Method for predicting concentration of dissolved gas in transformer oil based on EVO-VMD-LSSVM

Abstract

This paper presents a method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM. It combines the Energy Valley Optimization (EVO) algorithm, Variable Mode Decomposition (VMD), and Least Squares Support Vector Machine (LSSVM). The VMD method decomposes the original dissolved gas concentration signal into several Intrinsic Mode Functions (IMFs) with different frequency characteristics, extracting multi-scale features of the signal. Multiple factors, including transformer load, operating oil temperature, ambient temperature, and operating time, are incorporated into the model. Based on the correlation of feature parameters, an LSSVM prediction model is established for each IMF, and the EVO algorithm is used to optimize the LSSVM model parameters, improving prediction accuracy. The prediction results of each IMF are reconstructed to obtain the final predicted dissolved gas concentration value. The prediction performance is significantly better than traditional methods, with evaluation indicators such as mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²) all showing high prediction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring technology, and in particular to a method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM. Background Technology

[0002] In power systems, transformers, as key equipment, undertake important tasks such as voltage transformation, power distribution, and transmission. Their operating status directly affects the safety and stability of the entire power grid. To ensure reliable transformer operation, timely and accurate monitoring of their status is crucial. Dissolved Gas Analysis (DGA) of transformer oil is one of the most widely used transformer condition monitoring technologies. This technology analyzes dissolved gases (such as hydrogen) in transformer oil. methane ethylene Acetylene ethane carbon monoxide ,carbon dioxide Analyzing the concentration and composition of [a substance] can help infer the type and severity of internal faults in transformers, providing important information for transformer maintenance and repair.

[0003] However, traditional DGA methods rely heavily on empirical formulas and expert judgment, which inherently involve subjectivity and limitations. Because the dissolved gas concentration in transformer oil is influenced by various factors, its time-series data contains numerous components with different scales and trends, exhibiting strong nonlinear and non-stationary characteristics. This makes it difficult for traditional DGA methods to meet the high requirements of modern power systems for transformer fault prediction in terms of prediction accuracy and real-time performance.

[0004] To overcome the shortcomings of traditional DGA methods, data-driven prediction methods have gradually attracted researchers' attention in recent years. Among them, Least Squares Support Vector Machine (LSSVM), as an efficient machine learning algorithm, has been widely used in various prediction models due to its advantages in handling small samples and nonlinear problems. However, directly applying LSSVM to predict dissolved gas concentration in transformer oil still faces many challenges. On the one hand, dissolved gas concentration signals usually have strong non-stationarity and multi-scale characteristics, making it difficult to effectively model them using a single machine learning model. On the other hand, the performance of LSSVM models depends on the selection of parameters, and how to optimize model parameters to improve prediction accuracy is an urgent problem to be solved.

[0005] For example, CN117894389A discloses a method for predicting dissolved gas concentration data in transformer oil based on SSA-optimized VMD and LSTM, including the following steps: S1, using the SSA (Sparrow Search Algorithm) to optimize the number of decomposition modes K and the penalty parameter α of the VMD (Variational Mode Decomposition) model; S2, using the optimized VMD model to decompose the monitored dissolved gas concentration data in transformer oil into multiple subsequences; S3, using the SSA algorithm to optimize the number of hidden layer neurons, iteration count, and learning rate parameters of the LSTM (Long Short-Term Memory) network model; S4, using the optimized LSTM model to predict each subsequence, superimposing the results to obtain numerical prediction results and evaluating them; S5, constructing a set of prediction results at each time point, calculating the adaptive optimal window width for each set, and fitting the KDE to the predicted interval of dissolved gas concentration in transformer oil based on the optimal window width and Gaussian kernel function. This method optimizes variational mode decomposition parameters based on the sparrow search algorithm to perform mode decomposition on monitoring data, and uses SSA to optimize the long short-term memory network model, which can improve prediction accuracy. However, the structure is relatively complex, requiring high computational resources and exhibiting slow convergence speed. Especially when dealing with complex high-dimensional problems, the optimization efficiency may be insufficient. When data diversity is insufficient or the sample size is small, the generalization ability may be limited, and overfitting is prone to occur. At the same time, current models or methods for predicting dissolved gases in transformer oil do not consider the possible influence of other external factors such as transformer load, oil temperature, ambient temperature, and operating time on the concentration of dissolved gases in the oil, resulting in insufficient prediction stability and anti-interference ability.

