Method, system and equipment for detecting concentration of gas dissolved in oil based on improved extreme learning machine and medium

Abstract

The invention relates to the technical field of gas concentration detection of power equipment, and discloses a method and system for detecting the concentration of gas dissolved in oil based on an improved extreme learning machine, and the method comprises the steps: improving the extreme learning machine into a kernel extreme learning machine and a depth extreme learning machine, and respectively combining with a seagull optimization algorithm to obtain the concentration of gas dissolved in oil; constructing a model for predicting the concentration of the dissolved gas in the transformer oil, and training the model; and based on the transformer data, inputting the transformer data into the model for predicting the concentration of the dissolved gas in the transformer oil, and outputting a prediction result of the dissolved gas in the transformer oil and a fault type. According to the method, an intelligent optimization algorithm and a data processing method are combined, an extreme learning machine with high learning speed and good generalization performance is improved to form a deep extreme learning machine and kernel extreme learning machine model, an improved seagull algorithm is used for optimization, and a prediction model is formed for prediction. And the detection precision of the equipment gas concentration in the power system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of gas concentration detection of power equipment, and in particular to a method and system for detecting dissolved gas concentration in oil based on an improved extreme learning machine. Background Art

[0002] The frequency of power system equipment failures has increased significantly due to a combination of factors, including extreme weather, human error, and aging equipment. Although many power plants already have primary, backup, and auxiliary protection systems for key equipment, the rapid development of the power industry and continuous technological advancements have gradually exposed limitations in existing protection configurations. In the era of intelligent and digital technology, traditional protection systems are no longer able to meet the requirements for efficient and stable operation of modern power plants in terms of information processing capabilities, fault diagnosis accuracy, and rapid response capabilities.

[0003] During actual operation, electrical equipment is affected by numerous factors, including the surrounding environment, operating conditions, and device structure. This leads to complex internal insulation degradation mechanisms and uncertain failure occurrences. For example, the generation of dissolved gas in transformer oil is closely related to internal failures, resulting in complex and nonlinear time series. Research approaches can be broadly summarized into three directions: statistical methods, artificial intelligence methods, and combined prediction methods.

[0004] 1. Using a neural network model to predict the dissolved gas content in oil is one of the important methods for evaluating the operating status of power transformers. Data quality is a key factor affecting the accuracy of the neural network prediction model. Due to the complex operating environment of the 220kv transformer, the collected gas data contains various types of abnormal data, which in turn causes a decline in data quality and affects the prediction accuracy of the model.

[0005] 2. Parameter matching of the neural network model is also an important factor affecting its predictive performance. However, the traditional method of selecting parameters based on manual experience has shortcomings such as subjectivity, inefficiency and non-scalability, which also affects the predictive performance of the model to a certain extent.

[0006] 3. The problem of requiring a large number of training samples can be effectively solved by recurrent and convolutional neural networks in deep learning. However, traditional machine learning cannot analyze and resolve the characteristics hidden between time signals in the training data, making it difficult to improve the accuracy of predicting gas concentration in oil.

[0007] 4. Statistical prediction methods include grey theory and its improved models, fuzzy prediction models, time signal prediction models, etc. The final prediction accuracy of these models is mainly limited to the distribution law of test data. Summary of the Invention

[0008] In view of the above existing problems, the present invention is proposed. Therefore, the present invention provides a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine to solve the above problems.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In a first aspect, the present invention provides a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine, comprising: acquiring transformer data and performing data preprocessing;

[0011] By improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine, respectively, combined with the Seagull optimization algorithm, a prediction model for dissolved gas concentration in transformer oil was constructed and trained.

[0012] Based on the transformer data, the data is input into the transformer oil dissolved gas concentration prediction model, and the transformer oil dissolved gas concentration prediction result and fault type are output.

[0013] As a preferred solution of the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine described in the present invention, the transformer data includes gas concentration data, transformer operation data and environmental data, and the samples of historical transformer data and the corresponding simulation results are divided into training sets and test sets according to a preset ratio to train the prediction model for dissolved gas concentration in transformer oil.

[0014] As a preferred solution of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine of the present invention, improving the extreme learning machine to a kernel extreme learning machine includes:

[0015] By comparing the extreme learning machine with a support vector machine, a kernel function is introduced into the extreme learning machine, and the extreme learning machine is improved into a kernel extreme learning machine;

[0016] The activation function of the hidden layer is replaced by the kernel function, the number of neurons in the hidden layer, the input parameters and the system bias are not considered, the kernel function is used to replace the feature mapping function, and the kernel coefficient and the penalty coefficient in the generalized inverse matrix are set;

[0017] The beneficial effect of this preferred solution is that the kernel extreme learning machine maps data to a high-dimensional space through a kernel function, can automatically learn the complex features of the data, and has better generalization ability when dealing with nonlinear problems.

