Energy Agency, the average annual growth rate of global grid-connected new energy capacity has exceeded 20% in recent years, with distributed new energy accounting for over 35%, leading to a significant increase in the coupling between the power receiving and transmission ends. For example, a sudden drop in photovoltaic output at the power receiving end may cause excessive power flow deviations at the transmission end, increasing the risk of line overload; conversely, transmission line faults can cause the voltage at the power receiving end to drop below 80% of its rated value, triggering outages of sensitive equipment on the user side. Against this backdrop, higher requirements are placed on the accuracy and robustness of cross-side collaborative fault diagnosis in power systems, and traditional parameter design methods relying on unilateral optimization and experience-based adjustments can no longer meet engineering needs.

[0003] Neurofuzzy Petri nets, as a modeling tool integrating fuzzy mathematics and Petri nets, can effectively characterize uncertain states and deterministic logic in power systems and have been widely used in the field of power system state assessment. However, existing parameter optimization methods suffer from the critical drawback of over-reliance on human experience, which severely limits their engineering value. Summary of the Invention

[0004] The purpose of this invention is to provide a method for fault diagnosis of power distribution networks based on optimized bilateral neural fuzzy Petri nets, providing a high-precision and robust condition assessment tool for the bilateral coordinated operation of power systems.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets includes: S1. Screen suspected faulty components and adjust the initial fault confidence level according to the type of suspected component; S2. Divide the corresponding component states into storage locations and corresponding logic transitions, establish the association link from storage location to transition, construct a two-sided fuzzy Petri net adapted to the state coupling from the power receiving end to the power transmitting end, and input the corrected initial fault confidence into the two-sided fuzzy Petri net as the initial state. S3. The reinforcement learning gray wolf method is used to optimize the MLP, and the weight matrix of the bilateral fuzzy Petri net is obtained based on the optimized MLP. S4. Configure the parameters of the associated link according to the weight matrix, and calculate the failure probability of the suspected component through forward inference; S5. If the calculated failure probability is greater than the preset threshold, the suspected component is determined to be faulty; otherwise, return to S1 to re-screen suspected components until the fault diagnosis is completed and the process ends.

[0006] Optionally, screening for suspected faulty components and adjusting the initial fault confidence based on the type of suspected component includes: The grid topology is traversed using the connection analysis method, and the suspected faulty components are screened by combining the action signals of the protection devices. Combining the timing logic of the protection-circuit breaker operation rules, the timing constraints of the protection device operation and the circuit breaker tripping are used to correct the initial fault confidence for different types of suspicious components, including transformers, lines, busbars, and power sources.

[0007] Optionally, establishing the association link from the place to the transition includes: obtaining the driving strength of the place to the transition by inputting a weight matrix, obtaining the update strength of the transition to the place by outputting a weight matrix, and establishing the association link.

[0008] Optionally, optimizing the MLP using the reinforcement learning gray wolf method includes: S31. Initialize the parameters of the reinforcement learning gray wolf method and construct a gray wolf population, where each individual represents a complete set of fully connected weights and parameter vectors of the multilayer perceptron used to generate the weight matrix of the two-sided fuzzy Petri net. S32. Calculate the fitness of individuals in the population, select elite individuals, perceive the current optimization state through reinforcement learning, dynamically adjust the parameters of the gray wolf optimization algorithm, and update the position of non-elite individuals in the population. When updating the position of non-elite individuals in the population, perturbation is added to several individuals with the worst fitness. The fitness is obtained based on the negative of the sum of the minimum value and the weighted average value in the diagnostic confidence set of all working conditions. S33. Calculate the internal reward function of reinforcement learning, train DQN and soft update the target network, calculate the fitness value of the individual after position update and return to S32, until the termination condition is met, obtain the MLP hyperparameters, and then obtain the optimized MLP.

[0009] Optionally, the parameters of the gray wolf optimization algorithm can be dynamically adjusted by: dynamically updating the exploration parameters according to a decreasing pattern, and randomly updating the convergence parameters within a preset range.

[0010] Optionally, obtaining the weight matrix of the two-sided fuzzy Petri net based on the optimized MLP includes: The mean of the power receiving end sample set is input into the optimized MLP, and the residual vector is output. The residuals are allocated according to the parameter mask of the two-sided fuzzy Petri net to obtain the power transmitting end input weight matrix, the power transmitting end output weight matrix, the power receiving end input weight matrix, and the power receiving end input weight matrix.

[0011] Optionally, calculating the failure probability of the suspected component through forward reasoning includes: By performing a Cartesian product combination on the power receiving end samples and the power transmitting end samples, any set of bilateral operating condition samples can be obtained. By calculating the confidence levels of the receiving end and the transmitting end for each sample group through forward inference, a robust objective function is constructed. The objective is minimized to maximize the confidence level. The robust objective function is constructed using the average confidence level of all operating conditions.

[0012] Optionally, the forward inference computation includes: Based on the transition activation threshold, the effective driving signal and noise interference are determined, and the failure probability of the suspected component is calculated using a Gaussian compression function on the effective driving signal.

