Method and device for predicting tripping risk of power transformation equipment, computer equipment and medium

Abstract

The invention relates to the technical field of intelligent power grids, in particular to a power transformation equipment tripping risk prediction method and device, computer equipment and a medium, and the method comprises the steps: obtaining time sequence data of power transformation equipment, and carrying out the preprocessing of the time sequence data, and obtaining a preprocessed time sequence data set; based on the preprocessed time series data set, time domain features are extracted through mean value, variance, maximum value, minimum value, skewness and kurtosis methods, frequency features are extracted through a Fourier transform method, and time domain-frequency domain features are extracted through a short-time Fourier transform or wavelet transform method; screening the extracted features through principal component analysis or an LASSO method to obtain a feature data set; inputting the feature data set into a pre-trained deep learning model to obtain a tripping risk prediction result; and generating an emergency processing scheme based on the tripping risk prediction result, and notifying operation and maintenance personnel of the emergency processing scheme.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, and in particular to a method, device, computer equipment and medium for predicting the tripping risk of power equipment. Background Technology

[0002] Transformer equipment is a crucial component of the power system, primarily used for receiving, converting, and distributing electrical energy. Its main task is to transmit and regulate electrical energy through components such as transformers, electrical switches, and protection devices. Equipment tripping occurs frequently, causing losses not only to power supply companies but also significantly impacting daily life for residents and production for industrial enterprises, severely affecting customer experience. Dynamic and impact analysis and emergency decision-making after equipment tripping generally require the integration of multiple information sources and in-depth model mechanism research and data mining. Currently, digital technology still has shortcomings in the practical application of real-time perception and accurate judgment of equipment tripping, mainly reflected in: Research on the practical application of automatic data tracking and intelligent linkage based on equipment defects and anomalies within the station is still in its infancy, and the effectiveness of intelligent substation construction in improving the efficiency and accuracy of fault handling needs to be further enhanced. The heterogeneity of data from primary and secondary equipment and auxiliary equipment within the station still necessitates in-depth exploration of data unification, efficient integration, and utilization. The application of collaborative handling technology for equipment anomalies / failures remains at a relatively superficial level, failing to achieve the integration of fault handling strategies into different roles. Summary of the Invention

[0003] Based on this, it is necessary to address the technical problems of existing substation tripping early warning methods, which suffer from single feature extraction dimensions, feature redundancy, and lack of effective screening, resulting in incomplete capture of early fault features, low model prediction accuracy, and poor interpretability, making it difficult to achieve accurate early warning and rapid closed-loop handling. Therefore, a method, device, computer equipment, and medium for predicting the tripping risk of substation equipment are proposed.

[0004] A first aspect of this application provides a method for predicting the tripping risk of power equipment, the method comprising: Acquire time-series data of the power equipment and preprocess the time-series data to obtain a preprocessed time-series dataset; Based on the preprocessed time series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-frequency domain features are extracted using the short-time Fourier transform or wavelet transform methods. The extracted features are filtered using principal component analysis or LASSO methods to obtain a feature dataset; The feature dataset is input into a pre-trained deep learning model to obtain the trip risk prediction result; An emergency response plan is generated based on the tripping risk prediction results, and the emergency response plan is then communicated to the operation and maintenance personnel.

[0005] Preferably, the step of acquiring time-series data of the power equipment and preprocessing the time-series data to obtain preprocessed time-series data includes: The time series data is denoised using Kalman filtering and wavelet transform. Linear interpolation is performed on the denoised time-series data to fill in the missing time-series data and obtain a complete time-series dataset.

[0006] Preferably, the expression for Kalman filtering is: ; in, H is the Kalman gain, and H is the observation matrix. These are observed values. It is time-series data; The expression for the wavelet transform is: ; in, It is the original time-series signal to be analyzed. It is the mother wavelet. a It is a scale factor. a ≠0, b is the shift factor; It is a normalization factor; The expression for the linear interpolation is: ; Where t is the timestamp corresponding to the missing data point to be filled, satisfying... ≤t≤ It is the interpolation interval.

[0007] This refers to the timestamps corresponding to two known valid data points. .

[0008] Preferably, the step of inputting the feature dataset into a deep learning model to obtain the trip risk prediction result includes: The feature dataset is constructed as a feature sequence arranged by time steps; The feature sequence is input into the pre-trained deep learning model; wherein the deep learning model is a Long Short-Term Memory (LSTM) network model or a Gated Recurrent Unit (GRU) network model. The deep learning model is used to perform forward propagation calculation on the feature sequence, processing the input at each time step in sequence and updating its internal hidden state. The tripping risk prediction result is obtained by mapping the hidden state output by the deep learning model.

[0009] Preferably, the timing data of the power equipment is the electrical parameters of the power equipment collected by current sensors, voltage sensors, temperature sensors and vibration sensors.

[0010] Preferably, the step of generating an emergency response plan based on the tripping risk prediction result and notifying the operation and maintenance personnel of the emergency response plan includes: Obtain the fault type and risk level from the trip risk prediction results; Based on the fault type and risk level, a repair method is mapped to obtain an emergency response plan; The emergency response plan will be communicated to the operations and maintenance personnel.

[0011] Preferred options also include: Obtain repair execution information from the aforementioned maintenance personnel; Based on the repair execution information, the deep learning model is optimized using a weighted binary cross-entropy loss function.

[0012] A second aspect of this application provides a device for predicting the tripping risk of power equipment, the device comprising: The data acquisition module is used to acquire time-series data of the substation equipment, preprocess the time-series data, and obtain a preprocessed time-series dataset. The risk prediction module is used to extract time-domain features based on the preprocessed time-series dataset by means, variance, maximum, minimum, skewness and kurtosis methods, extract frequency features by Fourier transform method, and extract time-domain-frequency domain features by short-time Fourier transform or wavelet transform method. Based on the time-domain, frequency-domain, and time-frequency-domain features, feature selection is performed using the LASSO method to obtain a feature dataset. The feature dataset is input into a deep learning model to obtain the trip risk prediction result; Emergency response plan generation module: Generates an emergency response plan based on the trip risk prediction results and notifies the operation and maintenance personnel of the emergency response plan.

[0013] A third aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the substation tripping risk prediction method as described in any of the preceding claims.

[0014] A fourth aspect of this application provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of predicting the tripping risk of power equipment as described in any of the preceding claims.

[0015] The beneficial effects of this application.

[0016] By acquiring real-time time-series data of power equipment, the system covers all aspects of equipment operation. Advanced denoising techniques such as wavelet transform and Kalman filtering are used, combined with methods such as linear interpolation to handle missing and outlier values, effectively improving data quality and laying a solid foundation for subsequent analysis. Features are extracted from multiple dimensions, including time domain, frequency domain, and time-frequency domain, such as time domain features statistics such as mean, variance, maximum, minimum, skewness, and kurtosis, as well as frequency features and power spectral density extracted by root mean square value and Fourier transform. Feature selection algorithms such as PCA and LASSO are used to optimize the feature space and screen out features that have a significant impact on fault prediction, thereby enhancing the predictive ability of the model. By employing a deep learning model, the system enables real-time interaction with equipment status data, generates trip prediction results, and outputs fault risk warnings (such as low risk and high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, thus significantly improving the timeliness and accuracy of fault handling. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] in: Figure 1 This is a diagram illustrating the application environment of a power equipment tripping risk prediction method in one embodiment. Figure 2 A flowchart of a method for predicting the tripping risk of substation equipment in one embodiment; Figure 3 This is a structural block diagram of a power equipment tripping risk prediction device in one embodiment; Figure 4 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0019] 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, not all, of the embodiments of the present invention. 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.

[0020] The method for predicting the tripping risk of power equipment provided in this invention can be applied to, for example... Figure 1 In this application environment, the device communicates with the server via a network. The server can obtain time-series data of the substation equipment from the device and preprocess the time-series data to obtain a preprocessed time-series dataset. Based on the preprocessed time-series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-domain-frequency-domain features are extracted using the short-time Fourier transform or wavelet transform method. The extracted features are then filtered using principal component analysis or LASSO methods to obtain a feature dataset. The feature dataset is input into a pre-trained deep learning model to obtain tripping risk prediction results. Based on the tripping risk prediction results, an emergency response plan is generated and notified to the operation and maintenance personnel. In this application, a deep learning model is employed in the application scenario of power equipment to achieve real-time interaction with equipment status data, generate tripping prediction results, and output fault risk warnings (such as low risk, high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, significantly improving the timeliness and accuracy of fault handling. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0021] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the method for predicting the tripping risk of power equipment provided in this embodiment of the invention includes the following steps: S1: Obtain the time-series data of the power equipment, preprocess the time-series data, and obtain the preprocessed time-series dataset.

