Method and system for predicting service life of power transformer

Abstract

The invention discloses a service life prediction method and system for a power transformer, and belongs to the technical field of transformer health monitoring, and the method comprises the steps: obtaining the concentration of gas dissolved in oil, the load rate, the operation temperature and the service life data of the power transformer according to the historical operation condition of the power transformer, inputting the data into a pre-trained prediction model, and outputting a service life prediction result. According to the method, various factors influencing the service life of the power transformer can be comprehensively captured, the parameters reflect the running state of the transformer from different angles, for example, the concentration of gas dissolved in oil can indicate internal faults, and the load rate and the running temperature are related to thermal aging and mechanical stress of the transformer. By combining the local feature extraction of the CNN and the time sequence modeling capability of the LSTM, the method can more accurately capture the dynamic change features in the operation process of the power transformer. The accurate capture of the dynamic time sequence characteristics enables the model to better adapt to the real-time change of the operation state of the transformer, thereby improving the accuracy of life prediction.

Description

Technical Field

[0001] This invention relates to the field of transformer health monitoring technology, and more specifically to a method and system for predicting the lifespan of power transformers. Background Technology

[0002] Electricity enters the building's power system terminal equipment from the municipal power grid. Transformers are irreplaceable intermediate core equipment. Transformers are responsible for converting high-voltage electricity into low-voltage electricity that can be directly used by the building's internal equipment. At the same time, they play a key role in voltage stabilization, power distribution, and safety isolation. The stable operation of transformers ensures the safety and reliability of the building's power supply.

[0003] Transformer lifespan prediction is a key technology for shifting transformer maintenance from passive repair to proactive operation and maintenance. It not only mitigates power outage risks and reduces maintenance costs, but also provides data support for the intelligent and sustainable operation of building energy systems, making it a crucial component of modern building power system upgrades. Therefore, transformer lifespan prediction is a core element in ensuring building power supply reliability, reducing operating costs, and achieving intelligent operation and maintenance, and is of great significance.

[0004] Traditional transformer life prediction methods based on convolutional neural networks (CNNs) can extract spatial features and capture spatial correlation features between multi-dimensional parameters. However, they are essentially static feature extraction models that ignore the time-series characteristics of the dynamic changes in transformer operating parameters over time. They cannot capture the long-term dependencies of equipment operating status and cannot reflect the gradual and cumulative effects of equipment aging, thus affecting the accuracy of transformer life prediction. Summary of the Invention

[0005] To address the problems existing in the above-mentioned fields, this invention proposes a method and system for predicting the lifespan of power transformers. The constructed prediction model accurately compensates for the inherent defects of existing single deep learning models in feature extraction, realizes the collaborative mining of spatial features and temporal dynamic features, and significantly improves the accuracy and reliability of transformer lifespan prediction.

[0006] To address the aforementioned technical problems, this invention discloses a method for predicting the lifespan of a power transformer, comprising the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a time-series feature sequence. The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to the time series, and aggregates them as input to the fully connected layer.

[0007] Preferably, the step of inputting the insulating medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters into the pre-trained prediction model and outputting the power transformer life prediction result specifically includes: The prediction model uses a CNN convolutional neural network as the base network, and adds a feature sequence reshaping module and an LSTM temporal modeling module after the pooling layer of the original CNN convolutional neural network. The original CNN convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The LSTM timing modeling module includes multiple LSTM units arranged in a timing sequence; The input layer maps dissolved gas concentration, load rate, operating temperature, and lifespan data in the oil into a feature matrix that matches the feature input format of the convolutional layer. This feature matrix is ​​then input into the convolutional layer, where the local features of the power transformer are extracted by sliding the convolutional kernel against the feature matrix. An activation function is then used to perform a nonlinear transformation on the local spatial features of the power transformer to obtain a nonlinearly transformed feature matrix. Finally, the pooling layer downsamples the nonlinearly transformed feature matrix to obtain a dimensionally reduced feature matrix. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix according to the time dimension through a sliding window to obtain a time-series feature sequence. The time-series feature sequences are input into the LSTM unit respectively. Through the collaborative effect of the gating mechanism, each LSTM unit autonomously learns the long-short time-series correlation in the time-series feature sequences to capture the dynamic time-series features of transformer operation data and generate multiple time-series corresponding real-time health feature vectors of the transformer. The fully connected layer aggregates the real-time health feature vectors of multiple time-series transformers into a one-dimensional vector; and uses the Softmax activation function to classify the one-dimensional vector, mapping it to a probability distribution of different lifespan intervals. The output layer filters the lifetime interval with the highest probability value based on the probability distribution of different lifetime intervals and a confidence threshold, and outputs the final lifetime prediction result.

