A transformer oil dissolved gas concentration prediction method, medium and terminal

Abstract

The application is suitable for the technical field of power equipment, and relates to a transformer oil dissolved gas concentration prediction method, medium and terminal, comprising: constructing an initial sample data set; performing empirical mode decomposition on time series data to obtain a plurality of intrinsic mode function components; calculating the energy proportion of each component, and dividing high-frequency groups and low-frequency groups according to the energy proportion; constructing a joint prediction model for different frequency bands, including a first LSTM model and a second LSTM model, and performing joint training; and adaptively optimizing the time window size of the first LSTM model and the second LSTM model by using a particle swarm optimization algorithm. Through the framework of decomposition-classification-optimization-joint prediction, the application can effectively capture the change patterns of different time scales in the gas concentration sequence, and match the optimal modeling parameters for high-frequency details and low-frequency trends, thereby significantly improving the prediction accuracy and providing more reliable data support for early warning of transformer latent faults.

Description

Technical Field

[0001] This invention belongs to the field of power equipment technology, and in particular relates to a method, medium and terminal for predicting the concentration of dissolved gases in transformer oil. Background Technology

[0002] As a core piece of equipment in the power system, the reliable operation of transformers directly affects the safety and stability of the power grid. During long-term operation or when internal faults occur (such as localized overheating or arcing), the insulating oil and solid insulation materials inside the transformer undergo chemical decomposition, producing characteristic gases including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO), and carbon dioxide (CO2), which dissolve in the oil. Dissolved gas analysis (DGA) is a key technology for diagnosing latent faults inside transformers. Proactive prediction of dissolved gas concentrations in the oil enables early warning of potential transformer fault risks, providing a scientific basis for developing preventative maintenance strategies and effectively improving the operational reliability of the power system.

[0003] Transformer oil dissolved gas concentration sequence data exhibits strong non-stationarity, complex time dependence, and noise interference. Traditional prediction methods struggle to effectively process this data, failing to fully extract the details of short-term fluctuations and long-term evolutionary trends. For example, the traditional EMD-LSTM method typically applies the same modeling strategy and input window size to all intrinsic mode function (IMF) components obtained from decomposition, neglecting the differences in physical meaning, frequency of change, and noise levels among different components. High-frequency components change drastically and are highly susceptible to noise, making smaller windows suitable for capturing short-term fluctuations; low-frequency components change slowly and exhibit strong trends, making larger windows suitable for capturing long-term dependencies. A uniform window setting makes it difficult to accommodate features at different scales, limiting the model's predictive power. Furthermore, existing methods often rely on empirical judgments or simple rules for classifying IMF components, lacking objective and quantitative classification standards. This leads to subjectivity in the division between high-frequency and low-frequency components, affecting the stability and generalization ability of subsequent modeling. In addition, existing technologies typically train independent models for each component when building prediction models, failing to effectively utilize the complementary information between high and low frequency components. This limits the overall representational ability of the model, and there is still room for improvement in prediction accuracy.

[0004] Therefore, it is imperative to address the problems of insufficient feature utilization, simplistic modeling strategies, and low prediction accuracy in existing methods for predicting dissolved gas concentrations in transformer oil when processing non-stationary, multi-scale time series data, so as to improve the accuracy and reliability of predictions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for predicting dissolved gas concentration in transformer oil, thereby solving the problems of insufficient feature utilization, single modeling strategy, and low prediction accuracy in existing methods for predicting dissolved gas concentration in transformer oil when processing non-stationary, multi-scale time series data. In addition, this invention also provides a medium and terminal for predicting dissolved gas concentration in transformer oil.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the concentration of dissolved gases in transformer oil, comprising the following steps: S10. Obtain historical concentration time series data of dissolved gases in transformer oil and construct an initial sample dataset; S20. Based on the target gas concentration time series data in the initial sample dataset, perform empirical mode decomposition (EMD) to obtain multiple intrinsic mode function (IMF) components and one residual component. S30. Calculate the energy percentage of each IMF component, divide the multiple IMF components into high-frequency group and low-frequency group according to the cumulative energy percentage curve, and add the IMF components in the high-frequency group to obtain the high-frequency feature sequence, and add the IMF components in the low-frequency group to obtain the low-frequency feature sequence. S40. Construct a frequency band joint prediction model, which includes a first LSTM model and a second LSTM model. Jointly train the first LSTM model and the second LSTM model. The first LSTM model is used to receive the high-frequency feature sequence and output a first prediction component. The second LSTM model is used to receive the low-frequency feature sequence and output a second prediction component. Add the first prediction component and the second prediction component to obtain the final prediction value. S50. Adaptively optimize the first time window size of the first LSTM model and the second time window size of the second LSTM model using the particle swarm optimization algorithm; S60. Retrain or configure the frequency band joint prediction model using the window value optimized in S50 to predict the future dissolved gas concentration in transformer oil.