[0022] Specifically, time series data consists of a series of data points collected and recorded in chronological order, with each data point associated with a specific timestamp. This type of data naturally exhibits time dependence; correlations (such as autocorrelation, trends, and periodicity) often exist between consecutive observations, therefore it cannot be considered an independent random sample. It is usually expressed as: Where t is the time index and T is the sequence length. This is an ordered set representing all data points collected sequentially from time point t=1 to t=T. In a substation scenario, time-series data includes records of physical quantities such as voltage, current, power, temperature, and vibration sampled continuously or discretely over time.

[0023] Real-time monitoring data (time-series data) of power equipment is acquired and processed through cleaning, repair, and standardization to form a well-organized dataset suitable for subsequent analysis. Time-series data typically originates from sensors, protection devices, and other equipment, recording continuous changes in key parameters such as voltage, current, and temperature. Preprocessing is a fundamental step in data mining, aiming to eliminate noise interference, correct outliers, and fill in missing data, transforming the raw data into consistent and reliable structured information.

[0024] During data acquisition, raw time-series data is first obtained from monitoring systems or online monitoring devices to check its integrity and consistency. Next, data cleaning is performed to identify and process outlier data points that are outside the physical range or have remained unchanged for a long time. Then, methods such as linear interpolation or nearest-neighbor imputation are used to repair missing data. Finally, normalization techniques are used to standardize the data, eliminating the influence of different units of measurement and forming a high-quality time-series dataset.

[0025] Taking transformers as an example, we obtain continuous 24-hour temperature and load current data from the oil temperature monitoring system and current transformers. In the preprocessing stage, we first remove abrupt outliers caused by sensor failures (such as instantaneous temperatures exceeding 150°C), then use time series interpolation to supplement the 5-minute data segment missing due to communication interruption, and finally normalize all current data to the [0,1] interval to form a preprocessed dataset containing 12,000 regular time points, laying the foundation for feature extraction.

[0026] S2: Based on the preprocessed time series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-frequency domain features are extracted using the short-time Fourier transform or wavelet transform method.

[0027] Specifically, time-domain features are statistics or indicators calculated directly in the time dimension of the original signal. They describe the intuitive characteristics of how the signal amplitude changes over time, such as signal strength, fluctuation range, distribution shape, and rate of change. Frequency-domain features, on the other hand, are obtained by converting the signal from the time dimension to the frequency dimension and analyzing its frequency components. This conversion helps us answer which different frequencies and intensities of 'basic sine waves' combine to form this complex signal. Frequency-domain features reveal the periodicity, resonance, and structural information implicit in the signal, information that is often difficult to perceive directly in the time domain.

[0028] Through mathematical transformations, physically meaningful feature indicators are extracted from preprocessed time-series data. These include time-domain features reflecting statistical characteristics, frequency-domain features revealing periodic patterns, and time-frequency-domain features that simultaneously capture changes in time and frequency. These features constitute a multi-dimensional representation of the device's state. Time-domain features describe the overall signal distribution, frequency-domain features show the energy distribution across frequencies, and time-frequency-domain features can pinpoint frequency component changes at specific moments.

[0029] In the time domain, six basic statistics are calculated for the data within each time window: mean, variance, maximum, minimum, skewness, and kurtosis. The time window covers the typical development time of equipment failure (e.g., partial discharge takes minutes to develop, and mechanical wear takes hours). There are sufficient sampling points (generally ≥30 points) within the window to ensure statistical reliability. In the frequency domain, Fourier transform is used to convert the time domain signal into a frequency domain representation, and features such as the main frequency amplitude and harmonic components are extracted. In the time-frequency domain, short-time Fourier transform or wavelet transform is used to obtain a spectrum diagram that changes over time, from which feature parameters of energy concentration regions are extracted, and finally a set of feature vectors containing multiple dimensions is formed.

[0030] The Short-Time Fourier Transform (STFT) involves segmenting and windowing the signal before performing the Fourier transform. The specific expression for a continuous-time signal x(t) is as follows: Where x(t) is the original time-domain signal, such as current, voltage, or vibration signal; For integration, represents all historical time points that need to be considered when calculating the time-frequency representation at time t; It is a window function, a weighted function around time t; t is a time variable, representing the time point in the analysis; For frequency variables, it represents the frequency of the analysis; For time-frequency representation, it is a complex number containing amplitude and phase information at time t and frequency f.

[0031] For frequency domain analysis kernel function, the decomposition expression is: Where j is the imaginary unit, satisfying ; Real part matching: Detecting signals with Similarity; Imaginary part matching: Detecting signals with... πfτ) Similarity; Complex form Simultaneously acquire amplitude and phase information.

[0032] Taking the vibration signal of a circuit breaker operating mechanism as an example, vibration sensors (such as accelerometers) are used to collect the vibration signal x(t) of the circuit breaker operating mechanism during the closing and opening processes. The original signal undergoes preprocessing such as denoising and interpolation (e.g., Kalman filtering, wavelet denoising, linear interpolation) to obtain clean vibration time-series data. A window function w(t) is set, such as a Hanning window or a Hamming window, with the window length set according to the frequency resolution requirements of the signal (e.g., 10 ms ~ 100 ms). To improve time resolution, an overlap of 50% ~ 75% is typically set. The vibration signal x(t) is then subjected to STFT calculation to obtain the time-frequency matrix. From the time-spectrum obtained by STFT, various features can be extracted for fault diagnosis and risk prediction. For example, within a specific frequency band (e.g., 100 Hz ~ 500 Hz), the energy integral at each time point can be calculated to identify whether the mechanical resonant frequency is abnormal; the entropy or variance of the energy distribution in the time-spectrum can be calculated to determine whether the vibration is disordered (potentially indicating looseness or wear). The extracted time-frequency features (such as energy concentration, transient pulse energy, etc.) are combined with time-domain features (mean, variance, etc.) and frequency-domain features (main frequency amplitude extracted by Fourier transform, etc.) to form a high-dimensional feature vector.

[0033] Taking a specific application as an example, the vibration signal x(t) during the circuit breaker closing process was collected at a sampling frequency of 10 kHz, and the closing action lasted approximately 100 ms. The short-time Fourier transform employed a Hanning window with a window length of... =20ms (corresponding to 200 sampling points), overlap rate set to 75% (window shifted every 5 ms), frequency range: 0 ~ 2 kHz, time-domain-frequency domain analysis results are as follows: Normal closing: The time-spectrum graph shows a brief high-frequency energy pulse (approximately 800 Hz) at the moment of closing (t≈30 ms), which then decays rapidly. Abnormal closing (e.g., mechanism jamming): The time-spectrum graph shows a prolonged high-frequency energy duration, and multiple frequency components may appear (e.g., 500 Hz and 1 kHz appearing simultaneously). Calculate the energy integral in the 800 Hz band at the moment of closing (t=30 ms ± 5 ms) as the "closing impact energy" feature. Calculate the frequency entropy (measuring the degree of disorder in energy distribution) throughout the entire closing process as the "vibration disorder" feature. The above features, combined with electrical features such as current and voltage, can be directly used as input for subsequent deep learning models.

[0034] S3: The extracted features are filtered using principal component analysis or LASSO methods to obtain a feature dataset.

[0035] Specifically, the core idea of ​​Principal Component Analysis (PCA) for feature selection is to transform the original high-dimensional features into low-dimensional principal components while retaining most of the variance information, thereby achieving feature dimensionality reduction and redundancy removal. It can identify the most discriminative and representative feature subsets from a large number of extracted time-domain, frequency-domain, and time-frequency-domain features, while eliminating redundancy and multicollinearity among features.

