Oil chromatography online monitoring and early warning method and system based on trend prediction

Abstract

The invention discloses an oil chromatography online monitoring and early warning method and system based on trend prediction, and the method comprises the steps: obtaining continuous oil chromatography data from an online monitoring device of a target transformer, dividing the continuous oil chromatography data into a historical data set D1 and a current data sequence D2, and determining the historical data set D1 and the current data sequence D2 as original data; preprocessing the original data, and unifying the sampling frequency to a preset time interval; according to the preprocessed historical data set D1, constructing a mixed kernel function support vector regression SVR model to obtain a to-be-optimized parameter set, and performing parameter tuning on the to-be-optimized parameter set through a real number coding GA algorithm to obtain optimal parameters; the optimal parameters are substituted into a mixed kernel function support vector regression SVR model, an SVR model of the optimal parameters is obtained, the current data sequence D2 is input, and a gas concentration predicted value of the next period is obtained; and setting a trend early-warning threshold value, and when the gas concentration predicted value breaks through the early-warning threshold value, generating an alarm containing early-warning growth trend information.

Description

Technical Field

[0001] This invention relates to the field of online oil chromatography monitoring technology, and more specifically, to a trend prediction-based online oil chromatography monitoring and early warning method and system. Background Technology

[0002] Online oil chromatography monitoring technology has become one of the effective means of condition monitoring for large oil-filled equipment such as oil-immersed transformers and reactors. The basic principle of oil chromatography detection technology is that the insulating oil and solid insulating materials in the equipment will produce various gases under the action of thermal stress, electrical stress, and aging, which will dissolve in the mineral oil. The traditional analysis method is to use characteristic gas ratio analysis methods for further analysis and diagnosis when the concentration of gases such as hydrogen, acetylene, or total hydrocarbons exceeds the standard recommended threshold. However, in recent years, many cases of high-voltage oil-filled equipment failures have shown that although the oil chromatography concentration has not reached the recommended threshold in the early stage of the defect, the gas concentration is low but has a significant abnormal growth trend.

[0003] For example, in 2024, a 1000kV high-voltage reactor experienced a fault during operation. Prior to the fault, there was a sudden increase in the levels of acetylene and hydrogen, but these levels did not reach the alarm threshold. In 2025, a 500kV transformer experienced a bushing rupture fault during operation. Its online oil chromatography monitoring data showed a significant abnormal increase in hydrogen and ethylene before the fault occurred, but the overall concentration was far below the attention threshold. Therefore, if historical data can be used to predict short-term growth trends in oil chromatography monitoring data when equipment is operating normally or when fault symptoms are not obvious, it is possible to provide timely warnings of latent defects and deterioration trends before key parameters reach the fault critical point, ensuring the operational reliability of large oil-filled equipment.

[0004] Artificial intelligence techniques such as grey models, artificial neural networks, least squares support vector machines, and support vector regression can all be applied to predict the concentration of oil chromatographic gases in oil-immersed equipment. Taking artificial neural networks as an example, while backpropagation neural networks, radial basis function neural networks, and generalized regression neural networks possess excellent self-learning capabilities, good generalization performance, and powerful nonlinear data processing capabilities, their accurate predictions typically rely on large amounts of historical data to set the network structure and parameters, limiting their practicality in real-world applications. Grey models, although capable of obtaining ideal prediction results using small sample data, are only suitable for scenarios describing monotonically exponential changes in observed variables over time. However, the actual oil chromatographic concentration changes in oil-filled equipment such as transformers often do not conform to this premise, leading to inherent errors when using grey models for prediction. Furthermore, current technologies have not yet developed an executable prediction-based early warning mechanism. Summary of the Invention

[0005] According to the present invention, an online monitoring and early warning method and system for oil chromatography based on trend prediction is provided to solve the technical problem that the traditional threshold method cannot capture latent faults that show abnormal gas growth trends before reaching the concentration threshold.

