Distribution network automation equipment group fault diagnosis system method

Abstract

The invention provides a distribution network automation equipment group fault diagnosis system and method, and relates to the technical field of power distribution networks, and the method comprises the following steps: carrying out the feature extraction of fault data through employing fractional order Boltzmann-Shannon interaction entropy; the rational activation function is combined with an extreme learning machine algorithm, a rational extreme learning machine is constructed, and faults are diagnosed; the non-smooth optimization problem of the rational extreme learning machine is solved by adopting a novel umbrella-exponent optimization algorithm; according to the method, the fractional order Boltzmann-Shannon interaction entropy is adopted to carry out feature extraction on the fault data, the data does not need to be decomposed or transformed, and the complexity of the data is directly measured, so that the defect that hidden fault features cannot be effectively extracted by a traditional time-domain and frequency-domain feature extraction method is overcome; the parameter-free measurement can be accurately estimated in a calculation-efficient mode, fractional order information is further considered, and the dynamic change of a fractional order time sequence can be measured.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network, and in particular to a distribution network automation equipment group fault diagnosis system method. Background Art

[0002] As an important part of the power grid, the distribution network directly faces the vast number of electricity users. It is a key hub connecting energy production and consumption and an important part of the entire power grid. With the optimization and upgrading of the energy structure and the gradual deepening of the power system reform, the distribution network is about to enter a new stage of development. It is necessary to coordinate safety, quality and efficiency and benefits, and continuously develop into a safe, reliable, green, intelligent, friendly, interactive, economical and efficient smart distribution network. Distribution automation has become an important supporting link in the construction of smart grids. Distribution automation equipment plays an important role in monitoring the operating status of the distribution network and improving power supply reliability. The number of terminals involved in the distribution automation system is very large. The joint debugging of a large number of distribution automation terminal equipment to the main station system requires the cooperation of the main station and the site. There are many equipment in indoor and outdoor switchgear, and the communication lines involved in the distribution automation system are relatively complex. In the joint debugging of a large number of distribution automation terminals connected to the master station system, the master station and the site need to cooperate in real time. Among them, in response to the operation and maintenance needs in different distribution scenarios, fault detection of equipment groups such as switch cabinets and distribution automation terminals is a key problem that needs to be solved in the distribution detection process. At present, the fault detection method for distribution network automation equipment groups is still mainly manual inspection, which requires a lot of manpower and time. The human resource expenditure is large, the power outage time is uncontrollable, and the coordination burden between departments is heavy, which directly affects the power supply reliability. Artificial intelligence is rarely used in the field of fault diagnosis of distribution network automation equipment groups. Traditional neural networks are mainly used as the main classification tool. Considering the feasibility of the algorithm in terminal equipment, the limitations of the computing power of on-site equipment and the complexity of manual parameter adjustment are mainly considered. Deep learning is almost not used in this field;

[0003] There is a lack of cutting-edge methods in the fault feature extraction of distribution network automation equipment groups. The fault current and voltage data of distribution network automation equipment groups are nonlinear feature data. Traditional time domain feature extraction methods, such as mean, root mean square, kurtosis and skewness, and frequency domain feature extraction methods, such as power spectrum, envelope spectrum analysis and fast Fourier transform, are difficult to analyze such nonlinear signals.

[0004] Artificial intelligence is rarely used in the field of fault diagnosis of distribution network automation equipment groups. Traditional neural networks are mainly used as the main classification tool. Traditional neural network methods have the following disadvantages:

[0005] An ideal neural network activation function needs to meet many conditions, such as nonlinearity, non-saturation, and almost everywhere differentiability. Existing classic activation functions all have their shortcomings, such as saturation on both sides of Sigmoid; dead zone on the left side of ReLU; PReLU, ELU, and LReLU are not differentiable at 0. This makes the network's fit to the inherent rules of the data insufficient. In layman's terms, neurons and neural networks constructed by these traditional activation functions cannot learn the associations contained in the data through training, and cannot represent the inherent essence of things;

[0006] The number of neural network parameters is still too large for physical terminal devices. There is still a lack of lightweight models for actual application scenarios.

[0007] The traditional neural network model has nested hidden layer parameters, making its loss function a strongly nonlinear non-convex function, which is extremely difficult to globally optimize. Generally, only local suboptimal solutions can be solved using optimization methods that rely on gradient information, which will affect the network learning effect. In real scenarios, in order to suppress overfitting, non-smooth regularization terms are often added to the loss function. This term is to pursue parameter sparsity and is generally not differentiable, which makes the entire network parameter optimization problem equivalent to a non-convex non-smooth optimization problem. At this stage, this problem lacks a unified theoretical framework and technical path, making parameter optimization more difficult.

[0008] Therefore, the present invention proposes a distribution network automation equipment group fault diagnosis system method to solve the problems existing in the prior art. Summary of the invention

[0009] In view of the above problems, the present invention proposes a distribution network automation equipment group fault diagnosis system method, which uses fractional-order Boltzmann-Shannon interaction entropy to extract features from fault data. It does not need to decompose or transform the data, and directly measures the complexity of the data, thereby overcoming the shortcomings of traditional time and frequency domain feature extraction methods that often cannot effectively extract hidden fault features.

[0010] To achieve the purpose of the present invention, the present invention is implemented by the following technical solution: a distribution network automation equipment group fault diagnosis system method, comprising the following steps:

[0011] The fractional Boltzmann-Shannon interaction entropy is used to extract features from fault data;

[0012] Combine the rational activation function with the extreme learning machine algorithm to construct a rational extreme learning machine to diagnose faults;

[0013] A novel lizard optimization algorithm is used to solve the non-smooth optimization problem of rational extreme learning machines.

[0014] Further improvements are: Fractional Boltzmann-Shannon interaction entropy: Using the frequency and measurement-based entropy estimation method, the distribution of samples in the continuous state space is analyzed, where JDS is the JS divergence (Jensen–Shannon Divergence), which represents the standard asymmetric KL divergence (Kullbeck-LeiblerDivergence,

[0015] Before training the fault classification algorithm, the fault time signals of the distribution network automation equipment group are generally divided into several partitions according to the fault category. P and Q are different probability distributions extracted from the same partition based on frequency and measurement technology respectively. P is an equal-length histogram segmentation distribution, and Q is a quantile segmentation distribution.

[0016] The KL divergence is defined as follows:

[0017]

[0018] JS divergence JSD(P||Q) is defined as follows:

[0019]

[0020] in

[0021] The calculation formula of JSD(P||Q) is as follows:

[0022]

[0023] Furthermore, the Boltzmann-Shannon interaction entropy is defined as follows:

[0024]

[0025] Among them, X is the bounded region sample sequence;

[0026] The fractional JSD calculation formula is as follows:

[0027]

[0028] Where -1≤α≤1.

