10kv transformer operation state monitoring method based on time series analysis

Abstract

The application discloses a 10kV transformer operation state monitoring method based on time series analysis and relates to the technical field of transformer state monitoring.The method comprises the following steps: time series decomposition is performed on historical operation data of a 10kV transformer, trend components and periodic components are obtained, and an operation state baseline model is constructed; operation parameter sequences are collected in real time and mapped to a state space to generate a state deviation sequence; the state deviation sequence is scanned by using a sliding window, abnormal pattern matching results are obtained after deviation fluctuation characteristics are extracted and input into an abnormal pattern recognition network; an abnormal starting time is located and an abnormal precursor sequence is generated by tracing back; the abnormal precursor sequence and the state deviation sequence are time series aligned to generate a joint feature tensor, which is input into a fault evolution prediction model to obtain a fault evolution trajectory; and the remaining effective operation time is calculated according to the fault evolution trajectory, and a state monitoring alarm level is generated.

Description

Technical Field

[0001] This invention relates to the field of transformer condition monitoring technology, specifically a method for monitoring the operating status of a 10kV transformer based on time-series analysis. Background Technology

[0002] As a critical piece of equipment at the end of the distribution network, the operational reliability of 10kV distribution transformers directly impacts the power supply quality for end users. Existing 10kV transformer operational status monitoring technologies typically rely on preset single fixed thresholds or simple deviation statistics for anomaly detection. When real-time collected operating parameters exceed the set threshold, an alarm is triggered. This passive monitoring approach can only respond after the anomaly has developed to a certain extent. The drawback of existing technologies is that the alarm mechanism based on fixed thresholds is essentially a delayed detection, unable to capture weak signals indicating deterioration in the early stages of fault formation. The transformation of a transformer from normal operation to failure is a continuous degradation process. The small amplitude and varied patterns of fluctuations in operating parameters in the early stages of degradation are easily masked by noise, and traditional statistical extraction methods lack sensitivity to this time-series information. Furthermore, after issuing alarms for detected anomalies, existing solutions lack the ability to quantitatively predict the subsequent development of the anomaly. Maintenance personnel cannot know the time window between the current abnormal state and equipment failure, making it difficult to formulate targeted maintenance decisions.

[0003] The problems that need to be solved in order to achieve advanced fault prediction are how to proactively backtrack and extract early signs that were not detected before anomalies occurred from real-time transformer operation data, and how to perform in-depth time-series correlation modeling of these early signs with the deviation changes after the anomalies occurred. Summary of the Invention

[0004] This invention provides a method for monitoring the operating status of a 10kV transformer based on time-series analysis. It can actively extract the anomaly precursor sequence from real-time operating parameters and construct a joint feature tensor of anomaly precursors and deviation evolution, achieving complete coverage from anomaly detection to fault evolution trajectory prediction, and providing a quantitative basis for the remaining effective operating time for operation and maintenance decisions.

[0005] To achieve the above objectives, this invention provides the following technical solution: This invention provides a method for monitoring the operating status of a 10kV transformer based on time-series analysis. By performing time-series decomposition on the historical operating data of the 10kV transformer, trend components and periodic components at multiple time scales are obtained, and a baseline model of the operating status is constructed accordingly. Compared with single-scale feature extraction methods, multi-time-scale decomposition can more completely separate the long-term slowly changing trend components and the periodic components caused by factors such as periodic load fluctuations in the transformer operating data, making the constructed baseline model more reflective of the inherent operating laws of the transformer, thereby improving the accuracy of subsequent deviation analysis.

[0006] The operating parameter sequence of a 10kV transformer is acquired in real time and mapped to the state space defined by the baseline model of the operating state to generate a state deviation sequence. By comparing the real-time operating parameters with the normal operating state depicted by the baseline model point by point, the inherent parameter fluctuations under normal operating conditions can be effectively filtered out, making the deviations truly caused by abnormal states stand out, thus laying the foundation for reliable identification of abnormal modes.

[0007] A sliding window scan is performed on the state deviation sequence to extract deviation fluctuation features within the window. These features are then input into a pre-trained anomaly pattern recognition network to obtain anomaly pattern matching results. The sliding window scan captures local statistical characteristics segment by segment on the deviation sequence, generating multi-dimensional fluctuation features such as the window mean, variance, maximum deviation value, and deviation change rate, which can describe the dynamic change pattern of the deviation from different perspectives. The pre-trained anomaly pattern recognition network comprehensively judges these multi-dimensional fluctuation features, automatically identifying the corresponding anomaly type label and outputting the confidence score as the matching result, achieving rapid classification and accurate localization of various potential anomaly patterns.

[0008] As a technical solution of this invention, the anomaly initiation time is located based on the anomaly pattern matching result, and the operating parameter sequence is traced back for a preset time based on this anomaly initiation time to generate an anomaly precursor sequence. The anomaly initiation time is determined by using a target window where the anomaly confidence first exceeds a preset threshold, which can accurately capture the instant the anomaly state begins to appear. Based on this, tracing back the operating parameters for a certain period and performing time-series decomposition consistent with historical data allows for the extraction of parameter trend changes and periodic fluctuations before the anomaly fully manifests, forming an anomaly precursor sequence describing the anomaly gestation process and providing early warning information for fault evolution prediction.

[0009] By temporally aligning the anomaly precursor sequence and the state deviation sequence, a joint feature tensor is generated and input into the fault evolution prediction model to obtain the predicted fault evolution trajectory. The joint feature tensor, constructed by temporally aligning and fusing the anomaly precursor sequence (reflecting early-stage anomaly changes) with the subsequent deviation sub-sequence (reflecting post-anomaly deviation development), simultaneously contains both fault precursor and fault development information. Based on this joint feature tensor, the fault evolution prediction model predicts deviation values ​​at multiple future moments, forming a fault evolution trajectory showing the gradual expansion of deviations from the present to a future period, thus achieving proactive judgment of the fault development trend.

[0010] The remaining effective runtime is calculated based on the fault evolution trajectory, and a status monitoring alarm level is generated based on this remaining effective runtime. A sequence of predicted deviation values ​​is extracted from the predicted evolution trajectory and compared item by item with a preset failure deviation threshold to dynamically determine the specific moment when the deviation reaches the failure level. The difference between this failure arrival time and the current time is the transformer's remaining effective runtime, intuitively quantifying the equipment's safe operating time window. Matching the remaining effective runtime with multiple alarm duration intervals to determine the alarm level provides maintenance personnel with hierarchical status monitoring alarms, helping them take appropriate measures within the appropriate time window.

