Prediction method and device for abrupt change type signal of gas dissolved in oil

Abstract

The invention provides a prediction method for a sudden change type signal of gas dissolved in oil. The prediction method comprises the following steps: firstly, acquiring a time sequence signal of transformer gas monitoring; processing the time sequence signal according to a variational mode decomposition algorithm to obtain a plurality of mode components; according to the variational mode decomposition algorithm, the sum of confusion entropies of all mode components is used as an objective function of parameter optimization; and finally, according to a pre-established prediction model, performing prediction and superposition reconstruction on each modal component to obtain a prediction result of the abrupt change type signal of the gas dissolved in oil. According to the method, the VMD decomposition architecture optimized by the frost ice algorithm is introduced, the non-stationarity of the signal is quantified and obviously reduced through the chaos entropy index, higher prediction precision and robustness of the gas signal in the mutation type oil are realized, and the method has obvious innovativeness and superiority in the field of transformer fault trend prediction.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and in particular to a method and apparatus for predicting abrupt changes in dissolved gas signals in oil. Background Technology

[0002] During long-term operation, the oil-paper insulation material of power transformers gradually ages, potentially leading to faults such as overheating and discharge. Dissolved gas analysis in the oil is a core method for assessing the transformer's operating status and enabling predictive maintenance. However, when a transformer experiences a sudden fault or sensor malfunction, the time-series signal of dissolved gas concentration in the oil often exhibits abrupt changes and drastic fluctuations, displaying significant non-stationary characteristics. This poses a serious challenge to traditional predictive models, making it difficult for them to track signal changes in a timely manner and significantly reducing prediction accuracy.

[0003] In existing technologies, some methods predict gas concentrations by constructing combined models, but these are mainly applicable to stationary signals and are not well-suited for non-stationary signals containing abrupt changes. Other methods attempt to improve signal stationarity by removing outliers from the data, but this approach struggles to effectively distinguish between anomalies caused by real faults and noise interference, potentially leading to the accidental deletion of critical information or misjudgment of faults, thereby weakening the model's sensitivity to actual faults. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting abrupt changes in dissolved gas signals in oil, in order to solve the problem of inaccurate fault monitoring when the concentration time-series signals of dissolved gases in oil often show abrupt changes and drastic fluctuations.

[0005] In a first aspect, embodiments of the present invention provide a method for predicting abrupt changes in dissolved gas signals in oil, comprising: Acquire timing signals for transformer gas monitoring; The time-series signal is processed using the variational mode decomposition algorithm to obtain multiple mode components; the variational mode decomposition algorithm uses the sum of the disorder entropy of each mode component as the objective function for parameter optimization. Based on the pre-established prediction model, each modal component is predicted and superimposed to reconstruct the signal, thus obtaining the prediction results of the abrupt change signal of dissolved gas in oil.

[0006] In one possible implementation, the method further includes: The objective function is transformed into an unconstrained optimization function by using an augmented Lagrangian function and a quadratic penalty factor. Solve the unconstrained optimization function to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

[0007] In one possible implementation, the method further includes: The time-series signal is processed using the variational mode decomposition algorithm to obtain multiple initial components; The disorder entropy of each initial component is calculated, and the sum of the disorder entropy of all modal components is used as the objective function. The disorder entropy is used to measure the complexity and disorder of the time series data and is determined according to the probability of the occurrence of the arrangement pattern of the sequence segments. The sequence segments are obtained by embedding and reconstructing the time series signal in the phase space.

[0008] In one possible implementation, the unconstrained optimization function is solved to determine the optimized values ​​of the first parameter of the variational mode decomposition algorithm, including: The frost-ice optimization algorithm is adopted, with minimizing the objective function as the optimization objective. The first parameter is iteratively optimized in the parameter space to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

[0009] In one possible implementation, the modal components are predicted and reconstructed based on a pre-established prediction model to obtain the prediction results of the abrupt change signal of dissolved gas in oil, including: Based on the TCN-iTransformer prediction model, the modal components are predicted and superimposed to reconstruct the results, thus obtaining the prediction results of the abrupt change signal of dissolved gas in oil.

[0010] In one possible implementation, the method further includes: The frost-ice optimization algorithm is adopted, with the goal of minimizing the final prediction error. The second parameter of the TCN-iTransformer prediction model is iteratively optimized in the parameter space to determine the optimal value of the second parameter.

[0011] In one possible implementation, the mechanism of the frost optimization algorithm includes at least one of the following: Soft cream search mechanism, hard cream puncture mechanism, and active greedy selection mechanism.

[0012] In one possible implementation, the method further includes: An alert is triggered when the residual of the prediction result exceeds a preset threshold.

[0013] In one possible implementation, the method further includes: The characteristic segments of abrupt changes in a time series signal are identified using pre-defined expert rule criteria, and the trend change rate of the time series signal is calculated at multiple time scales; among them, the characteristic segments of abrupt changes and the trend change rate are used to determine the prior information in the parameter optimization process.

[0014] Secondly, embodiments of the present invention provide a predictive device for abrupt changes in dissolved gas signals in oil, characterized in that it comprises: The acquisition module is used to acquire the timing signals of transformer gas monitoring; The decomposition module is used to process the time-series signal according to the variational mode decomposition algorithm to obtain multiple mode components; wherein, the variational mode decomposition algorithm uses the sum of the chaos entropy of each mode component as the objective function for parameter optimization. The prediction module is used to predict and reconstruct each modal component based on a pre-established prediction model, so as to obtain the prediction results of the abrupt change signal of dissolved gas in oil.