[0018] As a preferred solution of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine of the present invention, improving the extreme learning machine to a deep extreme learning machine includes:

[0019] Multiple extreme learning machines are stacked to form a deep extreme learning machine, which maps data features by stacking multiple hidden layers and adopts a hierarchical unsupervised training model;

[0020] During the training process of the deep extreme learning machine network model, the training data is used as the output of the first layer of extreme learning machine, the output of the hidden layer in the deep extreme learning machine model is used as the input data of the second layer of extreme learning machine, and the training of each layer of extreme learning machine is completed;

[0021] The beneficial effect of this preferred solution is that it not only improves the model's prediction accuracy, but also increases the model's generalization ability. The deep extreme learning machine adopts layered unsupervised training, which can minimize the reconstruction error, so that the input data and output data in the model are as similar as possible, thereby improving the prediction accuracy.

[0022] As a preferred solution of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine of the present invention, the Seagull optimization algorithm includes:

[0023] The seagull optimization algorithm is divided into a migration phase and a predation phase;

[0024] During the migration phase, the latest position of the seagull is controlled by the convergence factor, which is a Sigmoid function. The value of the convergence factor is dynamically adjusted according to the current number of iterations, so that the seagull optimization algorithm performs a global search in the early stage and switches to a local search in the later stage. The convergence factor is used to control the seagull to fly to the optimal position.

[0025] During the described hunting phase, the gull attacks its prey in a spiral flight pattern;

[0026] The fitness function is used to update the fitness value of the current global optimal position at each iteration, and it is determined whether to update the current position and the global optimal position to obtain the optimal solution and make the population update continuous.

[0027] The beneficial effect of this preferred solution is that it can improve the algorithm's optimization ability and global optimal continuity.

[0028] As a preferred solution of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine of the present invention, constructing the prediction model for dissolved gas concentration in transformer oil includes:

[0029] Step 1: Select a machine learning model and improve the extreme learning machine to a kernel extreme learning machine and a deep extreme learning machine, set the number of variables, and set the upper and lower bounds of the search space;

[0030] Step 2: Initialize the fitness function, and set the fitness function to the minimum root mean square error obtained by training calculation;

[0031] Step 3, using the fitness as food, and sorting the sites from low to high according to the fitness, so as to obtain sites with more food sources;

[0032] Step 4: Calculate the value of the convergence factor to avoid collisions, update to the global optimal position, determine the food location, update the position of each individual, and calculate the updated seagull fitness;

[0033] Step 5: Obtain the training model of the training set and the test set, compare the fitness obtained in each iteration with the optimal fitness obtained in the previous iteration, and update the global optimal fitness;

[0034] According to a preset number of iterations, steps three to five are continuously repeated to obtain the prediction model for the concentration of dissolved gas in transformer oil.

[0035] As a preferred solution of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine of the present invention, the output of the prediction result of dissolved gas in transformer oil and the fault type includes:

[0036] The gas concentration data includes hydrogen, carbon monoxide, carbon dioxide, and acetylene; the transformer operation data includes load, temperature, and oil temperature; the environmental data includes ambient temperature and humidity;

[0037] The gas concentration data, transformer operation data, and environmental data are used as input data for a prediction model of dissolved gas concentration in transformer oil. The same gas concentration is predicted using multiple evaluation indicators, and the most accurate model output is selected. The predicted value of the gas concentration, concentration development trend, and fault type are also output.

[0038] The beneficial effect of this preferred solution is to construct a new type of intelligent predictive protection, which can provide early warning of possible problems in the equipment before an accident occurs.

[0039] In a second aspect, the present invention provides a system for detecting dissolved gas concentration in oil based on an improved extreme learning machine, comprising:

[0040] Acquisition module, used to obtain transformer data and perform data preprocessing;

[0041] A construction module is used to build a prediction model for dissolved gas concentration in transformer oil by improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine, and combining it with the Seagull optimization algorithm, and then train the model;

[0042] The output module is used to input the transformer oil dissolved gas concentration prediction model based on the transformer data, and output the transformer oil dissolved gas concentration prediction result and fault type.

[0043] In a third aspect, the present invention provides an electronic device, comprising:

[0044] memory and processor;

[0045] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine are implemented.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a gas concentration detection method for an improved extreme learning machine, combines an intelligent optimization algorithm with a data processing method, improves the extreme learning machine with fast learning speed and good generalization performance to form a deep extreme learning machine and a kernel extreme learning machine model, and uses an improved seagull algorithm for optimization to form a prediction model for prediction, thereby improving the detection accuracy of the gas concentration of equipment in the power system. Through simulation based on real data of transformers, the effectiveness and accuracy of the present method are proved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] Figure 1 This is a schematic diagram of the overall process of a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to one embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of an extreme learning machine structural model of a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to an embodiment of the present invention;

[0051] Figure 3 Schematic diagram of the DELM network structure of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine according to one embodiment of the present invention;

[0052] Figure 4 A schematic diagram of the migration and predation behavior of seagulls according to the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to one embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the structure of the ISOA-machine learning model of the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to one embodiment of the present invention;

[0054] Figure 6 A schematic diagram of a single model prediction error box of a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the prediction error box of the combined model of the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine according to one embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0057] Reference Figure 1-Figure 5 , which is an embodiment of the present invention, provides a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine, comprising:

[0058] S101, obtaining transformer data and performing data preprocessing;

[0059] S102: By improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine and combining them with the Seagull optimization algorithm, a prediction model for dissolved gas concentration in transformer oil is constructed and the model is trained;

[0060] S103: Based on the transformer data, the data is input into a prediction model for the concentration of dissolved gas in transformer oil, and the prediction result of the concentration of dissolved gas in transformer oil and the fault type are output.