[0007] Furthermore, the specific steps of S20 are as follows: S201. Input historical concentration time series data of dissolved gases in transformer oil. ; S202, Initialize residual components ; S203. Process the current residual components and extract the intrinsic mode function (IMF) components through an iterative filtering process. This continues until the current residual component satisfies the monotonicity condition or the number of extreme points does not exceed a preset number. S204, ultimately Decomposed into n IMF components and a residual component .

[0008] Furthermore, the specific steps of S30 are as follows: S301. Calculate the energy of each IMF component. ; S302. Calculate the total energy of all IMF components. ; S303. Calculate the energy percentage of each IMF component. ; S304. Calculate the cumulative energy percentage according to the IMF components from high frequency to low frequency. ; S305. Find the steep rise point on the cumulative energy percentage curve, divide all IMF components before the steep rise point into the high frequency group, and divide all IMF components after the steep rise point into the low frequency group. S306. Summate the IMF components in the high-frequency group and the low-frequency group respectively to obtain the corresponding high-frequency features and low-frequency features.

[0009] Furthermore, in step S40, the root mean square error (RMSE) between the final predicted value and the true value is used as the loss function, and the network parameters of the first LSTM model and the second LSTM model are optimized simultaneously.

[0010] Furthermore, during the joint training process, the early stopping method is used to monitor the RMSE size on the validation set to determine whether training has ended.

[0011] Furthermore, the specific steps of S50 are as follows: S501. Initialize the particle swarm, where the position of each particle represents a combination of a candidate first window value and a second window value; S502. Round down and limit the position of each particle, and use it as a model parameter. Calculate the RMSE of the corresponding joint training model on the validation set, and use it as the fitness of that particle. S503. Perform iterative updates, record the individual historical best position and the global historical best position, stop the search according to the preset iteration termination condition, and the window value corresponding to the final global best position is the optimal first window value and second window value.

[0012] Furthermore, different optimization boundary ranges are set for the first window value and the second window value, wherein the boundary range of the first window value is [1, 20] and the boundary range of the second window value is [5, 30].

[0013] Furthermore, prior to S10, the method includes a step of normalizing the collected dissolved gas concentration data in the transformer oil.

[0014] In a second aspect, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0015] Thirdly, the present invention also provides an electronic terminal, comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal to perform the method described above.

[0016] Compared with existing technologies, the method, medium, and terminal for predicting dissolved gas concentration in transformer oil provided by this invention have at least the following advantages: Existing methods for predicting dissolved gas concentration (IMF) in transformer oil suffer from insufficient feature utilization, simplistic modeling strategies, and low prediction accuracy when processing non-stationary, multi-scale time series data. This invention, based on Empirical Mode Decomposition (EMD), innovatively introduces energy density as an objective quantification standard. It automatically divides IMF components into high-frequency and low-frequency groups based on the steep rise point of the cumulative energy percentage curve, achieving accurate separation of high-frequency detail information (such as noise and rapid fluctuations) from low-frequency main trend information. This avoids the uncertainty caused by subjective classification, laying a solid foundation for subsequent frequency-band modeling and effectively improving the model's ability to represent complex DGA sequences. This invention employs a particle swarm optimization (PSO) algorithm to adaptively optimize the time window sizes of the high-frequency and low-frequency LSTM sub-models. For the high-frequency components, which change rapidly and have high noise levels, a smaller window is optimized to focus on short-term dynamic changes; for the low-frequency components, which change slowly and have strong trends, a larger window is optimized to capture long-term dependencies. This adaptive window mechanism can match the optimal time scale according to the intrinsic characteristics of different frequency components, significantly improving the model's ability to extract features at different scales from the data. This invention constructs independent LSTM prediction sub-models for high-frequency and low-frequency features respectively, and adopts a joint training method. During the training process, the two sub-models share information and optimize collaboratively. Finally, the outputs are added together for fusion prediction, which effectively improves the prediction accuracy and reliability. Attached Figure Description