[0006] According to a first aspect of the present invention, an online monitoring and early warning method for oil chromatography based on trend prediction is provided, comprising:

[0007] Continuous oil chromatography data are obtained from the online monitoring device of the target transformer and divided into historical dataset D1 and current data sequence D2. Historical dataset D1 and current data sequence D2 are then used as the original data.

[0008] The raw data is preprocessed, including outlier removal and cubic spline interpolation, and the sampling frequency is standardized to a preset time interval.

[0009] Based on the preprocessed historical dataset D1, a hybrid kernel function support vector regression (SVR) model is constructed to obtain the set of parameters to be optimized. The parameters are then tuned using the real-number encoded GA algorithm to obtain the optimal parameters.

[0010] The optimal parameters are substituted into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with the optimal parameters. The current data sequence D2 is then input to obtain the gas concentration prediction value for the next period.

[0011] Set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, generate an alarm containing warning growth trend information.

[0012] Optionally, the historical dataset D1 contains monitoring data for at least a predetermined number of days for optimizing model parameters, and the current data sequence D2 is used as input to the optimized model for trend prediction.

[0013] Optionally, based on the preprocessed raw data, a hybrid kernel function support vector regression (SVR) model is constructed to obtain a set of parameters to be optimized, including:

[0014] Construct a hybrid kernel function support vector regression (SVR) model based on the following formula:

[0015]

[0016] Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3).

[0017]

[0018] In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is:

[0019] K(x i x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4)

[0020] Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel.

[0021] In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

[0022] Optionally, the optimal parameters are obtained by tuning the set of parameters to be optimized using the real-number encoded GA algorithm, including:

[0023] A real number encoding method is used to randomly generate real numbers within a specific range as initial parameters;

[0024] The evaluation is performed by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective.

[0025] The evolutionary process terminates when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, thus obtaining the fine-tuning parameters {C,ε,γ,d,w} for the defined prediction model.

[0026] If the stopping condition is not met, the population evolution is controlled by operating genetic operators, including selection, crossover, and mutation. The operation will be repeated until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

[0027] According to another aspect of the present invention, an online oil chromatography monitoring and early warning system based on trend prediction is also provided, comprising:

[0028] The original data module is determined to acquire continuous oil chromatography data from the online monitoring device of the target transformer and divide it into historical dataset D1 and current data sequence D2. The historical dataset D1 and current data sequence D2 are determined as the original data.

[0029] The raw data preprocessing module is used to preprocess the raw data, including outlier removal and cubic spline interpolation, and to unify the sampling frequency to a preset time interval.

[0030] The module for obtaining optimal parameters is used to construct a hybrid kernel function support vector regression (SVR) model based on the preprocessed historical dataset D1, obtain the set of parameters to be optimized, and perform parameter tuning on the set of parameters to be optimized using the real number encoding GA algorithm to obtain the optimal parameters.

[0031] The module for obtaining gas concentration prediction values ​​is used to input the optimal parameters into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with optimal parameters, input the current data sequence D2, and obtain the gas concentration prediction values ​​for the next period.

[0032] A trend information alarm module is used to set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, an alarm containing warning growth trend information is generated.

[0033] Optionally, the historical dataset D1 contains monitoring data for at least a predetermined number of days for optimizing model parameters, and the current data sequence D2 is used as input to the optimized model for trend prediction.

[0034] Optionally, the module for obtaining optimal parameters includes:

[0035] Construct a hybrid kernel function SVR model submodule to build a hybrid kernel function support vector regression SVR model according to the following formula:

[0036] y i =f(x) i )=ω·φ(x i )+b (1)

[0037] Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3).

[0038]

[0039]

[0040] In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is:

[0041] K(x i x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4)

[0042] Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel.

[0043] In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

[0044] Optionally, the module for obtaining optimal parameters includes:

[0045] The initial parameter setting submodule is used to randomly generate real numbers as initial parameters within a specific range using real number encoding.

[0046] The fitness evaluation submodule is used to evaluate by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective.

[0047] The fine-tuning parameter submodule is used to terminate the evolutionary process when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, and to obtain the fine-tuning parameters {C,ε,γ,d,w} that define the prediction model.