[0029] The formula is further refined as follows:

[0030]

[0031] Finally, the fractional Boltzmann-Shannon interaction entropy formula is obtained as follows:

[0032] FrBSIE(X)=1-JSD α (p‖q). (6)

[0033] Among them, Γ(·) is the gamma function, ψ(·) is the digamma function, and α is the fractional order.

[0034] A further improvement is that the rational activation function consists of a rational function with trainable coefficients and, which has the following form:

[0035]

[0036] Where γ1 and γ2 are the polynomial degrees of the numerator and denominator, respectively, and are hyperparameters.

[0037] Further improvements are: using rational activation functions to replace the traditional activation functions of the extreme learning machine algorithm and constructing a rational extreme learning machine:

[0038] RELM algorithm principle:

[0039]

[0040] w is the weight parameter matrix of the input layer and the hidden layer, n is the number of nodes in the input layer, and l is the number of neurons in the hidden layer;

[0041]

[0042] β is the weight parameter matrix of the hidden layer and the output layer, m is the number of nodes in the output layer, and l is the number of neurons in the hidden layer;

[0043]

[0044] b is the hidden layer bias vector;

[0045]

[0046] X is the training input data matrix, that is, the feature matrix, each column is a set of training input data, a total of Q sets of training data;

[0047]

[0048] T is the training output target matrix, which is also the label matrix, that is, each column corresponds to the label vector of each set of training input data (a total of Q groups);

[0049]

[0050] where σ R (·) is a rational activation function, H is the hidden layer output matrix, and in the extreme learning machine algorithm, the parameter matrix or vector w, b, η, ξ is randomly generated during training, and the training input data (x1, x2, ..., x Q ) is a known quantity, so H is a fixed value during model training;

[0051] The goal is to multiply the hidden layer output matrix H by the hidden layer and output layer weight parameter matrix β, which is exactly equal to the transpose of the output target matrix T. This goal means that the algorithm has completely learned the intrinsic relationship between features and labels and has the ability to classify data:

[0052] Hβ=T T (14)

[0053] To require Hβ and T T Close to the condition, H and T T is a determined quantity, β is an unknown quantity, and this model is equivalent to solving the following optimization problem:

[0054]

[0055] It means to minimize the error between the model predicted label and the true label by solving the optimal β, where ||·|| is the Frobenius norm of the matrix elements.

[0056] Further improvement is to add L1 and L2 regularization terms to the above optimization problem in order to prevent overfitting and data collinearity.

[0057] L1 regularization means: it is achieved by adding a penalty term of the L1 norm (i.e. the sum of the absolute values ​​of each element in the vector) to the objective function; L2 regularization means: by adding a regularization term to the loss function, which is the sum of the squares of the model weight parameters;

[0058] Therefore, the original optimization problem is changed to:

[0059]

[0060] Among them, ||·||1 is the L1 norm, ||·||2 is the L2 norm, λ1 and λ2 are the L1 and L2 regularization coefficients respectively, both of which are hyperparameters provided manually. Since the L1 regularization term is a non-smooth term and cannot be differentiable, this optimization problem is a non-smooth optimization problem.

[0061] Further improvements are as follows: the novel lizard optimization algorithm includes population initialization, exploration phase and development phase. In the population initialization, an initial population uniformly distributed for the search space is designed by the principle of optimal point sets in number theory. In the exploration phase, an adaptive weight factor is added to promote a smooth transition between the global and local searches of the algorithm. In the development phase, a Gaussian mutation operator is introduced to perturb the global optimal solution obtained in each iteration.

[0062] A further improvement is that the population initialization includes the following steps:

[0063] The FLO population is formed by the frilled lizards, which is mathematically described as a matrix using equation (17). The initial positions of the frilled lizards are established in the problem solution space by random initialization using equation (18):

[0064]

[0065] In the standard frilled lizard optimization algorithm, the population is initialized randomly:

[0066] X i,d =lb d +r·(ub d -lb d ), (18)

[0067] r represents the number randomly obtained from the interval [0, 1], lb d andub d are the lower and upper bounds of the d-dimensional decision variable respectively;

[0068] The initial population of the search space is designed with uniform distribution through the good point set principle in number theory. The good point set theory assumes that H m is the unit cube of m-dimensional Euclidean space, in which there exists the point set:

[0069] P N (i)={({τ1·i},{τ2·i},...,{τ m ·i}),1≤i≤N}, (19)

[0070] Its deviation satisfies:

[0071]

[0072] where C(τ,ε) is a constant and is related to τ=(τ1,τ2,…,τ m ),ε(ε>0),P N (i) is called the good point set, τ is called the good point, and the value of the good point τ is:

[0073]

[0074] Where p is the smallest prime number satisfying (p-3) / 2≥m, {τ j ·i} represents τ j The decimal part of i, based on the good point set theory, the new initialization strategy is:

[0075] x i,d =(ub d -lb d ){τ d ·i}+lb d , (twenty two)

[0076] Where X represents the FLO overall matrix, X i represents the i-th frill (candidate solution), x i,d represents its dth dimension (decision variable) in the search space, N gives the number of frilled lizards, and m represents the number of decision variables;

[0077] Each parallax lizard represents a candidate solution to the problem. Corresponding to each candidate solution, the corresponding objective function value of the problem is calculated. The determined objective function value set is mathematically represented using the vector given in equation (23):

[0078]

[0079] F represents the vector of calculated objective function values, F i The evaluation objective function value corresponding to the i-th frilled lizard is given;

[0080] The determined objective function value is a measure of the quality of the individuals of the group (i.e., the candidate solutions). The best evaluation value of the objective function corresponds to the best individual of the population (i.e., the best candidate solution). Similarly, the worst evaluation value of the objective function corresponds to the worst individual of the population (i.e., the worst candidate solution). Since at each iteration of FLO the position of the frilled lizard in the solution space is updated, a new value is evaluated for the objective function under consideration. At each iteration, the best positioned individual (i.e., the best candidate solution) is updated. At the end of the implementation of the algorithm FLO, the best candidate solution obtained during the algorithm iteration is taken as the solution to the problem.