[0006] In summary, while existing LSSVM-based methods for predicting dissolved gas concentrations in transformer oil have improved prediction performance to some extent, they still suffer from insufficient prediction accuracy, difficulty in optimizing model parameters, and poor resistance to external interference. Therefore, there is an urgent need for a new method that can more effectively extract multi-scale features of dissolved gas concentration signals, integrate internal and external factors to construct and optimize LSSVM model parameters, thereby improving prediction accuracy and real-time performance. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM, which solves the technical problems existing in the prediction of dissolved gas concentration in transformer oil in the power system, especially the challenges faced by the traditional DGA method, which relies on empirical formulas and expert judgment, resulting in subjectivity and limitations, and the prediction model based on LSSVM when dealing with dissolved gas concentration signals with non-stationary and multi-scale characteristics.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM, comprising the following steps:

[0009] Step 1: Data preprocessing. Preprocess the dissolved gas concentration data in transformer oil, as well as parameters such as transformer load, operating oil temperature, ambient temperature, and operating time, including noise reduction and normalization.

[0010] Step 2: VMD decomposition. The preprocessed dissolved gas concentration signal is input into the VMD algorithm to decompose it into several intrinsic mode functions (IMFs) to extract the multi-scale features of the signal.

[0011] Step 3: LSSVM Modeling. For each IMF, select gas concentration parameters (IMF sequences) with a correlation greater than 0.6, along with external parameters such as transformer load, oil temperature, ambient temperature, and operating time, as inputs to the prediction model. Establish an LSSVM regression model, and include the penalty parameter C and kernel function parameters as the main parameters of the LSSVM model. ;

[0012] Step 4: EVO (Energy Valley Optimizer) optimization is a metaheuristic algorithm used to solve global optimization problems. It simulates the process by which unstable particles change the ratio of neutrons to protons in the particle through α, β, and γ decay and random movement, causing the unstable particles to move towards the stable zone or energy valley to increase their stability level. This solves the optimization problem. The EVO algorithm is used to optimize the parameters of the LSSVM model to improve the prediction accuracy.

[0013] Step 5: Reconstruct the prediction results. Reconstruct the prediction results of each IMF to obtain the final predicted value of dissolved gas concentration.

[0014] Step 6: Model validation, verify the model's predictive performance using real data.

[0015] In a preferred embodiment, the data preprocessing in Step 1 further includes filling in the missing data using linear interpolation.

[0016] In the preferred embodiment, the VMD decomposition in Step 2 specifically includes the following steps:

[0017] Step 2.1: Initialization, setting the initial frequency center and bandwidth parameters;

[0018] Step 2.2: Iterative optimization, by iteratively updating the frequency center and bandwidth parameters, to minimize the bandwidth of each mode;

[0019] Step 2.3: Modal reconstruction. Based on the optimized frequency center and bandwidth parameters, bandpass filtering is performed on the signal to obtain each IMF.

[0020] In the preferred embodiment, the EVO optimization in Step 4 specifically includes the following steps:

[0021] Step 4.1: Initial Population Generation. A set of initial solutions is randomly generated in the parameter space to form the initial population; the population containing N candidate solutions is initialized, with each candidate solution being randomly generated in the search space:

[0022] (1)

[0023] In the formula, For the first The first particle 3D spatial location, , The first The first particle Upper and lower bounds of a dimensional variable; A random number in the range [0,1]; The particle size in the search space; Let be the dimension of the decision variables.

[0024] Step 4.2: Calculate the fitness value, i.e., the neutron enrichment level of the particles, and the enrichment boundary of the particle swarm. Calculate the stability level of the particles :

[0025] (2)

[0026] (3)

[0027] In the formula, This represents the enrichment boundary of particles in the search space. For the first Neutron enrichment level of individual particles; and These represent the fitness function values ​​for the particles with the best and worst stability, respectively.

[0028] Step 4.3: and Decay position update, if the first Stability level of individual particles Then it will happen and decay, Location of the decaying new particle:

[0029] (4)

[0030] In the formula, In the search space The position of the new particle generated by decay; For the first in the search space The current position of each particle; The position of the particle with the best stability; For the first One decision variable.