[0013] The beneficial effects of this invention are as follows: This invention overcomes the limitations of traditional fuzzy Petri nets, such as one-sided adaptation, reliance on manual experience for adjustment, and insufficient robustness under extreme conditions. By optimizing the weights of the two-sided NFPN through a hybrid framework that integrates reinforcement learning and gray wolf optimization, and combining the two-sided topological coupling characteristics and multi-engineering constraints of the power system, it not only improves the fault diagnosis accuracy and robustness under all-cooperative operating conditions, but also realizes the automated optimization of weight parameters, significantly reducing manual costs and improving the practicality and reliability of two-sided collaborative fault diagnosis in the power system. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0015] Figure 1 This is a flowchart of a power distribution network fault diagnosis method based on an optimized bilateral neural fuzzy Petri net, according to an embodiment of the present invention. Figure 2This is a schematic diagram of the DS-NFPN structure of the combined diagnostic model according to an embodiment of the present invention; Figure 3 This is a flowchart of the RLGWO algorithm for the combined diagnostic model in an embodiment of the present invention; Figure 4 The following are the optimization results of the diagnostic confidence with each iteration in scenarios one and two of the present invention, where (a) is the iterative convergence graph of the diagnostic confidence in scenario one, and (b) is the fault in the iterative convergence graph of the diagnostic confidence in scenario two. Figure 5 The following are the optimization results of the diagnostic confidence with iteration rounds in scenarios three and four of the present invention, where (a) is the diagnostic confidence iteration convergence graph for scenario three and (b) is the diagnostic confidence iteration convergence graph for scenario four. Figure 6 The following are the optimization results of the diagnostic confidence as a function of iteration in scenarios five and six of the present invention, where (a) is the iterative convergence graph of the diagnostic confidence in scenario five, and (b) is the fault in the iterative convergence graph of the diagnostic confidence in scenario six. Figure 7 The following are the optimization results of the diagnostic confidence as a function of iteration in scenarios seven and eight of the present invention, where (a) is the iterative convergence graph of diagnostic confidence in scenario seven and (b) is the fault in the iterative convergence graph of diagnostic confidence in scenario eight. Figure 8 The following are the optimization results of the diagnostic confidence as a function of iteration in scenarios nine and ten of the present invention, where (a) is the iterative convergence graph of the diagnostic confidence in scenario nine and (b) is the fault in the iterative convergence graph of the diagnostic confidence in scenario ten. Figure 9 The following are the optimization results of the diagnostic confidence as a function of iteration in scenarios 11 and 12 of this invention, where (a) is the iterative convergence graph of the diagnostic confidence in scenario 11 and (b) is the fault in the iterative convergence graph of the diagnostic confidence in scenario 12. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1As shown, this embodiment proposes a distribution network fault diagnosis method based on an optimized bilateral neural fuzzy Petri net, including: S1. Screen suspected faulty components and adjust the initial fault confidence level according to the type of suspected component; S2. Divide the corresponding component states into storage locations and corresponding logic transitions, establish the association link from storage location to transition, construct a two-sided fuzzy Petri net adapted to the state coupling from the power receiving end to the power transmitting end, and input the corrected initial fault confidence into the two-sided fuzzy Petri net as the initial state. S3. The reinforcement learning gray wolf method is used to optimize the MLP. Based on the optimized MLP, the weight matrix of the dual-side fuzzy Petri Net (DS-NFPN) is obtained. The DS-NFPN parameter settings are shown in Table 1. S4. Configure parameters for the associated links based on the weight matrix, and calculate the failure probability of suspicious components through forward inference; S5. If the calculated failure probability is greater than the preset threshold, the suspected component is determined to be faulty; otherwise, return to S1 to re-screen suspected components until the fault diagnosis is completed and the process ends.

[0019] Table 1 Furthermore, the process involves screening suspected faulty components and refining the initial fault confidence based on the type of the suspected component, including: The grid topology is traversed using the connection analysis method, and the suspected faulty components are screened by combining the action signals of the protection devices. Combining the timing logic of the protection-circuit breaker operation rules, the timing constraints of the protection device operation and the circuit breaker tripping are used to correct the initial fault confidence for different types of suspicious components, including transformers, lines, busbars, and power sources.

[0020] Furthermore, establishing the association link from the repository to the change includes: obtaining the driving strength of the repository to the change by inputting the weight matrix, obtaining the update strength of the change to the repository by outputting the weight matrix, and establishing the association link.

[0021] Specifically, the topology of the two-sided hierarchical fuzzy Petri net satisfies: The DS-NFPN libraries at the power receiving and transmission ends are categorized by function: protection category corresponds to different levels of protection devices, execution category corresponds to execution equipment such as circuit breakers, monitoring category corresponds to fault and status monitoring nodes, and auxiliary category corresponds to correction nodes such as environment and time delay. The transitions are categorized by logical association: core action logic protection start-up, equipment opening and closing, and comprehensive judgment logic integrates multi-dimensional signals to correct action thresholds.