[0081] A further improvement is that the exploration phase includes the following steps:

[0082] In the first stage of FLO, the positions of population individuals in the problem solution space are updated according to the hunting strategy of the frilled lizard; in the design of FLO, for each frilled lizard, the positions of other population members with better objective function values ​​are regarded as prey positions, based on which the candidate prey position set of each frilled lizard is determined using equation (24):

[0083] CP i ={X k :F k <F i andk≠i},wherei=1,2,...,Nandk∈{1,2,...,N}, (24)

[0084] CP i is the candidate prey set of the i-th frilled lizard, X k is a population member with a better objective function value than the i-th frilled lizard, F k is its objective function value;

[0085] Assume that the frilled lizard randomly selects one of these candidate prey and attacks it. Based on the modeling of the frilled lizard's movement toward the selected prey, the new position of each individual in the population is calculated using equation (25). Then, when the objective function value is better, the previous position of the corresponding individual is replaced using equation (26):

[0086]

[0087] represents the new proposed position of the ith frilled lizard based on the first stage of FLO, represents its dth dimension, F i P1 represents the objective function value, r is a normally distributed random number in the interval [0, 1], SP i,d indicates the dth dimension of the selected prey of the i-th frilled lizard, I is a number randomly selected from the set {1, 2}, N represents the number of frilled lizards, and m gives the number of decision variables;

[0088] Based on the particle swarm algorithm, considering its inertia weight, when the inertia weight is small, the algorithm search range is finer, and when the inertia weight is large, the algorithm search range is larger. Based on the influence of the hunting strategy of the frilled lizard on the position update of the frilled lizard and the adaptive weight factor balancing algorithm's search ability at different stages, an adaptive weight factor is introduced to improve the standard frilled lizard optimization algorithm:

[0089]

[0090] Where t represents the iteration counter of the algorithm, and T describes the maximum number of iterations of the algorithm;

[0091] The new position of each individual in the improved population and the previous position of the corresponding individual are:

[0092]

[0093] Θ increases nonlinearly with the increase of iteration times. In the early stage of algorithm iteration, the value of the adaptive weight factor is small and the target position is fuzzy. In the middle and late stages of algorithm iteration, the target gradually becomes clear. At this time, the value of the adaptive weight factor gradually increases with the increase of iteration times, and the attraction of the target to the individual frilled lizard increases.

[0094] A further improvement is that the development phase includes the following steps:

[0095] After eating, the frilled lizards retreat to the treetops where they are located. In the second stage of FLO, the positions of the individuals in the population in space are updated according to the strategies of the frilled lizards when they retreat to the treetops after eating.

[0096] Based on the modeling of the movement of frilled lizards to nearby treetops, the new position of each population individual is calculated using equation (30), and if this new position improves the objective function value, the previous position of the corresponding individual is replaced using equation (31):

[0097]

[0098] represents the new proposed position of the i-th frilled lizard based on the second stage of FLO, represents its dth dimension, F i P2 Given its objective function value, t represents the iteration counter of the algorithm, and T describes the maximum number of iterations of the algorithm;

[0099] The Gaussian mutation operator is introduced to perturb the global optimal solution obtained in each iteration. The Gaussian mutation strategy is added after the standard lizard optimization algorithm formula (31):

[0100]

[0101] in, The Gaussian optimal solution is obtained after Gaussian mutation. Gaussian(α) is a random vector with a Gaussian distribution with a mean of 0 and a variance of 1.

[0102] The beneficial effects of the present invention are:

[0103] 1. The present invention uses fractional-order Boltzmann-Shannon interaction entropy to extract features from fault data without decomposing or transforming the data, and directly measures the complexity of the data, thereby overcoming the shortcomings of traditional time and frequency domain feature extraction methods that often cannot effectively extract hidden fault features. In addition, this parameter-free measure can be accurately estimated in a computationally efficient manner, further considering fractional-order information, and can measure the dynamic changes of fractional-order time series, so that more comprehensive fault features can be extracted from fault data.

[0104] 2. The present invention combines the rational activation function with the extreme learning machine algorithm, and proposes a rational extreme learning machine, a low-parameter, lightweight, and highly efficient artificial intelligence method. By adding a regularization term to its loss function, it overcomes the problems of overfitting and data collinearity and meets the requirements of sparsity of learning parameters.

[0105] 3. The present invention introduces the good point set principle, adaptive weight factor and Gaussian mutation strategy into the standard lizard optimization algorithm, and proposes a new lizard optimization to enhance the global search capability of the algorithm, avoid falling into the local optimum during the optimization process, thereby improving the convergence speed of the algorithm and ensuring the diversity of solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1 This is a flow chart of the fault diagnosis algorithm of the distribution network automation equipment group of the present invention. DETAILED DESCRIPTION

[0107] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with examples. The examples are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0108] Embodiment 1

[0109] according to Figure 1 As shown, this embodiment proposes a distribution network automation equipment group fault diagnosis system method, including the following steps:

[0110] The fractional Boltzmann-Shannon interaction entropy is used to extract features from fault data;

[0111] Combine the rational activation function with the extreme learning machine algorithm to construct a rational extreme learning machine to diagnose faults;

[0112] The umbrella lizard optimization algorithm is used to solve the non-smooth optimization problem of rational extreme learning machines.

[0113] The present invention uses fractional Boltzmann-Shannon interactive entropy to extract features from fault data. The introduced geometric partitioning entropy provides an alternative to differential Shannon entropy for quantifying the uncertainty of sample estimates extracted from a one-dimensional bounded continuous probability distribution. The appropriate combination of effectiveness for sparse samples and the impact on extreme outliers is used to define an entropy measure that is not biased towards sample size, which is defined based on the standard deviation between the measure-based and frequency-based distributions, which can be associated with these two historical numbers. This parameter-free measure can be accurately estimated in a computationally efficient manner, and the fractional-order calculation method is introduced into the Boltzmann-Shannon interactive entropy, which further considers the fractional-order information and can measure the dynamic changes of fractional-order time series.

[0114] The present invention improves the fault diagnosis model and overcomes the shortcomings of traditional neural networks that have too many parameters and are not conducive to the operation of terminal devices. Extreme learning machines, which are machine learning with few parameters but very efficient, are used as the basic model. This type of algorithm can be regarded as a special case of a neural network algorithm with only a single hidden layer, because all the parameters contained between its input layer and the hidden layer are randomly generated. When the program is running, this part of the parameters is determined and does not contain unknown parameters. However, the weight parameters between the hidden layer and the input layer need to be obtained through network training and then solved in combination with the optimization algorithm, and are unknown parameters. From this, it can be seen that the extreme learning machine has a greatly reduced number of unknown parameters that need to be solved compared to the traditional neural network model.