[0031] Location of the decaying new particle:

[0032] (5)

[0033] In the formula, In the search space The position of the new particle generated by decay; For the first The positions of the particles around each particle.

[0034] Step 4.4: Decay position update, if the first Stability level of individual particles Then it will happen decay, Location of the decaying new particle:

[0035] (6)

[0036] (7)

[0037] In the formula, In the search space The position of the new particle generated by decay; , Two random numbers in the range [0, 1]; This is the position of the particle's center.

[0038] use The decaying particle undergoes a process of updating its position at another location:

[0039] (8)

[0040] In the formula, This is the location of another new particle generated by decay in the search space; , For those in the range [0, 1] that are different , Two random numbers.

[0041] Step 4.5: Randomly move the position to update; if the neutron enrichment level is below the enrichment threshold... If the particle is captured by an electron or emitted by a positron, it will move towards a stable position, and its position will be updated as follows:

[0042] (9)

[0043] In the formula, To randomly generate new particle positions in the search space; It is a random number in the range [0, 1].

[0044] Step 4.6: Termination condition check. Determine if the termination condition is met. If the algorithm has reached the maximum number of iterations, exit the loop and output the result. Otherwise, return to step 4.2 and continue running.

[0045] In the preferred embodiment, the model validation in Step 5 uses mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) as evaluation indicators.

[0046] In a preferred embodiment, the mean squared error (MSE) is a measure of the average squared error between the predicted and actual values, and its expression is:

[0047] (10)

[0048] In the formula, The length of the data sequence; The sequence number position corresponding to each time step; For the first The actual values ​​of each parameter; For the first The predicted values ​​of each parameter.

[0049] In the preferred embodiment, the Mean Absolute Error (MAE) is a measure of the average absolute error between the predicted value and the actual value, and its expression is:

[0050] (11)

[0051] In the preferred embodiment, the coefficient of determination (R²) measures the goodness of fit of the model to the data, and its expression is as follows:

[0052] (12)

[0053] In the formula, This is the average of the measured values ​​of the parameter time series.

[0054] The method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM provided by this invention has the following beneficial effects:

[0055] 1. This invention solves the technical problems existing in the prediction of dissolved gas concentration in transformer oil in the current power system. In particular, it addresses the problem that the traditional DGA method mainly relies on empirical formulas and expert judgment, which has subjectivity and limitations and is difficult to meet the high accuracy and real-time requirements of modern power systems for transformer fault prediction.

[0056] 2. This invention proposes a prediction model that combines multi-factor analysis and feature fusion. It not only considers the time series characteristics of dissolved gas concentration in oil, but also introduces external factors such as transformer load, operating oil temperature, ambient temperature and operating time to conduct multivariate correlation analysis. This multi-factor fusion approach makes the model more comprehensively reflect the actual working conditions and improves the adaptability and accuracy of the prediction.

[0057] 3. This invention decomposes the original dissolved gas concentration signal into multiple IMFs through VMD decomposition, effectively extracting the multi-scale features of the signal. This solves the problem that dissolved gas concentration signals usually have strong non-stationarity and multi-scale features, making it difficult to effectively model them with a single machine learning model and difficult to handle nonlinear and non-stationary data.

[0058] 4. This invention utilizes the EVO algorithm to optimize the parameters of the LSSVM model, significantly improving the model's prediction accuracy and robustness, resulting in more accurate and reliable predictions; it also enhances the ability to handle nonlinear and high-dimensional complex relationships. LSSVM exhibits good generalization performance in small sample sizes, overcoming the shortcomings of neural network-based methods, such as long training times and high risk of overfitting.

[0059] 5. This invention organically combines VMD, LSSVM and EVO to form a highly efficient method for predicting dissolved gas concentration. It overcomes the technical bottlenecks faced by existing technologies when directly applying LSSVM to predict dissolved gas concentration in transformer oil, such as the difficulty in optimizing model parameters and the impact of external factors on prediction accuracy. It can effectively support real-time monitoring and fault prediction of dissolved gas concentration in transformer oil.

[0060] 6. This invention significantly improves the prediction accuracy of dissolved gas concentration by extracting multi-scale features through VMD decomposition and optimizing LSSVM model parameters through EVO.