[0022] Furthermore, the optimization of MLP using the reinforcement learning gray wolf method includes: S31. Initialize the parameters of the reinforcement learning gray wolf method and construct a gray wolf population, where each individual represents a complete set of fully connected weights and parameter vectors of the multilayer perceptron used to generate the weight matrix of the two-sided fuzzy Petri net. S32. Calculate the fitness of individuals in the population, select elite individuals, perceive the current optimization state through reinforcement learning, dynamically adjust the parameters of the gray wolf optimization algorithm, and update the position of non-elite individuals in the population. When updating the position of non-elite individuals in the population, add perturbation to several individuals with the worst fitness. S33. Calculate the internal reward function of reinforcement learning, train DQN and soft update the target network, calculate the fitness value of the individual after position update and return to S32, until the termination condition is met, obtain the MLP hyperparameters, and then obtain the optimized MLP.

[0023] Furthermore, dynamically adjusting the parameters of the Grey Wolf optimization algorithm includes: dynamically updating the exploration parameters through a decreasing pattern, and randomly updating the convergence parameters within a preset range.

[0024] Specifically, RLGWO collaborative optimization satisfies: GWO parameters a The dynamic adjustment logic is as follows a Updated according to a decreasing pattern, the formula is: ; in, To explore the initial maximum value of the parameters, To explore the terminating minimum value of the parameter, For the maximum number of iterations, This represents the current iteration number. , It adapts to the optimization needs of global exploration in the early stage of iteration, covering more parameter space, and local refinement in the later stage of iteration, converging to the optimal region. GWO parameters C Random characteristics: C The random numbers are uniformly distributed within the range [0,2]. Their randomness helps prevent the population from prematurely converging to a local optimum: when C When the value is greater than 1, the population's attractiveness to elite individuals increases, accelerating convergence; when... C When the value is less than 1, the attractiveness decreases, thus maintaining population diversity; The complementarity of the four dimensions of the DQN state vector: population diversity Reflects whether the exploration is sufficient; relative degree of improvement Reflecting the current optimization effect; iteration progress n The early stages of the optimization phase require extensive exploration; the current optimal target value... It reflects the quality of optimization; the four dimensions together constitute a comprehensive description of the optimization state, and the absence of any one of them can easily lead to one-sided decision-making in DQN.

[0025] Furthermore, based on the optimized MLP, the weight matrix of the two-sided fuzzy Petri net is obtained as follows: The mean of the power receiving end sample set is input into the optimized MLP, and the residual vector is output. The residuals are allocated according to the parameter mask of the two-sided fuzzy Petri net to obtain the power transmitting end input weight matrix, the power transmitting end output weight matrix, the power receiving end input weight matrix, and the power receiving end input weight matrix.

[0026] Specifically, the MLP residual adjustment DS-NFPN parameters satisfy: The construction logic of the adjustable element mask: Elements with an initial weight of 0 correspond to library transition pairs with no logical relationship. Not assigning residuals can avoid adding meaningless logical links and ensure that the DS-NFPN topology is consistent with the actual operation logic of the power system. The advantages of using tanh activation in the output layer of an MLP are: it enables bidirectional adjustment of the residuals, which is more suitable for the full range of weight optimization needs compared to only positive adjustment. Experiments show that bidirectional adjustment can improve the convergence speed of the objective function.

[0027] Furthermore, to ensure the physical meaning of the weight parameters of the two-sided fuzzy Petri net, the weight matrix needs to be constrained. For the original weight vector output by the MLP, a group-based fixed-sum projection algorithm is used for normalization to ensure that the weights of the associated links are non-negative and their sum is constant. This projection satisfies the standard non-negative simplex projection. The specific calculation steps are as follows: ① v Sort in descending order to get ; ②Calculate the cumulative ; ③ Find the largest ,satisfy ; ④ Calculate the threshold ; ⑤ Output projection results ; The time complexity of this algorithm is O(n). The main source of this is the sorting step, which can quickly project high-dimensional vectors and rapidly map the residual correction values ​​of the MLP output back to the effective Petri net weight space.

[0028] Furthermore, calculating the failure probability of the suspected component through forward reasoning includes: By performing a Cartesian product combination on the power receiving end samples and the power transmitting end samples, any set of bilateral operating condition samples can be obtained. By calculating the confidence levels of the receiving end and the transmitting end for each sample group through forward inference, a robust objective function is constructed. The objective is minimized to maximize the confidence level. The robust objective function is constructed using the average confidence level of all operating conditions.

[0029] Furthermore, forward reasoning computation includes: Based on the transition activation threshold, the effective driving signal and noise interference are determined, and the failure probability of the suspected component is calculated using a Gaussian compression function on the effective driving signal.

[0030] Specifically, the forward inference computation employs a two-stage processing mechanism of "threshold gating-Gaussian quantization" to balance the diagnostic robustness with the confidence level's discriminative power. First, effectiveness screening is performed based on transition activation thresholds: transition activation thresholds The engineering basis referenced the design principles of the protection device action threshold in the technical specifications for relay protection and automatic safety devices. To effectively drive the boundary between signal and noise interference, when When the fault signal or overload signal detected by the protection device is deemed a valid drive, it is determined that the device is validly activated; when When noise is detected, it is identified as noise such as line capacitance current or electromagnetic interference, thus avoiding false activation of transitions caused by noise. The transition input strength is a numerical value that reflects the ability of the current fault characteristics to drive the transition.