[0011] Preferably, when performing time-series decomposition on historical operating data, a preset sliding time window is used to segment the historical operating data to obtain multiple historical data segments. Empirical mode decomposition (EMD) is performed on each historical data segment to obtain multiple intrinsic mode functions (IMFs). The zero-crossing frequency of each IMF is calculated, and the data is divided into a low-frequency trend group and a high-frequency periodic group based on this frequency. The trend component and periodic component are then summed to obtain the latter. Using a sliding window combined with EMD, nonlinear and non-stationary transformer operating data can be adaptively decomposed into trend and periodic components with clear physical meanings. This avoids the insufficient adaptability caused by relying on fixed basis functions, making the baseline reconstruction sequence more closely match actual operating characteristics. After reconstructing the trend and periodic components, a long short-term memory (LSTM) network is trained as the operating state baseline model. This leverages the strong modeling capability of the LSM network for dynamic time-series behavior to accurately learn the evolution law of normal transformer operating conditions, improving the accuracy and robustness of baseline prediction.

[0012] As a preferred embodiment of the present invention, the mirror extension method is used to perform boundary processing on both ends of historical data segments during the empirical mode decomposition process, in order to suppress the endpoint effect during the decomposition process. This method can effectively prevent data distortion at the endpoints from propagating inward, ensuring the accuracy of the intrinsic mode functions obtained from the decomposition across the entire range of the data segment, thereby guaranteeing the fidelity of trend components and periodic components.

[0013] Preferably, when generating the state deviation sequence, the operating parameters at each acquisition time in the operating parameter sequence are input into the baseline prediction value output by the operating state baseline model. The difference between the operating parameters and the prediction value is calculated to obtain the original deviation value. The original deviation value is normalized and arranged in time sequence to obtain the state deviation sequence. By comparing the baseline model output with the real-time parameters point by point to obtain the deviation, and normalizing the deviation, the influence of differences in the dimensions and amplitudes of different operating parameters on subsequent analysis can be eliminated, so that the state deviation sequence reflects the change process of the purity of the anomaly relative to the normal baseline.

[0014] As a preferred embodiment of the present invention, the construction method of the anomaly pattern recognition network includes: collecting historical fault records of 10kV transformers; extracting a historical state deviation sequence within a preset time period before the fault occurs from each historical fault record; performing sliding window extraction on the sequence to obtain multiple historical deviation windows; labeling each window with a corresponding historical anomaly type label; and training an initial classification network based on a convolutional neural network using the deviation fluctuation features of the windows as input and the anomaly type label as output. By utilizing the deviation data before the fault occurs in the historical fault records to construct training samples, the network can learn the deviation fluctuation patterns exhibited by various faults during their incubation stage, thereby possessing the ability to identify early anomaly patterns during real-time monitoring.

[0015] As a preferred technical solution of the present invention, the construction method of the fault evolution prediction model includes: using multiple historical fault events from historical operational data as samples, extracting a pre-fault anomaly precursor subsequence and a post-fault deviation evolution subsequence from each historical fault event, concatenating the two in chronological order to form a sample joint sequence, and extracting multiple sample three-dimensional tensors. Using the sample three-dimensional tensors as input and the actual deviation change value with a preset step size as output, an initial prediction network based on gated recurrent units is trained. Training the gated recurrent unit network with joint data of fault precursors and fault development enables it to grasp the core laws of how anomaly precursor features gradually evolve into faults, thereby accurately predicting future fault evolution trajectories based on real-time joint features during the application stage.

[0016] Preferably, after generating the status monitoring alarm level, the current load rate of the 10kV transformer is also obtained. The remaining effective operating time is then adjusted based on the current load rate to obtain the adjusted effective operating time. This adjusted effective operating time is then used to re-match the alarm duration interval to obtain the updated alarm level. The higher of the two levels is then output as the final alarm level. By comprehensively considering the effect of load rate on accelerating or delaying fault development, and by adaptively adjusting the remaining effective operating time based on baseline deviation prediction according to actual load conditions, the final alarm level can better reflect the actual operating risk level of the transformer under the current operating conditions. Furthermore, a higher level output ensures the adequacy of the alarm and prevents delayed response due to underestimation of risk.

[0017] The technical effects and advantages provided by the present invention in the above technical solution are as follows: After locating the initial moment of the abnormal pattern, the system traces back a preset duration of operating parameter sequences from that moment. The traced parameter sequences are then subjected to the same temporal decomposition operation as historical operating data, yielding trend and periodic components, which are combined to form an anomaly precursor sequence. This scheme does not rely on alarm signals that only change significantly after an anomaly occurs. Instead, it traces back to the historical interval before the anomaly confidence level first exceeds the threshold, extracting early trend and periodic characteristics from this interval that have not yet triggered alarms. This precursor information contains the temporal patterns of subtle changes in operating parameters during the fault initiation stage, reflecting the gradual transition of the transformer's state from normal to abnormal. This allows the fault prediction model to extrapolate based on the actual state information before the anomaly truly forms, rather than simply extrapolating from deviation data after the anomaly occurs. The introduction of precursor information improves the lead time and accuracy of fault evolution prediction. The abnormal precursor sequence and the state deviation sequence are time-aligned according to the acquisition time. The value of each time step in the aligned abnormal precursor sequence is appended to the normalized deviation value of the same time step in the subsequent deviation subsequence, generating a multi-dimensional joint feature sequence. This sequence is then stacked into a three-dimensional tensor and used as input to the fault evolution prediction model. This approach achieves channel-level fusion of abnormal precursor information and deviation evolution information at the feature level, rather than simply splicing the two sequences end to end. The joint feature tensor carries both the deviation state at each time step and the corresponding precursor background. When extracting time dependencies, the fault evolution prediction model can simultaneously perceive the intrinsic relationship between the deviation change trend and its historical background before occurrence. Compared to using only the deviation sequence for unidirectional prediction, this joint modeling approach allows the model to learn the dynamic laws of transformer state degradation from a more complete causal chain, and the output predicted fault evolution trajectory is more consistent with the nonlinear temporal characteristics of the actual degradation process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of a 10kV transformer operation status monitoring method based on time series analysis; Figure 2 This is a flowchart of transformer condition monitoring and alarm process based on fault evolution prediction; Figure 3 This is a time series diagram of the trend and periodic components of the abnormal precursor sequence before a 10kV transformer fault. Figure 4 This is a comparison chart of subsequent deviation subsequences and fault evolution trajectories; Figure 5 It is a curve showing the prediction deviation of fault evolution and the remaining effective operating time of a 10kV transformer. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1 This invention provides a method for monitoring the operating status of a 10kV transformer based on time-series analysis. The method includes the following steps: First, time-series decomposition of historical operating data of the 10kV transformer is performed to obtain trend and periodic components at multiple time scales. Then, an operating status baseline model is constructed based on these trend and periodic components. Second, real-time acquisition of the operating parameter sequence of the 10kV transformer is performed, and the operating parameter sequence is mapped to the state space defined by the operating status baseline model to generate a state deviation sequence. Third, a sliding window scan is performed on the state deviation sequence to extract deviation fluctuation features within the window. These features are then input into a pre-trained anomaly pattern recognition network to obtain anomaly pattern matching results. Fourth, the anomaly initiation time is located based on the anomaly pattern matching results, and the operating parameter sequence is traced back for a preset time based on the anomaly initiation time to generate an anomaly precursor sequence. Fifth, the anomaly precursor sequence and the state deviation sequence are time-series aligned to generate a joint feature tensor. This joint feature tensor is then input into a fault evolution prediction model to obtain the predicted fault evolution trajectory. Sixth, the remaining effective operating time is calculated based on the fault evolution trajectory, and a status monitoring alarm level is generated based on the remaining effective operating time.