[0061] 7. The method of the present invention can effectively extract the multi-scale features of dissolved gas concentration signals. Through VMD decomposition, the original signal is decomposed into several IMFs, each IMF representing the components of the signal in different frequency ranges, thereby accurately reflecting the nonlinear and non-stationary characteristics of the signal.

[0062] 8. The method of this invention enables real-time monitoring and fault prediction of dissolved gas concentration in transformer oil, providing a strong guarantee for the safe and stable operation of the power system;

[0063] 9. The method of the present invention provides a new and effective means for transformer condition monitoring and fault prediction, and has broad application prospects and practical value;

[0064] 10. This invention also possesses high flexibility and scalability, and can be adapted to meet the needs of different application scenarios, further enhancing its competitiveness in the field of dissolved gas concentration prediction in transformer oil. Attached Figure Description

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0066] Figure 1 This is a schematic diagram of the technical process of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments:

[0068] Example 1

[0069] like Figure 1 As shown, the method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM includes the following steps:

[0070] Step 1: Data preprocessing. Preprocess the dissolved gas concentration data in transformer oil, as well as parameters such as transformer load, operating oil temperature, ambient temperature, and operating time, including noise reduction and normalization.

[0071] Step 2: VMD decomposition. The preprocessed dissolved gas concentration signal is input into the VMD algorithm to decompose it into several intrinsic mode functions (IMFs) to extract the multi-scale features of the signal.

[0072] Step 3: LSSVM Modeling. For each IMF, select gas concentration parameters (IMF sequences) with a correlation greater than 0.6, along with external parameters such as transformer load, oil temperature, ambient temperature, and operating time, as inputs to the prediction model. Establish an LSSVM regression model, and include the penalty parameter C and kernel function parameters as the main parameters of the LSSVM model. ;

[0073] Step 4: EVO (Energy Valley Optimizer) optimization, using the EVO algorithm to optimize the parameters of the LSSVM model to improve prediction accuracy;

[0074] Step 5: Reconstruct the prediction results. Reconstruct the prediction results of each IMF to obtain the final predicted value of dissolved gas concentration.

[0075] Step 6: Model validation, verify the model's predictive performance using real data.

[0076] In this embodiment, the data preprocessing in Step 1 also includes filling in the missing data using linear interpolation.

[0077] Furthermore, the VMD decomposition in Step 2 specifically includes the following steps:

[0078] Step 2.1: Initialization, setting the initial frequency center and bandwidth parameters;

[0079] Step 2.2: Iterative optimization, by iteratively updating the frequency center and bandwidth parameters, to minimize the bandwidth of each mode;

[0080] Step 2.3: Modal reconstruction. Based on the optimized frequency center and bandwidth parameters, bandpass filtering is performed on the signal to obtain each IMF.

[0081] Furthermore, the EVO optimization in Step 4 specifically includes the following steps:

[0082] Step 4.1: Initial population generation. A set of initial solutions is randomly generated in the parameter space to form the initial population.

[0083] Step 4.2: Energy calculation, calculate the energy value of each individual based on the prediction error;

[0084] Step 4.3: Individual update, generating new individuals through selection, crossover, and mutation operations;

[0085] Step 4.4: Energy Valley Search, through local search, further reduces the individual's energy value;

[0086] Step 4.5: Termination condition check. Determine if the termination condition is met. If it is met, terminate the algorithm.

[0087] Furthermore, the model validation in Step 5 uses mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) as evaluation metrics.

[0088] Furthermore, the mean squared error (MSE) measures the average squared error between the predicted and actual values, and its expression is:

[0089] (10)

[0090] In the formula, The length of the data sequence; The sequence number position corresponding to each time step; For the first The actual values ​​of each parameter; For the first The predicted values ​​of each parameter.

[0091] Furthermore, the Mean Absolute Error (MAE) measures the average absolute error between the predicted and actual values, and its expression is:

[0092] (11)

[0093] Furthermore, the coefficient of determination (R²) measures how well the model fits the data, and its expression is:

[0094] (12)

[0095] In the formula, This is the average of the measured values ​​of the parameter time series.