[0008] Preferably, the step of inputting the time-series feature sequences into the LSTM units respectively, and through the collaborative effect of the gating mechanism in each LSTM unit, autonomously learning the long-short time-series correlation in the time-series feature sequences to capture the dynamic time-series features of transformer operation data, and generating multiple time-series corresponding real-time health feature vectors of the transformer, specifically includes: In each LSTM unit, the input at the current time step in the temporal feature sequence is determined as... ; The hidden state at time t-1 is determined as follows This indicates the real-time health characteristics of the transformer at the previous moment; The output of the forget gate is: (7) in, The retention value output by the forget gate; This is the weight matrix. For the corresponding bias vector, This is the hidden state from the previous moment. The input for each LSTM unit at the current time; The cell state updated through the input gate is: (10) in, This represents the updated cell state. This represents the memory state of the previous moment, indicating the long-term insulation degradation trend of the transformer; The degree of importance added to the current state is indicated by a value closer to 1, which means the current feature is more important. This represents the intermediate state, indicating the nonlinear fault trend of the transformer at time t. This indicates element-wise multiplication; The hidden state at the current moment is generated through the output gate. This serves as the real-time health feature vector of the transformer at the current moment: in, The output level value of the output gate is represented as an output coefficient between 0 and 1; It is the hyperbolic tangent function; The real-time health feature vectors of the transformers corresponding to each time series in different time series feature sequences are obtained.

[0009] Preferably, the acquisition of the insulation dielectric state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted specifically includes: The characteristic parameters of the insulating medium of a power transformer include the concentration of dissolved gases in the oil; Obtain operational load characteristic parameters, including load rate; The thermal characteristic parameters of the equipment include operating temperature; It also includes data preprocessing of the obtained dissolved gas concentration, load rate and operating temperature in the oil, and using the preprocessed data as input to the prediction model.

[0010] Preferably, it also includes a power transformer life prediction system, comprising: The data acquisition module is used to acquire the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted. The life prediction module is used to input the insulation medium state characteristic parameters, operating load characteristic parameters and equipment thermal state characteristic parameters into the pre-trained prediction model, and output the life prediction results of the power transformer. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a time-series feature sequence. The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to the time series, and aggregates them as input to the fully connected layer.

[0011] Preferably, the device further includes a computer apparatus, the computer apparatus comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a time-series feature sequence. The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to the time series, and aggregates them as input to the fully connected layer.

[0012] Preferably, the system further includes a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a time-series feature sequence. The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to the time series, and aggregates them as input to the fully connected layer.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The proposed method for predicting the lifespan of power transformers in this invention constructs a prediction model that extracts feature matrices of insulation medium state parameters, operating load parameters, and equipment thermal state parameters using a CNN convolutional neural network. A feature sequence reshaping module arranges the dimensionality-reduced feature matrices output from the pooling layer by time dimension through a sliding window to obtain a time-series feature sequence. An LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple time-series corresponding to real-time health feature vectors of the transformer, and aggregates them as input to the fully connected layer.