[0036] First, construct a feature matrix from all time-domain, frequency-domain, and time-domain-frequency-domain features extracted in step S2. Calculate the mean and standard deviation of each feature to obtain a standardized matrix. Then, calculate the covariance matrix of the standardized feature matrix. The covariance matrix is ​​a symmetric square matrix, where the diagonal elements represent the variance of each feature, and the off-diagonal elements represent the pairwise covariance between features. Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The magnitude of the eigenvalue represents the variance contribution of the corresponding principal component, and the eigenvector defines the direction of the principal component. Calculate the variance contribution rate (eigenvalue of the principal component divided by the sum of all eigenvalues) and cumulative variance contribution rate for each principal component. Typically, a cumulative variance contribution rate of 85%-95% is set as a threshold, and the minimum number of principal components that can reach this threshold is selected. Only principal components with eigenvalues ​​greater than 1 are retained, because principal components with eigenvalues ​​less than 1 explain less variance than the original single feature. You can select features whose importance score exceeds a set threshold (such as 1.5 times the average importance score of all features); or select the top N features in terms of importance score (such as the top 20% or top 15 features). You can also sort the features by importance score from high to low and select features whose cumulative importance reaches 80%-90% of the total importance.

[0037] Taking feature selection for transformer tripping risk prediction as an example: Sixty features were extracted from transformer monitoring data, including: oil temperature-related features (15): mean, variance, and rate of change of top-layer oil temperature; mean and gradient of winding temperature; temperature difference (top-bottom layer), etc. Oil chromatographic features (12): content and growth rate of gases such as hydrogen, methane, ethane, ethylene, and acetylene. Electrical quantity features (18): mean three-phase current and unbalance; voltage harmonic distortion rate; power factor; reactive power, etc. Vibration features (10): mean core vibration acceleration, dominant frequency amplitude, and harmonic energy ratio, etc. Environmental features (5): ambient temperature, humidity, and load rate, etc.

[0038] The 60 features were standardized to eliminate the influence of dimensions. The covariance matrix was calculated, revealing high correlations (covariance > 0.8) among oil temperature-related features and among the gas contents of oil chromatograms. Eigenvalue decomposition yielded 60 eigenvalues, sorted by magnitude: the first 5 eigenvalues ​​were 12.5, 8.3, 6.1, 4.7, and 3.9, with subsequent eigenvalues ​​decreasing rapidly, the 15th eigenvalue being 0.8. The number of principal components to retain was determined: the first 10 principal components contributed 87.5% of the cumulative variance; 14 principal components had eigenvalues ​​greater than 1; the scree plot showed a clear inflection point at the 8th principal component; therefore, the first 10 principal components were retained.

[0039] Principal component load analysis identifies key original features: First principal component (variance contribution rate 20.8%): Main loads come from "acetylene growth rate" (0.92), "hydrogen growth rate" (0.85), and "top oil temperature change rate" (0.78), reflecting insulation thermal faults. Second principal component (variance contribution rate 13.8%): Main loads come from "current imbalance" (0.88), "negative sequence current" (0.82), and "vibration dominant frequency shift" (0.75), reflecting mechanical and electrical imbalances. Third principal component (variance contribution rate 10.2%): Main loads come from "ambient temperature" (-0.91), "load rate" (0.83), and "power factor" (0.79), reflecting the impact of operating conditions.

[0040] The overall importance score of 60 original features was calculated, and the 15 features with the highest importance scores were selected: acetylene gas growth rate (importance score: 0.185), top oil temperature change rate (0.162), current imbalance (0.148), hydrogen growth rate (0.142), winding temperature gradient (0.135), core vibration dominant frequency amplitude (0.128), negative sequence current percentage (0.121), methane / ethane ratio (0.115), total harmonic distortion rate of voltage (0.109), reactive power fluctuation variance (0.102), oil trace water content change rate (0.098), partial discharge pulse repetition rate (0.094), cooler operating rate (0.089), bushing dielectric loss factor (0.085), and neutral point current third harmonic (0.081). These 15 features constitute the final feature dataset, which is used as input to a deep learning model for trip risk prediction.

[0041] Specifically, the LASSO (Minimum Absolute Shrinkage and Selection Operator) regularization method automatically selects the subset of key features most relevant to tripping risk from a large number of extracted features. LASSO achieves both feature selection and model simplification by adding an L1 regularization term to the loss function, compressing some feature coefficients to zero. This method effectively handles the multicollinearity problem among high-dimensional features, improving the model's interpretability and generalization ability.

[0042] First, all time-domain, frequency-domain, and time-domain-frequency-domain features extracted in step S2 are constructed into a feature matrix, with historical tripping records of the equipment used as supervision labels. Then, a LASSO regression model is used, and the optimal regularization strength λ is determined through five-fold cross-validation to train and obtain a sparse coefficient vector. Finally, features with non-zero coefficients are selected to form a simplified feature set. This process typically reduces the number of features by 60%-80%, retaining the most discriminative feature combinations.

[0043] LASSO (Least Absolute Shrinkage and Selection Operator) is an improved method for linear regression. By adding an L1 regularization term (penalty term) to the loss function, it can compress the coefficients of some unimportant features to 0, thereby achieving automatic feature selection and realizing the dual goals of feature selection and model complexity control. In power industry scenarios, LASSO regularization can identify dozens of truly relevant indicators from hundreds of monitoring indicators, clarifying which sensors and indicators are most important. It is highly robust to irrelevant features and enables a real-time updated monitoring system.

[0044] For example, in a transformer fault early warning scenario, we extracted 150 initial features from oil chromatography data, electrical quantities, and vibration signals. After LASSO regularization, the model automatically selected 22 key features, including the acetylene growth rate in the oil, the third harmonic content of the core grounding current, and the gradient of the top oil temperature change, while eliminating redundant indicators that were highly correlated with these features. These selected features demonstrated stronger discriminative power in subsequent prediction models.

[0045] S4: Input the feature dataset into the deep learning model to obtain the trip risk prediction result.

[0046] Specifically, the LSTM (Long Short-Term Memory) model is a special type of recurrent neural network (RNN) designed to handle long-term dependencies in time-series data. The model as a whole adopts a sequence-to-class architecture, mapping the time-series feature sequences of power equipment to tripping risk prediction results. The first layer is the input layer, receiving preprocessed temporal feature sequences, including: batch size (the number of samples processed at one time during training); sequence length (the number of sampling points within the time window); and feature dimension (the number of features after filtering). The second layer is an LSTM layer (core layer), specifically 1-3 stacked LSTM layers, each containing 64-256 LSTM units. The forward LSTM processes forward temporal information, and the backward LSTM processes reverse temporal information. The forward and backward hidden states are concatenated along the feature dimension. The first LSTM layer outputs the hidden state at each time step, while the last layer only outputs the hidden state at the last time step. The third layer is a Dropout layer, randomly discarding some neurons to prevent overfitting. The fourth layer is a fully connected layer, consisting of 1-2 layers with 32-128 neurons, fusing the high-level temporal features extracted by the LSTM. The fifth layer is the output layer, using a binary classification sigmoid function to output probability values ​​between 0 and 1, generating the final trip risk prediction result.

[0047] The filtered feature dataset is concatenated into a multi-dimensional feature vector after being aligned according to time windows based on time domain, frequency domain, and time-frequency domain features. This vector is then input into a deep learning model for training and prediction, outputting the probability of future device tripping risks. Deep learning models can automatically learn the complex nonlinear relationship between features and risks, exhibiting stronger representational capabilities compared to traditional methods. Commonly used models include Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and their hybrid architectures. These models are particularly suitable for processing device state data with time-dependent dependencies. This application preferably uses Long Short-Term Memory (LSTM) networks.

[0048] The computation process of a deep learning model involves first dividing the feature dataset into training, validation, and test sets in chronological order to ensure temporal independence. Then, a deep learning network architecture is constructed, such as using LSTM layers to capture temporal dependencies, fully connected layers for feature fusion, and a sigmoid function in the output layer to generate risk probabilities between 0 and 1. Next, the model parameters are optimized using backpropagation, and an early stopping strategy is employed to prevent overfitting, ultimately resulting in a predictive model capable of accurately assessing tripping risks. Typically, deep learning models are pre-trained before being deployed on the system management end of substation equipment. During the use of the deep learning model, parameters are continuously optimized and updated.

[0049] The Sigmoid function is an S-shaped curve function that maps any real number to the interval (0,1). In this implementation, the deep learning model (LSTM / GRU) ultimately outputs a real number z, and the algorithm's goal is to obtain the probability of tripping risk. The Sigmoid function can perform this data transformation, and its expression is as follows: in, Enter a real number; e is the natural constant, approximately equal to 2.71828; The output value is in the range (0,1).

[0050] Assuming the output of the last layer of the model is z=1.2, then: the tripping probability = That is, the model predicts that there is a 76.8% probability of a power outage occurring at some point in the future.