[0048] The optimal parameter acquisition submodule is used to control population evolution by operating genetic operators when the stopping condition is not met. The genetic operators include selection, crossover, and mutation. The operation will be executed cyclically until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

[0049] According to another aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method as described in any one of the claims.

[0050] According to another aspect of the present invention, an electronic device is also provided, comprising:

[0051] The aforementioned computer-readable storage medium; and

[0052] One or more processors for executing a program in the computer-readable storage medium.

[0053] Therefore, by combining the convex combination structure of polynomial kernel and Gaussian kernel, the system effectively integrates global pattern capture and local feature learning capabilities. Based on real-number encoding, the genetic algorithm achieves intelligent search for SVR parameters (such as penalty factor and kernel weight) through Latin hypercube sampling initialization and adaptive Gaussian mutation mechanism. Cubic spline interpolation is used to unify the frequency of non-uniform sampling data. Combined with outlier removal mechanism, the system ensures the completeness and stability of model input data, enhances adaptability to the sampling frequency differences of monitoring equipment, sets graded early warning thresholds based on actual cases and industry standards, and combines dynamic prediction of absolute growth rate of gas concentration (ABR) to achieve early (4-hour) warning of faults, meet the requirements for safe operation of power equipment, improve early warning response efficiency, and effectively support operation and maintenance decisions. Attached Figure Description

[0054] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures:

[0055] Figure 1 This is a schematic flowchart of an online oil chromatography monitoring and early warning method based on trend prediction as described in this embodiment.

[0056] Figure 2 This is a schematic diagram of the specific process of an online monitoring and early warning method for oil chromatography based on trend prediction as described in this embodiment;

[0057] Figure 3 This is a schematic diagram of an online oil chromatography monitoring and early warning system based on trend prediction as described in this embodiment. Detailed Implementation

[0058] Exemplary embodiments of the invention will now be described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention. In the drawings, the same units / elements are referred to by the same reference numerals.

[0059] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0060] According to a first aspect of the present invention, an online monitoring and early warning method 100 for oil chromatography based on trend prediction is provided, with reference to... Figure 1 As shown, the method 100 includes:

[0061] S101: Obtain continuous oil chromatography data from the online monitoring device of the target transformer, divide it into historical dataset D1 and current data sequence D2, and determine historical dataset D1 and current data sequence D2 as the original data;

[0062] S102: Preprocess the raw data, including outlier removal and cubic spline interpolation, and unify the sampling frequency to a preset time interval;

[0063] S103: Based on the preprocessed historical dataset D1, construct a hybrid kernel function support vector regression (SVR) model to obtain the set of parameters to be optimized. Use the real number encoding GA algorithm to fine-tune the parameters to obtain the optimal parameters.

[0064] S104: Substitute the optimal parameters into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with optimal parameters, input the current data sequence D2, and obtain the gas concentration prediction value for the next period.

[0065] S105: Set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, generate an alarm containing warning growth trend information.

[0066] Specifically, this paper provides an early warning strategy for online monitoring data trends of oil chromatography based on a genetic algorithm-optimized SVR model with kernel function convex combination, in order to overcome the inability of traditional threshold methods to capture latent faults that show abnormal gas growth trends before reaching the concentration threshold.

[0067] To achieve the above objectives, this invention implements an early warning strategy for dissolved gas trends in mineral oil-immersed transformers based on a hybrid kernel function support vector regression model optimized by a genetic algorithm, referencing... Figure 2 As shown, it includes the following steps:

[0068] (1) The continuously monitored oil chromatographic data obtained from the target transformer are divided into a historical dataset D1 for training the prediction model and a latest time series D2 for prediction. The historical dataset D1 contains at least 7 days of monitoring data and is used to optimize the model parameters; the current data series D2 is input into the optimized model for trend prediction.