[0014] The prediction model uses a CNN as its base network, leveraging its powerful spatial feature extraction capabilities to fully explore the cross-correlation and spatial distribution patterns among the insulation medium state parameters, operating load parameters, and equipment thermal state parameters. Simultaneously, through the synergistic effect of the feature sequence reshaping module and the LSTM temporal modeling module, the static feature matrix output from the CNN pooling layer is transformed into a temporal feature sequence. Utilizing the LSTM's ability to capture long-term and short-term temporal dependencies, it accurately characterizes the dynamic evolution of transformer operating parameters over time and the gradual cumulative effect of equipment aging. This overcomes the shortcoming of traditional CNN models that ignore temporal characteristics, achieving full-dimensional coverage of both spatial and temporal features. The sliding window processing of the feature sequence reshaping module allows the dimensionality-reduced feature matrix to be arranged in an orderly manner along the time dimension, ensuring the integrity and coherence of the temporal feature sequence and laying the foundation for the LSTM module to efficiently capture dynamic temporal features. The multi-temporal real-time health feature vectors generated by the LSTM temporal modeling module are aggregated and input into the fully connected layer, integrating equipment health status information from different time points and accurately mapping the correspondence between equipment aging trends and lifespan intervals. Compared to existing single models, this method effectively avoids prediction bias caused by incomplete feature mining, and significantly improves the accuracy and stability of lifetime prediction results. Attached Figure Description

[0015] Figure 1 This is a flowchart of the power transformer life prediction method proposed in this invention; Figure 2 The network architecture diagram of CNN-LSTM deep learning constructed for this invention; Figure 3 A network architecture diagram of a traditional convolutional neural network (CNN) provided in an embodiment of the present invention; Figure 4 This is a network architecture diagram of a traditional Long Short-Term Memory (LSTM) network provided in an embodiment of the present invention; Figure 5 The model training curve provided in the embodiments of the present invention; Figure 6 A comparison chart of the model's predicted category and the true category provided in an embodiment of the present invention; Figure 7 Provided for embodiments of the present invention Figure 6 A schematic diagram of the confusion matrix of the corresponding test results. Detailed Implementation

[0016] The following will refer to the appendices in the embodiments of the present invention. Figures 1-7 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0017] Power transformers are core equipment in building power systems. A failure in a transformer can lead to a power outage, disrupting the entire power system and impacting residents' lives, business operations, and industrial production. Lifespan prediction allows for proactive maintenance or replacement, preventing power outages caused by transformer failures and ensuring a stable power supply. Transformer lifespan prediction primarily involves real-time monitoring of electrical parameters, oil temperature, and operating environment to identify potential risks and implement measures to eliminate them, ensuring electrical safety, reducing high maintenance costs due to equipment aging, and optimizing operation and maintenance costs.

[0018] This invention achieves transformer life prediction by constructing a life prediction model for variable-power transformers based on CNN-LSTM deep learning.

[0019] First, the training data is normalized to improve generalization performance. Then, data features are extracted using CNN and input into LSTM for classification to improve the overall network performance.

[0020] The steps of the CNN-LSTM algorithm are as follows: 1) After the input data is standardized by the input layer of the CNN, it is fed into the convolutional layer, and the local feature extraction is completed by taking advantage of the adaptive properties of the wide convolutional kernel.

[0021] 2) Input the local features extracted by the convolutional layer into the pooling layer for pooling operation, thereby reducing the data dimensionality while retaining key feature information.

[0022] 3) Subsequently, the dimensionality-reduced features are reshaped by the feature sequence reshaping module, arranged according to the time dimension, and transformed into time-series feature vectors that can be processed by the LSTM time-series modeling module. The time-series feature vectors are input into the LSTM unit, which learns the long-short-term time-series correlations in the time-series feature sequence to capture the dynamic time-series features of transformer operation data, generating multiple time-series corresponding to the real-time health feature vectors of the transformer.

[0023] 4) The Adam algorithm is used to backpropagate the training error and update the model parameters step by step according to the hierarchy.

[0024] 5) In the fully connected layer, the real-time health feature vectors of multiple time series transformers are aggregated into a one-dimensional vector, and the Softmax activation function is used to perform classification operations on the one-dimensional vector, thereby achieving the classification task of multi-feature input sequences and mapping them to the probability distribution of different lifetime intervals; the output layer selects the lifetime interval with the highest probability value based on the probability distribution of different lifetime intervals and outputs the final lifetime prediction result, thus achieving accurate model evaluation.