[0015] Compared to traditional technologies, this invention provides a method for predicting abrupt changes in dissolved gas signals in oil. First, a time-series signal from transformer gas monitoring is acquired. Then, the time-series signal is processed using a variational mode decomposition (VMD) algorithm to obtain multiple modal components. The VMD algorithm uses the sum of the chaos entropy of each modal component as the objective function for parameter optimization. Finally, each modal component is predicted and reconstructed based on a pre-established prediction model to obtain the prediction result for abrupt changes in dissolved gas signals in oil. This invention introduces a VMD decomposition architecture optimized by the frost-ice algorithm, quantifies and significantly reduces the non-stationarity of the signal through chaos entropy, achieving higher prediction accuracy and robustness for abrupt changes in dissolved gas signals in oil. This method demonstrates significant innovation and superiority in the field of transformer fault trend prediction. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the prediction method for abrupt signals of dissolved gases in oil provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the implementation of a method for predicting abrupt changes in dissolved gas signals in oil, provided in another embodiment of the present invention. Figure 3 This is a schematic diagram showing the ethane gas content. Figure 4 To optimize the process diagram; Figure 5 This is a schematic diagram of the modal components after decomposition of a time-series signal; Figure 6 This is a schematic diagram of the TCN-iTransformer model. Figure 7 The prediction results are for all modal components; Figure 8 The graph shows the prediction results for different models; Figure 9 This is a schematic diagram of the structure of the prediction device for abrupt changes in dissolved gas signals in oil provided in an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the implementation of a method for predicting abrupt changes in dissolved gas signals in oil, as provided in an embodiment of the present invention. Figure 1 As shown, the prediction method for abrupt changes in dissolved gas signals in oil includes: S110, acquire the timing signal for transformer gas monitoring; S120, The time-series signal is processed according to the variational mode decomposition algorithm to obtain multiple mode components; wherein, the variational mode decomposition algorithm uses the sum of the disorder entropy of each mode component as the objective function for parameter optimization. S130, based on the pre-established prediction model, predicts and superimposes each modal component to reconstruct the signal, thus obtaining the prediction result of the abrupt change signal of dissolved gas in oil.

[0019] This invention proposes a method for predicting non-stationary signals of dissolved gases in oil. First, the time signal is preprocessed, and the non-stationary characteristics of the signal to be predicted are weakened using a variational mode decomposition algorithm. The decomposition effect is determined by normalized entropy, and the parameters of the variational mode decomposition and prediction model are iteratively optimized using a frost-free algorithm. Next, a temporal convolutional network (TCN)-iTransformer prediction model is constructed to predict each modal component after decomposition. Finally, the prediction results of all modal components are reconstructed to obtain the final prediction result, achieving accurate prediction when non-stationary signals of dissolved gases exist in oil. In summary, this patent has good application potential in predicting sudden transformer faults.

[0020] Figure 2 This is a flowchart illustrating the implementation of a prediction method for abrupt changes in dissolved gas signals in oil, provided by another embodiment of the present invention. Figure 2 As shown, the data preprocessing of the time series signal includes two parts: feature interval extraction based on expert rules and multi-scale trend change rate extraction. The raw data of the abrupt change signal in oil is preprocessed, and feature interval extraction and multi-scale trend change rate extraction based on expert rules are performed. This preprocessing module uses preset expert rule criteria to identify the characteristic segments of the signal abrupt change and calculates the trend change rate of the signal at multiple time scales to quantify the abrupt change amplitude and rate of change characteristics of the non-stationary signal, thereby improving the targeting and accuracy of subsequent signal decomposition and prediction.

[0021] Perform feature interval extraction based on expert rules. Set a threshold for relative concentration abrupt changes. and duration threshold Point-by-point scanning of dissolved gas concentration sequence in oil If the conditions are met continuously during a certain period of time Da At each sampling time point, the segment is determined to be a sudden change characteristic interval. This rule allows for the identification of concentration spikes caused by faults without data rejection, avoiding the misinterpretation of valid fault information as noise. Next, multi-scale trend change rate extraction is performed: multiple time scales such as daily, weekly, and monthly are set. Calculate the rate of change of the signal at each scale, for example, for scale... Define the trend change rate indicator:

[0022] in, for The gas concentration at any given time. This is achieved by constructing multi-dimensional rate-of-change feature vectors for short-term, medium-term, and long-term changes. The preprocessing described above extracts and quantifies the abrupt change characteristics of the original signal at different time scales, providing richer input information and making the subsequent signal decomposition and prediction processes more targeted and robust when facing non-stationary abrupt change signals.

[0023] Figure 3 This is a schematic diagram showing the ethane gas content. (For example...) Figure 3 As shown, in the ethane gas signal, the daily rate of change significantly increases during the period of abrupt changes (up to 45%), and a synchronous upward trend also appears on the weekly and monthly scales. This rate of change, as an additional feature, is input into the model along with the original concentration values ​​to improve its ability to identify local drastic fluctuations and the evolution of potential risks.

[0024] In this embodiment of the invention, the preprocessed data is input into Variational Mode Decomposition (VMD) to decompose the non-stationary signal of dissolved gases in oil into multiple modal components. The preprocessed sequence is then input into the VMD algorithm to decompose the non-stationary signal of dissolved gases in oil into... The modal components are decomposed into the following form:

[0025] VMD introduces the Hilbert transform to transform each mode Represented as analytic signal And employing an augmented Lagrangian function and a penalty factor The constrained optimization of the original signal reconstructed by each mode is transformed into an unconstrained problem, and iterative solutions are used to minimize the sum of the bandwidths of all modes, ensuring that the total bandwidth of each mode is minimized.