[0115] However, the classic extreme learning machine algorithm also has its inherent bottlenecks, and its network activation function generally uses traditional activation functions. Traditional activation functions often have defects in non-saturation and differentiability, which limits the network's ability to fit the inherent rules of things. For example, Sigmoid is saturated on both sides; ReLU has a dead zone on the left; PReLU, ELU, and LReLU are all non-differentiable at 0. Moreover, these activation functions are all experience-driven improvement strategies, and the biggest drawback is the lack of rigorous statements and proofs of approximate theories;

[0116] Therefore, the present invention adopts the academic frontier rational activation function to replace the traditional activation function in the classic extreme learning machine. On the one hand, the rational activation function simultaneously satisfies the three properties of the ideal activation function: nonlinearity, non-saturation, and almost everywhere differentiability. On the other hand, many scholars, starting from the rigorous approximation theory, through a large number of theoretical deductions and practical applications, have proved that the neural network based on the rational activation function has a better approximation ability and a lower training cost than the neural network based on the traditional activation function.

[0117] The present invention combines the rational activation function with the extreme learning machine algorithm and is named the rational extreme learning machine. In order to couple the rational activation function with the underlying principle of the extreme learning machine, all parameters contained between the input layer and the hidden layer of the rational extreme learning machine, including the learnable parameters in the rational activation function, are set to be randomly generated.

[0118] The basic rational extreme learning machine loss function is a concise convex optimization problem form, and the global optimal solution can be obtained by using the optimization method that relies on gradient information commonly used in machine learning. However, considering the diversity and complexity of fault signals of distribution network automation equipment groups, in order to overcome the problems of overfitting and data collinearity and meet the sparsity requirements of learning parameters, the present invention adds a regularization term to the basic rational extreme learning machine loss function. Since the regularization term is a non-smooth term, it is not differentiable. Therefore, this optimization problem is a non-smooth optimization problem, and all optimization methods that rely on gradient information are invalid. In order to overcome this problem, the present invention proposes a derivative-free optimization method, which is also a meta-heuristic swarm intelligence method-a new type of umbrella lizard optimization algorithm.

[0119] In order to improve the optimization performance of the standard lizard optimization algorithm, the novel lizard optimization algorithm proposed in this invention makes the following innovations:

[0120] The random initialization method of the population in the standard lizard optimization algorithm is relatively blind, and some good points cannot be covered. The uniform design method is an effective method. For the initial population in the search space, it is neither roughly randomly generated nor traversing all conditions. The initial population uniformly distributed for the search space is designed through the principle of good point set in number theory to enhance the global search ability of the algorithm and avoid falling into the local optimum during the optimization process, thereby improving the convergence speed of the algorithm and ensuring the diversity of solutions.

[0121] In the exploration phase, the invention adds an adaptive weight factor to promote a smooth transition between the global and local searches of the algorithm, thereby improving the optimization performance and stability of the algorithm.

[0122] During the development phase, the present invention introduces a Gaussian mutation operator to perturb the global optimal solution obtained in each iteration, thereby preventing the algorithm from falling into a local optimum and premature phenomenon, while also maintaining the diversity of individuals in the population.

[0123] Embodiment 2

[0124] according to Figure 1 As shown, this embodiment proposes a distribution network automation equipment group fault diagnosis system method, including the following steps:

[0125] Through artificial intelligence technology, the distribution network automation equipment group can more accurately perceive and analyze the working status and health of the distribution equipment, thereby achieving more efficient and accurate data collection and analysis, timely discovering potential faults and abnormal conditions, and improving the accuracy of fault diagnosis and the efficiency of detection and operation and maintenance. Using a data-driven approach combined with artificial intelligence technology, the data of the distribution network automation equipment group is analyzed and processed. By establishing AI intelligent diagnosis and analysis algorithms and quantitative models, real-time monitoring, fault diagnosis and prediction of equipment operating status can be achieved. This can improve the operation and maintenance personnel's understanding of the equipment status so that they can quickly discover and solve problems, and also provide a basis for intelligent operation and maintenance. The following table shows common fault categories and their causes.

[0126]

[0127]

[0128]

[0129] The fractional Boltzmann-Shannon interactive entropy is used to extract comprehensive features of the distribution network automation equipment group fault data, and then the obtained features are input into the rational extreme learning machine for fault multi-classification processing. The optimizer of the algorithm adopts a new umbrella lizard optimization algorithm

[0130] Feature extraction

[0131] The fractional Boltzmann-Shannon interaction entropy is used to extract features from fault data without decomposing or transforming the data, directly measuring the complexity of the data. This overcomes the shortcomings of traditional feature extraction methods in the time and frequency domains that often cannot effectively extract hidden fault features. In addition, this parameter-free measure can be accurately estimated in a computationally efficient manner, further considering fractional-order information, and can measure the dynamic changes of fractional-order time series, so that more comprehensive fault features can be extracted from fault data.

[0132] Fractional Boltzmann-Shannon interaction entropy:

[0133] The fault signal of the distribution network automation equipment group is a nonlinear time series. Traditional time domain feature extraction methods, such as mean, root mean square, kurtosis and skewness, and frequency domain feature extraction methods, such as power spectrum, envelope spectrum analysis and fast Fourier transform, are difficult to analyze such nonlinear signals. Traditional time and frequency domain feature extraction methods often cannot effectively extract hidden fault features, resulting in a lower fault recognition rate. Due to the different system responses caused by fault shocks, the degree of confusion of monitoring data varies.

[0134] In recent years, the frequency-based Differential Shannon Entropy (DSE) analysis method (generally referred to as entropy analysis method) has been increasingly favored by many scholars and industries. The analysis method based on entropy theory can directly measure the complexity of data without decomposing or transforming the data, and can complete the identification of different faults. For example, sample entropy, permutation entropy and spread entropy have been proposed and used in fault diagnosis.

[0135] However, the frequency-based differential Shannon entropy obtains numerical results in the absence of any prior information about the underlying distribution. This has been criticized theoretically and rationally. And the DSE based on histogram estimation effectively ignores the true impact of outliers by creating ambiguity between sparsely sampled, low-probability measurement space regions and truly disallowed measurements. A more obvious drawback is that these methods rely heavily on frequency-based methods and require a large amount of data to obtain useful results, which can lead to long computational times.

[0136] The metric-based Geometric Partition Entropy (GPE) makes up for the shortcomings of DSE. The use of statistical averages for the entire sample based on the geometric quantile partition estimate of GPE also limits the impact of outliers, which can be seen as a new perspective on the foundation of continuous information theory.

[0137] The Boltzmann-Shannon Interaction Entropy (BSIE) is defined as the standard divergence between measurement-based and frequency-based distributions that can be associated with two historical data. This parameter-free metric can efficiently calculate the estimated value and can be seen as a complementary fusion of differential Shannon entropy and geometric partition entropy. This new metric provides an effective method for unbiased entropy estimation from samples in a one-dimensional continuous state space.

[0138] Furthermore, the present invention introduces the fractional-order calculus theory into the Boltzmann-Shannon interaction entropy, further considers the fractional-order information, and can measure the dynamic changes of the fractional-order time series, that is, proposes the concept of fractional-order Boltzmann-Shannon interaction entropy.