[0031] Secondly, Gaussian nonlinear quantization is performed on the selected effective signals: only when When a valid fault-driven signal is detected, a transition ignition process is triggered (i.e., the calculation process of activating the logic inference node and passing the state update to the subsequent library), and a Gaussian compression function is used. Mapping the state. Compared to linear compression. Additional restrictions required The Gaussian function can amplify the confidence gain in the high state intensity region, which is more in line with the engineering understanding that high state intensity in power systems corresponds to high confidence. and These are the standardized fault confidence after mapping and the intermediate state value output after transition activation, respectively.

[0032] Example 1: A method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets includes the following steps: S1. Construct a two-sided fuzzy Petri net and constraint system, define the initial input weights and output weight matrices of the power receiving end DS-NFPN, clarify the correspondence between the place and the transition, and set the constraint conditions; S2. Initialize the RLGWO hybrid optimization framework, generate the initial GWO population, with each individual corresponding to the parameter vector of the MLP, construct the DQN network, initialize the policy network and the target network, and set hyperparameters such as the experience replay buffer, learning rate, and discount factor. S3. MLP Residual Adjustment of DS-NFPN Parameters: An unsupervised parameter generation strategy based on swarm intelligence evolution is employed, eliminating the need for traditional gradient descent training. The mean of the historical fault sample set at the receiving end is used as the feature vector input to the MLP, and forward propagation outputs a residual vector containing parameter adjustment information. The specific weight determination process is as follows: (1) Residual mapping: A parameter mask is constructed using the topology of a bilateral neural fuzzy Petri net. This mask only identifies the logical connections that exist in the topology, ensuring that the optimization process does not change the network structure. (2) Overlay Update: The residual vector output by the MLP is mapped to the corresponding connection positions according to the mask order, and then overlaid onto the preset initial empirical weight matrix. The formula satisfies ,in r This is the residual scaling factor. The residual value output by the MLP; (3) Constraint Correction: Numerical truncation (limited to the [0,1] interval) and in-group summation projection normalization are performed on the superimposed weight matrix to finally obtain the power-transmitting end input / output weight matrix and the power-receiving end input / output weight matrix that satisfy the physical constraints. In this process, the full set of combined fault samples is used as the evolutionary evaluation set, and the optimal MLP network parameters are searched iteratively through the RLGWO algorithm, thereby indirectly realizing the adaptive optimization of the DS-NFPN weights.

[0033] S4. Full Combination Evaluation and Robust Objective Calculation: A Cartesian product full combination is performed on the power receiving end samples and the transmission end samples to obtain arbitrary sets of bilateral operating condition samples. The confidence levels of the power receiving end and the transmission end for each set of samples are calculated using NFPN forward inference to construct a robust objective function. , The lowest diagnostic confidence level across all operating conditions represents the worst-case performance. The average confidence level across all operating conditions represents the overall performance, where The weighting coefficient is set to 0.1 in this embodiment. The purpose of this step is to employ a robust optimization strategy of "maximizing minimum confidence" to minimize... This forces the algorithm to prioritize improving diagnostic accuracy under extreme fault conditions, i.e., improving the weakest link, while taking into account the average diagnostic accuracy of the overall sample, thereby ensuring that the model has high robustness and reliability in various complex scenarios with dual-side coupling.

[0034] S5, RLGWO iterative optimization, evaluate the target value of each individual in the GWO population, and sort them. (Optimal) (Second best) For the third-best individual, the position of non-elite individuals in the population is updated based on the action mapping GWO parameters output by DQN. Subsequently, a tail-end dynamic perturbation is performed: the bottom 10% of individuals with the worst fitness are identified, and random resets are introduced based on the current iteration round. If it is a specific period node (e.g., a multiple of 10 rounds), a high-intensity perturbation is applied centered on the globally optimal individual; otherwise, small-amplitude fine-tuning or a global space reset is performed according to a preset probability to enhance the algorithm's ability to escape local optima. Finally, reinforcement learning rewards are calculated, experience is stored in the replay buffer, DQN is trained, and the target network is softly updated, outputting the optimized weight matrix.

[0035] Relevant model descriptions: (1) Bilateral neural fuzzy Petri net: Bilateral neural fuzzy Petri nets are a modeling tool that introduces hierarchical and fuzzy dual extensions to the classic Petri net, used to solve the challenges of representation and reasoning in large-scale, uncertain knowledge systems. Its basic idea is to divide the complex network into several hierarchical subnets according to function or topology, with the output of the upper layer serving as the input of the lower layer, forming a progressive structure from macro to micro. Simultaneously, within each subnet, the confidence level and transition threshold of the library are labeled with fuzzy numbers [0,1], and approximate reasoning is achieved through fuzzy production rules, thus balancing the model's scalability with its tolerance for uncertain information.