[0051] For example, a neural network model consisting of a two-layer LSTM and a three-layer fully connected layer was constructed, with the input being a sequence of feature data every 15 minutes over the past 72 hours. After training on three months of historical data, the model was able to predict the tripping risk of circuit breakers two hours in advance. In actual testing, it achieved early warning in 13 out of 15 real tripping events, with the false alarm rate controlled below 5%, and the risk probability output value showed a significant positive correlation with the actual fault severity.

[0052] S5: Generate an emergency response plan based on the trip risk prediction results and notify the operation and maintenance personnel of the emergency response plan.

[0053] Specifically, based on the trip risk prediction results output by the deep learning model, corresponding preventative maintenance or control instructions are generated, and these corrective actions are promptly transmitted to relevant maintenance personnel or the automated systems associated with them. The generation of corrective actions requires comprehensive consideration of multiple factors, including risk level, equipment importance, and operational feasibility, to form a tiered response decision-making strategy. The transmission process must ensure the timeliness, accuracy, and security of the instructions, achieving a complete closed loop from risk prediction to risk handling.

[0054] First, a mapping rule is established between risk levels and repair actions. For example, low risk (<0.3) only logs are recorded, medium risk (0.3-0.7) sends an early warning notification, and high risk (>0.7) triggers automatic adjustment or emergency response. Then, the generated instructions are encapsulated into specific message formats using standard communication protocols. Finally, the transmission path is selected according to the action type: control instructions are sent directly to the protection device, while maintenance work orders are pushed to the operation and maintenance management system and mobile terminals.

[0055] For example, when the model predicts that the tripping risk of a 110kV line circuit breaker reaches 0.82 (high risk), the system automatically generates a three-level response strategy: immediately sends a load transfer instruction to the adjacent circuit breaker via message, automatically transferring 30% of the load of the line to the standby line; at the same time, it sends a red alarm to the monitoring center, highlights the device on the operation interface; and pushes an emergency maintenance work order to the smart terminal of the inspection personnel in the jurisdiction, including specific inspection items such as "inspecting the fatigue condition of the energy storage spring of the mechanism", forming a complete risk handling closed loop.

[0056] This application covers the full range of equipment operation status through real-time acquisition of time-series data of power equipment. It adopts advanced denoising techniques such as wavelet transform and Kalman filtering, combined with methods such as linear interpolation to handle missing and outlier values, effectively improving data quality and laying a solid foundation for subsequent analysis. Features are extracted from multiple dimensions such as time domain, frequency domain, and time-frequency domain, including time domain feature statistics such as mean, variance, maximum value, minimum value, skewness, and kurtosis, as well as frequency features and power spectral density extracted by root mean square value and Fourier transform. The feature space is optimized by feature selection algorithms such as PCA and LASSO to screen out features that have a significant impact on fault prediction, thereby enhancing the predictive ability of the model. By employing a deep learning model, the system enables real-time interaction with equipment status data, generates trip prediction results, and outputs fault risk warnings (such as low risk and high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, thus significantly improving the timeliness and accuracy of fault handling.

[0057] In some implementations, step S1 includes: S11: Denoise time series data using Kalman filtering and wavelet transform.

[0058] S12: Perform linear interpolation on the denoised time series data to fill in the missing time series data and obtain a complete time series dataset.

[0059] Specifically, time-series data refers to a sequence of data collected at fixed intervals at consecutive time points, reflecting the operating status of power equipment, such as records of parameters like current, voltage, and temperature changing over time. Its characteristics include time-series dependency and correlations between consecutive data points.

[0060] Preprocessing refers to a series of cleaning, repairing, and enhancement operations performed on data before it enters the core analysis stage. The aim is to improve data quality and provide an accurate and reliable foundation for subsequent analysis.

[0061] Kalman filtering is an optimal estimation algorithm applicable to linear dynamic systems. It recursively estimates the system's state. Combining the system's dynamic model with actual observations, it effectively tracks and smooths the true signal value by minimizing the covariance of the estimation error in the presence of random noise. Its core advantage lies in its ability to process data online and in real time.

[0062] Wavelet transform is a time-frequency analysis method that decomposes a signal into wavelet basis functions at different scales and locations, providing localized information about the signal in both the time and frequency domains. It excels at handling non-stationary signals, effectively separating abrupt changes, transient components, and background noise.

[0063] Linear interpolation is a simple method for imputing missing values. It assumes that the data changes linearly between two known data points, and thus calculates the estimated value at any intermediate time based on the values ​​of the known points.

[0064] This implementation aims to transform raw, noisy, and potentially incomplete sensor data into clean, complete, high-quality time-series data suitable for advanced analysis. The process follows a "denoise first, then fill in the gaps" logic. First, for the complex noise components in the time-series data (including Gaussian white noise, periodic interference, and impulse noise), a serial denoising strategy combining Kalman filtering and wavelet transform is employed.

[0065] By combining Kalman filtering and wavelet transform, effective filtering of mixed noise (stationary noise and non-stationary transient noise) in the time-series data of power equipment was achieved, significantly improving the signal-to-noise ratio while better preserving weak transient features characterizing early faults (such as partial discharge pulses and mechanical shocks), which is difficult to achieve with a single denoising method. Subsequently, linear interpolation ensured the continuity of the data in the time dimension, providing a solid foundation for subsequent feature extraction and model training that rely on complete sequences. Overall, this step transforms the original low-quality "dirty data" into high-quality "clean data," which is the first and crucial step in improving the accuracy and reliability of the model in the entire prediction process.

[0066] For example, when monitoring the top oil temperature of the main transformer: The optimal estimated oil temperature at the previous moment, such as 75℃; The infrared thermometer reading at the current time t is given, such as 76.5℃ (including error); the observation matrix is ​​H, where H=1 since the temperature is directly measured; the Kalman gain is dynamically calculated based on process noise (natural fluctuations in oil temperature) and observation noise (temperature measurement error), assuming the calculated value is... =0.6.

[0067] The calculation yielded: - =76.5 75 = 1.5℃; Updated estimate =75 + 0.6 × 1.5 = 75.9℃.

[0068] This 75.9℃ is the optimal estimate of the oil temperature at the current moment. It neither relies excessively on the noisy observation value of 76.5℃, nor completely trusts the previous estimate of 75℃, but rather makes an intelligent trade-off.

[0069] In some implementations, an adaptive hybrid denoising system can be employed. By calculating the signal's stationarity index (such as the stability of the mean and variance within a sliding window) and kurtosis (which measures the sharpness of the signal distribution; high values ​​often indicate the presence of impulses) in real time, a dynamic decision is made as to whether the current data block is more suitable for Kalman filtering or wavelet denoising, or to determine the fusion weight of the two.

[0070] In some implementations, outlier detection and correction are performed before interpolation to prevent outliers from being mistakenly used as valid points for interpolation and thus polluting the dataset.

[0071] Specifically, the denoised time-series data is obtained, buffered through a sliding window, and then detected and corrected by an anomaly detection algorithm (based on the 3σ principle). Interpolation is then performed to supplement the missing time-series data, resulting in a complete time-series dataset.

[0072] Specifically, a sliding window refers to the window at the previous detection point. Take M points before and after it (excluding potentially missing time periods) to form the reference window data W, the expression is: Window length Typically, M is set to 10-30, corresponding to several minutes of data.

[0073] Calculate the mean of the data within the window : ; Calculate the standard deviation of the data within the window. : ; in, This represents the number of valid data points within the reference window W. For the reference window W, the first i Data points.

[0074] Anomaly detection rules: If Then Marked as "statistical anomaly candidate points". To prevent genuine gradual faults (such as slow temperature rises) from being misclassified as anomalies, it is necessary to simultaneously calculate... Rather than the previous point Difference: Δ = | - | If Δ is greater than a certain empirical threshold (such as the maximum normal rate of change of the sensor), it is judged as abnormal.

[0075] For example, during normal operation, the current of a 110kV line fluctuates within the range of 280±5A. The reference window W has μw=282A and σw=1.5A. Current point =0A (Sensor momentary disconnection).

[0076] calculate: =∣0 282 | = 282A > 3 × 1.5 = 4.5A.

[0077] At the same time, the difference Δ=∣0 283 | = 283A (the normal value at the previous moment was 283A), which is much greater than the normal current fluctuation rate. Therefore, this point can be marked as a sensor anomaly. The calculation is simple and fast, meeting the timeliness requirements of online real-time processing.