[0139] Using the entropy estimation method based on frequency and measurement, a new tool Boltzmann-Shannon interaction entropy (BSIE) is provided for analyzing the distribution of samples in continuous state space. JDS is the JS divergence (Jensen-Shannon Divergence). KL (p||q) represents the standard asymmetric KL divergence (Kullbeck-LeiblerDivergence) between two distributions P and Q. Before training the fault classification algorithm, the fault time signals of the distribution network automation equipment group are generally divided into several partitions according to the fault category. P and Q are different probability distributions extracted from the same partition based on frequency and measurement technology, respectively.

[0140] P is the equal length histogram segmentation distribution, and Q is the quantile segmentation distribution.

[0141] The KL divergence is defined as follows:

[0142]

[0143] JS divergence JSD(P||Q) is defined as follows:

[0144]

[0145] in The calculation formula of JSD(P||Q) is as follows:

[0146]

[0147] Furthermore, the Boltzmann-Shannon interaction entropy is defined as follows:

[0148]

[0149] Where X is the bounded region sample sequence.

[0150] The fractional JSD calculation formula is as follows:

[0151]

[0152] Where -1≤α≤1.

[0153] The formula is further refined as follows:

[0154]

[0155] Finally, the fractional Boltzmann-Shannon interaction entropy formula is obtained as follows:

[0156] FrBSIE(X)=1-JSD α (p‖q). (6)

[0157] Among them, Γ(·) is the gamma function, ψ(·) is the digamma function, and α is the fractional order.

[0158] Rational Extreme Learning Machine

[0159] Extreme Learning Machine (ELM) is a single-layer feedforward neural network with the characteristics of fast learning speed and good generalization performance. The input weights and hidden layer biases of ELM can be randomly initialized and do not need to be adjusted during training. This makes ELM learn very quickly, especially when dealing with large-scale data sets. ELM randomly generates input weights and hidden layer biases, and then calculates the output weights using algorithms such as least squares. This mechanism enables ELM to approximate complex functions while maintaining a simple structure.

[0160] In the framework of artificial neural network neurons, the role of the activation function is to put the weighted sum into a nonlinear function for nonlinearization. Its essence is: the weighted sum can be regarded as a linear hyperplane, which divides the spatial data into two categories and can be regarded as a binary linear classifier. However, most real problems are nonlinear, and linear classifiers are not universally meaningful at the application level. Therefore, it is necessary to make the linear classifier nonlinear, and the activation function needs to be used. In particular, the choice of nonlinear activation function is crucial in deep learning architectures and seriously affects the performance of neural networks. An ideal activation function needs to meet many conditions, such as nonlinearity, non-saturation, and almost everywhere differentiability. Existing traditional activation functions all have their own shortcomings.

[0161] Common activation functions in the field of deep learning can be roughly divided into two categories: one is smooth activation functions, such as hyperbolic tangent, sigmoid, logistic, etc., but both sides of the function are saturated, that is, the derivatives on both sides tend to 0, so it is easy to cause the vanishing gradient problem during training; the other is polynomial activation functions, but they also have numerical instability and poor approximation of non-smooth functions. In recent years, the most popular choice of activation function is the rectified linear unit (ReLU), but the disadvantage of ReLU is also very obvious. There is a dead zone on the left side of its function. In order to overcome its shortcomings, many improved forms of ReLU have emerged recently, such as Parametric Linear Unit (PReLU), Exponential Linear Unit (ELU), Leaky ReLU, Scaled Exponential Linear Unit (SELU), etc. The biggest drawback of these experience-driven improvement strategies is the lack of rigorous statements and proofs of approximate theory.

[0162] Rational activation functions will be used in the neural networks used to replace traditional activation functions. Starting from rigorous approximation theory, many scholars have proved through a large number of theoretical deductions and practical applications that neural networks based on rational activation functions have better approximation capabilities and lower training costs than neural networks based on ReLU and its variants.

[0163] The rational activation function consists of a function with trainable coefficient η i and j The rational function of is as follows:

[0164]

[0165] Where γ1 and γ2 are the polynomial degrees of the numerator and denominator, respectively, and are hyperparameters. In view of the superiority of the rational activation function, the present invention will use it to replace the traditional activation function of the extreme learning machine and propose a rational extreme learning machine (RELM) algorithm.

[0166] RELM algorithm principle:

[0167]

[0168] w is the weight parameter matrix of the input layer and the hidden layer, n is the number of nodes in the input layer, and l is the number of neurons in the hidden layer.

[0169]

[0170] β is the weight parameter matrix of the hidden layer and the output layer, m is the number of nodes in the output layer, and l is the number of neurons in the hidden layer.

[0171]

[0172] b is the hidden layer bias vector.

[0173]

[0174] X is the training input data matrix, that is, the feature matrix. Each column is a set of training input data, with a total of Q sets of training data.

[0175]

[0176] T is the training output target matrix, which is also the label matrix, that is, each column corresponds to the label vector of each set of training input data (a total of Q groups).

[0177]

[0178] where σ R (·) is a rational activation function, and H is the hidden layer output matrix. In the extreme learning machine algorithm, the parameter matrix or vector w, b, η, ξ is randomly generated during training, and the training input data (x1, x2, ..., x Q ) is a known quantity, so H is determined during model training.

[0179] Ideally, we hope that the hidden layer output matrix H multiplied by the hidden layer and output layer weight parameter matrix β is exactly equal to the transpose of the output target matrix T, which means that the algorithm has fully learned the intrinsic relationship between features and labels and has the ability to classify data.

[0180] Hβ=T T (14)

[0181] But in practical applications, this is extremely rare, and generally only Hβ and T are required. T Close enough. Among them, H and T T is a determined quantity, and β is an unknown quantity. This model is equivalent to solving the following optimization problem:

[0182]

[0183] That is, by solving the optimal β, the error between the model predicted label and the true label is minimized. Among them, ||·|| is the Frobenius norm of the matrix elements. Furthermore, in order to prevent overfitting, data collinearity and other problems, the generalization ability of the model is improved. Inspired by the elastic network regression algorithm, the present invention adds L1 and L2 regularization terms to the above optimization problem.

[0184] L1 regularization is a regularization technique that is widely used in machine learning algorithms. Its core idea is to add a penalty term of the L1 norm (i.e., the sum of the absolute values ​​of each element in a vector) to the objective function. This penalty term helps to reduce the complexity of the model and prevent the model from overfitting the training data, thereby improving the generalization ability of the model. L1 regularization can reduce the complexity of the model and avoid overfitting the model to the training data, thereby improving the performance of the model on new data. L1 regularization tends to set the weight coefficients of irrelevant features to zero, so that the model only focuses on the most important features. This feature makes L1 regularization very useful when dealing with high-dimensional data sets, because it can automatically reduce the number of features and improve the efficiency of the model. When multiple features are highly correlated, L1 regularization can compress their weight coefficients to a smaller range, thereby reducing the collinearity between them and improving the stability of the model.