[0036] like Figure 2 As shown, three subnets—main protection, near backup, and far backup—are first established at both ends of the line. Then, a comprehensive diagnostic layer is constructed. Through confidence propagation and time-series cross-validation, the two layers can quickly locate faulty components and provide probabilistic assessments in the event of missing information or malfunctioning protection. The hierarchical strategy significantly compresses the dimension of the correlation matrix, mitigating state space explosion; the fuzzy mechanism allows the model to dynamically adjust weights based on evidence, making the inference results more robust.

[0037] The storage area corresponds to the power system state nodes, such as protection devices, fault monitoring nodes, and line monitoring nodes. The state vector is used... express( n (This represents the number of storage locations), with a value range of [0,1], and represents the confidence level of the status; The transitions correspond to logical transformation rules such as protection actions, overload judgment, and fault identification. The establishment of its interconnected links involves two dimensions: one is through the input weight matrix. I First, it characterizes the driving force of the library on the transition; second, it outputs the weight matrix. O This characterizes the intensity of the change's impact on the place after the transition is activated. The transition input intensity quantifies the driving force of the place state on the transition, and the formula is: ; In the formula, The current storage location state vector (power receiving end) Transmission end I is the input weight matrix (receiving end) Transmission end ), ( m (This represents the number of transitions), and a larger value indicates a stronger driving force.

[0038] Threshold filtering can filter out noise interference and retain only the valid drive signal. The formula is: ; In the formula, For the transition activation threshold, This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 otherwise. For element-wise multiplication, The effective driving strength after filtering.

[0039] The update calculation for the activation transition of the place state during the place state update is as follows: ; In the formula, O For the output weight matrix (power receiving end) Transmission end , for The transpose of the matrix, This is a temporary state vector.

[0040] Gaussian compression maps the temporary state vector to the confidence range of [0,1], as shown in the formula: ; In the formula, For the final state vector after compression, this function has nonlinear properties: when hour, The diagnosis is completely reliable; when hour, At this point, the diagnosis has only a very low confidence level; when hour, Follow The increase is consistent with the confidence level of power system condition assessment.

[0041] Forward inference employs a multi-layered iterative process, using transition activation, threshold filtering, state update, and Gaussian compression to derive the final confidence level from the initial state, ensuring the reliability of the evaluation.

[0042] (2) Multilayer perceptron: The Multi-Layer Perceptron (MLP) is the most classic feedforward neural network, consisting of an input layer, several hidden layers, and an output layer stacked together. Each layer contains several neurons, with full connectivity between layers and unidirectional information propagation: the input is weighted, summed, and biased, then passed to the next layer through non-linear activations (such as ReLU and Sigmoid), ultimately outputting the prediction result. Training uses the backpropagation algorithm: first, the loss is calculated forward, then the gradient is calculated backward along the chain rule, and the weights are updated using stochastic gradient descent or its variants to minimize the loss function. MLPs possess general approximation capabilities; theoretically, as long as the hidden layers are wide enough, they can fit any continuous function, thus being widely used in classification, regression, and feature extraction. Its advantages lie in its simple structure and ease of implementation. The MLP is a fully connected feedforward neural network, with its core used for residual adjustment of DS-NFPN weights. By learning the mapping relationship between the power receiving end's operating conditions and the weight correction amount, it achieves smooth and controllable updates of the DS-NFPN weights, avoiding logical conflicts caused by direct updates.

[0043] The internal calculations of the MLP model are shown in the following formula: ; ; ; ; This represents the mean of the power receiving end operating condition characteristic samples. That is, the number of power receiving terminals. For the first i layer to the first i The weight matrix of layer +1 For the i-th layer to the i-th layer i +1 layer bias vector This is the result of a linear transformation; This is the residual vector after tanh activation.

[0044] (3) Gray Wolf Optimization Algorithm: The Grey Wolf Optimization Algorithm (GWO), with its advantages of simple structure, strong global search capability, and easy parameter adjustment, becomes the core tool for optimal parameter search in MLP in this invention. In the DS-FPN weight optimization scenario, the parameter space of MLP has a high dimensionality, and traditional optimization algorithms such as PSO optimization are prone to getting trapped in local optima, leading to stagnation in weight optimization. GWO guides the population to search efficiently in a high-dimensional parameter space by simulating the leadership mechanism of α (optimal), β (second-best), and δ (third-best) wolves and the following behavior of ω (ordinary) wolves: elite individuals provide the optimal search direction, and ordinary individuals update their positions by following elites. Simultaneously, a random coefficient and a dynamic exploration parameter 'a' are introduced to balance global exploration and local refinement.

[0045] The internal iteration and optimization formula of RLGWO is shown below: ; ; ; ; X The random numbers are uniformly distributed in the interval [0,1]. Numerical randomness determines the diversity of search directions; These are the parameter vectors of the elite individuals in the population with the best, second-best, and third-best objective function values, respectively. For the current individual ( The parameter vector of the wolf; The parameter distance between the current individual and the elite individual; This represents the current individual's new position.

[0046] Furthermore, the newly added tail population perturbation mechanism in this invention further enhances the algorithm's global search capability, preventing the population from stagnating in local optima. GWO does not rely on the gradient information of the objective function, adapting to the black-box nature of MLP parameter optimization, and can quickly find the optimal combination of MLP parameters for the robust objective function, as shown in the following formula: ; The worst-case individual parameter vector, after perturbation correction, is the final parameter used for the next iteration. This represents the globally optimal individual parameter vector. k The disturbance coefficient; X It is a function of standard normal distribution random numbers.