[0061] It should be noted that the gas turbine unit in the present invention adopts a two-to-one gas-steam combined cycle mode. During the design process of the production system, full consideration is given to the safety and efficiency of production, and multiple sets of equipment with strict requirements on internal gas concentration are equipped. Among them, GIS, power transformers, etc. are core equipment with specific requirements on internal gas concentration. The present invention is based on the prediction model of the improved extreme learning machine (ELM), and improves the ELM with fast learning speed and good generalization performance to form a deep extreme learning machine (DELM) and a kernel-based extreme learning machine (KELM) model. The improved Seagull Optimization Algorithm (ISOA) is used for optimization to form a prediction model for dissolved gas concentration in transformer oil, namely, the Seagull-Deep Extreme Learning Machine (ISOA-DELM) model and the Seagull-Kernel Extreme Learning Machine (ISOA-KELM) model for training and prediction.

[0062] In a preferred embodiment, the transformer data includes gas concentration data, transformer operation data and environmental data, and the samples of historical transformer data and the corresponding simulation results are divided into training sets and test sets according to a preset ratio to train the prediction model of dissolved gas concentration in transformer oil.

[0063] In an optional implementation, in this embodiment, the transformer data may be divided into a training set and a test set in a ratio of 9:1.

[0064] In another optional implementation, the transformer data may be divided into a training set, a test set, and a validation set in a ratio of 7:2:1 to train and validate the prediction model.

[0065] In an optional embodiment, data preprocessing includes normalizing the transformer data. In order to avoid the problem of large prediction errors caused by the large difference in magnitude between the training set and the test set during network model prediction, the data is normalized. The normalization formula is as follows:

[0066]

[0067] Where: x n is the original data, x min is the minimum value in the data, x max is the maximum value in the data, X n is the normalized data.

[0068] In another optional embodiment, data preprocessing may also include data cleaning, data discretization, etc. Data cleaning includes processing missing values, processing noisy data, removing duplicate data, etc., and can be done by deleting records with missing values, filling them with specific values ​​(such as mean, median or mode), or using interpolation methods to fill missing values, smoothing outliers or noise in the data through binning, regression, clustering, etc., and identifying and removing duplicate records in the data set; data discretization is to divide continuous data into intervals or categories for easy classification.

[0069] In an optional embodiment, the extreme learning machine model ELM is a single hidden layer feedforward neural network, and its network structure is as follows: Figure 2 As shown in Figure 2, ELM has the advantages of strong generalization and simple structure, and its training speed is greatly improved compared to other gradient descent training methods. In recent years, various variants and improved models based on ELM have been widely used in fault diagnosis, parameter optimization, prediction, and other fields, and have achieved good results.

[0070] Given any N samples (x i ,t i ), where the input data is x i =[x i1 ,x i2 ,…,x iN ] T ∈R n , the desired output is t i =[t i1 ,t i2 ,...,t iN ] T ∈R m , defining g(x) as the activation function and the number of hidden layer neurons, the standard SLFNs output function can be expressed as:

[0071]

[0072] Among them, b i represents the bias of the i-th hidden layer neuron, ω i =[ω i1 ,ω i2 ,...,ω in ] T represents the weight vector connecting the i-th hidden layer neuron to the input layer neuron, β i =[β i1 ,β i2 ,...,β im ] T is the weight vector connecting the i-th hidden layer neuron to the output layer neuron, o jRepresents the output prediction value of the network for the jth sample, x j represents the jth input sample.

[0073] When the randomly selected ω and b are close to the number of training samples, SLFNs with hidden layer neurons o j The output function The activation function g(x) can approximate the training sample with zero error when ω and b are randomly selected, as shown in the following formula:

[0074]

[0075] Among them, t j represents the target output of the j-th sample above;

[0076] The ELM output function of N random samples is expressed as:

[0077]

[0078] Rewriting the above formula in matrix form can be expressed as:

[0079] Hβ=T

[0080] The neural network construction function of the ELM algorithm can also be expressed as:

[0081] f(x)=h(x)β

[0082] Where h(x) is the output of the i-th hidden layer neuron corresponding to the input x, β is the connection weight vector between the i-th hidden layer neuron and the output neuron, and H is the output matrix of the hidden layer in ELM, which is expressed as:

[0083]

[0084] Therefore, during the learning process, the network searches for the least squares solution of the linear system in the above formula β. The least squares solution is obtained by solving the connection weight β between the hidden layer and the output layer, which can be expressed as:

[0085] β min =||Hβ-T′||

[0086]

[0087] Among them, H + is the hidden layer output matrix, which is the Moore-Penrose generalized inverse matrix of H. H represents the output matrix of the hidden layer in ELM, and T' represents the corresponding real matrix of the hidden layer output matrix. Represents the output weight matrix (connecting the hidden layer and the output layer), expand the above formula We can get:

[0088]

[0089] Where O is the predicted target value vector. According to regression theory, by adding the positive constant 1 / C, the solution will be more stable and have better generalization ability. H T represents the transpose of the output matrix of the hidden layer in ELM.