[0096] Example 2

[0097] In another preferred embodiment, based on Embodiment 1, the remaining steps are the same as in Embodiment 1. This embodiment further refines and specifies the EVO optimization step in Step 4:

[0098] Step 4.1: Initial Population Generation. A set of initial solutions is randomly generated in the parameter space to form the initial population; the population containing N candidate solutions is initialized, with each candidate solution being randomly generated in the search space:

[0099] (1)

[0100] In the formula, For the first The first particle 3D spatial location, , The first The first particle Upper and lower bounds of a dimensional variable; A random number in the range [0,1]; The particle size in the search space; Let be the dimension of the decision variables.

[0101] Step 4.2: Calculate the fitness value, i.e., the neutron enrichment level of the particles, and the enrichment boundary of the particle swarm. Calculate the stability level of the particles :

[0102] (2)

[0103] (3)

[0104] In the formula, This represents the enrichment boundary of particles in the search space. For the first Neutron enrichment level of individual particles; and These represent the fitness function values ​​for the particles with the best and worst stability, respectively.

[0105] Step 4.3: and Decay position update, if the first Stability level of individual particles Then it will happen and decay, Location of the decaying new particle:

[0106] (4)

[0107] In the formula, In the search space The position of the new particle generated by decay; For the first in the search space The current position of each particle; The position of the particle with the best stability; For the first One decision variable.

[0108] Location of the decaying new particle:

[0109] (5)

[0110] In the formula, In the search space The position of the new particle generated by decay; For the first The positions of the particles around each particle.

[0111] Step 4.4: Decay position update, if the first Stability level of individual particles Then it will happen decay, Location of the decaying new particle:

[0112] (6)

[0113] (7)

[0114] In the formula, In the search space The position of the new particle generated by decay; , Two random numbers in the range [0, 1]; This is the position of the particle's center.

[0115] use The decaying particle undergoes a process of updating its position at another location:

[0116] (8)

[0117] In the formula, This is the location of another new particle generated by decay in the search space; , For those in the range [0, 1] that are different , Two random numbers.

[0118] Step 4.5: Randomly move the position to update; if the neutron enrichment level is below the enrichment threshold... If the particle is captured by an electron or emitted by a positron, it will move towards a stable position, and its position will be updated as follows:

[0119] (9)

[0120] In the formula, To randomly generate new particle positions in the search space; It is a random number in the range [0, 1].

[0121] Step 4.6: Termination condition check. Determine if the termination condition is met. If the algorithm has reached the maximum number of iterations, exit the loop and output the result. Otherwise, return to step 4.2 and continue running.

[0122] Example 3

[0123] In another preferred embodiment, based on embodiments 1 and 2, see [reference needed]. Figure 1This embodiment details the method for predicting dissolved gas concentration in transformer oil based on EVO-VMD-LSSVM, with the specific steps as follows:

[0124] 1. Data Preprocessing

[0125] First, the dissolved gas concentration data in transformer oil is preprocessed; since the original data may contain noise and outliers, denoising is required. Next, missing data is filled in. In this embodiment, linear interpolation is used, assuming that the data changes steadily near the missing points, and the missing values ​​are estimated through linear interpolation. Finally, in order to eliminate the dimensional differences between different indicators and improve the convergence speed and prediction accuracy of model training, the data is normalized using the min-max normalization method, mapping the data to the interval [0, 1].

[0126] 2. VMD decomposition

[0127] The preprocessed dissolved gas concentration signal is input into the VMD algorithm for decomposition. The VMD algorithm is a variable mode decomposition method that can effectively stabilize nonlinear and non-stationary data. Through VMD decomposition, the original signal can be decomposed into several intrinsic mode functions (IMFs) with different frequency characteristics. The specific decomposition steps are as follows:

[0128] 2.1 Initialization: Set the initial frequency center and bandwidth parameters;

[0129] 2.2 Iterative optimization: Minimize the bandwidth of each mode by iteratively updating the frequency center and bandwidth parameters;

[0130] 2.3 Modal Reconstruction: Based on the optimized frequency center and bandwidth parameters, the signal is bandpass filtered to obtain each IMF.

[0131] In this embodiment, the VMD algorithm is selected through EVO optimization to determine the number of decomposition modes and the penalty coefficient, so as to further optimize the decomposition effect.