[0026] In some embodiments, the method further includes: transforming the objective function into an unconstrained optimization function using an augmented Lagrangian function and a quadratic penalty factor; solving the unconstrained optimization function to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

[0027] In this embodiment of the invention, the objective of variational mode decomposition is to minimize the sum of the bandwidths of all modes. To satisfy the constraint that the sum of all modes can reconstruct the original signal, an augmented Lagrangian function is employed. and secondary penalty factor Transforming it into an unconstrained optimization form, the calculation formula is as follows:

[0028] in, u k ( t ) is the first k One modal component.

[0029] The objective function is solved iteratively using the alternating direction multiplier method. The iterative process is as follows: (1) Initialization , (No. k (central angular frequency of each mode) ; (2) Within the non-negative frequency range, update and The updated formula is:

[0030]

[0031] (3) Within the non-negative frequency range, update (Lagrange multipliers / dual variables), the update formula is:

[0032] in, The step size used to adjust the frequency improves convergence efficiency; (4) For a given error threshold If satisfied Then stop the loop. Perform an inverse Fourier transform and output the result. Otherwise, continue with step (2).

[0033] The above steps enable variational mode decomposition of non-stationary time-series signals.

[0034] To further improve the orderliness of the modal components, the frost-ice algorithm is used to optimize the decomposition number K and the penalty factor. The objective function is the sample entropy of the modal components, and its calculation formula is:

[0035]

[0036]

[0037] The smaller the sample entropy, the more regular the decomposition modulus, and the easier it is to predict.

[0038] In some embodiments, the method further includes: processing the time series signal according to the variational mode decomposition algorithm to obtain multiple initial components; calculating the disorder entropy of each initial component, and using the sum of the disorder entropies of all mode components as the objective function; wherein, the disorder entropy is used to measure the complexity and disorder of the time series data, and is determined according to the probability of the occurrence of the arrangement pattern of the sequence segments; the sequence segments are obtained by embedding and reconstructing the time series signal in the phase space.

[0039] In this embodiment of the invention, unlike traditional methods, the invention introduces permutation entropy as a quantitative indicator to measure the non-stationarity of the signal before and after VMD decomposition, and uses the RIME (Frost Ice Optimization) algorithm to adaptively optimize the parameters of VMD. Permutation entropy is used to measure the complexity and disorder of time-series data, and is defined as follows: [Embedding dimension selection...] The time series is reconstructed in phase space to obtain a length of The sequence fragments are processed and their permutation patterns are constructed. The probability of each permutation pattern occurring is calculated. The original entropy of a sequence is defined as:

[0040] in, The total number of all possible permutations.

[0041] Furthermore, the normalized entropy can be defined as:

[0042] Normalize the entropy value to the range of 0 to 1. The larger the entropy value, the more disordered the sequence (stronger non-stationarity), and the smaller the entropy value, the more ordered the sequence (higher stationarity).

[0043] This invention uses the sum of the disorder entropy of each modal component as the objective function for VMD parameter optimization, and utilizes the RIME algorithm to iteratively search for the optimal number of modes in the parameter space. and penalty factor By minimizing the sum of the entropy values ​​of each mode after decomposition, the RIME algorithm achieves adaptive optimization in the signal decomposition stage, making the fluctuations of the obtained modal components more regular and easier to predict. After VMD+RIME optimization, the chaotic entropy of each sub-signal obtained by decomposition is significantly reduced compared to the original signal, indicating that the nonlinear and non-stationary characteristics of the original gas signal have been effectively weakened, providing a more stable and reliable input source for the prediction model.

[0044] In some embodiments, solving the unconstrained optimization function to determine the optimized value of the first parameter of the variational mode decomposition algorithm includes: using the frost optimization algorithm, taking minimizing the objective function as the optimization objective, iteratively optimizing the first parameter in the parameter space to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

[0045] In this embodiment of the invention, the first parameters (K, α) of VMD directly affect the signal decomposition effect. The optimization objective is to minimize the objective function of each modal component after decomposition (in this paper, the sum of disorder entropy and normalized sample entropy is minimized; the smaller the entropy value, the more stable and predictable the modal component is), providing high-quality input for subsequent prediction models. After a preset number of iterations, when the objective function value reaches its minimum and stabilizes, the corresponding values ​​of K and α are the optimized values ​​of the first parameters of VMD, which can be used for the efficient decomposition of subsequent non-stationary signals (such as abrupt changes in dissolved gas signals in oil).

[0046] In some embodiments, the prediction results of the abrupt change signal of dissolved gas in oil are obtained by predicting each modal component according to a pre-established prediction model and superimposing and reconstructing them. This includes: predicting each modal component according to the TCN-iTransformer prediction model and superimposing and reconstructing them to obtain the prediction results of the abrupt change signal of dissolved gas in oil.

[0047] In some embodiments, the method further includes: employing a frost-ice optimization algorithm to iteratively optimize the second parameter of the TCN-iTransformer prediction model in the parameter space with the goal of minimizing the final prediction error, and determining the optimized value of the second parameter.

[0048] In this embodiment of the invention, a TCN-iTransformer prediction model is constructed. Based on the optimization of the variational mode decomposition (VMD) parameters, number of modes K and penalty factor α, using the RIME algorithm, the second parameters of the TCN-iTransformer prediction model (kernel size, attention dimension, dropout rate, and number of layers) are further incorporated into the optimization scope, achieving coordinated adaptive adjustment of parameters throughout the prediction process. Through this joint optimization strategy, the RIME algorithm simultaneously adjusts the parameters in both the signal decomposition stage and the prediction model stage during the global search process, iteratively optimizing with the goal of minimizing the final prediction error.