[0078] In some implementations, the Kalman filter is expressed as: in, It is the Kalman gain; H is the observation matrix, which maps the system state space to the observation space. For example, if we directly observe the state itself, H may be the identity matrix; if we observe a linear combination of states, H defines this combination relationship.

[0079] This represents the actual observation value obtained from the sensor at time t, which typically includes true state information and observation noise.

[0080] This indicates that at time t, based on the time up to t... The prior estimate of the system state is made based on all observations at time 1. It is derived from the system's state transition model. ( = From the optimal estimate of the previous moment It came as predicted.

[0081] It refers to the final posterior state estimate, which is the optimal estimate combining model predictions and current observation information, serving as the starting point for the filtered output and the prediction at the next time step.

[0082] For example, suppose we need to monitor the abnormal development of contact resistance in the high-voltage bushing joint of phase A of the main transformer. An infrared temperature sensor collects the bushing joint temperature every 5 minutes; the raw data contains a random temperature measurement error of ±2℃ and interference from environmental thermal radiation. When using Kalman filtering: State variable x t : The actual temperature observation value of the sleeve joint; Z t : Actual temperature value measured by infrared sensor; State transition matrix F t Set to 1 (assuming the temperature changes slowly); Observation matrix H: set to 1 (direct temperature measurement); Process noise covariance Q: A small value is set based on the heat conduction physical model; Observation noise covariance R: set according to sensor accuracy; After recursive calculation using Kalman filtering, the output temperature curve is smooth and stable, eliminating random fluctuations and accurately reflecting the temperature rise trend of the connector. When loose connectors lead to increased contact resistance, the filtered temperature curve will show a stable upward trend without being masked by noise.

[0083] The expression for wavelet transform is: in, It is the original time-series signal to be analyzed; The mother wavelet is an oscillation decay function with good locality in both the time and frequency domains (such as the Daubechies wavelet and the Morlet wavelet). It is the fundamental waveform for constructing all analytical wavelets.

[0084] a It is a scale parameter. a ≠0, scaling factor, controls the scaling of the wavelet. When | a When |>1, the wavelet is stretched and used to analyze the low-frequency components of the signal (large-scale features, such as trends); when 0< | a When |<1, wavelets are compressed and used to analyze the high-frequency components of the signal (small-scale features, such as details and abrupt changes).

[0085] It is a normalization factor that ensures the energy of wavelets remains constant at different scales.

[0086] This refers to the input timing signal to be analyzed.

[0087] b is the scale and shift factor, which controls the translation position of the wavelet on the time axis and is used to analyze the local characteristics of the signal at different time points. dt is the differential of the integral variable t.

[0088] For example, in the dissolved gas chromatography data of the same main transformer, there may be weak pulse characteristics caused by partial discharge. The total hydrocarbon content data sequence is collected once a day, but due to analytical errors and the intermittent nature of sampling, high-frequency noise exists in the data. A three-level decomposition is performed using the Db4 wavelet: Scale factor a: corresponds to different frequency bands; the wavelet coefficients of the first layer correspond to high-frequency noise. Shift factor b: Translation analysis along the time axis; Mother wavelet The wavelet basis function has good locality in both the time and frequency domains. In this embodiment, it is the Db4 wavelet, which is suitable for analyzing chromatographic data. By applying soft thresholding to the wavelet coefficients of the first layer to remove high-frequency noise, the signal is reconstructed. After processing, the small but continuous upward trend caused by partial discharge in the total hydrocarbon content sequence becomes more apparent, providing a clearer basis for judging insulation degradation.

[0089] The expression for linear interpolation is: Where t is the timestamp corresponding to the missing data point to be filled, satisfying... ≤t≤ It is the interpolation interval.

[0090] This refers to the timestamps corresponding to two known valid data points. .

[0091] It refers to time and The known data value at that location.

[0092] A scaling factor between 0 and 1, representing the distance of the missing point t from the starting point. The relative position.

[0093] When t= When t= When the factor is 1.

[0094] x(t) is the calculated interpolated estimate at time t. This formula geometrically represents the connection point ( ) and point ( The vertical coordinate value at the x-coordinate of the line t.

[0095] The main transformer's vibration monitoring system stopped collecting data for two hours during routine maintenance every Wednesday, resulting in missing data. For the missing vibration data point t (between known points...),... and between): Known The vibration acceleration at any given moment is 0.5 m / s². 2 ; Known The vibration acceleration at any given moment is 0.8 m / s². 2 ; Missing point t is located and The midpoint of the time was calculated to obtain: / s 2 This filling in of the gaps preserves the continuity of the vibration data and does not affect subsequent spectral analysis.

[0096] The Kalman filter formula clearly demonstrates how to achieve optimal recursive estimation of system state through dynamic tradeoff model prediction and real-time observation, making it particularly suitable for tracking trend changes in equipment operating parameters. The wavelet transform formula mathematically defines how to expand a signal into the time-frequency plane, providing a theoretical basis and computational tools for subsequent feature extraction from different scales (frequency) and time points. Its multi-resolution analysis characteristics are key to capturing transient features of equipment faults. While simple, the linear interpolation formula ensures the continuity of the data sequence along the time axis with minimal computational cost, avoiding analysis interruptions due to missing data. These three formulas together constitute the solid mathematical core of data preprocessing.

[0097] In some implementations, the timing data of the power equipment is the electrical parameters of the power equipment collected by current sensors, voltage sensors, temperature sensors, and vibration sensors.

[0098] Specifically, current sensors are devices used to measure the magnitude of current in a circuit, such as Hall effect current sensors and Rogowski coils. The timing data they collect reflects the equipment load, circuit integrity, and the presence of overcurrent or short circuits.

[0099] Voltage sensors are devices used to measure the potential difference (voltage) between two points in a circuit, such as voltage transformers (PTs). Their data reflects the system voltage level, stability, and the presence of overvoltage, undervoltage, or voltage distortion (harmonics). Temperature sensors are devices used to measure the local or ambient temperature of equipment, such as thermocouples, resistance temperature detectors (RTDs), and infrared thermometers. Their data directly reflects the equipment's heating status and is an important indicator for judging overload, poor contact, cooling system failure, and insulation aging. Vibration sensors are sensors used to measure the mechanical vibration acceleration, velocity, or displacement of equipment, such as piezoelectric accelerometers. Their data reflects the mechanical condition of the equipment, such as precursors to mechanical failures like loose transformer cores, stuck circuit breaker operating mechanisms, and worn bearings in rotating equipment. Electrical parameters broadly refer to a series of physical quantities describing the electrical performance and state of power systems and equipment. Besides current and voltage, they can also include, in a broader sense, power, frequency, power factor, harmonic content, insulation resistance, and partial discharge.

[0100] This application clarifies the physical source of data acquisition, forming the foundation of the perception layer for the entire prediction method. The implementation process begins with deploying the aforementioned four types of sensor networks at key monitoring points of the target power equipment (such as transformers, circuit breakers, GIS, etc.). Current sensors are typically installed in the primary circuit or secondary protection / measurement circuit of each phase, continuously acquiring current waveforms or RMS values ​​at a high sampling rate (e.g., thousands of times per second) to generate a current time-series data stream I(t). Voltage sensors are connected in parallel to the corresponding busbars or equipment terminals, synchronously acquiring voltage waveform data U(t), which, combined with current data, can be used to calculate other electrical quantities such as power and impedance. Temperature sensors are strategically placed in hotspot areas of the equipment, such as the top of transformer windings, circuit breaker contact shells, and cable joints, acquiring temperature data T(t) at minute or second intervals to monitor temperature rise trends. Vibration sensors are fixed to the equipment shell or foundation via magnetic bases or adhesive bonding, acquiring vibration signals A(t) generated during equipment operation or maintenance. Their data frequency range is wide, requiring a high sampling rate (e.g., above 10kHz) to capture high-frequency impact components. All these sensors are connected to the data acquisition unit via fieldbus or industrial Ethernet for synchronous or quasi-synchronous sampling, analog-to-digital conversion, and local buffering. The data is then uploaded to the monitoring master station or edge computing node via the station's communication network. At the monitoring master station, multiple time-series data streams from different sensors are aligned and packaged according to a unified timestamp, forming a device-centric multi-dimensional time-series dataset. For example, for a transformer monitoring point, at any time t, its data can be represented as a vector [ (t), , , , , , , , [,...], this vector is the original input for all subsequent analysis and processing.