[0017] To more clearly illustrate the solution of the present invention, a brief introduction will be given to the drawings used in the description of the embodiments below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for predicting dissolved gas concentration in transformer oil, provided in an embodiment of the present invention; Figure 2 This is a model framework diagram of a method for predicting dissolved gas concentration in transformer oil provided by an embodiment of the present invention; Figure 3 This is a decomposition diagram of training data in S20 of a method for predicting dissolved gas concentration in transformer oil provided in an embodiment of the present invention; Figure 4 This is a breakdown diagram of test data in S20 of a method for predicting dissolved gas concentration in transformer oil provided in an embodiment of the present invention; Figure 5 The figure shows the simulation results of a method for predicting dissolved gas concentration in transformer oil provided in an embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0021] This invention provides a method for predicting the concentration of dissolved gases in transformer oil, which includes the following steps: S10. Obtain historical concentration time series data of dissolved gases in transformer oil to construct an initial sample dataset; S20. Perform Empirical Mode Decomposition (EMD) based on the target gas concentration time series data in the initial sample dataset to obtain multiple intrinsic mode function (IMF) components and one residual component; S30. Calculate the energy proportion of each IMF component, divide the multiple IMF components into high-frequency and low-frequency groups according to the cumulative energy proportion curve, and sum the IMF components in the high-frequency group to obtain the high-frequency feature sequence, and sum the IMF components in the low-frequency group to obtain the low-frequency feature sequence; S40. Construct a frequency-band joint prediction model, which includes a first LST. The M-model and the second LSTM model are jointly trained. The first LSTM model is used to receive high-frequency feature sequences and output a first predicted component, and the second LSTM model is used to receive low-frequency feature sequences and output a second predicted component. The first predicted component and the second predicted component are added to obtain the final predicted value. S50: The first time window size of the first LSTM model and the second time window size of the second LSTM model are adaptively optimized using the particle swarm optimization algorithm. S60: The frequency band joint prediction model is retrained or configured using the window value optimized in S50 to predict the future dissolved gas concentration in transformer oil.

[0022] This invention, through a framework of decomposition-classification-optimization-joint prediction, can effectively capture the changing patterns of gas concentration sequences at different time scales, and specifically match the optimal modeling parameters for high-frequency details and low-frequency trends, thereby significantly improving prediction accuracy and providing more reliable data support for early warning of latent transformer faults.

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] This invention provides a method for predicting the concentration of dissolved gases in transformer oil, combined with... Figures 1 to 5 In this embodiment, the method for predicting the concentration of dissolved gases in transformer oil includes the following steps: S10. Obtain historical concentration time series data of dissolved gases in transformer oil and construct an initial sample dataset.

[0025] Specifically, in this embodiment, step S10 involves the following steps: S101. Obtain actual dissolved gas concentration data in transformer oil, such as: hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO), and carbon dioxide (CO2).

[0026] S102. The collected data is divided into two parts: 80% is used as the training set and 20% is used as the validation set. The training set is used to build a prediction model, and the validation set is used to select the best model and tune hyperparameters.

[0027] S20. Based on the target gas concentration time series data in the initial sample dataset, perform Empirical Mode Decomposition (EMD) to obtain multiple intrinsic mode function (IMF) components and one residual component.

[0028] Specifically, in this embodiment, step S20 consists of the following steps: S201. Input historical concentration time series data of dissolved gases (such as H2, CH4, C2H2, etc.) in transformer oil. .