[0049] This study utilizes causal convolution and dilated convolution in TCN to enhance local and long-term dependencies in time series data, respectively, to handle the extraction and prediction of non-stationary information before and after decomposition, thereby improving the accuracy and interpretability of the prediction results. For the training set data, the optimal mode components of variational mode decomposition are substituted into the prediction model for training and optimization. Mean relative error (MAPE) and root mean square error (RMSE) are selected as evaluation metrics for the model. The calculation formula is as follows:

[0050]

[0051] Where n is the number of samples in the test set. and Let be the actual and predicted values ​​of the dissolved gas concentration in the oil at time i.

[0052] The modal components of the test set data are substituted into the optimized prediction model, and the final prediction result of the time series signal is obtained by reconstructing based on the prediction results of each modal component.

[0053] The optimized prediction model is superimposed and reconstructed using the future trends obtained from each modal component to synthesize the final prediction result of dissolved gas concentration in oil. Through the above steps, the method of this invention can still provide a high-precision prediction curve even when there are drastic fluctuations in the gas signal. Furthermore, this method can introduce more physically relevant monitoring features as model inputs as needed, such as transformer hotspot temperature and acoustic emission signals. These auxiliary features can provide additional fault information, forming multi-source data fusion with the gas concentration signal, improving the model's ability to perceive fault precursors and the stability of the prediction results. This multi-parameter collaborative prediction approach has not been reported in existing patents, further highlighting the uniqueness of the technical solution of this invention.

[0054] In some embodiments, the mechanism of the frost optimization algorithm includes at least one of the following: a soft frost search mechanism, a hard frost puncture mechanism, and an active greedy selection mechanism.

[0055] In this embodiment of the invention, the implementation process of the frost-ice algorithm includes three mechanisms: ① Soft Frost Search Mechanism: The soft frost search mechanism relies on a random search formula. It simulates the attachment and expansion process of frost, fog, and ice under light wind conditions, enabling rapid coverage of the entire search space in the early stages of iteration. For each particle in the population, the update formula is:

[0056] in, This represents the position of the updated particle. Let be the optimal individual position in the j-th dimension of the population. For global exploration factors. The update method for the particle's direction of motion is as follows:

[0057] For environmental factors. The update method is:

[0058] These are the upper and lower bounds of the particle's motion space, respectively.

[0059] ② Hard Frost Piercing Mechanism: The hard frost piercing mechanism allows information exchange between different individuals, thereby improving the algorithm's global convergence ability and escaping local optima. Hard frosts increase in size during their growth, resulting in a greater probability of piercing. The position update formula for hard frost particles is:

[0060] in, This indicates the new position of the particle after the update. This represents the position of the j-th particle of the current best individual in the population. This represents the normalized fitness value of the current i-th individual, indicating the probability that the i-th individual will be selected. The disturbance control factor determines whether the position should be updated.

[0061] ③ Active Greedy Selection Mechanism: The active greedy selection mechanism is used to enhance the global exploration efficiency of the population. The process is as follows: compare the fitness value of each individual before and after the update. If the updated fitness value is better, replace the individual and record the updated state; simultaneously, compare the updated fitness value with the global optimum. If the updated fitness value is better, replace the global optimum.

[0062] Figure 4 To optimize the process diagram. For example... Figure 4 As shown, the above mechanisms can be used individually or in combination. For example, a combined strategy of "soft cream search mechanism + hard cream puncture mechanism + aggressive greedy selection mechanism" is employed, and the specific implementation process is as follows: Initialization and Soft Frost Search Initiation: First, set the initial parameters for the frost algorithm: population size is 8, initial parameter positions [K (number of modes), α (penalty factor)] are [5, 1000], total iterations are 50, and the range of K is determined to be 3-8, and the range of α is determined to be 1000-5000. In the initial iteration phase (iterations 1-20), only the soft frost search mechanism is activated: new parameter combinations are generated based on a random search formula. For example, for the K dimension of the 3rd individual in the population, using the current optimal K value of 4 as a benchmark, combined with the global exploration factor ř=0.8, random movement direction θ=1.2π, and environmental factor β=0.9 (dynamically adjusted with the number of iterations), a new K value of 5 is calculated; similarly, a new α value of 1200 is generated. This mechanism quickly covers the parameter space, avoids getting trapped in local optima, and initially selects 10 sets of relatively optimal parameter combinations (such as [5, 1500], [6, 2000], etc.).

[0063] Hard Frost Puncture Fine Optimization: Entering the mid-iteration stage (iterations 21-40), the hard frost puncture mechanism is superimposed on the soft frost search. For the previously selected optimal parameter regions, the normalized fitness value of each parameter combination is calculated (using the sum of the disorder entropy of the modal components as an indicator; the smaller the entropy value, the higher the fitness value). For example, the fitness value of the parameter combination [6,4000] is 0.92 (the highest), and it is selected as the current optimal individual. According to the hard frost position update formula, when the perturbation control factor r3 = 0.3 < the current individual's normalized fitness value of 0.92, the optimal individual [6,4000] is used as the benchmark, and other parameter combinations are adjusted in a targeted manner, such as adjusting [5,3800] to [6,4100], further narrowing the parameter search range and causing the parameter combinations to converge towards the optimal region. At this time, three core parameter combinations are selected ([6,4150], [6,4179], and [6,4200]).

