Arc prediction method and system for solid-state on-load tap changer of a transformer

Abstract

This application provides a method and system for predicting electric arcs in solid-state on-load tap changer of transformers, relating to the field of power technology. It identifies early insulation degradation characteristics by collecting oil insulation parameters and dissolved gas concentration data in the oil; it collects transient current waveforms during tap changer initiation and switching operations, obtains current waveform segments, and extracts features to form current feature vectors; it processes the oil insulation parameters, gas concentration data, early insulation degradation characteristics, and current feature vectors to obtain corresponding optimized data; it uses the convolutional layers and long short-term memory network of a pre-built deep learning model to process all optimized data, generating a time-series associated feature sequence; it performs dimensional matching and fusion processing on the feature array and the time-series associated feature sequence to generate an electric arc probability prediction result, improving the sensitivity of weak electric arc identification and the accuracy of early warning during solid-state on-load tap changer operation.

Description

Technical Field

[0001] This application relates to the field of power technology, and in particular to a method and system for predicting electric arcs in a transformer with solid-state on-load tap changer. Background Technology

[0002] In the operation of power systems, transformers, as one of the core pieces of equipment, directly affect the safe operation of the power grid due to their stability and reliability. Solid-state on-load tap-changing transformers, with their advantages such as fast response speed and no mechanical wear, are gradually becoming key equipment in high-reliability power supply scenarios. However, during the tap changer switching under load conditions, arcing is prone to occur. If it is not identified and intervened in time, the arc may cause the insulating oil to crack, abnormal gas release, or even lead to a sudden increase in internal pressure and insulation breakdown, seriously threatening equipment safety.

[0003] Current research has proposed an arc risk assessment method based on a combination of multi-source sensor fusion and shallow machine learning models. This method simultaneously collects dissolved gas content, oil dielectric strength parameters, and high-frequency current signals during tap changer operation in transformer oil. It then uses a support vector machine to jointly classify the extracted current frequency domain features and oil parameters to determine the presence of potential discharge risks. However, existing methods exhibit significant limitations in practical applications. For example, due to the use of static feature input and a linear classification model, they struggle to capture the local dynamic evolution of transient current waveforms during voltage regulation, particularly lacking sensitivity to non-stationary signal changes caused by weak arcs. Furthermore, the feature extraction process relies on manual experience to set windows and transformation methods, lacking adaptive learning capabilities and failing to cope with varying operating conditions under different load levels and environmental disturbances, thus limiting its generalization performance and early warning accuracy in complex operating environments. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for predicting electric arcs in solid-state on-load tap changer of transformers, so as to solve the problems in the prior art such as insufficient sensitivity to non-stationary signal changes caused by weak electric arcs and difficulty in coping with changes in operating conditions under different load levels and environmental disturbances.

[0005] To address the aforementioned technical problems, in a first aspect, this application provides a method for predicting arcing in a transformer with solid-state on-load tap changer, comprising: Collect data on oil insulation parameters and dissolved gas concentrations in the oil; Based on the gas concentration data, early insulation degradation characteristics were identified; The transient current waveform during the tap changer start-up switching operation is collected to obtain current waveform segments, and features are extracted from the current waveform segments to form current feature vectors. The oil insulation parameters, gas concentration data, early insulation degradation characteristic data, and current feature vector are spatiotemporally aligned and normalized to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data, and optimized current feature vector. Using the convolutional layer of a pre-built deep learning model, the optimized current feature vector is processed by sliding window to generate a feature array. Using the long short-term memory network of the pre-built deep learning model, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data are correlated to generate a time-series correlated feature sequence. The feature array and the time-series associated feature sequence are subjected to dimensional matching and fusion processing to generate an arc occurrence probability prediction result.

[0006] Optionally, the optimized current feature vector is processed by sliding windowing using the convolutional layers of a pre-built deep learning model to generate a feature array. Then, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data are correlated using the long short-term memory network of the pre-built deep learning model to generate a time-series correlated feature sequence, including: Set a sliding window and a window movement interval, and extract a sub-vector segment from the optimized current feature vector based on the sliding window and the window movement interval; By using the convolutional layers of a pre-built deep learning model, features are extracted from each sub-vector segment to obtain local feature values ​​for each sub-vector segment. These local feature values ​​include the range of numerical fluctuations, the frequency of numerical changes, and the number of times numerical peaks occur within the corresponding sub-vector segment. Arrange the local feature values ​​of all sub-vector segments to form a feature array. The row dimension of the feature array corresponds to the window truncation order, and the column dimension corresponds to the local feature values. The optimized oil insulation parameters, optimized gas concentration data, and optimized degradation characteristic data are divided into multiple time-series segments, and each time-series segment includes a preset number of data points. By using the long short-term memory network of a pre-built deep learning model, the associated feature values ​​corresponding to each time segment are generated, and the associated feature values ​​of all time segments are combined to form a time-series associated feature sequence.

[0007] Optionally, the associated feature values ​​corresponding to each temporal segment are generated through the long short-term memory network of a pre-built deep learning model, and the associated feature values ​​of all temporal segments are combined to form a temporal associated feature sequence, including: Based on the time interval order corresponding to each time segment, all time segments are numbered to form multiple numbered time segments; Based on the numerical distribution of the first time sequence segment with the first number, the associated feature value corresponding to the first time sequence segment is generated using the long short-term memory network of the pre-built deep learning model. Based on the numerical distribution of the first time sequence segment with the first number, calculate the similarity between the numerical difference magnitude and numerical change trend of each subsequent time sequence segment with a number greater than the first time sequence segment and the previously input time sequence segment, and convert the similarity of the numerical difference magnitude and numerical change trend into the corresponding associated feature value of the subsequent time sequence segment. Based on the numbering order, all associated feature values ​​are arranged to form a time-series associated feature sequence that includes time-dimensional correlation.

[0008] Secondly, this application provides an arc prediction system for solid-state on-load tap changer in transformers, comprising: The data acquisition module is used to collect data on oil insulation parameters and dissolved gas concentrations in the oil. The identification module is used to identify early insulation degradation characteristic data based on the gas concentration data; The extraction module is used to collect the transient current waveform when the tap changer starts switching, obtain current waveform segments, and extract features from the current waveform segments to form a current feature vector. The processing module is used to perform spatiotemporal alignment and normalization processing on the oil insulation parameters, the gas concentration data, the early insulation degradation characteristic data and the current feature vector to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data and optimized current feature vector. The correlation module is used to perform sliding window processing on the optimized current feature vector using the convolutional layer of the pre-built deep learning model to generate a feature array. It also uses the long short-term memory network of the pre-built deep learning model to perform correlation processing on the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data to generate a time-series correlation feature sequence. The generation module is used to perform dimensional matching and fusion processing on the feature array and the time-series associated feature sequence to generate an arc occurrence probability prediction result.

[0009] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of an arc prediction method for a transformer solid-state on-load tap changer as described in the first aspect above.

[0010] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the arc prediction method for a transformer solid-state on-load tap changer as described in the first aspect above.