[0029] S202. Initialize the residual components. .

[0030] Initialize the suspected modal components. .

[0031] S203, Calculation envelope mean ; ; in, and They are respectively The envelope curves of the maximum and minimum points are fitted by cubic spline functions; Subtract the mean of the envelope, = ; Using standard deviation Determining the number of zeros and extreme points Is it an IMF component, where: ; like Less than 0.2 or 0.3, and If the number of zeros and extreme points is the same or differs by at most one, it is denoted as an IMF component. , Proceed to step S203.

[0032] Subtracting the IMF component from the residual component yields the new residual component. .

[0033] If the new residual component If the condition is monotonic, or has no more than two extreme points, then it is impossible to construct an envelope further, and the decomposition ends; otherwise... Proceed to step S202.

[0034] S204, ultimately Decomposed into n IMF components and 1 residual component ,Right now: .

[0035] S30. Calculate the energy percentage of each IMF component. Based on the cumulative energy percentage curve, divide the multiple IMF components into high-frequency groups and low-frequency groups. Add the IMF components in the high-frequency group to obtain the high-frequency feature sequence, and add the IMF components in the low-frequency group to obtain the low-frequency feature sequence.

[0036] Specifically, in this embodiment, step S30 consists of the following steps: S301. Calculate the energy of each IMF component: For each decomposition, the resulting intrinsic mode function (IMF) components Calculate its energy: ; S302. Calculate the total signal energy: ; S303. Calculate the energy percentage of each component: ; S304. Calculate the cumulative energy percentage: The cumulative percentage is calculated by IMF serial number from highest to lowest (i.e., from highest to lowest frequency): ; Note: The last cumulative value .

[0037] S305, Find the cumulative energy percentage curve The steep rise point on the spectrum. IMF values ​​before this point are considered high-frequency components, while IMF values ​​after this point (including this point) are considered low-frequency components.

[0038] If there is no steep rise point, the last IMF is assumed to be low frequency and the other IMFs are assumed to be high frequency.

[0039] S306. The input features of the model are obtained by summing the features within each group: Training data: High frequency and: ; Low frequency and: ; Similarly, generate verification data.

[0040] S40. Construct a frequency band joint prediction model. The frequency band joint prediction model includes a first LSTM model and a second LSTM model. Jointly train the first LSTM model and the second LSTM model. The first LSTM model is used to receive high-frequency feature sequences and output a first prediction component. The second LSTM model is used to receive low-frequency feature sequences and output a second prediction component. Add the first prediction component and the second prediction component to obtain the final prediction value.

[0041] Specifically, in this embodiment, step S40 consists of the following steps: S401, Data Normalization; The data processed in step S306 is normalized, and the normalized dataset of the training data is defined as follows: ; The normalization method is as follows: ; in, x For the dataset, This is the minimum value of the data among all samples in the dataset. This represents the maximum value of the data among all samples in the dataset. These are the data values ​​after normalization.

[0042] S402. Train the normalized sample dataset from step S401. Divided into training set and test set; S403. Determine the window size of the model, and predict the next data point using the data from the first max(look_back_high, look_back_low) windows. High-frequency data window: look_back_high, low-frequency data window: look_back_low, both default to 10; S404. Construct a joint model, using the Long Short-Term Memory (LSTM) neural network model as the learning model, and set the structure of the LSTM model, including the number of layers, the number of neurons in each layer, and the optimizer type. Branch A: LSTM_a (LSTM structure is a 1-layer structure, 1 fully connected layer, the first LSTM structure has LTMCell units that are consistent with the feature dimension of the input, the hidden layer has 64 units, and the number of units in the fully connected layer is 1), input shape (look_back_high,1). Branch B: LSTM_b (LSTM structure is a 1-layer structure, 1 fully connected layer, the first LSTM structure has LTMCell units that are consistent with the feature dimension of the input, the hidden layer has 64 units, and the number of units in the fully connected layer is 1), input shape (look_back_low, 1) The final predicted value is obtained by adding the outputs of the two branches. ; by The labels are denoted by RMSE, the loss function is RMSE, and the optimizer is Adam. ; In the formula For predicted values, The true value is n, and the number of samples is n. S405, Training: Unified compilation and joint training (updating both LSTM weights simultaneously). Early stopping is used to monitor the validation set. The size of the value determines whether training has ended.