[0064] Active Greedy Selection to Determine the Optimal Solution: In the later stages of iteration (iterations 41-50), an active greedy selection mechanism is activated based on the two mechanisms mentioned above. For each new parameter combination generated by hard frost puncture, the sum of the disorder entropy of its corresponding modal components is calculated. For example, the newly generated parameter combination [6,4179] has a disorder entropy of 0.32 for each modal component of the ethane signal after decomposition, which is lower than 0.35 for the original core combination [6,4150] and 0.34 for [6,4200]. At this time, according to the active greedy selection rule, [6,4179] replaces the original better individual and updates the global optimal solution. In subsequent iterations, if the entropy value of the newly generated parameter combination is not lower than 0.32, [6,4179] is retained. Finally, after 50 iterations, the first parameter optimization value of VMD is determined to be K=6 and α=4179. The modal components obtained by decomposition under these parameters have the best stability, providing high-quality input for the subsequent TCN-iTransformer prediction model.

[0065] For example, when optimizing the second parameter (kernel size and attention dimension) of the TCN-iTransformer prediction model, the soft-snow search mechanism can be used separately: set the kernel size range to 2-6 and the attention dimension range to 64-256, and use the soft-snow random search formula to quickly explore the impact of different parameter combinations on the prediction error within 30 iterations. Finally, the parameter combination with kernel size = 4 and attention dimension = 128 is selected, and its corresponding prediction MAPE is only 2.761, which meets the accuracy requirements and significantly shortens the optimization time. It is suitable for scenarios with high requirements for parameter optimization efficiency.

[0066] Soft frost search is suitable for when the initial parameter space is unknown, the global scope needs to be covered quickly, or only pre-screening of parameters is required (such as quickly eliminating unreasonable values ​​when optimizing the TCN dropout rate); hard frost puncture is suitable for regions with known optimal parameters (such as determining the approximate range of K and α in VMD based on historical data), where local fine-tuning is required; aggressive greedy selection is suitable for when there are multiple sets of candidate parameters, where only the optimal needs to be selected (such as selecting TCN-iTransformer parameter combinations in the later stages of model training).

[0067] Combined use cases: The full combination of the three (soft defrost + hard defrost + aggressive greedy) is suitable for high-dimensional parameters (such as simultaneous optimization of VMD and TCN-iTransformer multiple parameters), strong signal mutations (such as sudden gas fault signals in oil), and high-precision optimization (such as transformer fault prediction). For example, in VMD parameter optimization, soft defrost covers the whole, hard defrost focuses on mutation areas, and aggressive greedy locks in the optimal K=6 and α=4179. Soft defrost + aggressive greedy is suitable for low-dimensional parameters (such as only optimizing the number of TCN layers and the drop rate), stable signals, and balancing accuracy and efficiency. Hard defrost + aggressive greedy is suitable for high-precision optimization in narrow intervals (such as α needing to be accurate to the single digit), based on directional optimization of known intervals.

[0068] Soft-frost search + hard-frost puncture: This approach is suitable for scenarios with high parameter space dimensions (such as simultaneously optimizing 1-2 parameters of VMD, K, α, and TCN), requiring global exploration followed by local fine-tuning, but with low requirements for the accuracy of selecting the best individual parameters. For example, for CO2 gas signals without clear abrupt changes but with strong non-stationarity, a soft-frost search (1-25 iterations) is first used to quickly cover the parameter space of K (3-8), α (1000-5000), and TCN kernel size (2-6), initially selecting 20 better combinations. Then, a hard-frost puncture is initiated, using these combinations as a basis to adjust the parameters in a targeted manner (e.g., adjusting K=5, α=3000, kernel=3 to K=6, α=3500, kernel=4), focusing on high-quality regions to narrow the scope. This provides parameters that meet basic accuracy requirements for subsequent predictions without additional screening, balancing global coverage and local convergence efficiency.

[0069] Soft Frost Search + Active Greedy Selection: Adaptable to low-dimensional parameter optimization (e.g., optimizing only the attention dimension of TCN from 64 to 256, and the dropout rate from 0.1 to 0.3), stable signals (e.g., H2 gas signals during normal transformer operation), and scenarios requiring a balance between accuracy and efficiency. For example, when optimizing TCN parameters, Soft Frost Search generates 15 parameter combinations using a random formula, each corresponding to a different attention dimension and dropout rate; it directly calculates the prediction error (e.g., MAPE) of each group through Active Greedy Selection, retaining the combination with the smallest error (e.g., attention dimension 128, dropout rate 0.1), eliminating the need for local fine-tuning through Hard Frost. While ensuring prediction accuracy (MAPE ≤ 3%), it reduces optimization time by 30%, adapting to the need for rapid parameter iteration.

[0070] Hard puncture combined with aggressive greedy selection: This approach is suitable for scenarios where the optimal range of known parameters is relatively high (e.g., based on the experience of the document's examples, VMD's K=6 and α=4000-4500), and high-precision optimization within a narrow range is required. For example, for VMD parameter optimization of ethane gas mutation signals, given that K=6 is fixed, only the α value needs to be precisely optimized: First, using the range of 4000-4500, candidate values ​​such as α=4100, 4150, and 4179 are generated directionally through hard puncture; then, aggressive greedy selection is used to calculate the disorder entropy of the modal components corresponding to each α, locking in the α=4179 with the smallest entropy value. Without global search, precise parameter positioning is achieved directly, meeting the high stability requirement of mutation signal decomposition.

[0071] In some embodiments, the method further includes triggering an early warning when the residual of the prediction result exceeds a preset threshold.