[0132] 3. LSSVM modeling

[0133] For each IMF (Integrated Gas Flow Factor), gas concentration parameters with a correlation greater than 0.6 and external parameters such as transformer load, oil temperature, ambient temperature, and operating time are selected as inputs to the prediction model to establish an LSSVM (Laser-Side Flow Vector Machine) regression model. The LSSVM model is an efficient machine learning algorithm suitable for handling small sample sizes and nonlinear problems. The main parameters of the LSSVM model include the penalty parameter C and the kernel function parameter. The model parameters are determined by training the model using historical data. In this embodiment, the radial basis function (RBF) is used as the kernel function.

[0134] 4. EVO optimization

[0135] The EVO algorithm is used to optimize the parameters of the LSSVM model to improve prediction accuracy. The EVO algorithm is an energy valley optimization algorithm that can find the global optimum. The specific steps are as follows:

[0136] 4.1 Initial Population Generation: A set of initial solutions is randomly generated in the parameter space to form the initial population;

[0137] 4.2 Energy Calculation: Calculate the energy value of each individual based on the prediction error. The smaller the energy value, the higher the fitness of the individual.

[0138] 4.3 Individual Position Update: New individuals are generated through selection, crossover, and mutation operations;

[0139] 4.4 Energy Valley Search: Further reduce the individual's energy value through local search;

[0140] 4.5 Termination Condition Judgment: Determine whether the termination condition is met. If it is met, terminate the algorithm; otherwise, return to the energy calculation step.

[0141] 4.6 The optimal penalty parameters and kernel function parameters of the LSSVM model can be found using the EVO algorithm. This significantly improves the model's prediction accuracy.

[0142] 5. Reconstruction of Prediction Results

[0143] The prediction results of each IMF are reconstructed to obtain the final predicted value of dissolved gas concentration; specifically, the prediction values ​​of each IMF are added together to obtain the reconstructed prediction value.

[0144] 6. Model Validation

[0145] The predictive performance of the model was verified using real data. The model's predictive accuracy and fitting ability were comprehensively evaluated using metrics such as mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). Experimental results show that the method proposed in this invention significantly outperforms traditional methods in terms of predictive accuracy.

[0146] Example 3

[0147] In another preferred embodiment, based on embodiments 1 and 2, see [reference needed]. Figure 1 To demonstrate the versatility and practicality of the prediction method of this invention, this embodiment provides another specific application scenario.

[0148] Unlike Example 2, this example uses different parameter settings and optimization strategies in the VMD decomposition and LSSVM modeling processes.

[0149] 1. Data Preprocessing

[0150] Similar to Example 2, the data is denoised, missing values ​​are filled, and normalized.

[0151] 2. VMD decomposition

[0152] The same VMD algorithm as in Example 2 was used for decomposition, but the selection of the number of decomposition modes and the penalty coefficient differed. Through multiple experiments, this example determined the optimal number of decomposition modes and the penalty coefficient to further improve the decomposition effect.

[0153] 3. LSSVM modeling

[0154] When building an LSSVM regression model, different penalty parameters C and kernel function parameters are used. Meanwhile, this embodiment also tried different kernel function types, such as polynomial kernel functions and sigmoid kernel functions, in order to find the kernel function type most suitable for the present invention.

[0155] 4. EVO optimization

[0156] Similar to Example 2, the parameters of the LSSVM model were optimized using the EVO algorithm; however, different strategies were employed in the individual update and energy valley search processes to further improve the optimization effect.

[0157] 5. Reconstruction of Prediction Results

[0158] The prediction results of each IMF are reconstructed to obtain the final predicted value of dissolved gas concentration.

[0159] 6. Model Validation

[0160] Similar to Example 2, the predictive performance of the model was verified using actual data, and the same evaluation metrics were used to comprehensively evaluate the model's prediction accuracy and fitting ability. The experimental results show that after adopting the parameter settings and optimization strategies of this example, the model's prediction accuracy and fitting ability were significantly improved, and the stability and reliability of the prediction results were also stronger, further verifying the diversity and practicality of the prediction method of this invention.

[0161] In a preferred embodiment, the data preprocessing in Step 1 further includes filling in missing data using linear interpolation. This setting can effectively reduce analysis errors caused by missing data and improve the accuracy of the model. At the same time, outliers are cleaned using the interquartile range method to ensure the quality of the input data and lay a solid foundation for subsequent steps.