[0043] S50. Adaptively optimize the first time window size of the first LSTM model and the second time window size of the second LSTM model using the particle swarm optimization algorithm.

[0044] Specifically, in this embodiment, the parameters sought by the PSO algorithm are look_back_high and look_back_low, which belong to a two-dimensional target search space. It can be assumed that there are m particles forming a swarm, where the position of the i-th particle (i=1, 2, ..., m) is represented as... =( , That is, the position of the i-th particle in the 2D search space is In other words, the position of each particle is a potential solution. X Substituting the values ​​into the objective function allows us to calculate the fitness value, which is used to measure the particle's quality. The best position a particle has ever reached is denoted as... =( , The best position experienced by all particles in the entire swarm is denoted as . =( , The velocity of particle i is denoted as... =( , ).

[0045] Furthermore, in this embodiment, the specific steps of step S50 are as follows: S501. Initialize the particle swarm (velocity and position), inertia factor w, acceleration constants c1 and c2, particle swarm size m, maximum number of iterations max_steps, iteration termination condition (set to terminate when there is no significant improvement in the global optimum after c consecutive iterations), boundary range of the solution space, and velocity limit.

[0046] S502. Round the particle positions down to the nearest integer and restrict them to the boundary of the solution space. Evaluate the initial fitness of each particle, and select the validation set for the evaluation function. Size.

[0047] S503. The initial fitness value is taken as the local optimum of each particle, and the position corresponding to each fitness value is taken as the position of the local optimum of each particle.

[0048] S504. Take the best initial fitness value as the current global optimum, and take the position corresponding to the best fitness value as the position of the global optimum.

[0049] S505. Update the current flight speed of each particle according to the following formula.

[0050] ; ; in, i =1, 2, ..., m; d =1,2; w It is a non-negative number, called the inertia factor; acceleration constant. c 1. c 2 is a nonnegative constant; r 1 and r 2 is a random number that varies within the range [0, 1].

[0051] also, That is, the flight speed of particle i V by a maximum speed The limitations. If the particle's velocity in a certain dimension at the current moment... After the update, the maximum flight speed exceeds that dimension. The velocity of this dimension at the current moment is limited to .

[0052] S506. Limit the flight speed of each particle to ensure it does not exceed the set maximum flight speed.

[0053] S507. Update the current position of each particle, round down the position of each particle, and restrict it to the range of the solution space.

[0054] S508. Compare the current fitness value of each particle with the historical local optimum. If it is higher, then take the current particle's fitness value as the local optimum of the particle, and take its corresponding position as the position of the local optimum of each particle.

[0055] S509. Find the global optimum in the current swarm and take the position corresponding to the current global optimum as the position of the global optimum of the particle swarm.

[0056] S510. Repeat steps S505 to S509 until the set maximum number of iterations or iterations is met. c The position corresponding to the suboptimal value remains unchanged.

[0057] S511. Output the global optimal value of the particle swarm and its corresponding position. This position is the optimal window value of the high-frequency component look_back_high and the low-frequency component look_back_low of the model.

[0058] S60, using the window value optimized by S50, retrain or configure the frequency band joint prediction model for predicting future dissolved gas concentrations in transformer oil.

[0059] Example 1 In this embodiment, methane (CH4) is used as an example. The dissolved gas data in the transformer oil obtained this time is the monitoring data of transformer oil in a substation, with a total of 1005 data points.

[0060] Step 1: Segment the acquired dissolved gas data in transformer oil without disrupting the time series. The first 804 data points are used as the training set, and the last 201 data points are used as the test set.

[0061] Step 2: Perform EMD decomposition on the training and test data respectively. The decomposition results are as follows: Figure 3 and Figure 4 As shown.