[0072] In this embodiment of the invention, the residual of the prediction result is the difference between the real-time measured value of the dissolved gas concentration in the oil and the model prediction value. The determination of the preset threshold needs to be combined with the industry standard for dissolved gases in transformer oil. For example, for ethane gas, the threshold can be set to 1.5-2 times the average residual value of its historical normal operation phase, or directly associated with the fault warning level.

[0073] When the residual calculated by the system in real time exceeds the preset threshold for multiple consecutive sampling periods (such as 3 consecutive sampling days), it is determined to be an abnormal signal prediction deviation. At this time, an early warning mechanism is automatically triggered: on the one hand, the operation and maintenance personnel are notified through pop-up windows, audible and visual alarms, etc. in the substation monitoring system; on the other hand, an early warning report is generated. The report includes the time period when the residual exceeds the threshold, the curve comparing the measured value and the predicted value of the corresponding gas concentration, and the potential fault types (such as overheating and partial discharge) analyzed in combination with expert rule criteria. This provides data support for operation and maintenance personnel to locate abnormal transformer conditions and formulate maintenance strategies, avoiding missed fault diagnosis or delayed processing due to signal changes.

[0074] In some embodiments, the method further includes: identifying characteristic segments of abrupt changes in a time series signal using preset expert rule criteria, and calculating the trend change rate of the time series signal on multiple time scales; wherein the characteristic segments of abrupt changes and the trend change rate are used to determine prior information in the parameter optimization process.

[0075] In this embodiment of the invention, the identified signal mutation feature segments and multi-scale trend change rates are used as prior information in the parameter optimization process. They are primarily used to guide the Frost Ice algorithm in optimizing VMD parameters (number of modes K, penalty factor α) and TCN-iTransformer model parameters (kernel size, attention dimension, etc.). For example, for the identified ethane signal mutation segment from day 120 to 150, in VMD parameter optimization, the prior information is used to narrow the search range of K (from the original 3-8 to 5-7) and the search range of α (from the original...). The range of parameters was adjusted from 1000-5000 to 3000-5000 to avoid ineffective searches in irrelevant parameter ranges. At the same time, for high trend change rates (such as daily change rate of 45%) corresponding to abrupt change segments, the TCN-iTransformer parameter optimization prioritizes exploring parameter combinations with larger convolution kernel size (such as 4-6) and higher attention dimension (such as 128-256) to ensure that the model can accurately capture signal abrupt change features, improve the targeting and efficiency of parameter optimization, and ultimately achieve efficient decomposition and high-precision prediction of gas signals in abrupt change oil.

[0076] By substituting the prediction results of all modal components into the signal reconstruction process and superimposing the prediction values ​​of each subsequence, the final prediction result of the dissolved gas concentration in the oil can be obtained. Through the above decomposition, prediction, and reconstruction steps, even if the original signal contains sudden spikes or violent fluctuations, the model of this invention can still provide a prediction curve that highly matches the actual trend, significantly improving the accuracy and reliability of gas concentration prediction under non-stationary operating conditions. Furthermore, this method is easily embedded into the online oil and gas monitoring system of power transformers: the model can be deployed and run on the edge computing device at the substation site, using the deviation between the real-time predicted value and the sensor measured value as a fault precursor criterion. When the prediction residual exceeds a preset threshold, an alarm can be triggered, thereby realizing a high real-time fault early warning function. In summary, this invention provides a complete solution from data preprocessing, signal decomposition, joint modeling to online prediction and early warning.

[0077] This invention innovatively introduces a VMD decomposition + TCN-iTransformer multi-scale prediction model architecture optimized by the frost-ice algorithm: on the one hand, it quantifies and significantly reduces the non-stationarity of the signal by using the chaos entropy index; on the other hand, it captures features at different scales through the structural synergy of TCN and Transformer models, achieving higher prediction accuracy and robustness for abrupt gas signals in oil, and has obvious innovation and superiority in the field of transformer fault trend prediction.

[0078] The following embodiment further illustrates the present invention, but is not intended to limit it. This embodiment uses online monitoring data from the oil chromatography of a 220kV transformer in a substation. The sampling period was from March 22, 2023 to November 1, 2024, totaling 540 data sets, with a sampling period of one day. Figure 3 Taking the ethane gas content as an example, the signal exhibits a jump phenomenon between day 120 and day 150, with sharp local peaks and high kurtosis. To improve the processing effect of non-stationary signals in subsequent decomposition and prediction, this invention first preprocesses the original gas signal, specifically including two stages: feature interval extraction based on expert rules and multi-scale trend change rate extraction. (1) Feature interval extraction based on expert rules. Based on engineering experience in the field of dissolved gas analysis (DGA) and relevant industry standards, a threshold for abrupt changes in relative concentration is set. And the duration threshold L, for the sequence Execution condition judgment at each time point: If the condition is met continuously during a certain period of time... This segment is then marked as a "mutation characteristic interval". In this case, we take... =15%, L=3, successfully marked the abrupt change segment of the ethane signal in the 120-150 day period, which can be used to guide subsequent decomposition and prediction models to focus more on such key fault symptom regions.