[0162] In the preferred embodiment, the model validation in Step 5 uses mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) as evaluation indicators. The above settings aim to comprehensively measure the model's predictive performance. MSE evaluates the sum of squares of the deviations between the predicted and actual values, MAE directly reflects the average level of the prediction error, and R² reveals the model's ability to explain the fluctuations of variables.

[0163] In summary, the EVO-VMD-LSSVM-based method for predicting dissolved gas concentration in transformer oil proposed in this invention solves the technical problems existing in the prediction of dissolved gas concentration in transformer oil in current power systems. Specifically, it addresses the issue that traditional DGA methods rely heavily on empirical formulas and expert judgment, which are subjective and limited, making it difficult to meet the high accuracy and real-time requirements of modern power systems for transformer fault prediction. This invention innovatively introduces the Variable Mode Decomposition (VMD) method, decomposing the original signal into several intrinsic mode functions (IMFs) with different frequency characteristics, thereby achieving accurate extraction of multi-scale features of the signal. This method not only improves the accuracy of signal processing, but also... This invention also provides richer feature information for subsequent prediction models. Addressing the challenge of optimizing Least Squares Support Vector Machine (LSSVM) model parameters, this solution employs the Energy Valley Optimization (EVO) algorithm. As an emerging optimization technique, the EVO algorithm offers significant advantages in parameter search and optimization. Optimizing LSSVM model parameters using the EVO algorithm significantly improves the model's prediction accuracy and generalization ability, thereby achieving more accurate predictions of dissolved gas concentrations in transformer oil. The prediction model is constructed by combining multi-factor analysis and feature fusion, considering not only the time-series characteristics of dissolved gas concentrations in the oil but also external factors such as transformer load, operating oil temperature, ambient temperature, and operating time. This invention employs multivariate correlation analysis, a multi-factor fusion approach that allows the model to more comprehensively reflect actual operating conditions, improving the adaptability and accuracy of predictions. Furthermore, this invention organically combines VMD, LSSVM, and EVO methods to construct a novel comprehensive prediction model. This model not only integrates the advantages of each method but also achieves real-time monitoring and fault prediction of dissolved gas concentrations in transformer oil through synergistic effects. This comprehensive prediction model possesses high novelty in both theory and practice. By introducing the VMD method and EVO algorithm, this invention achieves technological innovations in signal processing, model construction, and parameter optimization. These innovations not only improve the accuracy of predictions but also... This invention improves both accuracy and efficiency, and provides a new technical approach for transformer condition monitoring and fault prediction. It organically combines VMD, LSSVM, and EVO methods to form a novel comprehensive prediction method. This method not only overcomes the limitations of traditional methods but also achieves more accurate prediction of dissolved gas concentrations in transformer oil through innovative combination. This invention has broad application prospects in the field of transformer condition monitoring and fault prediction. By monitoring changes in dissolved gas concentrations in transformer oil in real time, potential fault hazards can be detected promptly, providing strong protection for the safe and stable operation of power systems. Furthermore, this solution can provide reference and guidance for condition monitoring and fault prediction in other fields.This invention proposes an efficient and accurate scheme for predicting the concentration of dissolved gases in transformer oil. This not only promotes the development of transformer condition monitoring technology but also provides important technical support for the construction and operation of smart grids, possessing significant scientific value and practical application value.