[0022] Example 1:

[0023] In practice, the operating data of a 10kV transformer within a preset historical time period is acquired to form a historical operating data sequence. This historical operating data includes at least one of the parameters such as voltage, current, and temperature, collected at fixed sampling intervals. The length of the sliding time window is set to L, and the sliding step size is S. Starting from the beginning of the historical operating data sequence, the sliding time window is moved sequentially according to the sliding step size S, cutting out multiple historical data segments. Each historical data segment contains L consecutive sampling points.

[0024] Empirical Mode Decomposition (EMD) is performed on each historical data segment. During EMD, boundary processing is first applied to both ends of the historical data segment. A mirror extension method is used: a segment of extended data mirror-symmetric to the data near the start of the historical data segment is copied symmetrically before the start of the segment, and another segment of extended data mirror-symmetric to the data near the end of the segment is copied symmetrically after the end of the segment. These extended data are appended to both ends of the historical data segment to form the extended sequence. EMD is then performed on the extended sequence. After decomposition, only the decomposition results corresponding to the original historical data segment positions are retained, and the extended portions at both ends are removed, thereby suppressing the influence of endpoint effects on the intrinsic mode functions.

[0025] The specific process of empirical mode decomposition is as follows: Find all local maxima and minima of the extended sequence. Fit the upper and lower envelopes using cubic spline interpolation, and calculate their mean curves. Subtract the mean curves from the extended sequence to obtain the intermediate signal. Determine if the intermediate signal satisfies the conditions for intrinsic mode functions (EMFs). These conditions include: the number of extrema and zero-crossings are equal or at most differ by one over the entire signal length; and the mean of the upper and lower envelopes is zero at any given time. If satisfied, the intermediate signal is an EMF. If not, repeat the steps of finding the envelope and subtracting the mean for the intermediate signal until the conditions are met, obtaining the first EMF. Subtract the first EMF from the extended sequence to obtain the residual signal. Repeat the above process for the residual signal to decompose multiple EMFs sequentially until the residual signal becomes a monotonic function or no further EMFs can be extracted. Perform the above operations on each historical data segment to obtain multiple EMFs.

[0026] Calculate the zero-crossing frequencies for each intrinsic mode function. The zero-crossing frequency is calculated by counting the number of times the signal value in the intrinsic mode function crosses a zero; this number is denoted as the zero-crossing count. Let the time length of the intrinsic mode function be... The zero-crossing frequency is calculated as follows:

[0027] in, This represents the total number of times the intrinsic mode function crosses a zero point during its time length, changing from a positive to a negative value or vice versa. This represents the actual time length of the historical data segment corresponding to the intrinsic mode function, in seconds. This represents the zero-crossing frequency of the intrinsic mode function, measured in Hertz.

[0028] Based on the zero-crossing frequencies, multiple eigenmode functions are divided into a low-frequency trend group and a high-frequency periodic group. A preset frequency threshold is used for this division. The zero-crossing frequency is less than The intrinsic mode functions are classified into the low-frequency trend group, and the zero-crossing frequencies are greater than or equal to The intrinsic mode functions are assigned to the high-frequency periodic group. Frequency thresholds are used for classification. The setting is based on the quasi-periodic characteristics of the 10kV transformer's operating state, and is selected as 0.01 Hz.

[0029] The trend component is obtained by summing the intrinsic mode functions (IMFs) of all data points within the low-frequency trend group in the time domain. The periodic component is obtained by summing the IMFs of all data points within the high-frequency periodic group in the time domain. The trend component and the periodic component are then added point-by-point at their corresponding time points to reconstruct the baseline reconstruction sequence. The baseline reconstruction sequence has a length of L sampling points, representing the baseline operating state corresponding to the historical data segment.

[0030] A Long Short-Term Memory (LSTM) network is trained using a baseline reconstruction sequence. The LSTM network comprises an input layer, two LSTM hidden layers, and a fully connected output layer. The input layer receives a baseline reconstruction sequence segment of length P time steps, where P is less than L. The two LSTM hidden layers have 128 and 64 units respectively. The fully connected output layer outputs the predicted value for the next time step. During training, a sequence segment of P consecutive time steps is extracted from the baseline reconstruction sequence as the input sample. The baseline reconstruction sequence value at the (P+1)th time step immediately following this segment is used as the training target. A mean squared error loss function is employed, and the network weight parameters are updated using the Adam optimizer. The learning rate is set to 0.001, and the number of training epochs is set to 200, until the loss function converges. The trained LSTM network serves as the baseline model for the running state.

[0031] Example 2:

[0032] In practice, the operating parameter sequence of the 10kV transformer is acquired in real time. This sequence contains operating parameter values ​​arranged chronologically by acquisition time, with a fixed time interval between adjacent acquisition moments. The operating parameters for each acquisition moment in the sequence are input into the operating state baseline model. The operating state baseline model is the Long Short-Term Memory (LSTM) network trained in Example 1. During the inference phase of this LSM network, for each acquisition moment, the operating state baseline model receives the operating parameters at the current moment and the hidden states passed within the network, generating and outputting the baseline prediction value for the current acquisition moment.