[0011] This application provides a method for predicting arcing in solid-state on-load tap changer of transformers. By collecting oil insulation parameters and dissolved gas concentration data, it achieves multi-dimensional perception of the transformer's internal insulation state and potential chemical degradation processes, providing a data foundation for identifying early fault signs. Based on gas concentration data, it identifies early insulation degradation characteristics to capture potential arcing conditions, improving the foresight of early warnings. Simultaneously, it collects transient current waveforms during tap changer initiation and switching operations, extracts current waveform segments to form current feature vectors, and focuses on the key stage of tap changer operation to obtain high-resolution signals reflecting the initial dynamics of the arc, providing direct evidence for capturing weak and transient discharge characteristics. Subsequently, it performs spatiotemporal alignment and normalization processing on oil insulation parameters, gas concentration data, early insulation degradation characteristic data, and current feature vectors to obtain optimized parameters and features, eliminating inconsistencies in time and scale between multi-source data and ensuring the accuracy of subsequent fusion analysis. Based on this, it utilizes deep... The convolutional layers of the deep learning model perform sliding window processing on the optimized current feature vector, automatically extracting local dynamic features within sub-segments and generating feature arrays. This overcomes the limitations of traditional methods that rely on manual experience for feature extraction, effectively enhancing the sensitivity to non-stationary changes in transient current. Simultaneously, the long short-term memory network of the deep learning model is used to correlate the optimized oil insulation parameters, gas concentration data, and degradation feature data, mining their evolution patterns and intrinsic correlations in the time dimension, generating a time-series correlated feature sequence, and improving the model's adaptability and generalization performance to different operating conditions and environmental disturbances. Finally, the feature array and the time-series correlated feature sequence are dimensionally matched and fused to achieve deep coupling between electrical transient characteristics and chemical / physical state information, comprehensively generating arc occurrence probability prediction results. This improves the sensitivity of weak arc identification and the reliability of early warning during solid-state on-load tap changer processes, breaking through the technical bottlenecks of existing methods in dynamic feature capture, adaptive learning, and early warning accuracy in complex environments.

[0012] Furthermore, by applying a sliding window to the optimized current feature vector and using convolutional layers to extract local dynamic features such as the range of numerical fluctuations, frequency of change, and number of peak occurrences within each sub-vector segment, an ordered feature array is formed. Simultaneously, the optimized oil parameters, gas concentration, and degradation features are divided into time-series sub-segments and input into a long short-term memory network to capture their cross-time-period dynamic correlations, generating a time-dependent correlation feature sequence, thus providing a structured multimodal feature representation for subsequent fusion. This addresses the problems of existing methods, which rely on static features and linear models, resulting in insufficient capture of the local evolution of transient currents and low sensitivity to weak non-stationary signals. Sliding window convolution enables adaptive perception of subtle dynamic changes in the current waveform during voltage regulation, avoiding the subjective limitations of manually setting feature windows. Combined with the long short-term memory network's ability to model the long-term evolution trends of oil and gas parameters, the model's adaptability to different load conditions and environmental disturbances is enhanced, improving the generalization ability and early warning accuracy of early arc risk under complex operating conditions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating an arc prediction method for solid-state on-load tap changer of a transformer provided in this application embodiment; Figure 2 A schematic diagram illustrating a specific implementation of an arc prediction method for solid-state on-load tap changer of a transformer provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a transformer solid-state on-load tap changer arc prediction system provided in an embodiment of this application. Detailed Implementation

[0015] To address the safety hazards posed by arc discharge during load switching in solid-state on-load tap-changing transformers, existing technologies rely on multi-source sensor fusion and shallow machine learning models for risk assessment. However, using static feature input and linear classification methods makes it difficult to capture the local dynamic evolution of the current waveform during transient voltage regulation, resulting in insufficient response to non-stationary signal changes caused by weak arcs. Furthermore, the feature extraction process relies on manually set windows and transformation methods, lacking adaptive capabilities, which limits generalization performance under different loads and environmental disturbances, leading to low early warning accuracy. To solve these problems, this application focuses on fusing oil parameters and gas concentrations reflecting insulation status with transient current signals reflecting electrical transient behavior. It achieves collaborative processing of multi-source heterogeneous data through spatiotemporal alignment and normalization, and introduces convolutional layers from a deep learning architecture to perform sliding window analysis on the optimized current feature vector, automatically extracting local fluctuation features. Simultaneously, it utilizes a long short-term memory network to model the temporal correlation between oil and gas parameters. Finally, it fuses the two types of features to output the arc occurrence probability, thereby improving the perception and prediction reliability of weak arcs under complex operating conditions.

[0016] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] The core of this application is to provide a method for predicting arcing in a transformer with solid-state on-load tap changer. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: Step 101: Collect data on oil insulation parameters and dissolved gas concentrations in the oil.

[0018] In this step, oil insulation parameters refer to parameters reflecting the insulation performance of transformer insulating oil. These parameters can include dielectric constant, volume resistivity, breakdown voltage, etc., and are obtained based on the detection of the physicochemical properties of the insulating oil. They are used to assess whether the oil's insulation capacity is normal. Dissolved gases in the oil refer to gases generated and dissolved in the insulating oil during the deterioration or failure of the transformer's insulating materials. These dissolved gases can include methane, ethane, acetylene, etc., and are obtained based on gas detection of the oil sample. Changes in their types and contents can reflect the internal condition of the equipment. Gas concentration data refers to the content data of various gases dissolved in the oil, obtained based on continuous monitoring of the oil sample by a gas detection device.

[0019] Step 102: Based on the gas concentration data, identify early insulation degradation characteristic data.

[0020] In this step, early insulation degradation characteristic data refers to characteristic information that reflects the early degradation state of the transformer insulation system.

[0021] In this embodiment of the application, step 102 specifically includes the following steps: Step 201: Integrate the concentration data of the target gas components related to insulation degradation in the gas concentration data to obtain the concentration time series of different target gas components, and mark the time nodes of concentration value sudden increase in each concentration time series. The concentration value sudden increase is when the concentration difference between adjacent collection times exceeds the preset increment.

[0022] In this step, the target gas component related to insulation degradation refers to the gaseous component specifically generated by transformer insulation materials during thermal aging, electrical aging, and other degradation processes. This target gas component can be acetylene, hydrogen, methane, carbon monoxide, etc., determined based on the insulation degradation mechanism and a large number of fault cases. The concentration time series refers to the sequence formed by arranging the concentration values ​​of a single insulation degradation-related target gas component at different acquisition times in chronological order. A sudden increase in concentration value refers to the phenomenon where the difference in the concentration values ​​of the target gas component at two adjacent acquisition times in the concentration time series exceeds a preset increment. The preset increment refers to a pre-set threshold for the concentration difference used to determine whether the concentration of the target gas component has an abnormal sudden increase, determined based on the natural variation range of the target gas component concentration during normal transformer operation.

[0023] In this embodiment, firstly, target gas components directly related to the degradation process of transformer insulation materials are screened from the collected gas concentration data. Then, the concentration values ​​of each target gas component at different collection times are arranged sequentially according to the collection time of the gas concentration data to form a concentration time series corresponding to each target gas component. Afterwards, the concentration values ​​of two adjacent collection times in each concentration time series are compared one by one, and the concentration difference between adjacent collection times is calculated. If the concentration difference exceeds a preset increment used to determine whether the concentration has increased abnormally, the next collection time is marked as the time node of the sudden increase in concentration value in the concentration time series.

[0024] Step 202: Calculate the concentration change rate at each time point, and integrate the concentration change rates at each time point into a concentration change rate sequence according to the order of the time points.

[0025] In this step, the concentration change rate sequence refers to the sequence formed by arranging the target gas component concentration change rate calculated at each time point of sudden increase in concentration value in chronological order.

[0026] In this embodiment, the concentration value of the target gas component corresponding to each time node is first determined and recorded as the current concentration value. The concentration value corresponding to the previous collection time of the time node is also determined and recorded as the preceding concentration value. Then, the difference between the current concentration value and the preceding concentration value is calculated, and the difference is divided by the time difference between adjacent collection times to obtain the concentration change rate of the target gas component at that time node. After that, the concentration change rates corresponding to all time nodes are arranged in chronological order to form a concentration change rate sequence.