Claims

1. A method for predicting the concentration of dissolved gases in transformer oil based on EVO-VMD-LSSVM, characterized in that, Includes the following steps: Step 1: Data preprocessing. The dissolved gas concentration data in transformer oil and external parameters are preprocessed, including noise reduction and normalization. External parameters include transformer load, operating oil temperature, ambient temperature, and operating time. Step 2: VMD decomposition. The preprocessed dissolved gas concentration signal is input into the VMD algorithm to decompose it into several intrinsic mode functions (IMFs) to extract multi-scale features of the signal. This includes the following steps: Step 2.1: Initialization, setting the initial frequency center and bandwidth parameters; Step 2.2: Iterative optimization. The VMD algorithm is selected through EVO optimization to determine the number of decomposed modes and the penalty coefficient, so as to further optimize the decomposition effect. The frequency center and bandwidth parameters are updated iteratively to minimize the bandwidth of each mode. Step 2.3: Mode reconstruction. Based on the optimized frequency center and bandwidth parameters, bandpass filtering is performed on the signal to obtain each IMF. Step 3: LSSVM Modeling. For each IMF, an LSSVM regression model is built. The main parameters of the LSSVM model include the penalty parameter C and the kernel function parameter. The specific process is as follows: for each IMF, select the gas concentration parameter IMF sequence with a correlation greater than 0.6 and external parameters as the input of the prediction model, and establish an LSSVM regression model. The external parameters include transformer load, operating oil temperature, ambient temperature, and operating time. Step 4: EVO optimization. The EVO algorithm is used to optimize the parameters of the LSSVM model to improve prediction accuracy. Specific steps include: Step 4.1: Initial population generation. A set of initial solutions is randomly generated in the parameter space to form the initial population. The population containing N candidate solutions is initialized, and each candidate solution is randomly generated in the search space. Step 4.2: Calculate the fitness value, i.e., the neutron enrichment level of the particles, and the enrichment boundary of the particle swarm. Calculate the stability level of the particles ; Step 4.3: and Decay position update, if the first Stability level of individual particles Then it will happen and decay, The location of the decaying new particle; Step 4.4: Decay position update, if the first Stability level of individual particles Then it will happen decay, The location of the decaying new particle; Step 4.5: Randomly move the position to update; if the neutron enrichment level is below the enrichment threshold... If the particle is captured by an electron or emitted as a positron, it will move toward a stable position. Step 4.6: Termination condition check. Determine if the termination condition is met. If the algorithm has reached the maximum number of iterations, exit the loop and output the result. Otherwise, return to step 4.2 and continue running; Step 5: Reconstruct the prediction results. Reconstruct the prediction results of each IMF to obtain the final predicted value of dissolved gas concentration. Step 6: Model validation, verify the model's predictive performance using real data.

2. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that: The data preprocessing in Step 1 also includes filling in missing data using linear interpolation.

3. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Step 4.1 initializes a population containing N candidate solutions, with each candidate solution randomly generated within the search space, as follows: （1）； In the formula, For the first The first particle 3D spatial location, , The first The first particle Upper and lower bounds of a dimensional variable; A random number in the range [0,1]; The particle size in the search space; Let be the dimension of the decision variables.

4. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Step 4.2 calculates the fitness value, which is the neutron enrichment level of the particles and the enrichment boundary of the particle swarm. Calculate the stability level of the particles The specific process is as follows: （2）； （3）； In the formula, This represents the enrichment boundary of particles in the search space. For the first Neutron enrichment level of individual particles; and These represent the fitness function values ​​for the particles with the best and worst stability, respectively.

5. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Step 4.3 and Decay position update, if the first Stability level of individual particles Then it will happen and decay, The positions of the decaying new particles are represented as follows: （4）； In the formula, In the search space The position of the new particle generated by decay; For the first in the search space The current position of each particle; The position of the particle with the best stability; For the first One decision variable; The positions of the decaying new particles are represented as follows: （5）； In the formula, In the search space The position of the new particle generated by decay; For the first The positions of the particles around each particle.

6. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Step 4.4 Decay position update, if the first Stability level of individual particles Then it will happen decay, The positions of the decaying new particles are represented as follows: （6）； （7）； In the formula, In the search space The position of the new particle generated by decay; , Two random numbers in the range [0, 1]; The location of the particle's center; use The decaying particle undergoes a process of updating its position at another location: （8）； In the formula, This is the location of another new particle generated by decay in the search space; , For those in the range [0, 1] that are different , Two random numbers.

7. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Step 4.5 involves random position updates; if the neutron enrichment level is below the enrichment threshold... If the particle is captured by an electron or emitted by a positron, it will move towards a stable position, and its position will be updated as follows: （9）； In the formula, To randomly generate new particle positions in the search space; It is a random number in the range [0, 1].

8. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that: The model validation in Step 5 uses mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) as evaluation metrics. MSE measures the average squared error between the predicted and actual values, and its expression is: （10）； In the formula, The length of the data sequence; The sequence number corresponding to each time step; For the first The actual values ​​of each parameter; For the first Predicted values ​​for each parameter; Mean Absolute Error (MAE) measures the average absolute error between predicted and actual values. Its expression is: （11）； The coefficient of determination, R², measures how well a model fits the data, and its expression is: （12）； In the formula, This is the average of the measured values ​​of the parameter time series.

Images