[0185] L2 regularization adds a regularization term to the loss function, which is the sum of the squares of the model weight parameters. This regularization method penalizes larger model parameters and forces the model weights to approach zero, making the model smoother and reducing the complexity of the model. The benefit of this is that it can improve the generalization ability of the model, that is, the performance of the model on new data. L2 regularization is particularly suitable for when the training data set is small or there is multicollinearity, and it can avoid overfitting by reducing the complexity of the model, thereby improving the performance of the model. In addition, L2 regularization also helps to stabilize the prediction results of the model and reduce the model's oversensitivity to small changes in the training data, thereby enhancing the robustness of the model. In the process of deep neural network training, L2 regularization is widely used to prevent the model from overfitting when the complexity increases, ensuring that the model minimizes the training error while keeping it simple, so that it has good generalization ability. L2 regularization is an effective technique that improves the generalization performance of the model by controlling the complexity of the model. It is one of the commonly used regularization methods in machine learning and deep learning.

[0186] Therefore, the present invention rewrites the original optimization problem as follows:

[0187]

[0188] Among them, ||·||1 is the L1 norm, and ||·||2 is the L2 norm. λ1 and λ2 are the coefficients of L1 and L2 regularization terms, respectively, and are both hyperparameters (provided manually). Since the L1 regularization term is a non-smooth term, it is not differentiable. Therefore, this optimization problem is a non-smooth optimization problem, and the optimization methods commonly used in machine learning that rely on gradient information are all invalid. Furthermore, the present invention will adopt the meta-heuristic swarm intelligence method in the derivative-free optimization technology to solve it.

[0189] New Lizard Optimization Algorithm

[0190] Population initialization

[0191] The standard Frilled Lizard Optimization (FLO) algorithm is a metaheuristic algorithm that considers frilled lizards as its members. FLO efficiently finds near-optimal solutions to optimization challenges by leveraging the search capabilities of its members within the problem-solving space. Each frilled lizard establishes a value assignment for the decision variables based on its specific position in the problem-solving space. Therefore, each frilled lizard represents a potential solution that can be mathematically interpreted by a vector. Collectively, the frilled lizards constitute the FLO population, which can be mathematically described as a matrix using equation (17). The initial position of the frilled lizard.

[0192] In the problem solution space, we establish by random initialization using equation (18):

[0193]

[0194] In the standard frilled lizard optimization algorithm, the population is initialized randomly:

[0195] X i,d =lb d +r·(ub d -lb d ), (18)

[0196] r represents the number randomly obtained from the interval [0, 1], lb d andub d are the lower and upper bounds of the d-dimensional decision variable, respectively. This random method causes the initialization to be relatively blind, and some good points cannot be covered. The uniform design method is an effective method. For the initial population in the search space, it is neither roughly randomly generated nor traversing all conditions. The good point set principle in number theory is used to design an initial population with uniform distribution for the search space to maintain good diversity of the population. The good point set theory assumes that H m is the unit cube of m-dimensional Euclidean space, in which there exists the point set:

[0197] P N (i)={({τ1·i},{τ2·i},…,{τ m·i}),1≤i≤N}, (19)

[0198] Its deviation satisfies:

[0199]

[0200] where C(τ,ε) is a constant that only depends on τ=(τ1,τ2,…,τ m ),ε(ε>0). Then P N (i) is called the good point set, and τ is called the good point. The value of the good point τ is:

[0201]

[0202] Where p is the smallest prime number satisfying (p-3) / 2≥m, {τ j ·i} represents τ j The fractional part of i. Therefore, based on the good point set theory, the new initialization strategy is:

[0203] x i,d =(ub d -lb d ){τ d ·i}+lb d , (twenty two)

[0204] Where X represents the FLO overall matrix, X i represents the i-th frill (candidate solution), x i,d represents its dth dimension (decision variable) in the search space, N gives the number of frilled lizards, and m represents the number of decision variables.

[0205] Considering that each fringed lizard represents a candidate solution to the problem, corresponding to each candidate solution, the corresponding objective function value of the problem can be calculated. The determined set of objective function values ​​can be mathematically represented using the vector given in equation (23):

[0206]

[0207] Here, F represents the vector of calculated objective function values, F i The evaluation objective function value corresponding to the i-th frilled lizard is given.

[0208] The determined objective function value is an appropriate measure of the quality of the individuals of the population (i.e., the candidate solutions). In particular, the best evaluation value of the objective function corresponds to the best individual of the population (i.e., the best candidate solution), and similarly, the worst evaluation value of the objective function corresponds to the worst individual of the population (i.e., the worst candidate solution). Since at each iteration of FLO the position of the frilled lizard in the solution space is updated, new values ​​of the objective function are also evaluated for the considered objective function. Therefore, at each iteration the best positioned individual (i.e., the best candidate solution) must also be updated. At the end of the implementation of the algorithm FLO, the best candidate solution obtained during the algorithm iterations is considered as the solution to the problem.

[0209] In each iteration of the FLO algorithm, the position of the frilled lizard in the problem solution space is updated in two different phases. First, the exploration phase simulates the frilled lizard moving towards its prey during hunting, aiming to diversify the search space and explore new potential solutions. This phase allows the algorithm to explore different areas of the problem space, thus helping to discover new areas that may contain optimal solutions.

[0210] Second, the development phase simulates the frilled lizard's movement toward the top of a tree after feeding. In this phase, the algorithm uses the knowledge gained during exploration to develop promising areas that were identified as potential optimal solutions. By focusing on refining these areas, the development phase aims to improve the quality of the solution and converge towards the global optimum.

[0211] Exploration Phase

[0212] One of the most characteristic natural behaviors of the frilled lizard is the hunting strategy of this animal. The frilled lizard is a sit-and-wait predator that attacks its prey upon sighting it. The simulation of the frilled lizard's movement toward its prey results in extensive changes in the positions of the population members in the problem-solving space, which improves the exploration capability of the global search algorithm. In the first stage of FLO, the update of the target position according to the hunting strategy of the frilled lizard is fuzzy, which has little effect on the update of the exploration and exploitation mode search routes in the standard frilled lizard optimization algorithm; in the middle and late stages of the algorithm iteration, the target gradually becomes clear, and at this time the value of the adaptive weight factor gradually increases with the increase in the number of iterations, and the target attracts the frilled lizard individuals more strongly. The adaptive weight factor enhances the ability of local exploitation, allowing the frilled lizard to find the target solution faster in the middle and late stages of the algorithm iteration. The addition of the adaptive weight factor can promote a smooth transition between the global and local searches of the algorithm, and improve the algorithm's optimization performance and stability.