[0079] (2) To further quantify the rate of change of the signal at different time scales, the rate of change of each gas signal at the daily, weekly, and monthly scales is calculated, as follows:

[0080] This is used to characterize the multidimensional trend evolution features of the signal. In ethane gas signals, the daily rate of change significantly increases during abrupt changes (up to 45%), and a synchronous upward trend also appears on weekly and monthly scales. This rate of change is used as an additional feature, inputting it into the model along with the original concentration values ​​to improve its ability to identify local drastic fluctuations and the evolution of potential risks. After completing the above preprocessing, the model enters the variational mode decomposition stage: To ensure that the sum of the estimated bandwidths of each mode is minimized, a Hilbert transform is introduced to transform the signal... Converted into an analytic signal, forming a one-sided spectrum. : ,in For the complex representation of the analytic signal, Let be the Dirac function. To satisfy the constraint that the sum of all modes can reconstruct the original signal, an augmented Lagrange function is used. and secondary penalty factor Transforming it into an unconstrained optimization form, the calculation formula is as follows:

[0081] The frost-ice algorithm is used to optimize the number of modes K and the penalty factor a. The optimization process is as follows: Figure 4 As shown, the initial population size of the frost algorithm is set to 8, the initial frost position [K, a] is [5, 1000], and the total number of iterations T is 50. This implementation uses normalized sample entropy to evaluate the orderliness of the decomposed modal components. The formula for calculating the normalized entropy of the decomposed modulus is:

[0082] ,

[0083] The smaller the entropy of the normalized sample, the more regular the decomposition modulus and the easier it is to predict. After 50 iterations, the optimal K value is 6 and the a value is 4179. Based on these parameters, Figure 3 The modal components and center frequency of a time-series signal after variational mode decomposition. For example... Figure 5 As shown, the time-series signal is decomposed into a stable low-frequency high-amplitude component and a relatively regular high-frequency low-amplitude component, and no spectral aliasing phenomenon is observed in each modal component.

[0084] Training and optimization of the TCN-iTransformer model. For example... Figure 6As shown, the optimal decomposed modal components are taken as input and sequentially fed into the normalization layer, TCN module, and iTransformer module before outputting the prediction results. The model uses the PyTorch framework to construct the training network. The optimal modal components obtained through VMD optimization are sequentially input into the normalization layer, TCN module, and iTransformer module, and the prediction results are output. The entire network is implemented using the PyTorch framework, and the core hyperparameters and VMD decomposition parameters are jointly optimized by the RIME algorithm, forming an adaptive parameter search mechanism for the entire "decomposition-prediction" process. After calculation by the RIME optimization algorithm, the TCN convolution kernel size is set to 4, and the number of hidden layers is set to 4. The attention dimension of the iTransformer model is 128, and the dropout rate is 0.1. In this implementation, the first 480 days of dissolved gas data in oil are used as the training set, and the last 60 days of data are used as the test set, employing a stepwise prediction method. The prediction process is as follows: the gas content on day 11 is predicted based on the gas content on days 1-10, the gas content on day 12 is predicted based on the gas content on days 2-11, and so on. The model is trained and optimized using the training set data, with mean relative error and root mean square error used as evaluation metrics. The calculation formula is as follows:

[0085]

[0086] Where n is the number of samples in the test set. and Let represent the true and predicted values ​​of the dissolved gas concentration in the oil at time i. Substituting the test set data into the optimized TCN-iTransformer model, the prediction results for all modal components are as follows: Figure 7 As shown, each modal component after decomposition exhibits good periodic stationarity, thus reducing the local spikes in the original signal. The predicted results of each modal component largely coincide with the true values. The prediction results demonstrate that variational mode decomposition based on the frost-ice algorithm can reduce the nonlinearity and nonstationarity of non-stationary signals, thereby improving the accuracy of the prediction results.

[0087] Signal reconstruction based on modal component prediction results. Substitute the prediction results of all modal components into step 1 to reconstruct the time-series signal, as shown in the following figure. Figure 8As shown in Table 1, compared to the original signal prediction model (TCN-iTransformer) and the variational mode decomposition signal prediction model (VMD-TCN-iTransformer), the predicted values ​​of the model in this invention largely coincide with the actual values. The MAPE (2.761) of the model in this invention is more than three times lower than that of the original signal MAPE (8.302), and the RMSE is also significantly reduced. Furthermore, for other dissolved gases in oil, the evaluation metrics of the prediction results are shown in Table 1. Compared to the TCN-iTransformer model and the VMD-TCN-iTransformer model, the model in this invention has the best error evaluation metrics for all dissolved gases in oil in the dataset, further demonstrating that the model in this invention has excellent prediction accuracy and good generalization ability.

[0088] Table 1 Evaluation of Prediction Results

[0089] This invention pioneered the use of multi-source data fusion for predicting non-stationary dissolved gases, enriching the model's perceptual dimensions. The incorporation of these physically relevant features significantly enhances the model's sensitivity to fault precursors and its ability to identify anomalies, avoiding potential misjudgments or delays that might occur when relying solely on a single gas concentration signal. Furthermore, this invention improves the reliability and stability of results through multivariate collaborative prediction, providing a more comprehensive and effective technical means for preventing sudden transformer failures.

[0090] Figure 2 This is a schematic diagram of the structure of a predictive device for abrupt changes in dissolved gas signals in oil, provided in an embodiment of the present invention. Figure 9 As shown, in some embodiments, the prediction device 9 for abrupt changes in dissolved gas signals in oil includes: The acquisition module 910 is used to acquire the timing signal of the transformer gas monitoring. The decomposition module 920 is used to process the time-series signal according to the variational mode decomposition algorithm to obtain multiple mode components; wherein, the variational mode decomposition algorithm uses the sum of the chaos entropy of each mode component as the objective function for parameter optimization. The prediction module 930 is used to predict and reconstruct each modal component based on a pre-established prediction model to obtain the prediction results of the abrupt change signal of dissolved gas in oil.