[0069] First, a preprocessing module is used to eliminate outliers (including null values ​​and abnormal zero values) from the raw data stream. Given that the high-frequency sampling frequency of online oil chromatography monitoring varies depending on demand, and can typically be set to once every 2 hours, 4 hours, 8 hours, or 24 hours based on the differentiated monitoring requirements of the monitored equipment, cubic spline interpolation is used to interpolate and complete the dataset after outlier removal, unifying it to a uniform time frequency. Optionally, this technical solution unifies the processing to a 4-hour sampling frequency.

[0070] (2) Constructing an SVR model framework with a convex combination of kernel functions. Support Vector Regression (SVR) is a support vector method specifically used for prediction and regression analysis. The basic principle of SVR is to map a low-dimensional time space to a high-dimensional feature space, and construct a linear function that approximates the data trend in this transformed space to characterize the relationship between input and output variables, as shown in Equation (1), thereby realizing the prediction and fitting of output under new input conditions.

[0071] y i =f(x) i )=ω·φ(x i )+b 1)

[0073] Among them, y i x represents the output, and x represents the input. This represents the mapping from the input space to the high-dimensional feature space. ω and b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3).

[0074]

[0075] In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function.

[0076] Kernel functions are a core component of SVR (Separate Classification and Reduction). Essentially, they are mapping functions that calculate the inner product of two vectors in a high-dimensional space, transforming nonlinearly inseparable data in the original low-dimensional space into linearly separable distributions in the high-dimensional space through appropriate mapping. This transformation enables the construction of an optimal classification hyperplane in the projected space to solve classification problems. Different kernel functions possess unique properties: polynomial kernels allow distant data points to influence the kernel function value, exhibiting strong feature extraction capabilities; conversely, Gaussian kernels are primarily sensitive to local regions near the observation point, with negligible influence from distant data on the kernel value. Therefore, these local kernels excel at learning local patterns but have weak extrapolation capabilities. To synergistically utilize the complementary advantages of different kernel types and overcome their respective limitations, this paper proposes a convex combination method of polynomial and Gaussian kernels. This design integrates the advantages of global and local kernels, thereby enhancing the model's learning ability and predictive performance. Its expression is:

[0077] K(x i x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4)

[0078] Where γ represents the kernel bandwidth, d represents the polynomial degree, and w is the kernel weight coefficient with a value range of [0,1]. When w=0, the hybrid kernel degenerates into a polynomial kernel, and when w=1, it degenerates into a Gaussian kernel.

[0079] This model architecture integrates global and local kernel characteristics through convex combination, providing a basic structure for subsequent parameterization.

[0080] In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

[0081] (3) Parameter tuning using the real-number encoding GA algorithm. This includes the following sub-steps:

[0082] 1) The first step is to set the initial parameters (initial population). The algorithm uses real number encoding, which can randomly generate real numbers within a specific range as initial parameters. The real number encoded GA algorithm is configured as follows: population size = 100, maximum number of iterations = 100, crossover probability = 0.8, mutation probability = 0.02.

[0083] 2) Fitness Evaluation. Fitness is evaluated by minimizing the Mean Absolute Percentage Error (MAPE). The algorithm starts with a randomly generated initial population, representing a set of candidate solutions to the problem (called chromosomes or individuals). Each chromosome includes all parameters to be optimized. Chromosome quality is evaluated using a fitness function constructed based on the optimization objective.

[0084] 3) Stopping Criteria. The preset stopping criterion is: the evolutionary process is terminated when the optimal fitness improvement is below 0.2 for three consecutive generations or when the maximum number of iterations is reached, and finally the fine-tuning parameters {C, ε, γ, d, w} of the defined prediction model are obtained.

[0085] 4) Population evolution is executed cyclically. When the stopping condition is not met, population evolution is controlled by manipulating genetic operators, including selection, crossover, and mutation. a) Selection: Individuals are selected for reproduction based on fitness, ensuring that high-quality individuals have a higher probability of survival and propagation to the next generation, while controlling the population size. b) Mutation: Random changes are introduced into the genetic information of a chromosome to generate new individuals. Its main function is to maintain population diversity and prevent convergence to a local optimum. c) Mutation: Random changes are introduced into the genetic information of a chromosome to generate new individuals. Its main function is to maintain population diversity and prevent convergence to a local optimum. These operations will be executed cyclically until the preset stopping criterion is met, ultimately obtaining the fine-tuning parameters {C, ε, γ, d, w} for defining the prediction model.