[0213] Development Phase

[0214] After feeding, the frilled lizards retreat to the top of a tree near their location. Simulating the frilled lizard's movement to the treetops results in small changes in the position of the population individuals in the problem solution space, thereby improving the algorithm's ability to exploit local search. In the second stage of FLO, the position of the population individuals in space is updated according to the frilled lizard's strategy when retreating to the treetop after feeding. Based on modeling the frilled lizard's movement to the nearby treetops, the new position of each population individual is calculated using equation (30). Then, if this new position improves the objective function value, the previous position of the corresponding individual is replaced using equation (31):

[0215]

[0216] here, represents the new proposed position of the i-th frilled lizard based on the second stage of FLO, represents its dth dimension, F i P2 Given its objective function value, t represents the iteration counter of the algorithm, and T describes the maximum number of iterations of the algorithm.

[0217] The Gaussian mutation operator is introduced to perturb the global optimal solution obtained in each iteration to avoid the algorithm falling into the local optimum and the disadvantage of premature maturity, while also maintaining the diversity of individuals in the population. We add the Gaussian mutation strategy after the standard lizard optimization algorithm formula (31):

[0218]

[0219] in, The Gaussian optimal solution is obtained after Gaussian mutation. Gaussian(α) is a random vector with a Gaussian distribution with a mean of 0 and a variance of 1.

[0220] The method for the fault diagnosis system of a distribution network automation equipment group adopts the fractional Boltzmann-Shannon interactive entropy to extract features from fault data, without decomposing or transforming the data, and directly measures the complexity of the data, thereby overcoming the shortcomings that the traditional time and frequency domain feature extraction methods often cannot effectively extract hidden fault features. In addition, this parameter-free measurement can be accurately estimated in a computationally efficient manner, further considering the fractional order information, and can measure the dynamic changes of the fractional order time series, so as to extract more comprehensive fault features from the fault data. In addition, the present invention combines the rational activation function with the extreme learning machine algorithm, proposes a rational extreme learning machine, a low-parameter, lightweight, and high-efficiency artificial intelligence method, and adds a regularization term to its loss function to overcome the problems of overfitting and data collinearity, meet the requirements of learning parameter sparsity, and the like. At the same time, the present invention introduces the good point set principle, adaptive weight factor, and Gaussian mutation strategy into the standard umbrella lizard optimization algorithm, and proposes a new type of umbrella lizard optimization to enhance the global search ability of the algorithm, avoid falling into the local optimum during the optimization process, thereby improving the convergence speed of the algorithm and ensuring the diversity of solutions.

[0221] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A distribution network automation equipment group fault diagnosis system method, characterized in that: The following steps are involved: The fractional Boltzmann-Shannon interaction entropy is used to extract features from fault data; Combine the rational activation function with the extreme learning machine algorithm to construct a rational extreme learning machine to diagnose faults; A novel lizard optimization algorithm is used to solve the non-smooth optimization problem of rational extreme learning machines.

2. A distribution network automation equipment group fault diagnosis system method according to claim 1, characterized in that: Fractional Boltzmann-Shannon Interaction Entropy: The frequency- and measurement-based entropy estimation method is used to analyze the distribution of samples in the continuous state space. JDS is the JS divergence (Jensen–Shannon Divergence), which represents the standard asymmetric KL divergence (Kullbeck-Leibler Divergence) between two distributions P and Q. Before training the fault classification algorithm, the fault time signals of the distribution network automation equipment group are generally divided into several partitions according to the fault category. P and Q are different probability distributions extracted from the same partition based on frequency and measurement technology, respectively. P is an equal-length histogram segmentation distribution, and Q is a quantile segmentation distribution. The KL divergence is defined as follows: JS divergence JSD(P||Q) is defined as follows: in The calculation formula of JSD(P||Q) is as follows: Furthermore, the Boltzmann-Shannon interaction entropy is defined as follows: Among them, X is the bounded region sample sequence; The fractional JSD calculation formula is as follows: Where -1≤α≤1. The formula is further refined as follows: Finally, the fractional Boltzmann-Shannon interaction entropy formula is obtained as follows: FrBSIE(X)=1-JSD α (p‖q). (6) Among them, Γ(·) is the gamma function, ψ(·) is the digamma function, and α is the fractional order.

3. A distribution network automation equipment group fault diagnosis system method according to claim 1, characterized in that: Rational activation functions consist of rational functions with trainable coefficients and , and have the following form: Where γ1 and γ2 are the polynomial degrees of the numerator and denominator, respectively, and are hyperparameters.

4. A distribution network automation equipment group fault diagnosis system method according to claim 3, characterized in that: The rational activation function is used to replace the traditional activation function of the extreme learning machine algorithm to construct a rational extreme learning machine: RELM algorithm principle: w is the weight parameter matrix of the input layer and the hidden layer, n is the number of nodes in the input layer, and l is the number of neurons in the hidden layer; β is the weight parameter matrix of the hidden layer and the output layer, m is the number of nodes in the output layer, and l is the number of neurons in the hidden layer; b is the hidden layer bias vector; X is the training input data matrix, that is, the feature matrix, each column is a set of training input data, a total of Q sets of training data; T is the training output target matrix, which is also the label matrix, that is, each column corresponds to the label vector of each set of training input data (a total of Q groups); where σ R (·) is a rational activation function, H is the hidden layer output matrix, and in the extreme learning machine algorithm, the parameter matrix or vector w, b, η, ξ is randomly generated during training, and the training input data (x1, x2, ..., x Q ) is a known quantity, so H is a fixed value during model training; The goal is to multiply the hidden layer output matrix H by the hidden layer and output layer weight parameter matrix β, which is exactly equal to the transpose of the output target matrix T. This goal means that the algorithm has completely learned the intrinsic relationship between features and labels and has the ability to classify data: Hβ=T T (14) To require Hβ and T T Close to the condition, H and T T is a determined quantity, β is an unknown quantity, and this model is equivalent to solving the following optimization problem: It means to minimize the error between the model predicted label and the true label by solving the optimal β, where ||·|| is the Frobenius norm of the matrix elements.