[0047] (4) DQN decision: In reinforcement learning, an agent needs to learn a policy to maximize cumulative rewards in the long run. Q-learning is a value function-based approach that guides decision-making by learning the value of each action pair. However, in real-world problems with high-dimensional or continuous state spaces, the traditional Q-learning table-based storage and updating of Q-values ​​becomes infeasible. DQN addresses this problem by leveraging the powerful fitting capabilities of deep neural networks to map states to the Q-values ​​of all actions.

[0048] Traditional single GWO algorithms, relying on fixed exploration and convergence parameters, cannot adapt to the entire iterative process; while single reinforcement learning lacks an efficient global search mechanism, making it difficult to quickly locate the optimal parameter region. This hybrid framework achieves a collaborative closed loop between DQN intelligent decision-making and GWO efficient search: DQN is responsible for real-time perception of the optimization state and dynamically outputting GWO core parameters adapted to the current stage; GWO, guided by DQN, uses elite individuals to guide population iteration, combined with a population perturbation mechanism to break local optima. The specific implementation process is as follows: Figure 3 As shown.

[0049] DQN state vectors contain four dimensions: population diversity Reflects whether the exploration is sufficient and the degree of relative improvement. Reflecting the current optimization results and iteration progress n The optimization phase requires extensive exploration in its early stages; the current optimal target value is... Reflecting optimization quality; the internal reward function of DQN is shown below: ; ; in, The relative improvement reward weight amplifies the positive incentive for target improvement; Population diversity reward weights encourage populations to maintain their exploratory capabilities; Non-degradable reward weighting, triggered when the target is improved. This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise. The penalty weight deteriorates and is triggered when the target deteriorates. The target Q-value and loss function within this architecture are shown below: ; ; in, Discount factors, emphasizing long-term optimization rewards; The next state output by the target network s' of Q value; Maximum expected reward for the next state; Batch training size; Current action output by the policy network a of Q value; No. i The target of each sample Q value.

[0050] The fault diagnosis process for the RLGWO-NFPN combined model includes: The fault diagnosis flowchart of the RLGWO-NFPN combined model is as follows: Figure 1 As shown, the process is divided into three core stages: initial screening of faulty components, initial confidence correction, and hierarchical neurofuzzy comprehensive diagnosis. The specific steps are as follows: 1. Initial screening and type determination of suspected faulty components: 1) Step 1: Initial screening of suspected faulty components. The connection analysis method is used to traverse the power grid topology and combine the action signals of the protection device to screen out the set of components that may be faulty, thus completing the initial narrowing of the fault range.

[0051] 2) Step Two: Fault Component Type Determination. For the suspicious components obtained from the initial screening, their types are determined to be transformers, lines, busbars, and power supplies. This is the logical rule corresponding to the subsequent confidence correction and NFPN inference matching.

[0052] 2. Initial confidence level correction: By combining the timing logic of the protection-circuit breaker action rules and the time constraints of the protection device action and the circuit breaker tripping, the initial fault confidence of different types of components is modified and assigned to the corresponding input layer library in the bilateral neural fuzzy Petri net as the initial state, thereby providing a basic input vector that conforms to the actual operation logic for subsequent NFPN forward inference.

[0053] 3. Layered neurological fuzzy syndrome diagnosis: 1) Step 1: Static data integration of three types of static data: protection equipment configuration information, power grid topology, and protection circuit breaker operation rules, which serve as the basis for the topology framework and logical association of NFPN.

[0054] 2) Step Two: DS-NFPN Topology Construction. Based on static data and component types, construct a DS-NFPN topology that adapts to the state coupling from the power receiving end to the power transmitting end: divide the corresponding component state into storage locations, the corresponding logic transitions, and establish the initial association link from storage locations to transitions.

[0055] 3) Step 3: RLGWO optimizes the DS-NFPN weight matrix. The MLP parameters are optimized through the RLGWO hybrid framework to determine the DS-NFPN weight matrix. The MLP network parameter settings are shown in Table 3. ① Initialize the RLGWO parameters. The RLGWO optimization parameter settings are shown in Table 2. Construct a gray wolf population, where each individual represents a complete set of fully connected weights and parameter vectors of the multilayer perceptron used to generate the bilateral fuzzy Petri net weight matrix. Calculate the fitness of the population individuals. Specifically, decode the individual parameters into a weight matrix and substitute it into the bilateral neural fuzzy Petri net (DS-NFPN) model. Perform forward inference on the bilateral full-combination fault sample set to obtain the diagnostic confidence set for all working conditions. To adapt to the characteristic of finding the minimum value in the gray wolf optimization algorithm, construct a fitness function that is the negative of the sum of the minimum value and the weighted average value in this set. Select the top three individuals with the best fitness as elite individuals. α, β, δ ; ② Reinforcement learning (RL) perceives the current optimization state and dynamically adjusts the exploration parameters of the gray wolf optimization. a Convergence parameters C The combination of the GWO exploration parameter a and the convergence parameter C is shown in Table 4. ③ Iteratively update the population position, and finally output the DS-NFPN weight matrix from the optimized MLP to complete the parameter configuration of the topology link.