[0086] (4) Trend Prediction. The optimal parameters {C,ε,γ,d,w} obtained after training are substituted into the optimized SVR model, and the gas concentration prediction value for the next period is obtained by inputting the real-time time series D2. The absolute growth rate (ABR) of key gases (hydrogen H2, total hydrocarbons TH, acetylene C2H2) in the 4-hour, daily, and weekly timeframes is estimated using the least squares method.

[0087] (5) Trend Warning. A trend warning threshold is set, referencing Q / GDW 10536—2021 "Technical Specification for Online Monitoring Devices of Dissolved Gases in Transformer Oil," as well as the requirements for handling abnormalities in oil chromatography of UHV transformers and converter transformers, and actual fault cases. Gas concentration rise limits are established, with three levels of warnings. Recommended warning rise limits for 500 kV transformers and reactors are shown in Table 1. When the warning threshold is exceeded, the system automatically generates an alarm containing information on the warning growth trend.

[0088] Table 1 Recommended Trend Warning Values ​​for 500 kV Voltage Level Transformers and Reactors

[0089]

[0090] Optionally, the historical dataset D1 contains monitoring data for at least a predetermined number of days for optimizing model parameters, and the current data sequence D2 is used as input to the optimized model for trend prediction.

[0091] Optionally, based on the preprocessed raw data, a hybrid kernel function support vector regression (SVR) model is constructed to obtain a set of parameters to be optimized, including:

[0092] Construct a hybrid kernel function support vector regression (SVR) model based on the following formula:

[0093] y i =f(x) i )=ω·φ(x i )+b (1)

[0094] Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3).

[0095]

[0096] In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is:

[0097] K(x i x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4)

[0098] Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel.

[0099] In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

[0100] Optionally, the optimal parameters are obtained by tuning the set of parameters to be optimized using the real-number encoded GA algorithm, including:

[0101] A real number encoding method is used to randomly generate real numbers within a specific range as initial parameters;

[0102] The evaluation is performed by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective.

[0103] The evolutionary process terminates when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, thus obtaining the fine-tuning parameters {C,ε,γ,d,w} for the defined prediction model.

[0104] If the stopping condition is not met, the population evolution is controlled by operating genetic operators, including selection, crossover, and mutation. The operation will be repeated until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

[0105] Therefore, by combining the convex combination structure of polynomial kernel and Gaussian kernel, the system effectively integrates global pattern capture and local feature learning capabilities. Based on real-number encoding, the genetic algorithm achieves intelligent search for SVR parameters (such as penalty factor and kernel weight) through Latin hypercube sampling initialization and adaptive Gaussian mutation mechanism. Cubic spline interpolation is used to unify the frequency of non-uniform sampling data. Combined with outlier removal mechanism, the system ensures the completeness and stability of model input data, enhances adaptability to the sampling frequency differences of monitoring equipment, sets graded early warning thresholds based on actual cases and industry standards, and combines dynamic prediction of absolute growth rate of gas concentration (ABR) to achieve early (4-hour) warning of faults, meet the requirements for safe operation of power equipment, improve early warning response efficiency, and effectively support operation and maintenance decisions.

[0106] According to another aspect of the present invention, an online oil chromatography monitoring and early warning system 300 based on trend prediction is also provided, with reference to... Figure 3 As shown, the system 300 includes:

[0107] The original data module 310 is used to obtain continuous oil chromatography data from the online monitoring device of the target transformer and divide it into historical dataset D1 and current data sequence D2. The historical dataset D1 and current data sequence D2 are determined as the original data.

[0108] The raw data preprocessing module 320 is used to preprocess the raw data, including outlier removal and cubic spline interpolation, and to unify the sampling frequency to a preset time interval.