[0090] It should be noted that the derivation steps of the above formula show that ELM consists of matrix multiplication. Compared with other algorithms, it has small computational complexity and fast learning speed. Therefore, based on ELM and improved, the ELM model is applied to the detection of dissolved gas concentration in oil. Transformer data is used as an example to verify the effectiveness of the improved model.

[0091] In a preferred embodiment, improving the extreme learning machine to a kernel extreme learning machine includes:

[0092] By comparing the extreme learning machine with the support vector machine, the kernel function is introduced into the extreme learning machine, and the extreme learning machine is improved into the kernel extreme learning machine;

[0093] The activation function of the hidden layer is replaced by the kernel function, the number of neurons in the hidden layer, the input parameters and the system bias are not considered, the kernel function is used to replace the feature mapping function, and the penalty coefficient in the kernel coefficient and the generalized inverse matrix are set.

[0094] It should be noted that the present invention can further enhance the generalization ability and stability of ELM. By comparing the principles of ELM and support vector machine (SVM), the kernel function is introduced into ELM to obtain better regression prediction accuracy, and the extreme learning machine is improved into a kernel extreme learning machine.

[0095] In an optional embodiment, a kernel function is introduced into the ELM to obtain better regression prediction accuracy, which is expressed as:

[0096]

[0097] Ω ELM =exp(-γ||x i -x j || 2 )

[0098] Among them, ΩELM is the selected kernel function and N is the dimension of the input layer.

[0099] By applying the kernel function instead of the activation function of the hidden layer, there is no need to consider the number of neurons in the hidden layer, input parameters and system bias. The ELM implemented based on the kernel function in the first formula above is called the kernel extreme learning machine (KELM), which has better generalization performance and stability than the original extreme learning machine. KELM uses the kernel function instead of the feature mapping function, so the output result of the neural network is independent of the input weights of the hidden layer and the hidden neuron threshold, and only depends on the kernel function.

[0100] The kernel function in this embodiment adopts the Gaussian kernel function as shown in the second formula above. For the KELM algorithm, although the dimension of the high-dimensional space does not need to be set manually, the kernel coefficient γ and the penalty coefficient C in the generalized inverse matrix will affect the performance of the model. The kernel coefficient determines the effect of the nonlinear mapping of the input sample to the high-dimensional feature space, and the penalty coefficient represents the balance between the minimum fitting error and the complexity of the model.

[0101] In another optional implementation, the kernel function type also includes a linear kernel, a Gaussian RBF kernel, a polynomial kernel, etc. The choice of the kernel function has a direct impact on the performance of the model.

[0102] In a preferred embodiment, improving the extreme learning machine to a deep extreme learning machine includes:

[0103] Multiple extreme learning machines are stacked to form a deep extreme learning machine, which maps data features by stacking multiple hidden layers and adopts a hierarchical unsupervised training model;

[0104] During the training process of the deep extreme learning machine network model, the training data is used as the output of the first layer of extreme learning machine, the output of the hidden layer in the deep extreme learning machine model is used as the input data of the second layer of extreme learning machine, and the training of each layer of extreme learning machine is completed.

[0105] In an optional embodiment, the deep extreme learning machine is a deep learning network composed of multiple ELM stacks. Due to the stacking of multiple hidden layers, DELM can better map data features, improve the model prediction accuracy, and increase the generalization ability of the model. DELM uses hierarchical unsupervised training to minimize reconstruction errors, so that the input data and output data in the model are as similar as possible. The deep extreme learning machine model is as follows: Figure 4 shown.

[0106] During the training process of the deep extreme learning machine network model, the training data is used as the output of the first layer ELM, so that the weight β1 can be output. Then the output of the hidden layer in the DELM model is used as the input data of the second layer ELM. The training of each layer of ELM is completed according to this method, which greatly improves the prediction accuracy of the DELM model.

[0107] In another optional embodiment, the method for detecting dissolved gas concentration in oil can also be solved by using other network models, for example, a particle swarm optimization-long short-term memory network (PSO-LSTM) model, a deep belief network (DBN) model, or a convolutional neural network (CNN) model. For example, a particle swarm optimization-long short-term memory network model takes a characteristic gas concentration sequence as input, and uses a particle swarm optimization algorithm to iteratively optimize the relevant hyperparameters of the LSTM network.