[0056] Table 2 Table 3 Table 4 4) Step Four: Based on the NFPN forward inference process, ignite via Gaussian function transition. Calculate the failure probability of the suspected component, where, The temporary state vector of the place is obtained by updating the output weight matrix after the transition activation.

[0057] 5) Step five: If the calculated failure probability is greater than the preset threshold, the component is determined to be faulty; otherwise, suspicious faulty components are re-screened and the above process is repeated until the fault diagnosis is completed and the process ends.

[0058] For the example of this invention, a simulation experiment was conducted with a fault occurring on line L1009 of the IEEE 14-node system. The received fault signals were (1009m, 5), (CB1009, 35), and (0910m, 10), all in milliseconds. The calculation process is as follows: 5.1) First, reverse temporal reasoning is performed to obtain the transition time constraint of L1009: ; ; ; ; The time backwards from the occurrence of the fault event is the intersection of the above times, indicating that the fault occurred... During this period of time.

[0059] 5.2) Through forward timing reasoning, the time constraints for each protection and circuit breaker are obtained as follows: ; ; ; Comparing the time constraints with the warnings received by the scheduling center, it was found that: L1213m, L1312m, and CB1312m meet their time constraints, so there is no need to adjust the confidence level of the library; CB1312p does not meet its time constraints, so the confidence level is adjusted to 0.1.

[0060] 5.3) For the input weight matrix I and O The initial inputs are all divided by the mean to reduce the distance to reaching the optimal parameters: ; 5.4) Based on the initial confidence levels for protection and circuit breakers set in Table 5, the input vector for the power supply end is: The diagnostic result obtained after ignition transition according to the weight matrix is ​​as follows: , The weight matrix is ​​obtained after iterating through the diagnostic architecture: ; .

[0061] Table 5 5.5) Twelve different fault scenarios were simulated for the line. At the transmission end: ① Main protection and main protection circuit breaker operate correctly; ② Main protection operates, circuit breaker fails to operate, near backup operates, near backup circuit breaker operates; ③ Both main protection and near backup operate, but circuit breaker signal is lost; ④ Neither main protection nor near backup operates and the corresponding circuit breaker opens, far backup operates and the corresponding circuit breaker opens. At the receiving end: ① Main protection and main protection circuit breaker operate correctly; ② Main protection operates, circuit breaker fails to operate, near backup operates, near backup circuit breaker operates; ③ Neither main protection nor near backup operates and the corresponding circuit breaker opens, far backup operates and the corresponding circuit breaker opens; The corresponding Cartesian combination yields the full combination. Types of faults: Scenario 1: A combination of power transmission end condition ① and power receiving end condition ①. Scenario 2: A combination of transmission end condition ① and receiving end condition ②. Scenario 3: A combination of transmission end condition ① and receiving end condition ③. Scenario 4: A combination of transmission end condition ② and receiving end condition ①. Scenario 5: A combination of transmission end condition ② and receiving end condition ②. Scenario 6: A combination of transmission end condition ② and receiving end condition ③. Scenario 7: A combination of transmission end condition ③ and receiving end condition ①. Scenario 8: A combination of transmission end condition ③ and receiving end condition ②. Scenario 9: A combination of transmission end condition ③ and receiving end condition ③. Scenario 10: A combination of transmission end condition ④ and receiving end condition ①. Scenario 11: A combination of transmission end condition 4 and receiving end condition 2. Scenario 12: A combination of transmission end condition ④ and receiving end condition ③.

[0062] Table 6 shows the diagnostic results of the new weight matrix for 12 different fault scenarios. Before optimization, the diagnostic accuracy of the 12 fault scenarios fluctuated greatly, ranging from 0.0971 to 0.9787. Among them, the accuracy of 3 scenarios was below 0.4, and the worst scenario was only 0.0971. The diagnostic reliability was seriously insufficient in extreme scenarios, with an average accuracy of about 0.642. After optimization, the diagnostic accuracy of all scenarios was significantly improved and highly concentrated, narrowing the range to 0.9826 to 0.9960. The accuracy of the worst scenario was more than 9 times higher than before optimization, and the average accuracy reached 0.987, an improvement of about 53.7% compared to before optimization. At the same time, the accuracy difference between the scenarios was significantly reduced after optimization, with a fluctuation range of less than 1.4%. This completely solved the problem of the original model's unbalanced adaptability to different bilateral collaborative scenarios, ensuring that highly reliable diagnostic results can be output stably under various fault scenarios.