[0062] Step 3: Based on the energy proportion of IMF classification, construct the model input features, as shown in Tables 1 and 2 below. These are the results calculated through steps S301 to S304. Table 1 shows the cumulative energy proportion of different IMF components (training data), and Table 2 shows the cumulative energy proportion of different IMF components (test data).

[0063] Table 1 Table 2 By taking the cumulative energy percentage of each IMF component pairwise, we can see that in the training data, IMF8-IMF7=89.53, which is the largest value, meaning IMF8 is the steepest point. IMF1-IMF7 are the high-frequency group, and IMF8 is the low-frequency group. In the test data, IMF5-IMF4=99.12, which is the largest value, meaning IMF5 is the steepest point. IMF1-IMF4 are the high-frequency group, and IMF5 is the low-frequency group.

[0064] Within the training set, the high-frequency group and the low-frequency group are summed to obtain two columns of features: high-frequency feature train_hf and low-frequency feature train_lf. The test set is processed similarly.

[0065] Step 4: Frequency-band LSTM modeling and joint training. Specifically: The normalized dataset after training data is denoted as The normalized dataset for verification is denoted as . ; training dataset The data is further divided, with 80% used as the training set and 20% as the test set. The training set is used to build the prediction model, and the test set is used to select the best model and fine-tune hyperparameters. look_back_high and look_back_low are parameters for particle swarm optimization (PSO). Construct the model and training strategy according to steps S404 to S405.

[0066] Step 5: Particle Swarm Optimization (PSO) Adaptive Window Selection. Specifically: Initialize the particle swarm (velocity and position), inertia factor w =0.6, acceleration constant c 1= c 2=2, number of particles m =20, maximum number of iterations max_steps=100, iteration termination condition ( c =30).

[0067] The boundary range of the solution space is: high frequency component [1,20], low frequency component [5,30]; Speed ​​limit: High-frequency components [-0.2] 19, 0.2 19], Low-frequency components [-0.2] 25, 0.2 25]; By iteratively solving according to steps S502 to S511, the optimal window values ​​for the high-frequency and low-frequency components of the model are 7 and 10, respectively.

[0068] Table 3 below shows a comparison of prediction results using different models. The window value for the LSTM model, the EMD-LSTM model, and the embodiment of this invention (fixed window) is set to 10. Figure 5 The figure shows the simulation results of the model. It can be seen that, compared with other algorithms, the method disclosed in this embodiment of the invention (adaptive window) improves the accuracy of the algorithm for the online monitoring device of dissolved gases in transformer oil, and provides more accurate data support for transformer condition assessment and preventive maintenance.

[0069] Table 3 This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods in this embodiment.

[0070] This invention also provides an electronic terminal, including: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal performs any of the methods in this embodiment.

[0071] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0072] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0073] Compared with existing technologies, the method, medium, and terminal for predicting dissolved gas concentration in transformer oil described in the above embodiments suffer from problems such as insufficient feature utilization, simplistic modeling strategies, and low prediction accuracy when processing non-stationary, multi-scale time series data. This invention, based on Empirical Mode Decomposition (EMD), innovatively introduces energy density as an objective quantification standard. It automatically divides the IMF components into high-frequency and low-frequency groups based on the steep rise point of the cumulative energy percentage curve, achieving accurate separation of high-frequency detail information (such as noise and rapid fluctuations) from low-frequency main trend information. This avoids the uncertainty caused by subjective classification, laying a solid foundation for subsequent frequency band modeling and effectively improving the model's ability to represent complex DGA sequences. This invention uses the Particle Swarm Optimization (PSO) algorithm to adaptively optimize the time window sizes of the high-frequency and low-frequency LSTM sub-models. To address the characteristics of rapid changes and high noise in high-frequency components, a smaller window is optimized to focus on short-term dynamic changes; conversely, to address the characteristics of slow changes and strong trends in low-frequency components, a larger window is optimized to capture long-term dependencies. This adaptive windowing mechanism matches the optimal time scale based on the intrinsic characteristics of different frequency components, significantly improving the model's ability to extract features at different scales from the data. This invention constructs independent LSTM prediction sub-models for high-frequency and low-frequency features respectively, and employs a joint training method. During training, the two sub-models share information and collaborate in optimization. Finally, the outputs are added together for fusion prediction, effectively improving prediction accuracy and reliability.