[0108] In a preferred embodiment, the Seagull optimization algorithm includes:

[0109] The seagull optimization algorithm is divided into the migration phase and the predation phase;

[0110] During the migration phase, the convergence factor controls the latest position of the seagull. The convergence factor is a Sigmoid function. The value of the convergence factor is dynamically adjusted according to the current number of iterations, so that the seagull optimization algorithm performs a global search in the early stage and switches to a local search in the later stage. The convergence factor is used to control the seagull to fly to the optimal position.

[0111] During the hunting phase, the gulls attack their prey using spiral flight patterns;

[0112] The fitness function is used to update the fitness value of the current global optimal position at each iteration, and it is determined whether to update the current position and the global optimal position to obtain the optimal solution and make the population update continuous.

[0113] In order to ensure the stability of the Seagull algorithm during global search, the following three points still need to be noted during the migration phase:

[0114] (1) Collision avoidance: To avoid collisions between seagulls, the latest position of the seagull is controlled by the convergence factor A, which is expressed as:

[0115]

[0116] in, represents the individual position of the seagull after avoiding collision s times; Indicates the position after x iterations; f c It is a hyperparameter with a default value of 2. The convergence factor A decreases as the number of iterations x increases. iteration is the maximum number of iterations.

[0117] (2) Moving towards the global optimal position: Under the premise of ensuring that there is no collision between adjacent seagulls, the seagulls all fly towards a better position, which can be expressed as:

[0118]

[0119] B=2·A2 ·rd

[0120] in, is the global optimal position direction; is the global optimal position; B is the number of random checks, which is used to adjust the relationship between migration and predation; rd∈[0,1] obeys uniform distribution.

[0121] (3) Approaching the global optimal position: After an individual seagull finds the global optimal position, the remaining seagulls will also move towards the optimal position, which can be expressed as:

[0122]

[0123] in: It represents the comprehensive displacement of the seagull after the collision avoidance process and the process of approaching the optimal position, and it represents the latest position of the seagull.

[0124] During the migration phase, seagulls attack their prey in a spiral flight. This small-scale hunting behavior has a strong local optimization ability. The specific position of the seagull in the Cartesian coordinate system is expressed as:

[0125] x=r·cos(k)

[0126] y=r·sin(k)

[0127] z=r·k

[0128] r=u·e kv

[0129] Among them, r represents the spiral radius of the seagull; k is a random number between [0, 2π]; the two hyperparameters v and u are used to control the spiral flight radius of the seagull, and the value is usually 1.

[0130] According to the above formula, the latest position of the whole seagull is finally obtained, which is expressed as:

[0131] P s =(D s ·x·y·z)+P bs (x)

[0132] As can be seen from the formulas derived for the migration and predation phases, the Seagull algorithm shifts from global search to local search as the linear convergence factor A decreases linearly. However, in reality, the transition between global and local search is not a simple linear one. Therefore, to improve the algorithm's optimization capabilities and maintain a continuous global optimum, this embodiment improves the traditional Seagull algorithm.

[0133] In an optional implementation, in order to improve the optimization ability of the algorithm, this embodiment uses the Sigmoid function as the convergence factor A s, which enables the algorithm to smoothly transition from the early global search to the later local search, and also speeds up the algorithm's optimization speed, which can be expressed as:

[0134]

[0135] P s (x) = A s ·Z+P bs (x)

[0136] Z=|P bs (x)+2A·rd·(P bs (x)-P b (x))|·x·y·z

[0137] To avoid the optimal solution appearing during the iteration process, the Seagull algorithm needs to perform a global search after all population iterations are completed, resulting in discontinuous global updates and reducing the algorithm's optimization speed. The improved Seagull algorithm uses the following formula to update the global optimum, making the population update continuous, expressed as:

[0138]

[0139] The improved heuristic intelligent Seagull algorithm is introduced into the deep extreme learning machine (DELM) and the kernel extreme learning machine (KELM) to form the Seagull-deep extreme learning machine (ISOA-DELM) model and the Seagull-kernel extreme learning machine (ISOA-KELM) model.

[0140] In a preferred embodiment, ISOA is applied to optimize the single models ELM, DELM and KELM in combination with the gas dissolution data in transformer oil of the present invention, and a basic process is established, such as Figure 5 As shown in Figure 2, the prediction model for dissolved gas concentration in transformer oil includes:

[0141] Step 1: Select a machine learning model and improve the extreme learning machine to a kernel extreme learning machine and a deep extreme learning machine, set the number of variables, and set the upper and lower bounds of the search space;

[0142] Step 2: Initialize the fitness, and set the fitness function to the minimum root mean square error (RMSE) obtained by training;

[0143] Step 3: Use fitness as food and sort the sites from low to high according to fitness to obtain sites with more food sources;

[0144] Step 4: Calculate the convergence factor A through step S102 sThe value of is used to avoid collision and update to the global optimal position. After the food position is determined, the position of each individual is updated through the improved seagull algorithm formula, and the updated seagull fitness is calculated;

[0145] Step 5: Obtain the training set and test set training model, compare the fitness obtained in each iteration with the optimal fitness obtained in the previous iteration, and update the global optimal fitness;

[0146] According to the preset number of iterations, steps three to five are repeated continuously to obtain a prediction model for the concentration of dissolved gas in transformer oil.