[0033] Calculate the difference between the operating parameters and the baseline prediction value at each acquisition time to obtain the raw deviation value at each acquisition time. Let the acquisition time be... The operating parameters are The baseline prediction value output by the running state baseline model is Then the original deviation value The raw deviation values ​​at all acquisition times are used to form a raw deviation value sequence, and the mean of the raw deviation value sequence is calculated. and standard deviation The Z-score normalization method is used to normalize each original deviation value, resulting in a normalized deviation value. The normalized deviation values ​​are arranged in chronological order of acquisition time to generate a state deviation sequence. ,in This represents the total length of the state deviation sequence.

[0034] A sliding window scan is performed on the state deviation sequence. Preset window length. and sliding step size ,in and All are positive integers, and Starting from the beginning of the state deviation sequence, according to the sliding step size... Move the window sequentially backward to extract multiple deviation windows. The starting index of each deviation window in the state deviation sequence is: The deviation window contains continuous The normalized bias values ​​are denoted as follows: .

[0035] Statistical calculations are performed on all normalized deviation values ​​within each deviation window to extract deviation fluctuation characteristics. These characteristics are composed of four features: window mean, window variance, maximum deviation value within the window, and window deviation change rate.

[0036] Window mean The calculation method is as follows:

[0037] in, This indicates the preset deviation window length. This indicates the starting index of the current deviation window in the state deviation sequence. , Indicates the number of deviations within the current deviation window. One normalized bias value, This represents the window mean of the current deviation window.

[0038] Window variance The calculation method is as follows:

[0039] in, This indicates the preset deviation window length. This represents the window mean of the current deviation window. Indicates the number of deviations within the current deviation window. One normalized bias value, This represents the window variance of the current deviation window.

[0040] Maximum deviation of window It is the maximum absolute value among all normalized deviation values ​​within the current deviation window, i.e. .

[0041] Window bias change rate The calculation method is as follows: ,in The first normalized deviation value within the current deviation window. This is the last normalized deviation value within the current deviation window. The time interval between adjacent data collection moments is expressed in seconds.

[0042] Window mean of the same deviation window Window variance Maximum deviation of the window and window deviation rate of change Combined into a four-dimensional feature vector This serves as a characteristic of deviation fluctuation.

[0043] The deviation fluctuation features of each deviation window are input into a pre-trained anomaly pattern recognition network. The pre-trained anomaly pattern recognition network uses a convolutional neural network structure, including an input layer, a first convolutional layer, a second convolutional layer, a global average pooling layer, a fully connected layer, and an output layer. The input layer receives a four-dimensional feature vector. The first convolutional layer contains 32 one-dimensional convolutional kernels with a kernel size of 3, and uses the ReLU activation function. The second convolutional layer contains 64 one-dimensional convolutional kernels with a kernel size of 3, and uses the ReLU activation function. The output of the second convolutional layer is connected to the global average pooling layer, which averages the values ​​of each feature map. After the global average pooling layer, a fully connected layer is connected, containing 128 neurons, using the ReLU activation function. The output layer uses the Softmax activation function, and the number of output nodes is set to the total number of anomaly types to be identified. Each output node corresponds to one anomaly type and outputs a value between 0 and 1 as the anomaly confidence score for the anomaly type label. The anomaly pattern recognition network outputs the anomaly type label with the highest anomaly confidence score and its corresponding anomaly confidence score.

[0044] The anomaly confidence score output by the anomaly pattern recognition network is compared with a preset threshold of 0.85. When the anomaly confidence score is greater than 0.85, the anomaly type label corresponding to the deviation window is taken as the anomaly pattern matching result. When the anomaly confidence score is less than or equal to 0.85, it is determined that the deviation window has not matched a clear anomaly pattern, and no anomaly pattern matching result is output. As the sliding window traverses all state deviation sequences, a series of anomaly pattern matching results are obtained, each indicating the specific anomaly type that appeared within the corresponding time window.

[0045] Example 3:

[0046] In the specific implementation, historical fault records of the 10kV transformer are collected over multiple historical time periods. Each historical fault record includes at least the fault occurrence time, fault type label, and the time series of operating parameters before and after the fault. The operating parameter sequence within a preset time period (30 minutes) before the fault is extracted from each historical fault record. For the extracted operating parameter sequence, the same time-series decomposition and baseline modeling process as in Example 1 is used to obtain the baseline predicted value sequence corresponding to the operating parameter sequence. Then, the difference between the operating parameter values ​​and the baseline predicted values ​​is calculated in the same way as in Example 2, and Z-score normalization is performed to obtain the historical state deviation sequence.

[0047] Perform the same window truncation operation as the sliding window scan in Example 2 on the historical state deviation sequence. Preset window length. and sliding step size The values ​​used are consistent with those used in Example 2. Starting from the beginning position of the historical state deviation sequence, the values ​​are adjusted according to the sliding step size. Multiple deviation windows are captured sequentially, each containing a continuous [value / section]. There are normalized bias values. Let the total length of a historical state bias sequence be... The number of historical deviation windows obtained is ,in This indicates the floor function.

[0048] For each historical deviation window, the window mean, window variance, maximum deviation value, and window deviation change rate are calculated according to the statistical calculation method in Example 2, and combined into a four-dimensional feature vector as the deviation fluctuation feature of the historical deviation window.

[0049] Each historical deviation window is labeled with a corresponding historical anomaly type label. The labeling of historical anomaly type labels is based on the fault types recorded in the original historical fault records, converting the fault types into coded category identifiers. If a historical deviation window is located in the precursor stage of a certain fault in time, then the label of that historical deviation window is set to the anomaly type label corresponding to that fault type; if a historical deviation window does not overlap with any known precursor stage of a fault in time, then the label of that historical deviation window is set to the label corresponding to "normal". The deviation fluctuation characteristics of all historical deviation windows and their corresponding historical anomaly type labels constitute the training sample set.

[0050] Construct an initial classification network based on a convolutional neural network. The input layer of the initial classification network receives a 4-dimensional bias fluctuation feature vector. After the input layer, a first convolutional layer is connected. This first convolutional layer uses one-dimensional convolution with 64 kernels, each with a size of 3, a stride of 1, and the ReLU activation function. The output of the first convolutional layer is connected to a second convolutional layer, which also uses one-dimensional convolution with 128 kernels, each with a size of 3, a stride of 1, and the ReLU activation function. After the second convolutional layer, a global average pooling layer is connected. This layer averages all elements of each feature map output from the second convolutional layer, outputting a 128-dimensional feature vector. After the global average pooling layer, a fully connected layer is connected with 64 neurons, using the ReLU activation function. Finally, a Dropout layer is connected, with a dropout probability set to 0.5. The dropout layer is followed by an output layer. The number of neurons in the output layer is equal to the total number of anomaly type labels. The activation function is the Softmax function, and the output layer shows the predicted probability for each anomaly type.