[0091] Optionally, the prediction device 9 for abrupt changes in dissolved gas signals in oil also includes an optimization module for: transforming the objective function into an unconstrained optimization function using an augmented Lagrangian function and a quadratic penalty factor; solving the unconstrained optimization function to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

[0092] Optionally, an optimization module is used to process the time series signal according to the variational mode decomposition algorithm to obtain multiple initial components; calculate the disorder entropy of each initial component, and use the sum of the disorder entropy of all mode components as the objective function; wherein, the disorder entropy is used to measure the complexity and disorder of the time series data, and is determined according to the probability of the occurrence of the arrangement pattern of the sequence segments; the sequence segments are obtained by embedding and reconstructing the time series signal in the phase space.

[0093] Optionally, an optimization module is used to employ the frost-ice optimization algorithm to iteratively optimize the first parameter in the parameter space with the goal of minimizing the objective function, thereby determining the optimized value of the first parameter of the variational mode decomposition algorithm.

[0094] Optionally, the prediction module 930 is used to predict and superimpose the modal components according to the TCN-iTransformer prediction model to obtain the prediction results of the abrupt change signal of dissolved gas in oil.

[0095] Optionally, the optimization module is also used to employ the frost-ice optimization algorithm to iteratively optimize the second parameter of the TCN-iTransformer prediction model in the parameter space with the goal of minimizing the final prediction error, and to determine the optimized value of the second parameter.

[0096] Optionally, the mechanism of the frost optimization algorithm includes at least one of the following: soft frost search mechanism, hard frost puncture mechanism, and active greedy selection mechanism.

[0097] Optionally, the prediction device 9 for abrupt changes in dissolved gas signals in oil also includes an early warning module for triggering an early warning when the residual of the prediction result exceeds a preset threshold.

[0098] Optionally, the prediction device 9 for abrupt changes in dissolved gas signals in oil also includes a preprocessing module, used to: identify characteristic segments of abrupt changes in time series signals using preset expert rule criteria, and calculate the trend change rate of time series signals on multiple time scales; wherein, the characteristic segments of signal abrupt changes and the trend change rate are used to determine prior information in the parameter optimization process.

[0099] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0100] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting abrupt changes in dissolved gas signals in oil, characterized in that, include: Acquire timing signals for transformer gas monitoring; The time-series signal is processed by the variational mode decomposition algorithm to obtain multiple modal components; wherein, the variational mode decomposition algorithm uses the sum of the disorder entropy of each modal component as the objective function for parameter optimization. Based on the pre-established prediction model, each modal component is predicted and superimposed to reconstruct the signal, thus obtaining the prediction results of the abrupt change signal of dissolved gas in oil.

2. The prediction method for abrupt changes in dissolved gas signals in oil according to claim 1, characterized in that, The method further includes: The objective function is transformed into an unconstrained optimization function by using an augmented Lagrangian function and a quadratic penalty factor. Solve the unconstrained optimization function to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

3. The prediction method for abrupt changes in dissolved gas signals in oil according to claim 2, characterized in that, The method further includes: The time-series signal is processed using a variational mode decomposition algorithm to obtain multiple initial components; The disorder entropy of each initial component is calculated, and the sum of the disorder entropies of all the modal components is used as the objective function; wherein, the disorder entropy is used to measure the complexity and disorder of the time series data, and is determined according to the probability of the occurrence of the arrangement pattern of the sequence segments; the sequence segments are obtained by embedding and reconstructing the time series signal in the phase space.

4. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 2, characterized in that, Solving the unconstrained optimization function to determine the optimized values ​​of the first parameter of the variational mode decomposition algorithm includes: The frost-ice optimization algorithm is adopted, with minimizing the objective function as the optimization objective. The first parameter is iteratively optimized in the parameter space to determine the optimized value of the first parameter of the variational mode decomposition algorithm.

5. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 1, characterized in that, Based on a pre-established prediction model, the modal components are predicted and reconstructed to obtain the prediction results of abrupt changes in dissolved gas signals in oil, including: Based on the TCN-iTransformer prediction model, the modal components are predicted and superimposed to reconstruct the results, thus obtaining the prediction results of the abrupt change signal of dissolved gas in oil.

6. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 5, characterized in that, The method further includes: The frost-ice optimization algorithm is adopted, with the goal of minimizing the final prediction error. The second parameter of the TCN-iTransformer prediction model is iteratively optimized in the parameter space to determine the optimal value of the second parameter.

7. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 1, characterized in that, The mechanism of the frost optimization algorithm includes at least one of the following: Soft cream search mechanism, hard cream puncture mechanism, and active greedy selection mechanism.

8. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 1, characterized in that, The method further includes: An alert is triggered when the residual of the prediction result exceeds a preset threshold.

9. The method for predicting abrupt changes in dissolved gas signals in oil according to claim 1, characterized in that, The method further includes: The characteristic segments of abrupt changes in a time series signal are identified using preset expert rule criteria, and the trend change rate of the time series signal is calculated at multiple time scales; wherein, the characteristic segments of the signal abrupt changes and the trend change rate are used to determine prior information in the parameter optimization process.

10. A predictive device for abrupt changes in dissolved gas signals in oil, characterized in that, include: The acquisition module is used to acquire the timing signals of transformer gas monitoring; The decomposition module is used to process the time-series signal according to the variational mode decomposition algorithm to obtain multiple mode components; wherein, the variational mode decomposition algorithm uses the sum of the chaos entropy of each mode component as the objective function for parameter optimization. The prediction module is used to predict and reconstruct each modal component based on a pre-established prediction model, so as to obtain the prediction results of the abrupt change signal of dissolved gas in oil.

Images