[0027] Step 203: Obtain the spectral characteristic peak intensity of the target gas component at each time point, and associate the spectral characteristic peak intensity with the concentration change rate sequence to form a joint feature group, which includes the concentration change amplitude and spectral characteristics.

[0028] In this step, the intensity of the spectral characteristic peak refers to the intensity value of the spectral characteristic peak exhibited by the target gas component related to insulation degradation at a specific wavelength, obtained based on the spectral detection of the target gas component. The joint characteristic set refers to the set formed by correlating the concentration change rate corresponding to each time point of sudden concentration increase with the intensity of the spectral characteristic peak corresponding to that time point. The concentration change amplitude refers to the degree of concentration change of the target gas component related to insulation degradation at the moment of sudden concentration increase, directly reflected by the concentration change rate at that time point. Spectral characteristics refer to the unique properties exhibited by the target gas component related to insulation degradation in spectral analysis, directly reflected by the intensity of the spectral characteristic peak, obtained based on spectral detection.

[0029] In this embodiment, at each marked time point, an oil sample containing the target gas component is collected synchronously. The target gas component in the oil sample is analyzed by a spectral detection device to identify the unique spectral characteristic peaks of the target gas component. For example, the spectral characteristic peak can be an absorption peak or an emission peak at a specific wavelength. The intensity value of the spectral characteristic peak is recorded to obtain the intensity of the spectral characteristic peak. Then, the intensity of the spectral characteristic peak corresponding to each time point is correlated one-to-one with the concentration change rate corresponding to that time point in the concentration change rate sequence to form a joint feature group including the concentration change rate and the intensity of the spectral characteristic peak at each time point.

[0030] Step 204: Extract the target abnormal segment from the joint feature group where the concentration change rate and the intensity of the spectral characteristic peak are synchronously abnormal, and integrate the concentration change rate, spectral characteristic peak intensity and time node corresponding to the target abnormal segment to form early insulation degradation characteristic data.

[0031] In this step, the target anomaly segment refers to the joint feature segment corresponding to consecutive time nodes in the joint feature group where the concentration change rate exceeds the concentration anomaly threshold and the intensity of the spectral feature peak simultaneously exceeds the spectral feature anomaly threshold.

[0032] In this embodiment, anomaly judgment thresholds for concentration change rate and spectral characteristic peak intensity are first set. Then, the concentration change rate and spectral characteristic peak intensity of each time node in the joint feature group are checked one by one. If the concentration change rate of a certain time node exceeds the concentration change rate anomaly threshold, and the spectral characteristic peak intensity of that time node also exceeds the spectral characteristic peak intensity anomaly threshold, then the joint feature segment corresponding to that time node and the adjacent consecutive time nodes that also meet the synchronous anomaly conditions are marked as target anomaly segments. Finally, all target anomaly segments are collected, and the concentration change rate, spectral characteristic peak intensity, and corresponding time node information corresponding to each target anomaly segment are integrated to form early insulation degradation characteristic data that can reflect the early degradation state of the transformer insulation system.

[0033] This application embodiment verifies through two dimensions of concentration change and spectral characteristics, which can accurately capture subtle signals of early insulation degradation, providing a reliable early degradation feature basis for subsequent arc prediction, reducing the risk of false alarms and missed alarms in arc warnings caused by inaccurate judgment of insulation degradation trends, while avoiding the limitations of manual experience in screening features, and improving the objectivity and accuracy of early insulation degradation identification.

[0034] Step 103: Collect the transient current waveform when the tap changer starts switching operation, obtain the current waveform segment, and extract features from the current waveform segment to form a current feature vector.

[0035] In this step, the tap changer initiation switching operation refers to the tap position switching action initiated by the transformer tap changer to adjust the output voltage. This is a critical operation in the transformer voltage regulation process, during which arcing can easily occur due to poor contact. The transient current waveform refers to the curve showing the brief, rapidly changing current in the circuit over time during the tap changer initiation switching operation. It is obtained in real-time from a current sensor, and its shape reflects the electrical state during switch switching. A current waveform segment refers to the portion of the transient current waveform directly related to the tap changer switching operation. The current feature vector refers to the vector formed by combining features extracted from the current waveform segment that reflect the characteristics of current changes in a preset order.

[0036] Step 104: Perform spatiotemporal alignment and normalization on the oil insulation parameters, the gas concentration data, the early insulation degradation characteristic data, and the current characteristic vector to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data, and optimized current characteristic vector.

[0037] In this step, the optimized oil insulation parameters refer to the parameters obtained after spatiotemporal alignment, data supplementation, and normalization of the original oil insulation parameters. These parameters are used to eliminate differences caused by different acquisition times and numerical ranges, facilitating subsequent model processing. The optimized gas concentration data refers to the data obtained after spatiotemporal alignment, data supplementation, and normalization of the original gas concentration data. The optimized degradation characteristic data refers to the data obtained after spatiotemporal alignment, data supplementation, and normalization of the early insulation degradation characteristic data. The optimized current feature vector refers to the vector obtained after normalizing the original current feature vector.

[0038] In this embodiment of the application, step 104 specifically includes the following steps: Step 401: Using the operation time of the tap changer switching operation as the reference point, determine the first time interval of the oil insulation parameter acquisition time relative to the reference point, the second time interval of the gas concentration data relative to the reference point, and the third time interval of the early insulation degradation characteristic data relative to the reference point.

[0039] In this step, the reference point refers to the specific moment when the tap changer begins its switching operation, determined based on the tap changer's action command trigger time. The first time interval refers to the range of oil insulation parameter acquisition time relative to the reference point, set based on the correlation between oil parameters and tap changer switching operations. The second time interval refers to the range of gas concentration data acquisition time relative to the reference point, set based on the lag in gas concentration changes. The third time interval refers to the time range of early insulation degradation characteristic data relative to the reference point, set based on the gradual nature of insulation degradation.

[0040] In this embodiment, the moment when the tap changer begins to perform the switching action is set as the reference point. Then, based on the transformer operating characteristics and data acquisition requirements, the time range for collecting oil insulation parameters relative to the reference point is defined as the first time interval. Similarly, the time range for collecting gas concentration data relative to the reference point is defined as the second time interval. The time range for early insulation degradation characteristic data relative to the reference point is defined as the third time interval, ensuring that the time range of each data is related to the tap changer switching operation.

[0041] Step 402: Based on the first time interval, the second time interval, and the third time interval, extract the first data segment of oil insulation parameters within a preset time span, the second data segment of gas concentration data within a preset time span, and the third data segment of early insulation degradation characteristic data within a preset time span.

[0042] In this step, the preset time span refers to a fixed time length pre-set for extracting each data segment, determined based on the duration of the impact of the tap changer switching operation. The second data segment refers to the portion of data extracted from the gas concentration data that belongs to the second time interval and has a length of the preset time span. The third data segment refers to the portion of data extracted from the early insulation degradation characteristic data that belongs to the third time interval and has a length of the preset time span.

[0043] In this embodiment, a fixed time length is first set as a preset time span. Then, a portion of the data belonging to the first time interval and with a length of the preset time span is extracted from the complete data of the oil insulation parameters to form a first data segment. A portion of the data belonging to the second time interval and with a length of the preset time span is extracted from the complete data of the gas concentration data to form a second data segment. A portion of the data belonging to the third time interval and with a length of the preset time span is extracted from the complete data of the early insulation degradation characteristic data to form a third data segment.