[0101] The IEEE 1588 precision clock protocol is used to perform high-precision time synchronization of all sensors in the station, ensuring strict alignment of different physical quantity data on the time axis, which is crucial for analyzing the causal relationship between current, vibration, and temperature.

[0102] By embedding a simple algorithm at the acquisition end, the effective value range, rate of change, signal-to-noise ratio, etc. of each channel data are calculated in real time. This enables preliminary diagnosis and alarm for faults such as sensor disconnection, saturation, and drift, ensuring data reliability from the source.

[0103] In some implementations, step S4 includes: S41: Construct the feature dataset into a feature sequence arranged by time steps; Specifically, a feature sequence refers to a sequence formed by arranging multidimensional feature data, after feature extraction and selection, according to their acquisition time order. Each time step corresponds to a feature vector, and the entire sequence constitutes an L×D matrix, where L is the sequence length (number of time steps) and D is the feature dimension.

[0104] The multidimensional feature vectors extracted from each device and each time window are sorted according to the end time of their corresponding time windows, forming a feature sequence arranged by time step. For example, to predict the risk "in the next hour", the 288 feature vectors extracted in 5-minute windows over the past 24 hours (assuming the windows do not overlap) are arranged in chronological order into a sequence of length 288. This sequence is then input into a pre-trained LSTM or GRU model.

[0105] S42: Input the feature sequence into the pre-trained deep learning model; Among them, the deep learning model is either the Long Short-Term Memory (LSTM) network model or the Gated Recurrent Unit (GRU) network model; Specifically, a deep learning model refers to a machine learning model composed of multiple layers of nonlinear processing units, capable of automatically learning high-level feature representations from data. In this embodiment, it specifically refers to a variant of a recurrent neural network suitable for processing sequential data.

[0106] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that addresses the vanishing or exploding gradient problems faced by standard RNNs when processing long sequences by introducing "cell states" and "gating mechanisms" (input gate, forget gate, output gate), thus effectively capturing long-term dependencies in time series. The Gated Recurrent Unit (GRU) is a simplified variant of LSTM that combines the input and forget gates of LSTM into an "update gate," and merges the cell state and hidden state. GRUs have a simpler structure, fewer parameters, and faster training speed, achieving performance comparable to LSTM on many tasks. This application preferably uses LSTM as the deep learning model.

[0107] S43: Utilize a deep learning model to perform forward propagation calculations on the feature sequence, process the input at each time step sequentially, and update its internal hidden state; based on the hidden state finally output by the deep learning model, map to obtain the trip risk prediction result.

[0108] Forward propagation computation refers to the computational process where input data starts from the input layer of a neural network, undergoes weighted summation, nonlinear activation, and other operations layer by layer, and finally reaches the output layer to produce a prediction result. For recurrent neural networks, the computation at each time step depends on the hidden state of the previous time step. The hidden state is a state vector maintained internally by the recurrent neural network, which encapsulates the "memory" of all historical sequence information processed by the network up to the current time step. At each time step, the hidden state is updated based on the current input and the hidden state of the previous time step.

[0109] The feature sequence is input into a pre-trained LSTM or GRU model. The model begins forward propagation computation, starting from the first time step of the feature sequence. The model reads the feature vector at that time step, combines it with the internally initialized hidden state (usually a zero vector), and calculates a new hidden state. This new hidden state contains the information from the first time step. Next, the model moves to the second time step, reads the second feature vector, and combines it with the updated hidden state from the previous step to calculate and update the hidden state again. This time, the hidden state incorporates information from the first two time steps. This process iterates until the last time step of the sequence is processed. During each time step, the LSTM / GRU's "gating" mechanism determines which historical information needs to be retained, which needs to be forgotten, and how much of the current input needs to be added to memory, thus intelligently managing the flow of information. After all time steps of the sequence have been processed, the model's final output hidden state (i.e., the hidden state of the last time step) is considered to encode a comprehensive, high-level representation of the device's operating state during the time period represented by the entire input sequence. Finally, this final hidden state is fed into a fully connected output layer (usually a linear layer with a sigmoid activation function), which maps this high-dimensional state vector to a scalar value between 0 and 1. This value is the tripping risk probability predicted by the model, completing the mapping from time-series features to risk quantification.

[0110] Taking the tripping risk prediction of 220kV GIS equipment as an example: (1) Construction of characteristic sequences: For partial discharge risk prediction in GIS: Feature dimension D selects 20 key features, including: Ultra-high frequency signal amplitude, pulse repetition rate, SF6 gas pressure, trace moisture content, and decomposition product concentration. Shell vibration energy, acoustic emission signal intensity, three-phase current imbalance, and voltage harmonic distortion rate; Sequence length L: set to 72 (data from the past 6 hours, one point every 5 minutes). Prediction step size H: Set to 12 (for predicting the next hour). The constructed feature sequence is .

[0111] (2) Forward propagation via LSTM model: The model uses a two-layer LSTM, with a hidden unit count of... =64; ① Initial time step: Input the feature vector of the first time step (20-dimensional), LSTM calculates the hidden state. ② Gradual spread: Time step 2: Input , combined calculate ; Time step 3: Input , combined calculate ;Calculate the final hidden state until time step 72 is completed. .

[0112] During propagation, the forget gate of the LSTM determines how much historical discharge pattern memory is retained, the input gate determines the importance of the current discharge feature, and the output gate controls which information is used for risk assessment.

[0113] (3) Output the mapping instance through the Sigmoid function Final hidden state (128-dimensional) data is fed into the output layer. weight matrix ; bias It is a scalar; The calculation formula is: Where z is a real number representing the model's original score for tripping risk; T is the transpose sign; assuming the model calculates z = 1.586, then: Map the original score z to a probability value and activate it using the Sigmoid function: =0.83; This indicates that the probability of the GIS tripping due to partial discharge within the next hour is 83%.

[0114] If the GRU model is used, its simplified structure can also handle this sequence: The update gate determines how to combine new inputs with historical memories. Reset Gate: Determining How Much Historical Information to Ignore For the same GIS partial discharge sequence, GRU may converge faster and perform better with less training data. In practical applications, such as deploying this model on the GIS equipment of a converter station, it successfully predicted the risk of flashover on the insulator surface 48 hours before a fault occurred, with a prediction accuracy of 92%, thus avoiding DC blockage accidents caused by GIS faults. Partial discharge in GIS refers to a non-penetrating discharge phenomenon that occurs in a localized area of ​​the insulating medium within a gas-insulated sealed switchgear (GIS) due to uneven electric field distribution, insulation defects, or contaminants, causing the electric field strength to exceed its tolerance limit.

[0115] In some implementations, step S5 includes: S51: Obtain the fault type and risk level from the trip risk prediction results; S52: Map repair methods based on fault type and risk level to obtain emergency response plans; S53: Notify the maintenance personnel of the emergency response plan.

[0116] Among them, fault type refers to the specific category of equipment abnormality or defect predicted by the deep learning model that may lead to tripping, such as "transformer winding overheating", "circuit breaker mechanical jamming", "GIS partial discharge", "insulation moisture", etc. It is a qualitative description of the nature of the risk.

[0117] Risk level: A quantitative classification of the predicted probability of a power outage and / or the severity of its consequences, typically categorized as "low risk," "medium risk," "high risk," or more detailed levels (e.g., 1-5). The risk level determines the urgency and intensity of the emergency response.

[0118] Repair method mapping refers to the process of searching and matching the most appropriate repair, maintenance or operation adjustment measures from a predefined "fault-handling" knowledge base or rule base based on the fault type and risk level.

[0119] An emergency response plan is a set of specific and actionable instructions, which may include: recommended inspection items, equipment that needs to be isolated, switching operation procedures to be performed, recommended maintenance measures, safety precautions, and required spare parts, personnel and tools.

[0120] The generated emergency response plan will be promptly and accurately communicated to relevant maintenance personnel, repair teams, or dispatch centers via human-machine interface or communication system.