5. A distribution network automation equipment group fault diagnosis system method according to claim 4, characterized in that: In order to prevent overfitting and data collinearity, L1 and L2 regularization terms are added to the above optimization problem. L1 regularization means: it is achieved by adding a penalty term of the L1 norm (i.e. the sum of the absolute values ​​of each element in the vector) to the objective function; L2 regularization means: by adding a regularization term to the loss function, which is the sum of the squares of the model weight parameters; Therefore, the original optimization problem is changed to: Among them, ||·||1 is the L1 norm, ||·||2 is the L2 norm, λ1 and λ2 are the L1 and L2 regularization coefficients respectively, both of which are hyperparameters provided manually. Since the L1 regularization term is a non-smooth term and cannot be differentiable, this optimization problem is a non-smooth optimization problem.

6. A distribution network automation equipment group fault diagnosis system method according to claim 5, characterized in that: The novel lizard optimization algorithm includes population initialization, exploration phase and development phase. In the population initialization, the optimal point set principle in number theory is used to design an initial population with uniform distribution for the search space. In the exploration phase, an adaptive weight factor is added to promote a smooth transition between the global and local searches of the algorithm. In the development phase, a Gaussian mutation operator is introduced to perturb the global optimal solution obtained in each iteration.

7. A distribution network automation equipment group fault diagnosis system method according to claim 6, characterized in that: The population initialization includes the following steps: The FLO population is formed by the frilled lizards, which is mathematically described as a matrix using equation (17). The initial positions of the frilled lizards are established in the problem solution space by random initialization using equation (18): In the standard frilled lizard optimization algorithm, the population is initialized randomly: X i,d =lb d +r·(ub d -lb d ), (18) r represents the number randomly obtained from the interval [0, 1], lb d andub d are the lower and upper bounds of the d-dimensional decision variable respectively; The initial population of the search space is designed with uniform distribution through the good point set principle in number theory. The good point set theory assumes that H m is the unit cube of m-dimensional Euclidean space, in which there exists the point set: P N (i)={({τ1·i},{τ2·i},...,{τ m ·i}),1≤i≤N}, (19) Its deviation satisfies: where C(τ,ε) is a constant and is related to τ=(τ1,τ2,…,τ m ),ε(ε>0),P N (i) is called the good point set, τ is called the good point, and the value of the good point τ is: Where p is the smallest prime number satisfying (p-3) / 2≥m, {τ j ·i} represents τ j The decimal part of i, based on the good point set theory, the new initialization strategy is: x i,d =(ub d -lb d ){τ d ·i}+lb d , (22) Where X represents the FLO overall matrix, X i represents the i-th frill (candidate solution), x i,d represents its dth dimension (decision variable) in the search space, N gives the number of frilled lizards, and m represents the number of decision variables; Each parallax lizard represents a candidate solution to the problem. Corresponding to each candidate solution, the corresponding objective function value of the problem is calculated. The determined objective function value set is mathematically represented using the vector given in equation (23): F represents the vector of calculated objective function values, F i The evaluation objective function value corresponding to the i-th frilled lizard is given; The determined objective function value is a measure of the quality of the individuals of the group (i.e., the candidate solutions). The best evaluation value of the objective function corresponds to the best individual of the population (i.e., the best candidate solution). Similarly, the worst evaluation value of the objective function corresponds to the worst individual of the population (i.e., the worst candidate solution). Since at each iteration of FLO the position of the frilled lizard in the solution space is updated, a new value is evaluated for the objective function under consideration. At each iteration, the best positioned individual (i.e., the best candidate solution) is updated. At the end of the implementation of the algorithm FLO, the best candidate solution obtained during the algorithm iteration is taken as the solution to the problem.

8. A distribution network automation equipment group fault diagnosis system method according to claim 7, characterized in that: The exploration phase includes the following steps: In the first stage of FLO, the positions of population individuals in the problem solution space are updated according to the hunting strategy of the frilled lizard; in the design of FLO, for each frilled lizard, the positions of other population members with better objective function values ​​are regarded as prey positions, based on which the candidate prey position set of each frilled lizard is determined using equation (24): CP i ={X k :F k <F i and k≠i},where i=1,2,...,N and k∈{1,2,...,N}, (24) CP i is the candidate prey set of the i-th frilled lizard, X k is a population member with a better objective function value than the i-th frilled lizard, F k is its objective function value; Assume that the frilled lizard randomly selects one of these candidate prey and attacks it. Based on the modeling of the frilled lizard's movement toward the selected prey, the new position of each individual in the population is calculated using equation (25). Then, when the objective function value is better, the previous position of the corresponding individual is replaced using equation (26): represents the new proposed position of the ith frilled lizard based on the first stage of FLO, represents its dth dimension, F i P1 represents the objective function value, r is a normally distributed random number in the interval [0, 1], SP i,d indicates the dth dimension of the selected prey of the i-th frilled lizard, I is a number randomly selected from the set {1, 2}, N represents the number of frilled lizards, and m gives the number of decision variables; Based on the particle swarm algorithm, considering its inertia weight, when the inertia weight is small, the algorithm search range is finer, and when the inertia weight is large, the algorithm search range is larger. Based on the influence of the hunting strategy of the frilled lizard on the position update of the frilled lizard and the adaptive weight factor balancing algorithm's search ability at different stages, an adaptive weight factor is introduced to improve the standard frilled lizard optimization algorithm: Where t represents the iteration counter of the algorithm, and T describes the maximum number of iterations of the algorithm; The new position of each individual in the improved population and the previous position of the corresponding individual are: Θ increases nonlinearly with the increase of iteration times. In the early stage of algorithm iteration, the value of the adaptive weight factor is small and the target position is fuzzy. In the middle and late stages of algorithm iteration, the target gradually becomes clear. At this time, the value of the adaptive weight factor gradually increases with the increase of iteration times, and the attraction of the target to the individual frilled lizard increases.

9. A distribution network automation equipment group fault diagnosis system method according to claim 8, characterized in that: The development phase includes the following steps: After eating, the frilled lizards retreat to the treetops where they are located. In the second stage of FLO, the positions of the individuals in the population in space are updated according to the strategies of the frilled lizards when they retreat to the treetops after eating. Based on the modeling of the movement of frilled lizards to nearby treetops, the new position of each population individual is calculated using equation (30), and if this new position improves the objective function value, the previous position of the corresponding individual is replaced using equation (31): represents the new proposed position of the i-th frilled lizard based on the second stage of FLO, represents its dth dimension, F i P2 Given its objective function value, t represents the iteration counter of the algorithm, and T describes the maximum number of iterations of the algorithm; The Gaussian mutation operator is introduced to perturb the global optimal solution obtained in each iteration. The Gaussian mutation strategy is added after the standard lizard optimization algorithm formula (31): in, The Gaussian optimal solution is obtained after Gaussian mutation. Gaussian(α) is a random vector with a Gaussian distribution with a mean of 0 and a variance of 1.

Images