[0044] Step 403: Perform parameter supplementation processing on the sub-segments in the first data segment, the second data segment, and the third data segment whose data length is less than the preset length, to obtain the supplemented first data segment, the supplemented second data segment, and the supplemented third data segment.

[0045] In this step, the preset length refers to the pre-defined standard for the number of data points that each data segment should include, calculated based on a preset time span and data acquisition frequency. A sub-segment refers to a continuous portion of the first, second, or third data segment where the number of data points is less than the preset length. This is due to potential interruptions or delays in data acquisition and requires supplementation to ensure it meets the preset length. The supplemented first data segment refers to the data segment obtained after parameter supplementation of the sub-segments in the first data segment that are shorter than the preset length. The number of data points equals the preset length, used to eliminate the impact of incomplete data acquisition. The supplemented second data segment refers to the data segment obtained after parameter supplementation of the sub-segments in the second data segment that are shorter than the preset length. The number of data points equals the preset length, ensuring the integrity of the gas concentration data. The supplemented third data segment refers to the data segment obtained after parameter supplementation of the sub-segments in the third data segment that are shorter than the preset length. The number of data points equals the preset length, ensuring the continuity of early insulation degradation characteristic data.

[0046] In this embodiment, sub-segments in the first, second, and third data segments whose length is less than a preset length are supplemented with parameters to obtain supplemented data segments. Specifically, a standard number of data points is first set as the preset length, and then the first, second, and third data segments are checked one by one: if the total number of data points in a data segment is less than 120, or if there are two or more consecutive missing data points forming a sub-segment, then three valid data points before and after the missing position of the segment are extracted; if these data points show a linear trend, specifically by calculating the fluctuation range of the difference between adjacent data points, if the fluctuation range is less than 5%, it is determined to be linear. If the trend is considered, the value of the missing position is calculated based on the average rate of change of the first 3 data points. If the data points show a smooth trend, for example, the difference between adjacent data points fluctuates by more than 5% but the overall trend is a gradual increase or decrease, then the moving average method is used for parameter supplementation. Specifically, the average value of the first 3 data points is taken as the starting value for the missing segment, and the value of each subsequent missing point is slightly adjusted based on the average slope of the first 3 data points on the basis of the previous value. After the filling is completed, the number of data points is checked again to ensure that the final supplemented first data segment, supplemented second data segment, and supplemented third data segment all include 120 data points, and the changing trend of the filled value is consistent with the change trend of the effective data before and after.

[0047] Step 404: Set the first reference range for oil insulation parameters, the second reference range for gas concentration data, the third reference range for early insulation degradation characteristic data, and the fourth reference range for current characteristic vector under normal operating conditions of the transformer.

[0048] In this step, normal operating condition refers to the transformer's state of being fault-free, with stable load and no insulation system deterioration, defined based on the transformer's factory standards and historical data from long-term fault-free operation. The first reference interval refers to the range of oil insulation parameters under normal operating conditions. The second reference interval refers to the range of gas concentration data under normal operating conditions. The third reference interval refers to the range of early insulation deterioration characteristic data under normal operating conditions. The fourth reference interval refers to the range of current characteristic vector values ​​during tap changer switching under normal operating conditions.

[0049] In this embodiment, first, second, third, and fourth reference intervals are defined. Specifically, historical data of the transformer under normal operating conditions for three consecutive months are collected. The sample size for oil insulation parameters is no less than 500 sets, the sample size for gas concentration data is no less than 1000 sets, the sample size for early insulation degradation characteristic data is no less than 300 sets, and the sample size for current characteristic vector during normal tap changer switching is no less than 200 sets. For oil insulation parameters, 5% of the extreme values ​​are removed, and the minimum and maximum values ​​of the remaining 95% of the data are taken as the lower and upper limits of the first reference interval. Similarly, after removing 5% of the extreme values ​​for gas concentration data, the range of the maximum and minimum values ​​of the remaining data is taken as the second reference interval. After removing 5% of the extreme values ​​for early insulation degradation characteristic data, the range of the maximum and minimum values ​​of the remaining data is taken as the third reference interval. For each feature dimension of the current characteristic vector, 5% of the extreme values ​​are removed, and the combination of the maximum and minimum values ​​of each dimension forms the fourth reference interval. All reference intervals must be repeatedly statistically verified at least three times to ensure coverage of more than 95% of parameter fluctuations under normal operating conditions.

[0050] Step 405: The supplemented first data segment, the supplemented second data segment, the supplemented third data segment, and the current feature vector are respectively converted into the first reference interval, the second reference interval, the third reference interval, and the fourth reference interval to obtain the converted first data segment, the converted second data segment, the converted third data segment, and the converted current feature vector. In this step, the transformed first data segment refers to the data segment obtained by mapping the supplemented first data segment to the first reference interval through a linear transformation. The transformed second data segment refers to the data segment obtained by mapping the supplemented second data segment to the second reference interval through a linear transformation. The transformed third data segment refers to the data segment obtained by mapping the supplemented third data segment to the third reference interval through a linear transformation. The transformed current characteristic vector refers to the vector obtained by mapping the current characteristic vector to the fourth reference interval through a linear transformation.

[0051] In this embodiment of the application, for each oil insulation parameter data point in the supplemented first data segment, the minimum and maximum values ​​of the original data of the dielectric constant of the segment, as well as the upper and lower limits of the first reference interval, are first determined. The converted value of each data point is obtained by calculating the product of the difference between the original data value and the lower limit of the original data, the quotient of the first reference interval span and the original data span, and then summing it with the lower limit of the first reference interval. In this way, all data points are mapped to the first reference interval to form the converted first data segment.

[0052] Using the same method, the transformed value of the supplemented second data segment is obtained by multiplying the difference between the original concentration value and the original minimum concentration value, the quotient of the second reference interval span and the original concentration span, and then summing it with the lower limit of the second reference interval. This maps the supplemented second data segment to the second reference interval, resulting in the transformed second data segment. The transformed value of the supplemented third data segment is obtained by multiplying the difference between the original concentration change rate value and the original minimum concentration change rate value, the quotient of the third reference interval span and the original concentration change rate span, and then summing it with the lower limit of the third reference interval. This maps the supplemented third data segment to the third reference interval, resulting in the transformed third data segment. The transformed values ​​of each feature dimension of the current feature vector are obtained by multiplying the difference between the original feature value and the original minimum feature dimension value, the quotient of the fourth reference interval span and the original feature dimension span, and then summing it with the lower limit of the fourth reference interval. This maps the current feature vector to the fourth reference interval, resulting in the transformed current feature vector. This ensures that the values ​​of all transformed data fall within the corresponding reference intervals, and that the scale of different types of data is uniform.

[0053] Step 406: Rearrange all the converted first data segment, converted second data segment, converted third data segment, and converted current feature vector to form optimized oil insulation parameters, optimized gas concentration data, optimized degradation feature data, and optimized current feature vector.

[0054] In this embodiment, a corresponding timestamp is added to each data point of the first data segment after conversion, and the data are reordered in ascending order of timestamp. If data with the same timestamp exists, the average value is taken. The sorted data points correspond sequentially to the switching time 60 minutes before the tap changer operation, 59 minutes before the operation, ..., 1 minute after the operation, ..., 30 minutes after the operation, forming optimized oil insulation parameters. For the second data segment after conversion, the data is also sorted in ascending order of timestamp, with a 1-minute interval between timestamps, forming optimized gas concentration data. For the third data segment after conversion, the data is sorted in ascending order of timestamp to ensure that each time point corresponds to a unique degradation feature data, forming optimized degradation feature data. For the converted current feature vector, the order of each feature value is rearranged according to the feature type to ensure that it matches the input dimension when the subsequent convolutional layer extracts local features, forming optimized current feature vector.