[0051] When training the initial classification network, the bias fluctuation feature vector from the training sample set is used as the input to the initial classification network, and the corresponding historical anomaly type label is used as the target output. The loss function is the classification cross-entropy loss function. The optimizer is the Adam optimizer, with a learning rate of 0.0005 and a batch size of 32. Early stopping is used during training; the training sample set is divided into a training subset and a validation subset in a 4:1 ratio. Training stops when the loss function value on the validation subset no longer decreases within 15 consecutive training epochs. The trained initial classification network serves as the pre-trained anomaly pattern recognition network. The pre-trained anomaly pattern recognition network takes a bias fluctuation feature vector from a bias window as input and outputs a vector containing the predicted probabilities of each anomaly type label. The maximum predicted probability is used as the anomaly confidence, and the anomaly type label corresponding to this maximum value is used as the anomaly type label for that bias window.

[0052] During the real-time monitoring phase, the anomaly initiation time is located based on the anomaly pattern matching results output by the anomaly pattern recognition network in Example 2. The target deviation window, whose anomaly confidence first exceeds a preset threshold, is extracted from the anomaly pattern matching results. The preset threshold is set to 0.85. The target deviation window has a definite starting index position in the state deviation sequence, and the acquisition time corresponding to this starting index position is the starting acquisition time of the target deviation window. This starting acquisition time of the target deviation window is taken as the anomaly initiation time.

[0053] Using the anomaly initiation time as the endpoint, historical operating parameters are traced back for a preset duration. The tracing duration is determined based on the precursor development time of a typical 10kV transformer fault, with a range of 10 to 60 minutes; in specific applications, the tracing duration is set to 20 minutes. Operating parameters from all data collected within 20 minutes prior to the anomaly initiation time are extracted from the historical operating parameter records and arranged chronologically to obtain the tracing parameter sequence.

[0054] The same time-series decomposition operation as that performed on historical operational data is applied to the traceability parameter sequence. The specific method of the time-series decomposition operation is consistent with that in Example 1: the traceability parameter sequence is divided into multiple traceability data segments using a preset sliding time window; empirical mode decomposition is performed on each traceability data segment to obtain multiple traceability intrinsic mode functions (IMFs); the zero-crossing frequency of each IMF is calculated; based on the zero-crossing frequency, the traceability IMFs are divided into a low-frequency trend group and a high-frequency periodic group; the traceability trend component is obtained by summing all traceability IMFs within the low-frequency trend group; and the traceability periodic component is obtained by summing all traceability IMFs within the high-frequency periodic group. The traceability trend component and the traceability periodic component are combined into an anomaly precursor sequence. The combination method involves taking the traceability trend component value and the traceability periodic component value at the same acquisition time as two elements of a two-dimensional vector, and arranging all two-dimensional vectors into a sequence according to the acquisition time order to obtain the anomaly precursor sequence. Each acquisition time in the anomaly precursor sequence corresponds to a two-dimensional feature vector containing trend information and periodic information.

[0055] See Figure 3 In the figure, the horizontal axis represents the tracing time in minutes, ranging from -30 minutes to 0 minutes, with 0 minutes corresponding to the start of the anomaly; the vertical axis represents the normalized amplitude. The figure contains two curves: the dashed curve represents the tracing period component, and the broken line curve represents the tracing trend component. Both are time-series decomposition results of different frequency components obtained by performing empirical mode decomposition on the tracing parameter sequence.

[0056] The overall trend component curve shows a slow changing trend, with amplitude fluctuating between -0.6 and -0.3, indicating a relatively stable negative trend. This reflects the overall trend changes of transformer operating parameters in the 30 minutes prior to the onset of the anomaly. The periodic component curve exhibits high-frequency fluctuations, with amplitude ranging from approximately -1.2 to 1.2, indicating that the operating parameters exhibit significant periodic fluctuations within the same time period.

[0057] As shown in the figure, as the tracing time gradually approaches the anomaly initiation time of 0 minutes from -30 minutes, the fluctuation amplitude of the tracing period component increases, and the periodic characteristics become more obvious, indicating that the periodic fluctuations of the operating parameters show an increasing trend before the anomaly. The tracing trend component, on the other hand, maintains a relatively stable change without abrupt changes, indicating that the overall trend responds relatively mildly to the anomaly.

[0058] Example 4:

[0059] In the specific implementation, the sequence index position of the abnormal start time in the state deviation sequence is determined. The state deviation sequence is a one-dimensional sequence arranged in chronological order of acquisition times, with each sequence position corresponding to an acquisition time and a normalized deviation value. The abnormal start time is the starting acquisition time of the target deviation window obtained in Example 3. This abnormal start time is compared one by one with each acquisition time in the state deviation sequence to find the acquisition time that matches it. The position number of this acquisition time in the state deviation sequence is the sequence index position, denoted as […]. .

[0060] Extracting from the state deviation sequence at the sequence index position The subsequent deviation subsequence from the beginning to the end of the state deviation sequence. Let the total length of the state deviation sequence be... Sequence index position satisfy Then the subsequent biased subsequence for:

[0061] in, This indicates the position in the sequence index of the state deviation sequence. Normalized deviation value at, This represents the normalized deviation value at the end of the state deviation sequence.

[0062] The length of the subsequent biased subsequence is denoted as .

[0063] The anomaly precursor sequence is time-aligned with subsequent deviation subsequences according to the acquisition time. The anomaly precursor sequence is generated by Example 3. Each acquisition time in the anomaly precursor sequence corresponds to a two-dimensional feature vector, which consists of a tracing trend component value and a tracing period component value. The acquisition time range of the tracing parameter sequence covers all acquisition times before the anomaly start time with a preset tracing duration. The last acquisition time of the anomaly precursor sequence is the anomaly start time.

[0064] The first acquisition time of the subsequent deviation subsequence is also the anomaly initiation time, and the subsequent deviation subsequence and the anomaly precursor sequence are naturally aligned at the anomaly initiation time. Starting from the anomaly initiation time, each acquisition time of the subsequent deviation subsequence corresponds to the feature vector of the corresponding acquisition time in the anomaly precursor sequence. If the anomaly precursor sequence has no value at a future acquisition time corresponding to a certain subsequent deviation subsequence, then no additional operation is performed at that time.