[0074] Obviously, the embodiments described above are merely preferred embodiments of the present invention, and not all embodiments. The accompanying drawings illustrate preferred embodiments of the present invention, but do not limit the scope of the patent. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this invention.

Claims

1. A method for predicting the concentration of dissolved gases in transformer oil, characterized in that, Includes the following steps: S10. Obtain historical concentration time series data of dissolved gases in transformer oil and construct an initial sample dataset; S20. Based on the target gas concentration time series data in the initial sample dataset, perform empirical mode decomposition (EMD) to obtain multiple intrinsic mode function (IMF) components and one residual component. S30. Calculate the energy percentage of each IMF component, divide the multiple IMF components into high-frequency group and low-frequency group according to the cumulative energy percentage curve, and add the IMF components in the high-frequency group to obtain the high-frequency feature sequence, and add the IMF components in the low-frequency group to obtain the low-frequency feature sequence. S40. Construct a frequency band joint prediction model, which includes a first LSTM model and a second LSTM model. Jointly train the first LSTM model and the second LSTM model. The first LSTM model is used to receive the high-frequency feature sequence and output a first prediction component. The second LSTM model is used to receive the low-frequency feature sequence and output a second prediction component. Add the first prediction component and the second prediction component to obtain the final prediction value. S50. Adaptively optimize the first time window size of the first LSTM model and the second time window size of the second LSTM model using the particle swarm optimization algorithm; S60. Retrain or configure the frequency band joint prediction model using the window value optimized in S50 to predict the future dissolved gas concentration in transformer oil.

2. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, The specific steps of S20 are as follows: S201. Input historical concentration time series data of dissolved gases in transformer oil. ; S202, Initialize residual components ; S203. Process the current residual components and extract the intrinsic mode function (IMF) components through an iterative filtering process. This continues until the current residual component satisfies the monotonicity condition or the number of extreme points does not exceed a preset number. S204, ultimately Decomposed into n IMF components and a residual component .

3. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, The specific steps of S30 are as follows: S301. Calculate the energy of each IMF component. ; S302. Calculate the total energy of all IMF components. ; S303. Calculate the energy percentage of each IMF component. ; S304. Calculate the cumulative energy percentage according to the IMF components from high frequency to low frequency. ; S305. Find the steep rise point on the cumulative energy percentage curve, divide all IMF components before the steep rise point into the high frequency group, and divide all IMF components after the steep rise point into the low frequency group. S306. Summate the IMF components in the high-frequency group and the low-frequency group respectively to obtain the corresponding high-frequency features and low-frequency features.

4. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, In step S40, the root mean square error (RMSE) between the final predicted value and the true value is used as the loss function, and the network parameters of the first LSTM model and the second LSTM model are optimized simultaneously.

5. The method for predicting dissolved gas concentration in transformer oil according to claim 4, characterized in that, During the joint training process, the early stopping method is used to monitor the RMSE size on the validation set to determine whether training has ended.

6. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, The specific steps of S50 are as follows: S501. Initialize the particle swarm, where the position of each particle represents a combination of a candidate first window value and a second window value; S502. Round down and limit the position of each particle, and use it as a model parameter. Calculate the RMSE of the corresponding joint training model on the validation set, and use it as the fitness of that particle. S503. Perform iterative updates, record the individual historical best position and the global historical best position, stop the search according to the preset iteration termination condition, and the window value corresponding to the final global best position is the optimal first window value and second window value.

7. The method for predicting dissolved gas concentration in transformer oil according to claim 6, characterized in that, Different optimization boundary ranges are set for the first window value and the second window value, wherein the boundary range of the first window value is [1, 20] and the boundary range of the second window value is [5, 30].

8. The method for predicting dissolved gas concentration in transformer oil according to claim 1, characterized in that, Before step S10, the method further includes the step of normalizing the collected dissolved gas concentration data in the transformer oil.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

10. An electronic terminal, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal to perform the method as described in any one of claims 1 to 8.

Images