[0055] The embodiments of this application solve the problem of asynchronous acquisition time of multi-source data; by normalizing and converting to the normal operating reference range, the influence of the difference in parameter value range under different operating conditions is eliminated, the limitations of inconsistent scale of multi-source data and difficulty in direct fusion in the prior art are solved, the adaptability of the model to complex operating conditions is improved, and the arc prediction error caused by data inconsistency is reduced.

[0056] Step 105: Using the convolutional layer of the pre-built deep learning model, perform sliding window processing on the optimized current feature vector to generate a feature array. Using the long short-term memory network of the pre-built deep learning model, perform correlation processing on the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data to generate a time-series correlated feature sequence.

[0057] In this step, the feature array refers to the array formed by extracting local feature values ​​within each window and arranging them in window order through sliding window processing of the optimized current feature vector using convolutional layers. The temporal correlation feature sequence refers to the feature value sequence that reflects the correlation between each time segment by performing correlation processing on the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data through a long short-term memory network.

[0058] This application provides a specific embodiment, such as Figure 2 Step 105: Using the convolutional layer of the pre-built deep learning model, the optimized current feature vector is processed by sliding window to generate a feature array. Using the long short-term memory network of the pre-built deep learning model, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data are correlated to generate a time-series correlated feature sequence. Specifically, this includes the following steps: Step 501: Set the sliding window and the window movement interval, and extract sub-vector segments from the optimized current feature vector according to the sliding window and the window movement interval.

[0059] In this step, the sliding window refers to a fixed-length data window used to extract the optimized current feature vector, the size of which is set based on the fluctuation period of the current signal and the requirements of local feature analysis. The window movement interval refers to the number of overlapping data points between adjacent sliding windows, set based on the principle of balancing feature coverage and computational efficiency. A sub-vector segment refers to a continuous data segment extracted from the optimized current feature vector through the sliding window, and each sub-vector segment includes data points the size of the sliding window.

[0060] In one embodiment of this application, a specific implementation method may be: based on the length of the optimized current feature vector and the requirements of local feature analysis, set the sliding window size to 10 and the window movement interval to 5, and then starting from the starting position of the optimized current feature vector, sequentially extract data segments of length 10, moving 5 points to the right each time to form 23 overlapping sub-vector segments, ensuring that the entire optimized current feature vector is covered.

[0061] Step 502: Through the convolutional layer of the pre-built deep learning model, feature extraction is performed on each sub-vector segment to obtain the local feature value of each sub-vector segment. The local feature value includes the numerical fluctuation range, numerical change frequency and the number of times the numerical peak occurs within the corresponding sub-vector segment.

[0062] In this step, local feature values ​​refer to the quantitative indicators that reflect the local characteristics of the current signal extracted from the sub-vector segment through the convolutional layer. These local feature values ​​include the range of numerical fluctuations, the frequency of numerical changes, and the number of times the numerical peaks occur.

[0063] In this embodiment, for each sub-vector segment, three different one-dimensional convolution kernels are used to perform cross-correlation operations. The fluctuation detection kernel calculates the difference between the maximum and minimum values ​​within the sub-vector segment to obtain the numerical fluctuation range. The frequency detection kernel counts the number of times the difference between adjacent data points within the sub-vector segment exceeds a preset threshold to obtain the numerical change frequency. The peak detection kernel identifies the number of data points within the sub-vector segment that exceed the peak threshold to obtain the number of times the numerical peak occurs. Finally, the outputs of the three convolution kernels are concatenated to form the local feature value of each sub-vector segment.

[0064] Step 503: Arrange the local feature values ​​of all sub-vector segments to form a feature array. The row dimension of the feature array corresponds to the window truncation order, and the column dimension corresponds to the local feature values.

[0065] In this step, the window capture order refers to the sequence number of all sub-vector segments arranged according to the time order in which the sliding window moves.

[0066] In this embodiment, the local feature values ​​of the 23 sub-vector segments are arranged sequentially according to the window truncation order to form a 23-row, 3-column feature array. The row dimension of this array corresponds to the window truncation order, and the column dimension corresponds to the local feature value type, which facilitates subsequent fusion with other data.

[0067] Step 504: Divide the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation characteristic data into multiple time-series segments, each time-series segment including a preset number of data points.

[0068] In this step, a time-series segment refers to a fixed-length data block into which the optimized oil insulation parameters, gas concentration data, and degradation characteristic data are divided in chronological order. Each time-series segment includes a preset number of data points. The associated feature value refers to the feature vector generated after performing time-series modeling on the time-series segment using a long short-term memory network, reflecting the trend, periodicity, and abnormal fluctuations of the data in the time dimension.

[0069] In this embodiment of the application, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation characteristic data are divided into multiple time-series segments in chronological order. Each time-series segment includes a preset number of data points. For example, the optimized oil insulation parameters, which include 120 data points, are divided into 6 time-series segments, and each time-series segment includes 20 consecutive data points.

[0070] Step 505: Generate the associated feature value corresponding to each temporal segment through the long short-term memory network of the pre-built deep learning model, and combine the associated feature values ​​of all temporal segments to form a temporal associated feature sequence.

[0071] Optionally, step 505 involves generating the associated feature values ​​corresponding to each temporal segment using the pre-built long short-term memory network of the deep learning model, and combining the associated feature values ​​of all temporal segments to form a temporal associated feature sequence, specifically including the following steps: Step 511: Based on the time interval order corresponding to each time segment, number all time segments to form multiple numbered time segments.

[0072] In this step, the time interval order refers to the sequential order of the acquisition times corresponding to the time segments on the tap changer switching operation timeline. For example, the first time segment corresponds to the data from 3 hours to 2 hours and 40 minutes before the switch, the second corresponds to the data from 2 hours and 39 minutes to 2 hours and 20 minutes before the switch, and so on, ensuring that the numbering matches the time order. A numbered time segment refers to an ordered set formed by numbering all time segments according to the time interval order.

[0073] In this embodiment, if the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation characteristic data are divided into 6 time-series segments, each corresponding to a 10-minute acquisition time interval, then according to the order of the tap changer switching operation time axis, the earliest acquisition time segment is numbered S1, the next adjacent time segment is numbered S2, and so on, until the latest acquisition time segment is numbered S6, ultimately forming 6 numbered time-series segments from S1 to S6, ensuring that the numbering order is completely consistent with the acquisition time interval order.

[0074] Step 512: Based on the numerical distribution of the first time sequence segment with the first number, use the long short-term memory network of the pre-built deep learning model to generate the associated feature value corresponding to the first time sequence segment.

[0075] In this step, the first time segment refers to the time segment numbered S1, which corresponds to the earliest time interval. Its numerical distribution includes the statistical characteristics of oil insulation parameters, gas concentration, or degradation characteristics data within that time period.

[0076] In this embodiment, the numerical distribution information of the first time-series segment S1 is extracted, including the mean, variance, and extreme values ​​of the oil insulation parameters during this time period, the fluctuation range of gas concentration data, and statistical characteristics such as the statistical peak value of degradation feature data. This numerical distribution information is organized into an input vector according to a preset dimension and input into the long short-term memory network of a pre-constructed deep learning model. The long short-term memory network filters irrelevant and redundant information through a forget gate, selects effective features through an input gate, and integrates key information through an output gate to perform time-series feature modeling on the input vector, and finally generates the associated feature value corresponding to S1. For example, the output feature vector with a dimension of 2 [0.85, -0.28] quantifies and reflects the core time-series features of the data in the S1 time period.