[0063] Figure 4 (a) to (b) Figure 9 Figures (a)-(b) show the diagnostic confidence curves of the proposed method under 12 different fault conditions in the full combination of the power receiving end and the power transmitting end, as well as the iteration rounds of the RLGWO algorithm. It can be seen from the figures that: First, strong convergence consistency. Although different fault combinations, especially complex coupled scenarios such as scenarios three, six, and nine, showed low diagnostic confidence and significant fluctuations in the early stages of optimization, thanks to the global optimization capability of the RLGWO algorithm, the curves for all scenarios achieved stable convergence to the high confidence interval (>0.98) within 500 iterations, verifying the model's high robustness in both extreme and normal scenarios. Second, escape capability. The step-like jumps in the curves intuitively reflect the process of the reinforcement learning agent dynamically adjusting the GWO exploration and convergence parameters according to the population state, effectively assisting the algorithm in escaping local optima and ensuring continuous evolution towards the global optimum. Third, high learning efficiency. The curves for each scenario showed a rapid upward trend in the early stages of iteration, indicating that the parameter optimization strategy proposed in this invention can quickly lock the optimal weight space of the bilateral neurofuzzy Petri net, meeting the timeliness requirements of engineering applications.

[0064] Table 6 In summary, the RLGWO-NFPN combined model for fault diagnosis optimization achieves the goal of maximizing worst-case performance while maintaining mean accuracy, effectively compensating for the robustness shortcomings of traditional models and providing more reliable technical support for dual-side collaborative fault diagnosis in power systems.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets, characterized in that, include: S1. Screen suspected faulty components and adjust the initial fault confidence level according to the type of suspected component; S2. Divide the corresponding component states into storage locations and corresponding logic transitions, establish the association link from storage location to transition, construct a two-sided fuzzy Petri net adapted to the state coupling from the power receiving end to the power transmitting end, and input the corrected initial fault confidence into the two-sided fuzzy Petri net as the initial state. S3. The reinforcement learning gray wolf method is used to optimize the MLP, and the weight matrix of the bilateral fuzzy Petri net is obtained based on the optimized MLP. S4. Configure the parameters of the associated link according to the weight matrix, and calculate the failure probability of the suspected component through forward inference; S5. If the calculated failure probability is greater than the preset threshold, the suspected component is determined to be faulty; otherwise, return to S1 to re-screen suspected components until the fault diagnosis is completed and the process ends.

2. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 1, characterized in that, Screening suspected faulty components and revising the initial fault confidence based on the type of suspected component includes: The connection analysis method is used to traverse the power grid topology and, combined with the action signals of protection devices, to screen suspected faulty components. Combining the timing logic of the protection-circuit breaker operation rules, the timing constraints of the protection device operation and the circuit breaker tripping are used to correct the initial fault confidence for different types of suspicious components, including transformers, lines, busbars, and power sources.

3. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 1, characterized in that, Establishing the association link from the place to the change includes: obtaining the driving strength of the place to the change by inputting a weight matrix, obtaining the update strength of the change to the place by outputting a weight matrix, and establishing the association link.

4. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 1, characterized in that, The optimization of the MLP using the reinforcement learning gray wolf method includes: S31. Initialize the parameters of the reinforcement learning gray wolf method and construct a gray wolf population, where each individual represents a complete set of fully connected weights and parameter vectors of the multilayer perceptron used to generate the weight matrix of the two-sided fuzzy Petri net. S32. Calculate the fitness of individuals in the population, select elite individuals, perceive the current optimization state through reinforcement learning, dynamically adjust the parameters of the gray wolf optimization algorithm, and update the position of non-elite individuals in the population. When updating the position of non-elite individuals in the population, perturbation is added to several individuals with the worst fitness. The fitness is obtained based on the negative of the sum of the minimum value and the weighted average value in the diagnostic confidence set of all working conditions. S33. Calculate the internal reward function of reinforcement learning, train DQN and soft update the target network, calculate the fitness value of the individual after position update and return to S32, until the termination condition is met, obtain the MLP hyperparameters, and then obtain the optimized MLP.

5. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 4, characterized in that, The parameters of the Grey Wolf optimization algorithm are dynamically adjusted by: dynamically updating the exploration parameters according to a decreasing pattern, and randomly updating the convergence parameters within a preset range.

6. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 1, characterized in that, Based on the optimized MLP, the weight matrix of the bilateral fuzzy Petri net is obtained as follows: The mean of the power receiving end sample set is input into the optimized MLP, and the residual vector is output. The residuals are allocated according to the parameter mask of the two-sided fuzzy Petri net to obtain the power transmitting end input weight matrix, the power transmitting end output weight matrix, the power receiving end input weight matrix, and the power receiving end input weight matrix.

7. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 1, characterized in that, The calculation of the failure probability of the suspected component through forward reasoning includes: By performing a Cartesian product combination on the power receiving end samples and the power transmitting end samples, any set of two-sided operating condition samples can be obtained. By calculating the confidence levels at the receiving end and the transmitting end of each sample group through forward inference, a robust objective function is constructed. The objective is minimized to maximize the confidence level. The robust objective function is constructed using the average confidence level of all operating conditions.

8. The method for fault diagnosis of distribution networks based on optimized bilateral neural fuzzy Petri nets according to claim 7, characterized in that, The forward inference calculation includes: Based on the transition activation threshold, the effective driving signal and noise interference are determined, and the failure probability of the suspected component is calculated using a Gaussian compression function on the effective driving signal.

Images