[0065] The two-dimensional feature vector of the anomaly precursor sequence at the onset of the anomaly is appended to the normalized deviation value at the same time in the subsequent deviation subsequence, generating a three-dimensional vector. The first element of the three-dimensional vector is the normalized deviation value, the second element is the traceability trend component value, and the third element is the traceability period component value. The same operation is performed for each acquisition time in the subsequent deviation subsequence, expanding each normalized deviation value in the subsequent deviation subsequence into a three-dimensional vector containing the normalized deviation value, the traceability trend component value, and the traceability period component value. The three-dimensional vectors of all acquisition times are arranged in chronological order to generate a multi-dimensional joint feature sequence.

[0066] In the multidimensional joint feature sequence, the first The feature vector at each acquisition time is denoted as . ,in For the subsequent biased subsequence Normalized deviation value at each acquisition time. For the corresponding number in the abnormal precursor sequence The retrospective trend component value at each acquisition time. For the corresponding number in the abnormal precursor sequence The traceability period component value at each acquisition time. .

[0067] The multidimensional joint feature sequence is stacked into a three-dimensional tensor along the time dimension. The multidimensional joint feature sequence contains a total of There are three-dimensional feature vectors, with the time dimension arranged in chronological order from the acquisition time to the end, and the feature channel dimension arranged according to the element numbers of the three-dimensional vectors. The shape of the stacked three-dimensional tensor is as follows: The first dimension The first dimension represents the time step, the second dimension (3) represents the number of feature channels at each time step, and the third dimension (1) is used to adapt to the input format of the subsequent fault evolution prediction model. The stacked three-dimensional tensor is used as the joint feature tensor.

[0068] The joint feature tensor is input into the fault evolution prediction model. The fault evolution prediction model employs an initial prediction network built upon gated recurrent units (GRUs), which is then used as the fault evolution prediction model after training. The fault evolution prediction model structure includes an input layer, two stacked GRU layers, a fully connected layer, and an output layer. The input layer receives signals of shape... The joint feature tensor is the two-dimensional form after removing the third single-channel dimension. The first gated recurrent unit layer contains 128 hidden units. With the return_sequences parameter set to True, the output shape is... The sequence. The second gated recurrent unit layer contains 64 hidden units. Setting the return_sequences parameter to False, the output shape is... The vector. The output of the second gated recurrent unit layer is connected to a fully connected layer, which contains 32 neurons and uses the ReLU activation function. After the fully connected layer, an output layer is connected, which contains... One neuron, Indicates the number of moments in the future forecast. Setting it to 24 means predicting the deviation values ​​at the next 24 acquisition times. The output layer does not use an activation function and directly outputs the values. The fault evolution prediction model receives the joint feature tensor and outputs a numerical value. A vector of values, representing the future starting from the current moment. The prediction deviation value at each moment. (The future...) The prediction deviation values ​​at each moment are arranged in chronological order to form the fault evolution trajectory.

[0069] The fault evolution prediction model is constructed through a training process. Training samples are obtained from multiple historical fault events in the historical operation data of a 10kV transformer. The operation time period recorded for each historical fault event is divided into a pre-fault stage and a post-fault stage, based on the fault occurrence time recorded in the historical fault record. A pre-fault anomaly precursor subsequence is extracted from the pre-fault stage, using the same extraction method as in Example 3. The operating parameter sequence is traced back for a preset traceability time based on the fault occurrence time. Through time-series decomposition, the traceability trend component and traceability cycle component are obtained and combined to form the pre-fault anomaly precursor subsequence. A post-fault deviation evolution subsequence is extracted from the post-fault stage. The extraction method is as follows: using the fault occurrence time as the starting time, and following the same baseline model and deviation calculation process as the real-time monitoring stage, the normalized deviation value sequence of multiple acquisition times after the fault occurrence time is calculated to obtain the post-fault deviation evolution subsequence.

[0070] The pre-fault anomaly precursor subsequence and the post-fault deviation evolution subsequence from each historical fault event are concatenated chronologically to form a joint sample sequence. The concatenation point is the fault occurrence time, and the last acquisition time of the pre-fault anomaly precursor subsequence and the first acquisition time of the post-fault deviation evolution subsequence are temporally continuous. Multiple sample 3D tensors are extracted from the joint sample sequence using a sliding time window method: a sample input duration window and a sample output duration window are set, with the sample input duration window length set to the same value as the input duration window. With the same adaptive length, the sample output duration window length is set to... Starting from the beginning of the sample joint sequence, the time window is moved sequentially with a step size of 1. Each time, a continuous input segment and a corresponding continuous output segment are extracted. The input segment is converted into a three-dimensional tensor format according to the method for generating the joint feature tensor in Example 4, becoming a sample three-dimensional tensor. The corresponding output segment is the segment immediately following the input segment with a length of... The actual deviation change value sequence.

[0071] The entire sample 3D tensor is used as the initial input to the prediction network, and a preset step size is applied to each sample's 3D tensor. The actual deviation change value is used as the output target of the initial prediction network, and the initial prediction network based on gated recurrent units is trained. The mean squared error loss function is used as the loss function, the Adam optimizer is used as the optimizer, the learning rate is set to 0.0003, the batch size is set to 16, and the maximum number of training epochs is set to 300. During training, a validation set early stopping mechanism is adopted. The validation set consists of 20% of the samples randomly divided from the training sample set. Training is terminated when the validation set loss function value no longer decreases within 20 consecutive training epochs. The initial prediction network after training is saved as a fault evolution prediction model and used to predict the fault evolution trajectory in the real-time monitoring stage.

[0072] See Figure 4 In the graph, the horizontal axis represents time in minutes, and the vertical axis represents the normalized deviation value. The solid curve represents the subsequent deviation subsequence, depicting the trend of the normalized deviation value from the start of the anomaly to the end of the monitoring; the dashed curve represents the fault evolution trajectory, showing the trend of the deviation value predicted by the fault evolution prediction model for the next 24 data collection times; the dotted horizontal line represents the failure deviation threshold, set at 3.0, used to determine whether the deviation value reaches the equipment failure threshold.

[0073] As time progressed, the subsequent deviation subsequences showed a gradually increasing trend. The deviation value started from near 0 and slowly increased. After about 300 minutes, the deviation value fluctuated frequently and continued to climb, reaching nearly 5.5 by the 500th minute, indicating that the equipment status was deviating from the normal baseline more significantly. After 300 minutes, the subsequent deviation subsequences repeatedly fluctuated by 1 to 2 times the initial value of the failure deviation threshold, showing an intensifying trend of the abnormal state.