[0077] Step 513: Based on the numerical distribution of the first time sequence segment with the first number, calculate the similarity between the numerical difference magnitude and numerical change trend of each subsequent time sequence segment with a number greater than the first time sequence segment and the preceding input time sequence segment, and convert the similarity of the numerical difference magnitude and numerical change trend into the associated feature value of the corresponding subsequent time sequence segment.

[0078] In this step, the subsequent time series segment refers to all time series segments numbered greater than S1, and the generation of its associated feature value depends on the numerical distribution of the preceding input time series segments. The numerical difference magnitude refers to the absolute difference between the subsequent time series segment and the preceding input time series segments in terms of statistical measures such as mean and variance. The preceding input time series segment refers to a set of time series segments with numbers less than that subsequent time series segment that have been pre-input into the Long Short-Term Memory (LSTM) network, based on the numbering order of the time series segments and the input order of the LSM network. For example, when processing the subsequent time series segment numbered S3, the preceding input time series segments are time series segments numbered S1 and S2, including the numerical distribution of S1 and S2. By calculating the numerical difference magnitude and the similarity of numerical change trends between S3 and S1 and S2, historical reference is provided for the conversion of the associated feature value of S3, ensuring that the LSM network can capture long-term dependencies in the time series data. The similarity of numerical change trends refers to the degree of matching between the subsequent time series segment and the preceding input time series segment in terms of trend characteristics such as data fluctuation direction and slope.

[0079] In this embodiment, taking the subsequent time series segment S2 as an example, its preceding input time series segment is S1. First, the absolute difference in the mean and the absolute difference in variance between S2 and S1 are calculated, and the numerical difference magnitude is obtained by integrating them. Then, the slope correlation coefficient of the data fluctuation curves of the two is calculated to quantify the similarity of the numerical change trend. The above numerical difference magnitude and trend similarity are converted into a standardized input vector and input into a long short-term memory network to generate the associated feature value corresponding to S2. For S3, its preceding input time series segments are S1 and S2. The average numerical difference magnitude and comprehensive trend similarity between S3 and S1 and S2 are calculated, converted into an input vector, and processed by a long short-term memory network to obtain the associated feature value of S3. The associated feature values ​​of S4 to S6 are calculated sequentially according to the same logic to ensure that the feature value of each subsequent time series segment is integrated with the numerical distribution association information of all preceding time series segments.

[0080] Step 514: Arrange all associated feature values ​​according to the numbering order to form a time-series associated feature sequence that includes time dimension correlation.

[0081] In this embodiment, the associated feature values ​​of all time-series segments are arranged sequentially according to their numbers to form a time-series associated feature sequence. This sequence fully preserves the dependencies and evolution patterns of each time-series segment in the time dimension, providing a time-series association basis for subsequent feature fusion.

[0082] This application's embodiments, through the combination of sliding windows and convolutional layers, decompose the current feature vector into overlapping sub-vector segments, enabling the capture of local abnormal fluctuations that are easily overlooked by traditional methods. By using a long short-term memory network to model the temporal sequence of the time-series segments, it can uncover the potential correlations between data such as oil parameters and gas concentrations in the time dimension, overcoming the limitations of traditional methods that rely solely on static features. This improves the ability to identify the trend of transformer insulation degradation, reduces false alarms and missed alarms in arc warnings caused by data bias, and provides more reliable technical support for the safe operation of the power system.

[0083] Step 106: Perform dimensional matching and fusion processing on the feature array and the time-series associated feature sequence to generate an arc occurrence probability prediction result.

[0084] In this step, the arc occurrence probability prediction result refers to the sequence of arc occurrence probability values ​​corresponding to each time node arranged in chronological order.

[0085] In this embodiment of the application, step 106 specifically includes the following steps: Step 601: Determine the number of first time nodes in the row dimension of the feature array and the number of second time nodes in the time dimension of the time-series associated feature sequence, so as to adjust the feature array and the time-series associated feature sequence to obtain a feature array and a time-series associated feature sequence with the same dimension length. If the number of first time nodes is less than the number of second time nodes, virtual feature values ​​are added to the end of the feature array; if the number of first time nodes is greater than the number of second time nodes, the first segment of the feature array with the same number of second time nodes is extracted.

[0086] In this step, the first time node count refers to the total number of time nodes included in the row dimension of the feature array, obtained by counting the number of rows in the row dimension of the feature array. The second time node count refers to the total number of time nodes included in the time dimension of the time-series associated feature sequence, obtained by counting the number of associated feature values ​​in the time-series associated feature sequence. Virtual feature values ​​refer to feature values ​​added at the end of the feature array to match the second time node count, calculated based on the changing trend of local feature values ​​at the end of the feature array. The changing trend of local feature values ​​refers to the direction and magnitude of the change of local feature values ​​at the end of the feature array over time, calculated based on the changing slope of multiple local feature values ​​at the end. The preceding feature segment refers to the feature segment with the same number of rows as the second time node count, extracted from the beginning of the feature array when the first time node count is greater than the second time node count.

[0087] In this embodiment, the number of row dimensions of the feature array is first counted, that is, the number of rows included in the array. Each row corresponds to a time node, and this number is the first time node number. Then, the number of time dimensions of the time-series associated feature sequence is counted, that is, the number of feature values ​​included in the sequence. Each feature value corresponds to a time node, and this number is the second time node number. If the number of first time nodes is less than the number of second time nodes, the last three local feature values ​​at the end of the feature array are extracted first, and the slope of change of these three values ​​is calculated to determine the trend of change of local feature values. Then, virtual feature values ​​are added to the end of the array according to the slope of change. For example, one way to add virtual feature values ​​is: if the last value at the end is 20 and the slope of change is 1, then the first virtual feature value added is 21, the second is 22, and so on, until the number of first time nodes is the same as the number of second time nodes. If the number of first time nodes is greater than the number of second time nodes, starting from the beginning of the feature array, a feature segment with the same number of rows as the number of second time nodes is extracted. This segment is the first feature segment, and finally, a feature array and a time-series associated feature sequence with the same dimension length are obtained.

[0088] Step 602: Arrange the local feature values ​​of the feature array with the same dimension length at the same time node and the associated feature values ​​of the time-series associated feature sequence to obtain a comprehensive feature set.

[0089] In this step, the integrated feature set refers to the joint feature set obtained by integrating the feature array with consistent dimensions and the time-series associated feature sequence according to the same time node.

[0090] In this embodiment, a time node mapping relationship is established so that each row dimension time node of the feature array after dimension consistency corresponds one-to-one with each time dimension time node of the time-series associated feature sequence. For example, the first row of the feature array corresponds to the first associated feature value of the time-series associated feature sequence, which corresponds to the time node 10 minutes before the tap changer switching. Then, for each time node, the local current feature value and the time-series associated feature value corresponding to the node are arranged in the order of [numerical fluctuation range, numerical change frequency, number of occurrences of numerical peak, numerical difference amplitude, and similarity of numerical change trend] to form the feature combination of the time node. Finally, the feature combinations of all time nodes are integrated in the order of time nodes to obtain a comprehensive feature set including the joint features of all nodes.

[0091] Step 603: Use the output layer of the deep learning model to perform feature assignment processing on the comprehensive feature set to calculate the arc occurrence probability value corresponding to each time node, and arrange all arc occurrence probability values ​​according to the time node order to form the arc occurrence probability prediction result.