[0121] This implementation defines a crucial step from intelligent prediction to decision support, enabling the "implementation" of prediction results. The entire process begins with parsing the output of the deep learning model. The deep learning model not only outputs a risk probability value but may also output labels or probability distributions of fault types through multi-task learning or cascaded classifiers. Simultaneously, based on the risk probability value and thresholds set by expert experience (e.g., probability <0.3 for low risk, 0.3-0.7 for medium risk, >0.7 for high risk), or a more complex assessment matrix based on the severity of consequences, the current risk level is determined. Subsequently, the system activates the remediation method mapping engine. At the core of this engine is a contingency plan knowledge base, which stores best practice handling solutions for various faults throughout history, using "fault type" and "risk level" as primary keys. For example, when the mapping engine receives the input "Fault Type: Transformer Top Oil Temperature Too High" and "Risk Level: High Risk," it retrieves the corresponding emergency response plan from the knowledge base. This plan might include: "1. Immediately check the cooler's operating status and activate all standby coolers; 2. Remotely monitor the load current and recommend that the dispatcher reduce the load by 10% within 30 minutes; 3. Arrange for infrared thermometers to conduct precise on-site temperature measurement; 4. Prepare for transformer oil chromatography analysis." After generating the plan, the system notifies the operator through multiple channels: it displays a prominent pop-up window and audible / visual alarm on the substation's local monitoring system's human-machine interface; simultaneously, it automatically dispatches work orders containing plan details to the smartphones of relevant maintenance personnel via production management systems (such as PMS) or mobile applications; for extremely high risks, the system can also automatically trigger SMS or telephone notifications to the on-duty supervisor. The notification content is clearly structured, facilitating quick understanding and execution by maintenance personnel.

[0122] In some implementations, digital twin technology can be combined to generate and optimize dynamic emergency response plans. That is, the development of a fault and the effects of different handling measures are simulated in a virtual model. When the effect meets expectations, the emergency response plan is communicated to maintenance personnel, or a warning message is generated and pushed to maintenance system nodes to remind them. The notification system can identify the recipient's role (e.g., shift worker, maintenance foreman, technical expert) and current location, pushing information summaries most relevant to their responsibilities and the on-site situation. For example, specific operating procedures and safety measures are pushed to on-site personnel, while detailed fault characteristic analysis data is pushed to technical experts.

[0123] The system should provide an interface for operations and maintenance personnel to provide feedback on the actual measures taken, the time taken, and the final results after implementing the plan. This feedback information is valuable data for optimizing the model and knowledge base.

[0124] In some implementations, after the repair is completed, repair execution information is obtained from the operations and maintenance personnel; based on the repair execution information, the deep learning model is optimized using a weighted binary cross-entropy loss function.

[0125] Specifically, repair execution information refers to the real data about the handling process reported by the system after maintenance personnel receive an alert and execute the emergency response plan. This typically includes: the confirmed actual fault type (which may be consistent with or inconsistent with the predicted type), the actual final state of the equipment after handling (whether a trip or other consequences actually occurred), the timestamp of the handling, and the specific measures taken.

[0126] The weighted binary cross-entropy loss function is an improved form of the standard binary cross-entropy loss function, which aims to solve the adverse effects of a severe imbalance in the number of positive and negative samples in classification problems (such as tripping events being far fewer than normal events) on model training.

[0127] The standard binary cross-entropy loss function is expressed as follows: Where i is the weighting coefficient set for positive samples (tripping / fault); ∈{0,1} represents the true label (0=normal, 1=risk of tripping); ∈(0,1) is the tripping probability predicted by the model; N is the number of loss samples; log: natural logarithm, converts probability into a logarithmic space.

[0128] Weighted binary cross-entropy loss function To address class imbalance, weights are introduced for positive samples. w : .

[0129] Where w>1, the weights based on the sample size ratio (a common method) are calculated as follows: By adding weights, the "voice" of positive samples in the loss function becomes equivalent to that of negative samples. The penalty for each positive sample misclassified (false negative) is equivalent to the penalty for misclassifying w negative samples (false positive).

[0130] In some implementations, the weights are not fixed values, but are dynamically adjusted according to the training process, as expressed by: in, Let be the weights for the t-th training cycle; Basic weights (e.g., calculated based on the proportion of sample size); This is an adjustment factor (e.g., 0.1). The target recall rate is (e.g., 0.9). This represents the recall rate on the validation set during the current training cycle. When the recall rate falls below the target, the weights are automatically increased, forcing the model to focus more on positive samples. The full recall rate, also known as the detection rate, represents the percentage of actual faults successfully detected by the model and is one of the core metrics for evaluating the performance of binary classification models.

[0131] Weighted cross-entropy loss for predicted values The gradient is: When y=1 (positive sample), the gradient magnitude is amplified by a factor of w. This means that the model parameters will be adjusted more significantly based on the amplified gradient.

[0132] When a false negative occurs, if the model predicts that the error is normal ( If the value is close to 0, but it is actually a fault (y=1), then weight adjustment is performed. Since y=1, the gradient is calculated as follows: If it is very small (close to 0), then the gradient The absolute value is large (the negative value is large), and a negative gradient means that it is necessary to increase... After backpropagation, all parameters of the model will receive a signal that "the output probability needs to be increased," with the signal strength being w times that of the standard case. The model will learn that "although this pattern is rare, it is very important, and a higher probability should be output next time."

[0133] For example, early partial discharge signals in transformers; model prediction of tripping probability. =0.1 (low risk), but it actually developed into a power outage.

[0134] The gradient of the weighted loss (w=100) is then calculated as follows: (Magnified 100 times). The model parameters will be significantly adjusted to make similar discharge patterns more likely to be predicted in the future.

[0135] When a correct prediction occurs but the confidence level is insufficient, i.e. Too low. For example, the model correctly predicts a fault ( >0.5), but the confidence level is not high enough (e.g., =0.6), which may lead to insufficient attention from operations and maintenance. In this case, a correction mechanism will be activated: The gradient is calculated as follows: = ≈-1.67w, the result is still a negative gradient (needs to be increased). Furthermore, amplified by w, the model learns that "this pattern should not only be judged as positive, but also given a higher confidence level."

[0136] When the prediction is correct and the confidence level is high, the model only needs fine-tuning to maintain training stability.

[0137] This implementation method addresses a problem specific to power systems, specifically handling early warning scenarios characterized by "extreme class imbalance." Through an interpretable parameter system, weights w are assigned clear business meanings (the trade-off between the costs of missed and false alarms) and can be dynamically adjusted based on operational feedback, enabling continuous optimization.

[0138] The closed-loop optimization and self-learning mechanism of the model is the core of the system's continuous evolution and long-term effectiveness. The process begins after the maintenance personnel execute the emergency response plan. The system will proactively prompt or provide a convenient interface for maintenance personnel to provide feedback on the repair execution information. Key feedback items include at least: 1) the actual fault verification result, i.e., the actual fault type confirmed after on-site inspection (e.g., "confirmed as overheating of the B-phase bushing joint"), or confirmed as a "false alarm" (no fault); 2) the event result, i.e., whether the equipment ultimately tripped. This feedback information is correlated and packaged with the original time-series data and its characteristics that triggered the warning, forming a batch of new sample data with high-quality real labels. For example, a new sample contains the feature sequence of the past 24 hours. The corresponding real tags (0 indicates no tripping / false alarm, 1 indicates tripping / correct warning), and possible sub-categorization labels (fault type). After accumulating a certain number of new samples (e.g., dozens to hundreds), the system initiates the model optimization process. During optimization, a weighted binary cross-entropy loss function is used. Among them, the weight w is a key hyperparameter, and its value is usually set as the ratio of the number of negative samples to the number of positive samples in the dataset, i.e. This is because tripping events (positive samples) are extremely rare. Without weighting, the loss function would be dominated by a massive amount of normal samples (negative samples), causing the model to tend to predict "normal" scenarios indefinitely, becoming insensitive to genuine faults. Weighting amplifies the cost of misclassifying a positive sample (missed detection), forcing the model to pay more attention to rare fault modes during learning. This weighted loss function, along with new feedback data, allows for supervised fine-tuning of the original model. During fine-tuning, a smaller learning rate can be used, and only some layers of the model (such as the final fully connected layer) may be updated to prevent catastrophic forgetting (i.e., losing extensive previously learned knowledge) when new data is scarce. This process can be performed periodically, such as monthly, or automatically when a certain amount of feedback data accumulates.

[0139] Please see Figure 3As shown, in one embodiment, a device for predicting the tripping risk of power equipment is provided, the device comprising: The data acquisition module is used to acquire time-series data of the substation equipment, preprocess the time-series data, and obtain a preprocessed time-series dataset. The risk prediction module is used to extract time-domain features based on preprocessed time-series datasets using the mean, variance, maximum, minimum, skewness, and kurtosis methods; extract frequency features using the Fourier transform method; and extract time-frequency domain features using the short-time Fourier transform or wavelet transform method. Based on time-domain, frequency-domain, and time-frequency-domain features, feature selection is performed using the LASSO method to obtain a feature dataset. Input the feature dataset into the deep learning model to obtain the trip risk prediction result; Emergency response plan generation module: Generates emergency response plans based on trip risk prediction results and notifies maintenance personnel of the emergency response plans.