[0147] In another optional embodiment, the combined model is used to solve the prediction of dissolved gas concentration in transformer oil. However, there are still aspects that can be improved for the optimization algorithm in the first half of the combined model. For example, the cuckoo optimization algorithm, the gray wolf optimization algorithm, etc., these algorithms will inevitably fall into the problem of local optimal solution when switching from global search to local search, but in fact, the conversion between global search and local search is not a simple linear conversion. Therefore, in order to improve the algorithm's optimization ability and global optimal continuous optimization, the algorithm in the combined model can be improved in a targeted manner for different problems or research directions.

[0148] In an optional embodiment, the prediction results of the model are analyzed and discussed from different perspectives, and the feasibility and effectiveness of each model are analyzed using the following evaluation criteria. The evaluation criteria are divided into the coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), training time, and test time, expressed as:

[0149]

[0150] Among them, n is the number of prediction samples; P Mi and are the average values ​​of the parameters of the i-th experimental sample, P Pi The network model estimates the parameters of the i-th experimental sample, and the coefficient of determination R 2 Evaluate the goodness of fit of the model, R 2 The closer it is to 1, the better the fitting effect.

[0151] In a preferred embodiment, outputting the prediction result of dissolved gas in transformer oil and the fault type includes:

[0152] Gas concentration data includes hydrogen, carbon monoxide, carbon dioxide, and acetylene; transformer operation data includes load, temperature, and oil temperature; environmental data includes ambient temperature and humidity;

[0153] Gas concentration data, transformer operation data, and environmental data are used as input data for the prediction model of dissolved gas concentration in transformer oil. The same gas concentration is predicted using multiple evaluation indicators, and the most accurate model output is selected. The predicted value of the gas concentration, concentration development trend, and possible fault type are also output.

[0154] It should be noted that because the transformer fault type, fault location, and energy density at the fault location are closely related to the composition and concentration of dissolved gases in the oil, potential transformer faults can be characterized by analyzing the changing trends of the dissolved gas composition, concentration, and concentration ratio in the oil. The improved prediction model is applied to gas concentration detection, using historical gas concentration data (such as hydrogen (H2), carbon monoxide (CO), carbon dioxide (CO2), and acetylene (C2H2) concentrations), transformer operating data (load, temperature, oil temperature, etc.), and environmental data (ambient temperature, humidity) as input data for the model. The predicted gas concentration, concentration development trend, and possible fault type are used as output. This constructs a new type of intelligent predictive protection that provides early warning of potential equipment problems before an accident occurs.

[0155] Specifically, this embodiment establishes ELM, DELM, KELM, SOA-DELM, SOA-KELM, ISOA-DELM and ISOA-KELM models to predict gas concentrations respectively. Reasonable parameter setting is very important for obtaining an accurate gas prediction model.

[0156] The Seagull optimization algorithm uses a population size of 20 and an iteration count of 100. The number of hidden layer neurons must be determined in the underlying ELM and DELM models. The number of hidden layer neurons is selected through trial and error and, during testing, is set at 100. In the hidden layer of the basic KELM model, the input dataset samples have a high dimensionality, so the kernel parameter γ to be optimized needs to be expanded to the range [0.1, 10] to obtain a more accurate prediction model.

[0157] During the prediction process, several different models predict the same gas concentration according to the six evaluation indicators in step S102 (coefficient of determination, root mean square error, mean absolute error and mean absolute percentage error, training time and test time). The model output with the most accurate prediction is selected as the main model, and the gas concentration trend results are output.

[0158] During normal operation of a forced oil circulation air-cooled transformer, a small amount of gas is dissolved in the transformer oil. Once a fault occurs, the composition of the dissolved gas in the oil will change significantly. Transformer oil contains a variety of gases, such as hydrogen (H2), carbon monoxide (CO), carbon dioxide (CO2), methane (CH4), ethane (C2H6), acetylene (C2H2) and ethylene (C2H4). Accurately predicting the development trend of dissolved gas concentration in transformer oil can provide an important basis for production maintenance and insulation status assessment of transformers.

[0159] During the hydrogen and acetylene prediction process, if ISOA-KELM or ISOA-DELM demonstrates higher accuracy and lower error than other models, the model with higher accuracy will be selected as the primary output model. Subsequently, when predicting gas concentration trends, if hydrogen and acetylene concentrations are rising above normal, early warning of gas concentrations will be issued, and internal transformer problems will be analyzed and diagnosed. Typically, the gases produced by transformer oil under arcing are primarily hydrogen and acetylene. The increasing concentration trend of these two gases in the predictions further indicates an internal transformer problem, namely, excessive arcing. In this case, human intervention can be implemented to control the development of transformer failures. The goal is to detect and address faults before they become accidents, thereby enhancing power system stability.