[0092] In this step, the arc occurrence probability value refers to the quantitative value of the possibility of an arc occurring at each time point. The value ranges from 0 to 1. The closer the value is to 1, the higher the possibility of an arc occurring at that time point.

[0093] In this embodiment of the application, step 603 specifically includes the following steps: Step 611: Based on the degree of correlation between various feature values ​​and the occurrence of electric arc, the feature association rules of the output layer of the deep learning model are used to assign corresponding influence coefficients to various feature values, so as to calculate the corresponding comprehensive calculation value based on various feature values ​​and corresponding influence coefficients. The various feature values ​​include local feature values ​​and associated feature values ​​in the comprehensive feature set.

[0094] In this step, the degree of correlation with arc occurrence refers to the magnitude and contribution of various feature values ​​in the comprehensive feature set to the occurrence of the arc. Feature association rules refer to the rule system built into the output layer of the deep learning model used to assign influence coefficients. These rules are formulated based on the degree of correlation between various feature values ​​and arc occurrence, including the core logic that the higher the degree of correlation, the larger the influence coefficient. Influence coefficients refer to the weight values ​​assigned to various feature values ​​in the comprehensive feature set, ranging from 0 to 1, with the sum of all coefficients being 1. The comprehensive calculated value refers to the value obtained by weighting and summing the various feature values ​​at each time point according to their corresponding influence coefficients.

[0095] In this embodiment, by comparing and statistically analyzing historical arc fault data and normal operation data of transformers, the correlation between various feature values ​​in the comprehensive feature set and the occurrence of arcs is determined. Among them, the frequency of numerical peaks will show significant anomalies before the occurrence of arcs, indicating the highest correlation; the difference in degradation feature data can directly reflect the inducing effect of insulation degradation on arcs, indicating the second highest correlation; the difference in numerical fluctuation range and gas concentration data has a moderate correlation; the similarity of numerical change frequency, oil insulation parameter difference, and change trend has a low correlation. The output layer of the deep learning model has built-in feature association rules based on this correlation, and the output layer calls the rules to assign corresponding influence coefficients to various feature values. Then, for the feature combination at each time node in the comprehensive feature set, each feature value is multiplied by the corresponding influence coefficient to obtain the weighted value of each feature value. Finally, all weighted values ​​at the same time node are summed to obtain the comprehensive calculated value corresponding to that time node.

[0096] Step 612: Convert each comprehensive calculation value into the arc occurrence probability value at the corresponding time node using the preset numerical conversion rules.

[0097] In this step, the preset numerical conversion rule refers to the pre-defined standardized rule that maps the comprehensive calculated value to the 0-1 probability range.

[0098] In this embodiment, the range of values ​​for the comprehensive calculation at all time points is statistically analyzed, and the maximum and minimum values ​​are determined. A preset numerical conversion rule is then invoked. For example, a preset numerical conversion rule may be: the probability value of arc occurrence is obtained by multiplying the difference between the current comprehensive calculation value and the minimum value of all comprehensive calculation values, the quotient of the probability interval span and the range of comprehensive calculation values, and then summing it with the lower limit of the probability interval. This rule is used to uniformly map comprehensive calculation values ​​of different sizes to the probability interval, realizing the standardized conversion from numerical value to probability. Then, the comprehensive calculation value of each time point is substituted into this rule for calculation, and the result obtained is the arc occurrence probability value corresponding to that time point.

[0099] For example, one feasible implementation method of this application is as follows: A HY-8000 oil insulation characteristic comprehensive tester is used to collect oil insulation parameters. This instrument has a detection accuracy of ±0.01, supports online continuous monitoring, and has a data sampling frequency of 1 time / minute. A GC-9860 gas chromatograph paired with a hydrogen flame ionization detector is used to collect dissolved gas concentration data in the oil. Its detection limit is as low as 0.1 μL / L, and the sampling cycle is 5 minutes / time. A CTB-1000 high-frequency current sensor is used to collect transient current waveforms during tap changer switching. This sensor has a bandwidth range of 1 kHz to 1 MHz and a sampling rate of 100,000 sampling points / second, which can accurately capture instantaneous current fluctuation signals. Simultaneously, a DAQ-6009 data acquisition module is used to integrate the output data of the three types of equipment, an SP-2000 signal preprocessing module performs filtering and amplification processing, and an ECU-5000 edge computing unit carries a pre-built deep learning model. The oil insulation parameters, dissolved gas concentration data, and transient current waveforms collected by the above-mentioned equipment are summarized by the data acquisition module and optimized by the signal preprocessing module. They are then used to identify early insulation degradation characteristics, extract current feature vectors, and input into the deep learning model of the edge computing unit to finally generate arc occurrence probability prediction results.

[0100] Figure 3 This is a schematic diagram of a specific embodiment of a transformer solid-state on-load tap changer arc prediction system provided in this application, referring to... Figure 3 The system may include: The acquisition module 21 is used to acquire oil insulation parameters and gas concentration data of dissolved gases in the oil; The identification module 22 is used to identify early insulation degradation characteristic data based on the gas concentration data; Extraction module 23 is used to collect the transient current waveform when the tap changer starts switching operation, obtain current waveform segments, and extract features from the current waveform segments to form a current feature vector; Processing module 24 is used to perform spatiotemporal alignment and normalization processing on the oil insulation parameters, the gas concentration data, the early insulation degradation characteristic data and the current characteristic vector to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data and optimized current characteristic vector. The association module 25 is used to perform sliding window processing on the optimized current feature vector using the convolutional layer of the pre-built deep learning model to generate a feature array, and to perform association processing on the optimized oil insulation parameters, optimized gas concentration data and optimized degradation feature data using the long short-term memory network of the pre-built deep learning model to generate a time-series association feature sequence. The generation module 26 is used to perform dimensional matching and fusion processing on the feature array and the time-series associated feature sequence to generate an arc occurrence probability prediction result.

[0101] An arc prediction system for solid-state on-load tap changer of a transformer according to an embodiment of this application is used to implement the aforementioned arc prediction method for solid-state on-load tap changer of a transformer. Therefore, the specific implementation of the arc prediction system for solid-state on-load tap changer of a transformer can be found in the embodiment section of the aforementioned arc prediction method for solid-state on-load tap changer of a transformer. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0102] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described transformer solid-state on-load tap changer arc prediction method.

[0103] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for predicting arcing in a transformer solid-state on-load tap changer.

[0104] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0105] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the transformer solid-state on-load tap changer arc prediction method.

[0106] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0107] The above provides a detailed description of the arc prediction method and system for solid-state on-load tap changer in transformers provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for predicting arcing in a transformer with solid-state on-load tap changer, characterized in that, include: Collect data on oil insulation parameters and dissolved gas concentrations in the oil; Based on the gas concentration data, early insulation degradation characteristics were identified; The transient current waveform during the tap changer start-up switching operation is collected to obtain current waveform segments, and features are extracted from the current waveform segments to form current feature vectors. The oil insulation parameters, gas concentration data, early insulation degradation characteristic data, and current feature vector are spatiotemporally aligned and normalized to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data, and optimized current feature vector. Using the convolutional layer of a pre-built deep learning model, the optimized current feature vector is processed by sliding window to generate a feature array. Using the long short-term memory network of the pre-built deep learning model, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data are correlated to generate a time-series correlated feature sequence. The feature array and the time-series associated feature sequence are subjected to dimensional matching and fusion processing to generate an arc occurrence probability prediction result. Based on the gas concentration data, early insulation degradation characteristics are identified, including: The concentration data of the target gas components related to insulation degradation in the gas concentration data are integrated to obtain the concentration time series of different target gas components, and the time nodes of the concentration value sudden increase in each concentration time series are marked. The concentration value sudden increase is when the concentration difference between adjacent collection times exceeds the preset increment. Calculate the concentration change rate at each time point, and integrate the concentration change rates at each time point into a concentration change rate sequence according to the chronological order of the time points; The intensity of the spectral characteristic peak of the target gas component at each time point is obtained, and the intensity of the spectral characteristic peak is correlated with the concentration change rate sequence to form a joint feature group, which includes the concentration change amplitude and spectral characteristics. Target abnormal segments with synchronously abnormal concentration change rate and spectral characteristic peak intensity are extracted from the joint feature group. The concentration change rate, spectral characteristic peak intensity and time node corresponding to the target abnormal segment are integrated to form early insulation degradation characteristic data.