[0140] This application covers the full range of equipment operation status through real-time acquisition of time-series data of power equipment. It adopts advanced denoising techniques such as wavelet transform and Kalman filtering, combined with methods such as linear interpolation to handle missing and outlier values, effectively improving data quality and laying a solid foundation for subsequent analysis. Features are extracted from multiple dimensions such as time domain, frequency domain, and time-frequency domain, including time domain feature statistics such as mean, variance, maximum value, minimum value, skewness, and kurtosis, as well as frequency features and power spectral density extracted by root mean square value and Fourier transform. The feature space is optimized by feature selection algorithms such as PCA and LASSO to screen out features that have a significant impact on fault prediction, thereby enhancing the predictive ability of the model. By employing a deep learning model, the system enables real-time interaction with equipment status data, generates trip prediction results, and outputs fault risk warnings (such as low risk and high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, thus significantly improving the timeliness and accuracy of fault handling.

[0141] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps: Acquire time-series data of power equipment and preprocess the time-series data to obtain a preprocessed time-series dataset; Based on the preprocessed time series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-frequency domain features are extracted using the short-time Fourier transform or wavelet transform methods. The extracted features are filtered using principal component analysis or LASSO methods to obtain a feature dataset; The feature dataset is input into a pre-trained deep learning model to obtain the trip risk prediction result; An emergency response plan is generated based on the tripping risk prediction results, and the emergency response plan is then communicated to the operation and maintenance personnel.

[0142] By acquiring real-time time-series data of power equipment, the system covers all aspects of equipment operation. Advanced denoising techniques such as wavelet transform and Kalman filtering are used, combined with methods such as linear interpolation to handle missing and outlier values, effectively improving data quality and laying a solid foundation for subsequent analysis. Features are extracted from multiple dimensions, including time domain, frequency domain, and time-frequency domain, such as time domain features statistics such as mean, variance, maximum, minimum, skewness, and kurtosis, as well as frequency features and power spectral density extracted by root mean square value and Fourier transform. The feature space is optimized by principal component analysis and feature selection algorithms such as LASSO, and features that have a significant impact on fault prediction are selected, enhancing the predictive ability of the model. By employing a deep learning model, the system enables real-time interaction with equipment status data, generates trip prediction results, and outputs fault risk warnings (such as low risk and high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, thus significantly improving the timeliness and accuracy of fault handling.

[0143] In one embodiment, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, performs the following steps: Acquire time-series data of power equipment and preprocess the time-series data to obtain a preprocessed time-series dataset; Based on the preprocessed time series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-frequency domain features are extracted using the short-time Fourier transform or wavelet transform methods. The extracted features are filtered using principal component analysis or LASSO methods to obtain a feature dataset; The feature dataset is input into a pre-trained deep learning model to obtain the trip risk prediction result; An emergency response plan is generated based on the tripping risk prediction results, and the emergency response plan is then communicated to the operation and maintenance personnel.

[0144] By acquiring real-time time-series data of power equipment, the system covers all aspects of equipment operation. Advanced denoising techniques such as wavelet transform and Kalman filtering are used, combined with methods such as linear interpolation to handle missing and outlier values, effectively improving data quality and laying a solid foundation for subsequent analysis. Features are extracted from multiple dimensions, including time domain, frequency domain, and time-frequency domain, such as time domain features statistics such as mean, variance, maximum, minimum, skewness, and kurtosis, as well as frequency features and power spectral density extracted by root mean square value and Fourier transform. Feature selection algorithms such as PCA and LASSO are used to optimize the feature space and screen out features that have a significant impact on fault prediction, thereby enhancing the predictive ability of the model. By employing a deep learning model, the system enables real-time interaction with equipment status data, generates trip prediction results, and outputs fault risk warnings (such as low risk and high risk). For high-risk warnings, the system can trigger emergency measures or alarm notifications, allowing maintenance personnel to handle the situation promptly and avoid equipment failures or power system outages, thus significantly improving the timeliness and accuracy of fault handling.

[0145] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0147] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0148] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting the tripping risk of power equipment, characterized in that, The method includes: Acquire time-series data of the power equipment and preprocess the time-series data to obtain a preprocessed time-series dataset; Based on the preprocessed time series dataset, time-domain features are extracted using the mean, variance, maximum, minimum, skewness, and kurtosis methods; frequency features are extracted using the Fourier transform method; and time-frequency domain features are extracted using the short-time Fourier transform or wavelet transform methods. The extracted features are filtered using principal component analysis or LASSO methods to obtain a feature dataset; The feature dataset is input into a pre-trained deep learning model to obtain the trip risk prediction result; An emergency response plan is generated based on the tripping risk prediction results, and the emergency response plan is then communicated to the operation and maintenance personnel.

2. The method for predicting the tripping risk of substation equipment according to claim 1, characterized in that, The steps of acquiring time-series data of the power equipment and preprocessing the time-series data to obtain preprocessed time-series data include: The time series data is denoised using Kalman filtering and wavelet transform. Linear interpolation is performed on the denoised time-series data to fill in the missing time-series data and obtain a complete time-series dataset.

3. The method for predicting the tripping risk of substation equipment according to claim 2, characterized in that, The expression for Kalman filtering is: ； in, H is the Kalman gain, and H is the observation matrix. These are observed values. It is time-series data; The expression for the wavelet transform is: ； in, It is the original time-series signal to be analyzed. It is the mother wavelet. a It is a scale factor. a ≠0, b is the shift factor; It is a normalization factor; The expression for the linear interpolation is: ； Where t is the timestamp corresponding to the missing data point to be filled, satisfying... ≤t≤ It is the interpolation interval. This refers to the timestamps corresponding to two known valid data points. .

4. The method for predicting the tripping risk of substation equipment according to claim 1, characterized in that, The step of inputting the feature dataset into the deep learning model to obtain the trip risk prediction result includes: The feature dataset is constructed as a feature sequence arranged by time steps; The feature sequence is input into the pre-trained deep learning model; wherein the deep learning model is a Long Short-Term Memory (LSTM) network model or a Gated Recurrent Unit (GRU) network model. The deep learning model is used to perform forward propagation calculation on the feature sequence, processing the input at each time step in sequence and updating its internal hidden state. The tripping risk prediction result is obtained by mapping the hidden state output by the deep learning model.

5. The method for predicting the tripping risk of power equipment according to claim 1, characterized in that, The timing data of the power equipment consists of electrical parameters of the power equipment collected by current sensors, voltage sensors, temperature sensors, and vibration sensors.

6. The method for predicting the tripping risk of substation equipment according to claim 1, characterized in that, The steps of generating an emergency response plan based on the tripping risk prediction results and notifying the operation and maintenance personnel of the emergency response plan include: Obtain the fault type and risk level from the trip risk prediction results; Based on the fault type and risk level, a repair method is mapped to obtain an emergency response plan; The emergency response plan will be communicated to the operations and maintenance personnel.

7. The method for predicting the tripping risk of substation equipment according to claim 1, characterized in that, Also includes: Obtain repair execution information from the aforementioned maintenance personnel; Based on the repair execution information, the deep learning model is optimized using a weighted binary cross-entropy loss function.

8. A device for predicting the tripping risk of power equipment, characterized in that, The device includes: The data acquisition module is used to acquire time-series data of the substation equipment, preprocess the time-series data, and obtain a preprocessed time-series dataset. The risk prediction module is used to extract time-domain features based on the preprocessed time-series dataset by means, variance, maximum, minimum, skewness and kurtosis methods, extract frequency features by Fourier transform method, and extract time-domain-frequency domain features by short-time Fourier transform or wavelet transform method. Based on the time-domain, frequency-domain, and time-frequency-domain features, feature selection is performed using the LASSO method to obtain a feature dataset. The feature dataset is input into a deep learning model to obtain the trip risk prediction result; Emergency response plan generation module: Generates an emergency response plan based on the trip risk prediction results and notifies the operation and maintenance personnel of the emergency response plan.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the substation tripping risk prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the substation tripping risk prediction method as described in any one of claims 1 to 7.

Images