[0109] The optimal parameter acquisition module 330 is used to construct a hybrid kernel function support vector regression (SVR) model based on the preprocessed historical dataset D1, obtain the set of parameters to be optimized, and perform parameter tuning on the set of parameters to be optimized using the real number encoding GA algorithm to obtain the optimal parameters.

[0110] The gas concentration prediction module 340 is used to input the optimal parameters into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with optimal parameters, input the current data sequence D2, and obtain the gas concentration prediction value for the next period.

[0111] The trend information alarm module 350 is used to set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, an alarm containing warning growth trend information is generated.

[0112] Optionally, the historical dataset D1 contains monitoring data for at least a predetermined number of days for optimizing model parameters, and the current data sequence D2 is used as input to the optimized model for trend prediction.

[0113] Optionally, the module for obtaining optimal parameters includes:

[0114] Construct a hybrid kernel function SVR model submodule to build a hybrid kernel function support vector regression SVR model according to the following formula:

[0115] y i =f(x) i )=ω·φ(x i )+b (1)

[0116] Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3).

[0117]

[0118]

[0119] In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is:

[0120] K(x i x j )=w·exp(-γ||x i-x j ||) 2 +(1-w)·(x i ·x j +j) d (4)

[0121] Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel.

[0122] In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

[0123] Optionally, the module for obtaining optimal parameters includes:

[0124] The initial parameter setting submodule is used to randomly generate real numbers as initial parameters within a specific range using real number encoding.

[0125] The fitness evaluation submodule is used to evaluate by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective.

[0126] The fine-tuning parameter submodule is used to terminate the evolutionary process when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, and to obtain the fine-tuning parameters {C,ε,γ,d,w} that define the prediction model.

[0127] The optimal parameter acquisition submodule is used to control population evolution by operating genetic operators when the stopping condition is not met. The genetic operators include selection, crossover, and mutation. The operation will be executed cyclically until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

[0128] The trend-prediction-based online oil chromatography monitoring and early warning system 300 of this invention corresponds to the trend-prediction-based online oil chromatography monitoring and early warning method 100 of another embodiment of this invention, and will not be described again here.

[0129] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0130] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0131] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0134] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A trend-prediction-based online monitoring and early warning method for oil chromatography, characterized in that, include: Continuous oil chromatography data are obtained from the online monitoring device of the target transformer and divided into historical dataset D1 and current data sequence D2. Historical dataset D1 and current data sequence D2 are then used as the original data. The raw data is preprocessed, including outlier removal and cubic spline interpolation, and the sampling frequency is standardized to a preset time interval. Based on the preprocessed historical dataset D1, a hybrid kernel function support vector regression (SVR) model is constructed to obtain the set of parameters to be optimized. The parameters are then tuned using the real-number encoded GA algorithm to obtain the optimal parameters. The optimal parameters are substituted into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with the optimal parameters. The current data sequence D2 is then input to obtain the gas concentration prediction value for the next period. Set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, generate an alarm containing warning growth trend information.

2. The method according to claim 1, characterized in that, The historical dataset D1 contains monitoring data for at least a predetermined number of days, which is used to optimize model parameters. The current data sequence D2 is used as input to the optimized model for trend prediction.

3. The method according to claim 1, characterized in that, Based on the preprocessed raw data, a hybrid kernel function support vector regression (SVR) model is constructed to obtain the set of parameters to be optimized, including: Construct a hybrid kernel function support vector regression (SVR) model based on the following formula: y i =f(x i )=ω·φ(x i )+b (1) Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3). In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is: K(x i ,x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4) Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel. In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

4. The method according to claim 3, characterized in that, The optimal parameters are obtained by tuning the set of parameters to be optimized using the real-number encoded GA algorithm, including: A real number encoding method is used to randomly generate real numbers within a specific range as initial parameters; The evaluation is performed by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective. The evolutionary process terminates when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, thus obtaining the fine-tuning parameters {C,ε,γ,d,w} for the defined prediction model. If the stopping condition is not met, the population evolution is controlled by operating genetic operators, including selection, crossover, and mutation. The operation will be repeated until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