[0160] It should be noted that the present invention has the following effects:

[0161] 1. To avoid large prediction errors caused by large differences in the magnitude of the training and test sets during network model prediction, the applied data is normalized. The main purpose is to eliminate differences in the numerical size and range of different features, so that different features have a relatively consistent scale during model training, which helps improve model training results and convergence speed.

[0162] 2. The Deep Extreme Learning Machine is a deep learning network composed of multiple extreme learning machines. Due to the stacking of multiple hidden layers, the Deep Extreme Learning Machine can better map data features, improve the model's prediction accuracy, and increase the model's generalization ability. The Deep Extreme Learning Machine uses layered unsupervised training, which can minimize reconstruction errors, thereby making the input data and output data in the model as similar as possible, thereby improving prediction accuracy.

[0163] 3. The kernel extreme learning machine maps data to a high-dimensional space through the kernel function, can automatically learn the complex characteristics of the data, and has better generalization ability when dealing with nonlinear problems. The selection of the kernel function and parameter adjustment can control the complexity of the model to a certain extent, so that the kernel extreme learning machine can achieve better generalization performance and computational efficiency under different data distributions.

[0164] 4. The proposal of the Seagull Optimization Algorithm (SOA) is more inspired by the migration and attack behavior of seagulls in nature. The traditional Seagull algorithm will linearly progress with the linear convergence factor. When the algorithm switches from global search to local search, it will inevitably fall into the problem of local optimal solution. However, in fact, the conversion between global search and local search is not a simple linear conversion. Therefore, in order to improve the algorithm's optimization ability and global optimal continuous optimization, improvements have been made to the convergence function and parameter iteration of the Seagull algorithm. Therefore, an Improved Seagull Optimization Algorithm (ISOA) is proposed.

[0165] Neural networks can effectively handle complex nonlinear relationships and are suitable for detecting changing trends in gas concentrations within equipment. The proposed improved extreme learning machine model reduces overfitting and enhances the model's ability to process complex data, thereby improving the accuracy of gas concentration detection.

[0166] The improved extreme learning machine model is derived and improved based on the ELM model. Therefore, the improved model has the advantages of short training time and fast response speed of the ELM model, and is suitable for subsequent online monitoring systems based on the improved extreme learning machine model.

[0167] Before inputting the actual data of the transformer into the improved extreme learning machine model, applying normalization processing can suppress noise and enhance the improved model's ability to handle noise and abnormal data.

[0168] The improved Seagull optimization algorithm can dynamically adjust model parameters according to new data, is not easily trapped in local optimality and thus affects detection accuracy, effectively adapts to changes in gas concentration in oil, and maintains the accuracy of long-term monitoring, which is a manifestation of strong adaptive ability.

[0169] The above is a schematic diagram of a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to this embodiment. It should be noted that the technical solution of this system for detecting dissolved gas concentration in oil based on an improved extreme learning machine and the technical solution of the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine share the same concept. For details not described in detail in the technical solution of the system for detecting dissolved gas concentration in oil based on an improved extreme learning machine in this embodiment, please refer to the description of the technical solution of the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine.

[0170] Example 2

[0171] This embodiment provides a system for detecting dissolved gas concentration in oil based on an improved extreme learning machine, comprising:

[0172] Acquisition module, used to obtain transformer data and perform data preprocessing;

[0173] A construction module is used to build a prediction model for dissolved gas concentration in transformer oil by improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine, and combining it with the Seagull optimization algorithm, and then train the model;

[0174] The output module is used to input the dissolved gas concentration prediction model in the transformer oil based on the transformer data, and output the prediction results of the dissolved gas in the transformer oil and the fault type.

[0175] Example 3

[0176] This embodiment provides an electronic device suitable for detecting dissolved gas concentration in oil based on an improved extreme learning machine, including:

[0177] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the method for detecting dissolved gas concentration in oil based on the improved extreme learning machine as proposed in the above embodiment.

[0178] This embodiment further provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine as proposed in the above embodiment is implemented.

[0179] The storage medium proposed in this embodiment and the method for detecting the concentration of dissolved gas in oil based on the improved extreme learning machine proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0180] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer's floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0181] Example 4

[0182] Reference Figure 6-Figure 7 , which is an embodiment of the present invention, provides a cumulative climbing height calculation method. In order to verify its beneficial effects, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0183] Since there are many dissolved gases in oil, the prediction effect of the model is demonstrated by taking the C02 content as an example, and the existing model is used to predict the C02 content.

[0184] The error box plots of predicting C02 content using a single model and a combined model are shown in the figure below. Figure 6 、 7 As shown in the figure, the horizontal axis represents the model name and the vertical axis represents the prediction error. The absolute errors of the ISOA-DELM and ISOA-KELM models are mostly within [-30, 30]. The accuracy of the prediction directly affects the prediction results. The ISOA-DELM and ISOA-KELM models have good prediction accuracy and can meet the higher requirements for dissolved gas concentration detection accuracy in oil.