[0074] The fault evolution trajectory lies to the right of the subsequent deviation subsequence, with its starting point aligned with the failure deviation threshold. The predicted deviation value curve starts at approximately 3.0 and rapidly rises to over 8.0, indicating that the future deviation value will significantly exceed the failure threshold, suggesting that the equipment is about to enter a severe failure stage. The volatility of the predicted trajectory indicates that the deviation value may fluctuate in the short term during the fault evolution process, but the overall deviation level remains in the high-risk range.

[0075] Example 5:

[0076] In specific implementation, please refer to Figure 2 The prediction deviation value sequence is extracted from the fault evolution trajectory. The fault evolution trajectory is output by the fault evolution prediction model in Example 4 and includes future... The prediction deviation value at each acquisition time. Set to 24. Let the prediction bias value sequence be... ,in This indicates the number of data points collected from the current data collection time. The prediction deviation value at each future acquisition time. The value ranges from 1 to 24.

[0077] Each predicted deviation value in the predicted deviation value sequence is compared item by item with a preset failure deviation threshold. The failure deviation threshold is set according to the allowable deviation range of the 10kV transformer operating state. Based on the voltage deviation limit requirements in the power equipment condition assessment guidelines and the corresponding relationship after normalization, the failure deviation threshold is set to 3.0. During the item-by-item comparison, starting from... Begin, Compared with the failure deviation threshold of 3.0, if Then continue the comparison. ;like If the comparison fails, stop the comparison and record the current sequence number. .

[0078] The prediction time corresponding to the first prediction deviation value in the prediction deviation value sequence that reaches or exceeds the failure deviation threshold is determined as the failure arrival time. Let the first value satisfy... The prediction deviation value is numbered as follows The time interval between adjacent acquisition times is , If a fixed value is used, for example, 1 minute, then the time difference between the time the failure occurred and the current acquisition time is [value missing]. Calculate this time difference to obtain the remaining effective runtime. If all 24 prediction bias values ​​in the prediction bias value sequence are less than 3.0, then... Set to a value greater than The default value is 720 minutes, which means that no failure trend has occurred within the predictable range.

[0079] The remaining valid runtime is matched against multiple preset alarm duration intervals. There are three levels of alarm duration intervals: the first interval is for remaining valid runtime less than or equal to 30 minutes, corresponding to a level 1 alarm; the second interval is for remaining valid runtime greater than 30 minutes but less than or equal to 120 minutes, corresponding to a level 2 alarm; and the third interval is for remaining valid runtime greater than 120 minutes, corresponding to a level 3 alarm. During matching, the following checks are performed... The alarm level corresponding to the alarm duration range that falls within the range will be determined as the status monitoring alarm level.

[0080] After generating the status monitoring alarm level, the current load rate of the 10kV transformer is obtained. The current load rate is obtained from the real-time collected operating parameters and is calculated based on the current measured current value. With the rated current value of a 10kV transformer The ratio, expressed as a percentage, is... The remaining effective runtime is adjusted based on the current load rate to obtain the adjusted effective runtime. The load adjustment formula is:

[0081] in, Indicates the effective runtime after correction, in minutes; Indicates the remaining valid runtime, in minutes; This indicates the current load factor of a 10kV transformer, expressed as a percentage, ranging from 0% to 200%. This represents the rated load rate of a 10kV transformer, with a value of 100%, which is the benchmark value of the load rate under rated operating conditions. Indicates the load correction factor. The value is determined based on the insulation aging acceleration characteristic curve of a 10kV transformer under overload operation, and is obtained by fitting offline thermal aging test data. Set to 1.2.

[0082] Based on the corrected effective runtime Rematch the multiple alarm duration ranges. Compare with the threshold values ​​of 30 minutes and 120 minutes for the alarm duration range to determine The updated status monitoring alarm level is obtained based on the alarm duration range it falls within. The updated status monitoring alarm level is then compared with the previous one. The matched status monitoring alarm levels are then selected, with the higher level being the final alarm level output. Level 1 is the highest, followed by Level 2, and Level 3 is the lowest. The final alarm level is sent to the monitoring system as an alarm signal.

[0083] See Figure 5 In the graph, the horizontal axis represents time in minutes, and the vertical axis represents the normalized deviation value. The solid curve in the legend represents the prediction deviation sequence, indicating the trend of the prediction deviation value over the next 24 data collection moments from the current moment. This sequence starts at 0 and shows an overall upward trend over time, reflecting a gradual increase in prediction deviation. The horizontal dashed line represents the failure deviation threshold, with a value of 3.0. This indicates that exceeding the normalized deviation value is considered a critical threshold for equipment failure. The vertical dashed line marks the time point at which the prediction deviation sequence first reaches the failure deviation threshold, approximately 60 minutes in, indicating the time of failure arrival.

[0084] The two vertical gray dashed lines in the figure correspond to the start and end times of the remaining effective runtime calculation, respectively. The short gray dashed line represents the original value of the remaining effective runtime, which is the time difference between when the predicted deviation sequence first reaches the failure deviation threshold and the current time, approximately 60 minutes. The gray dotted line represents the load correction result of the remaining effective runtime based on the current load rate. The corrected effective runtime is slightly less than the original remaining effective runtime, showing the impact of load correction on the remaining lifetime assessment.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring the operating status of a 10kV transformer based on time-series analysis, characterized in that, The method includes: The historical operating data of a 10kV transformer is decomposed into time series to obtain trend components and periodic components at multiple time scales, and an operating status baseline model is constructed based on the trend components and periodic components. The operating parameter sequence of a 10kV transformer is collected in real time, and the operating parameter sequence is mapped to the state space defined by the operating state baseline model to generate a state deviation sequence. A sliding window scan is performed on the state deviation sequence to extract the deviation fluctuation features within the window, and the deviation fluctuation features are input into a pre-trained abnormal pattern recognition network to obtain abnormal pattern matching results. Based on the abnormal pattern matching result, the anomaly start time is located, and the running parameter sequence of a preset duration is traced back based on the anomaly start time to generate an anomaly precursor sequence. The abnormal precursor sequence and the state deviation sequence are time-aligned to generate a joint feature tensor, and the joint feature tensor is input into the fault evolution prediction model to obtain the predicted fault evolution trajectory. The remaining effective runtime is calculated based on the fault evolution trajectory, and a status monitoring alarm level is generated based on the remaining effective runtime.

2. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 1, characterized in that, Historical operating data of a 10kV transformer is decomposed into time series to obtain trend components and periodic components at multiple time scales. Based on these trend and periodic components, an operating state baseline model is constructed, including: The historical running data is segmented using a preset sliding time window to obtain multiple historical data segments, and empirical mode decomposition is performed on each historical data segment to obtain multiple intrinsic mode functions; Calculate the zero-crossing frequency of each intrinsic mode function, and divide the multiple intrinsic mode functions into a low-frequency trend group and a high-frequency periodic group based on the zero-crossing frequency; The trend component is obtained by summing all intrinsic mode functions within the low-frequency trend group, and the period component is obtained by summing all intrinsic mode functions within the high-frequency period group. The trend component and the periodic component are added together to reconstruct a baseline reconstruction sequence, and the long short-term memory network is trained using the baseline reconstruction sequence. The long short-term memory network is used as the baseline model of the running state.

3. The method for monitoring the operating status of a 10kV transformer based on time-series analysis according to claim 2, characterized in that, The empirical mode decomposition process employs the mirror extension method to perform boundary processing on both ends of the historical data segments in order to suppress the endpoint effect during the decomposition process.

4. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 2, characterized in that, The operating parameter sequence is mapped to the state space defined by the operating state baseline model to generate a state deviation sequence, including: The operating parameters at each acquisition time in the operating parameter sequence are input into the operating state baseline model, and the operating state baseline model outputs the baseline prediction value at each acquisition time. Calculate the difference between the operating parameters and the baseline prediction value at each acquisition time to obtain the original deviation value at each acquisition time; The original deviation values ​​are normalized to obtain normalized deviation values, and the normalized deviation values ​​are arranged in chronological order of acquisition time to generate the state deviation sequence.

5. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 4, characterized in that, A sliding window scan is performed on the state deviation sequence to extract deviation fluctuation features within the window. These features are then input into a pre-trained anomaly pattern recognition network to obtain anomaly pattern matching results, including: According to the preset window length and sliding step size, multiple deviation windows are sequentially extracted from the starting position of the state deviation sequence, and each deviation window contains multiple consecutive normalized deviation values. Statistical calculations are performed on all normalized deviation values ​​within each deviation window to obtain the window mean, window variance, maximum window deviation value, and window deviation change rate for each deviation window. The window mean, window variance, maximum window deviation value, and window deviation change rate are then combined to form the deviation fluctuation feature. The deviation fluctuation features of each deviation window are input into the pre-trained anomaly pattern recognition network. The anomaly pattern recognition network outputs the anomaly type label and anomaly confidence score corresponding to each deviation window, and the anomaly type label with an anomaly confidence score exceeding a preset threshold is taken as the anomaly pattern matching result.

6. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 5, characterized in that, The pre-trained anomaly pattern recognition network is constructed in the following ways: Collect historical fault records of 10kV transformers and extract the historical state deviation sequence within a preset period before the fault occurred from each historical fault record; Perform the same window truncation operation as the sliding window scan on the historical state deviation sequence to obtain multiple historical deviation windows, and label each historical deviation window with the corresponding historical anomaly type label; Using the deviation fluctuation features of each historical deviation window as input and the corresponding historical anomaly type label as output, an initial classification network based on a convolutional neural network is trained, and the trained initial classification network is used as the pre-trained anomaly pattern recognition network.

7. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 6, characterized in that, Based on the abnormal pattern matching results, the anomaly initiation time is located, and an operational parameter sequence of a preset duration is traced back from the anomaly initiation time to generate an anomaly precursor sequence, including: Extract the target deviation window from the abnormal pattern matching results when the abnormal confidence first exceeds the preset threshold, and take the starting acquisition time of the target deviation window as the abnormal start time; Using the start time of the anomaly as the endpoint, trace back the historical operating parameters for the preset duration to obtain the traced parameter sequence; Perform the same time-series decomposition operation as the historical running data on the traceability parameter sequence to obtain the traceability trend component and the traceability cycle component, and combine the traceability trend component and the traceability cycle component into the abnormal precursor sequence.

8. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 7, characterized in that, The abnormal precursor sequence and the state deviation sequence are time-aligned to generate a joint feature tensor. The joint feature tensor is then input into the fault evolution prediction model to obtain the predicted fault evolution trajectory, including: Determine the sequence index position of the anomaly start time in the state deviation sequence, and extract the subsequent deviation subsequence from the state deviation sequence starting from the sequence index position to the end of the state deviation sequence; The abnormal precursor sequence is time-aligned with the subsequent deviation subsequence according to the acquisition time, and the value of each time step of the aligned abnormal precursor sequence is appended to the normalized deviation value of the same time step in the subsequent deviation subsequence to generate a multidimensional joint feature sequence. The multidimensional joint feature sequence is stacked into a three-dimensional tensor according to the time dimension. The three-dimensional tensor is used as the joint feature tensor and input into the fault evolution prediction model. The fault evolution prediction model outputs the prediction deviation values ​​at multiple future time points. The prediction deviation values ​​at multiple future time points are arranged in chronological order to form the fault evolution trajectory.

9. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 8, characterized in that, The fault evolution prediction model is constructed in the following ways: Using multiple historical fault events from the historical operation data of 10kV transformers as samples, the pre-fault anomaly precursor subsequence and the post-fault deviation evolution subsequence are extracted from each historical fault event; The pre-fault anomaly precursor subsequence and post-fault deviation evolution subsequence in each historical fault event are concatenated in chronological order to form a sample joint sequence, and multiple sample three-dimensional tensors are extracted from the sample joint sequence. Using the multiple sample 3D tensors as input and the actual deviation change value after a preset step size for each sample 3D tensor as output, an initial prediction network based on gated recurrent units is trained, and the trained initial prediction network is used as the fault evolution prediction model.

10. The method for monitoring the operating status of a 10kV transformer based on time-series analysis as described in claim 1, characterized in that, The remaining effective runtime is calculated based on the fault evolution trajectory, and a status monitoring alarm level is generated based on the remaining effective runtime, including: Extract the predicted deviation value sequence from the fault evolution trajectory, and compare each predicted deviation value in the predicted deviation value sequence with a preset failure deviation threshold item by item; The prediction time corresponding to the first prediction deviation value in the prediction deviation value sequence that reaches or exceeds the failure deviation threshold is determined as the failure arrival time, and the difference between the failure arrival time and the current acquisition time is calculated, and the difference is used as the remaining effective running time. The remaining valid runtime is matched with multiple preset alarm duration intervals, and the alarm level corresponding to the matched alarm duration interval is determined as the status monitoring alarm level.

Images