5. An online oil chromatography monitoring and early warning system based on trend prediction, characterized in that, include: The original data module is determined to acquire continuous oil chromatography data from the online monitoring device of the target transformer and divide it into historical dataset D1 and current data sequence D2. The historical dataset D1 and current data sequence D2 are determined as the original data. The raw data preprocessing module is used to preprocess the raw data, including outlier removal and cubic spline interpolation, and to unify the sampling frequency to a preset time interval. The module for obtaining optimal parameters is used to construct a hybrid kernel function support vector regression (SVR) model based on the preprocessed historical dataset D1, obtain the set of parameters to be optimized, and perform parameter tuning on the set of parameters to be optimized using the real number encoding GA algorithm to obtain the optimal parameters. The module for obtaining gas concentration prediction values ​​is used to input the optimal parameters into the hybrid kernel function support vector regression (SVR) model to obtain the SVR model with optimal parameters, input the current data sequence D2, and obtain the gas concentration prediction values ​​for the next period. A trend information alarm module is used to set a trend warning threshold. When the predicted gas concentration exceeds the warning threshold, an alarm containing warning growth trend information is generated.

6. The system according to claim 5, characterized in that, include: The historical dataset D1 contains monitoring data for at least a predetermined number of days, which is used to optimize model parameters. The current data sequence D2 is used as input to the optimized model for trend prediction.

7. The system according to claim 5, characterized in that, The module for obtaining optimal parameters includes: Construct a hybrid kernel function SVR model submodule to build a hybrid kernel function support vector regression SVR model according to the following formula: y i =f(x i )=ω·φ(x i )+b (1) Among them, y i x represents the output, and x represents the input. Let ω represent the mapping from the input space to the high-dimensional feature space, and let b represent the regression coefficient and the bias, respectively. By introducing Lagrange multipliers, the inequality-constrained optimization problem is transformed into an equality-constrained problem to solve for ω and b, as shown in equations (2) and (3). In equations (2) and (3), α and α* are Lagrange multipliers, C represents the penalty factor used to balance empirical risk and confidence range, ε represents the loss factor, and K(x) i ,y j ) is the kernel function, and its expression is: K(x i ,x j )=w·exp(-γ||x i -x j ||) 2 +(1-w)·(x i ·x j +1) d (4) Where γ represents the kernel bandwidth, d represents the polynomial degree, w is the kernel weight coefficient and its value ranges from [0,1]. When w = 0, the hybrid kernel degenerates into a polynomial kernel, and when w = 1, it degenerates into a Gaussian kernel. In this model, the parameters to be optimized are the penalty factor C, loss factor ε, kernel bandwidth γ, polynomial degree d, and kernel weight coefficient w, which are defined as the set of parameters to be optimized {C,ε,γ,d,w}.

8. The system according to claim 7, characterized in that, The module for obtaining optimal parameters includes: The initial parameter setting submodule is used to randomly generate real numbers as initial parameters within a specific range using real number encoding. The fitness evaluation submodule is used to evaluate by minimizing the mean absolute percentage error (MAPE). It starts with a randomly generated initial population, which represents a set of candidate solutions to the problem, called a chromosome or individual. Each chromosome includes the set of all parameters to be optimized {C, ε, γ, d, w}. The quality of the chromosome is evaluated by a fitness function, which is constructed according to the optimization objective. The fine-tuning parameter submodule is used to terminate the evolutionary process when the optimal fitness improvement is below the threshold for three consecutive generations or when the maximum number of iterations is reached, and to obtain the fine-tuning parameters {C,ε,γ,d,w} that define the prediction model. The optimal parameter acquisition submodule is used to control population evolution by operating genetic operators when the stopping condition is not met. The genetic operators include selection, crossover, and mutation. The operation will be executed cyclically until the preset stopping criterion is met to obtain the optimal parameters {C,ε,γ,d,w}.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-4.

10. An electronic device, characterized in that, include: The computer-readable storage medium as described in claim 9; as well as One or more processors for executing a program in the computer-readable storage medium.

Images