[0185] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting dissolved gas concentration in oil based on an improved extreme learning machine, characterized in that: include: Obtain transformer data and perform data preprocessing; By improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine, respectively, combined with the Seagull optimization algorithm, a prediction model for dissolved gas concentration in transformer oil was constructed and trained. Based on the transformer data, the data is input into the transformer oil dissolved gas concentration prediction model, and the transformer oil dissolved gas concentration prediction result and fault type are output.

2. The method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to claim 1, wherein: The transformer data includes gas concentration data, transformer operation data and environmental data, and samples of historical transformer data and corresponding simulation results are divided into training sets and test sets according to a preset ratio to train the prediction model for dissolved gas concentration in transformer oil.

3. The method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to claim 1, wherein: Improving the extreme learning machine to the kernel extreme learning machine includes: By comparing the extreme learning machine with a support vector machine, a kernel function is introduced into the extreme learning machine, and the extreme learning machine is improved into a kernel extreme learning machine; The activation function of the hidden layer is replaced by the kernel function, the number of neurons in the hidden layer, input parameters and system bias are not considered, the kernel function is used to replace the feature mapping function, and the kernel coefficient and the penalty coefficient in the generalized inverse matrix are set.

4. The method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to claim 3, wherein: Improving the extreme learning machine to a deep extreme learning machine includes: Multiple extreme learning machines are stacked to form a deep extreme learning machine, which maps data features by stacking multiple hidden layers and adopts a hierarchical unsupervised training model; During the training process of the deep extreme learning machine network model, the training data is used as the output of the first layer of extreme learning machine, the output of the hidden layer in the deep extreme learning machine model is used as the input data of the second layer of extreme learning machine, and the training of each layer of extreme learning machine is completed.

5. The method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to claim 4, wherein: The Seagull optimization algorithm includes: The seagull optimization algorithm is divided into a migration phase and a predation phase; During the migration phase, the latest position of the seagull is controlled by the convergence factor, which is a Sigmoid function. The value of the convergence factor is dynamically adjusted according to the current number of iterations, so that the seagull optimization algorithm performs a global search in the early stage and switches to a local search in the later stage. The convergence factor is used to control the seagull to fly to the optimal position. During the described hunting phase, the gull attacks its prey in a spiral flight pattern; The fitness function is used to update the fitness value of the current global optimal position at each iteration, and it is determined whether to update the current position and the global optimal position to obtain the optimal solution and make the population update continuous.

6. The method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to claim 5, wherein: Constructing the prediction model of dissolved gas concentration in transformer oil includes: Step 1: Select a machine learning model and improve the extreme learning machine to a kernel extreme learning machine and a deep extreme learning machine, set the number of variables, and set the upper and lower bounds of the search space; Step 2: Initialize the fitness function, and set the fitness function to the minimum root mean square error obtained by training calculation; Step 3, using the fitness as food, and sorting the sites from low to high according to the fitness, so as to obtain sites with more food sources; Step 4: Calculate the value of the convergence factor to avoid collisions, update to the global optimal position, determine the food location, update the position of each individual, and calculate the updated seagull fitness; Step 5: Obtain the training model of the training set and the test set, compare the fitness obtained in each iteration with the optimal fitness obtained in the previous iteration, and update the global optimal fitness; According to a preset number of iterations, steps three to five are continuously repeated to obtain the prediction model for the concentration of dissolved gas in transformer oil.

7. A method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to any one of claims 3 to 6, characterized in that: Output transformer oil dissolved gas prediction results and fault types include: The gas concentration data includes hydrogen, carbon monoxide, carbon dioxide, and acetylene; the transformer operation data includes load, temperature, and oil temperature; the environmental data includes ambient temperature and humidity; The gas concentration data, transformer operation data and environmental data are used as input data of the prediction model for dissolved gas concentration in transformer oil. The same gas concentration is predicted using multiple evaluation indicators, and the most accurate model output is selected. The predicted value of the gas concentration, concentration development trend and fault type are also output.

8. A system for detecting dissolved gas concentration in oil based on an improved extreme learning machine, applying the method for detecting dissolved gas concentration in oil based on an improved extreme learning machine according to any one of claims 1 to 7, characterized in that: include, Acquisition module, used to obtain transformer data and perform data preprocessing; A construction module is used to build a prediction model for dissolved gas concentration in transformer oil by improving the extreme learning machine into a core extreme learning machine and a deep extreme learning machine, and combining it with the Seagull optimization algorithm, and then train the model; The output module is used to input the transformer oil dissolved gas concentration prediction model based on the transformer data, and output the transformer oil dissolved gas concentration prediction result and fault type.

9. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for detecting the concentration of dissolved gas in oil based on an improved extreme learning machine as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that It stores computer-executable instructions, which, when executed by a processor, implement the steps of a method for detecting dissolved gas concentration in oil based on an improved extreme learning machine as described in any one of claims 1 to 7.

Images