2. The method according to claim 1, characterized in that, Using the convolutional layers of a pre-built deep learning model, a sliding window process is applied to the optimized current feature vector to generate a feature array. Then, using the long short-term memory network of the pre-built deep learning model, the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data are correlated to generate a time-series correlated feature sequence, including: Set a sliding window and a window movement interval, and extract a sub-vector segment from the optimized current feature vector based on the sliding window and the window movement interval; By using the convolutional layers of a pre-built deep learning model, features are extracted from each sub-vector segment to obtain local feature values ​​for each sub-vector segment. These local feature values ​​include the range of numerical fluctuations, the frequency of numerical changes, and the number of times numerical peaks occur within the corresponding sub-vector segment. Arrange the local feature values ​​of all sub-vector segments to form a feature array. The row dimension of the feature array corresponds to the window truncation order, and the column dimension corresponds to the local feature values. The optimized oil insulation parameters, optimized gas concentration data, and optimized degradation characteristic data are divided into multiple time-series segments, and each time-series segment includes a preset number of data points. By using the long short-term memory network of a pre-built deep learning model, the associated feature values ​​corresponding to each time segment are generated, and the associated feature values ​​of all time segments are combined to form a time-series associated feature sequence.

3. The method according to claim 2, characterized in that, Using a pre-built long short-term memory network of a deep learning model, associated feature values ​​are generated for each temporal segment. These associated feature values ​​are then combined to form a temporal associated feature sequence, including: Based on the time interval order corresponding to each time segment, all time segments are numbered to form multiple numbered time segments; Based on the numerical distribution of the first time sequence segment with the first number, the associated feature value corresponding to the first time sequence segment is generated using the long short-term memory network of the pre-built deep learning model. Based on the numerical distribution of the first time sequence segment with the first number, calculate the similarity between the numerical difference magnitude and numerical change trend of each subsequent time sequence segment with a number greater than the first time sequence segment and the previously input time sequence segment, and convert the similarity of the numerical difference magnitude and numerical change trend into the corresponding associated feature value of the subsequent time sequence segment. Based on the numbering order, all associated feature values ​​are arranged to form a time-series associated feature sequence that includes time-dimensional correlation.

4. The method according to claim 1, characterized in that, The oil insulation parameters, gas concentration data, early insulation degradation characteristic data, and current feature vector are spatiotemporally aligned and normalized to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data, and optimized current feature vector, including: Using the tap changer switching operation time as a reference point, the first time interval of the oil insulation parameter acquisition time relative to the reference point, the second time interval of the gas concentration data relative to the reference point, and the third time interval of the early insulation degradation characteristic data relative to the reference point are determined. Based on the first time interval, the second time interval, and the third time interval, extract the first data segment of oil insulation parameters within a preset time span, the second data segment of gas concentration data within a preset time span, and the third data segment of early insulation degradation characteristic data within a preset time span. The sub-segments whose data length is less than a preset length in the first data segment, the second data segment, and the third data segment are subjected to parameter supplementation processing to obtain the supplemented first data segment, the supplemented second data segment, and the supplemented third data segment. The first reference range for oil insulation parameters, the second reference range for gas concentration data, the third reference range for early insulation degradation characteristic data, and the fourth reference range for current characteristic vector are set under normal operating conditions of the transformer. The supplemented first data segment, the supplemented second data segment, the supplemented third data segment, and the current feature vector are respectively converted into the first reference interval, the second reference interval, the third reference interval, and the fourth reference interval to obtain the converted first data segment, the converted second data segment, the converted third data segment, and the converted current feature vector. All the converted first data segment, converted second data segment, converted third data segment, and converted current feature vector are rearranged to form optimized oil insulation parameters, optimized gas concentration data, optimized degradation feature data, and optimized current feature vector.

5. The method according to claim 1, characterized in that, The feature array and the time-series associated feature sequence are subjected to dimensional matching and fusion processing to generate an arc occurrence probability prediction result, including: The number of first time nodes in the row dimension of the feature array and the number of second time nodes in the time dimension of the time-series associated feature sequence are determined to adjust the feature array and the time-series associated feature sequence to obtain a feature array and a time-series associated feature sequence with the same dimension length. If the number of first time nodes is less than the number of second time nodes, virtual feature values ​​are added to the end of the feature array; if the number of first time nodes is greater than the number of second time nodes, the first segment of the feature array with the same number of second time nodes is extracted. Arrange the local feature values ​​of the feature array with the same dimension length at the same time point and the associated feature values ​​of the time-series associated feature sequence to obtain a comprehensive feature set; The output layer of the deep learning model is used to perform feature assignment processing on the comprehensive feature set to calculate the arc occurrence probability value corresponding to each time node. All arc occurrence probability values ​​are arranged according to the time node order to form the arc occurrence probability prediction result.

6. The method according to claim 5, characterized in that, The output layer of the deep learning model is used to perform feature assignment processing on the comprehensive feature set to calculate the arc occurrence probability value corresponding to each time node, including: Based on the degree of correlation between various feature values ​​and the occurrence of electric arc, the feature association rules of the output layer of the deep learning model are used to assign corresponding influence coefficients to various feature values, so as to calculate the corresponding comprehensive calculation value based on various feature values ​​and corresponding influence coefficients. The various feature values ​​include local feature values ​​and related feature values ​​in the comprehensive feature set. By using preset numerical conversion rules, each comprehensive calculation value is converted into the probability value of arc occurrence at the corresponding time point.

7. An arc prediction system for solid-state on-load tap changer in a transformer, used to execute the arc prediction method for solid-state on-load tap changer in a transformer as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to collect data on oil insulation parameters and dissolved gas concentrations in the oil. The identification module is used to identify early insulation degradation characteristic data based on the gas concentration data; The extraction module is used to collect the transient current waveform when the tap changer starts switching, obtain current waveform segments, and extract features from the current waveform segments to form a current feature vector. The processing module is used to perform spatiotemporal alignment and normalization processing on the oil insulation parameters, the gas concentration data, the early insulation degradation characteristic data and the current feature vector to obtain optimized oil insulation parameters, optimized gas concentration data, optimized degradation characteristic data and optimized current feature vector. The correlation module is used to perform sliding window processing on the optimized current feature vector using the convolutional layer of the pre-built deep learning model to generate a feature array. It also uses the long short-term memory network of the pre-built deep learning model to perform correlation processing on the optimized oil insulation parameters, optimized gas concentration data, and optimized degradation feature data to generate a time-series correlation feature sequence. The generation module is used to perform dimensional matching and fusion processing on the feature array and the time-series associated feature sequence to generate an arc occurrence probability prediction result.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the arc prediction method for a transformer solid-state on-load tap changer as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of an arc prediction method for solid-state on-load tap changer of a transformer as described in any one of claims 1 to 6.

Images