[0025] As an indispensable part of the power system, the lifespan prediction of power transformers is crucial for maintaining the stability of the power system.

[0026] Example To verify the feasibility of the transformer lifetime prediction method based on CNN-LSTM deep learning proposed in this invention, such as... Figure 1 The figure shows a method for predicting the lifespan of a power transformer proposed in this invention, which includes the following steps: S1: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; S2: Input the insulation medium state characteristic parameters, operating load characteristic parameters and equipment thermal state characteristic parameters into the pre-trained prediction model, and output the power transformer life prediction results; The prediction model uses a CNN (Convolutional Neural Network) as its base network, and adds a feature sequence reshaping module and an LSTM (Laser-Based Temporal Modeling) module after the pooling layers of the original CNN. Figure 2 As shown.

[0027] The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer by time dimension through a sliding window to obtain the time-series feature sequence. The feature matrix is ​​obtained by mapping the insulation medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to multiple time series, and aggregates them as input to the fully connected layer.

[0028] Specifically, in step S1, the data collection for predicting the lifespan of power transformers involves multiple factors, requiring comprehensive consideration of data such as the transformer's operating status, environmental impacts, internal faults, and external power grid fluctuations. Through scientific data collection and analysis, the lifespan of transformers can be extended, ensuring the stable operation of the power system.

[0029] The insulation medium state characteristic parameters of the power transformer obtained by this invention include the concentration of dissolved gas in the oil, the operating load characteristic parameters include the load rate, and the equipment thermal state characteristic parameters include the operating temperature.

[0030] Data is collected using sensors (temperature sensor, humidity sensor, current and voltage sensor, oil level and oil temperature sensor), online monitoring system, and periodic inspections. The collected data include transformer operation data (voltage, current, power, frequency, and load data), operating environment data (humidity, air quality, and vibration data), electrical insulation monitoring data (dielectric loss tangent (Tanδ), insulation resistance, and partial discharge), and oil quality monitoring data (oil temperature and level, dissolved gas analysis (DGA), oil acid value, viscosity, and dielectric strength), as shown in Table 1.

[0031] Table 1 Data Collection Based on the historical operating conditions of power transformers, oil quality monitoring data, electrical insulation monitoring data, and operating condition data are mapped to variables such as dissolved gas concentration in oil, load rate, and operating temperature.

[0032] It also includes data preprocessing of the obtained dissolved gas concentration, load rate and operating temperature in the oil, and using the preprocessed data as input to the prediction model.

[0033] The final dataset contains four core parameters: dissolved gas concentration in oil, load rate, operating temperature, and lifespan. This dataset provides crucial and effective data support for the training and validation of the subsequent lifespan prediction model, as shown in Table 2 below.

[0034] Table 2 Dataset Description Before building a predictive model, the obtained data needs to be cleaned to remove missing values ​​and redundant data.

[0035] Data standardization and normalization are performed. Standardization transforms the data into a distribution with a mean of 0 and a standard deviation of 1, accelerating the convergence speed of model training. The standardization formula is as follows: (1) in, This represents the mean of the sample data. The standard deviation (std) of the sample data.

[0036] Normalization scales the input data to the range [0,1], improving model convergence speed and ensuring better training and optimization. Min-max normalization is used to normalize the data; the normalization formula is as follows: (2) in, This is the data after normalization. The original data, The maximum value in the dataset. It is the minimum value in the dataset.

[0037] To ensure the accuracy and convergence speed of the model predictions, the dataset was divided into training, validation, and test sets in a 6:2:2 ratio for model training, tuning, and evaluation.

[0038] For spatial data within a single time step, standardization and format conversion are performed to adapt to the input requirements of CNN convolutional neural networks.

[0039] In step S2 of this invention, a power transformer life prediction method based on CNN-LSTM deep learning is proposed, which combines the advantages of CNN and LSTM to establish a transformer life prediction model based on CNN-LSTM deep learning.

[0040] Among them, CNN is a model widely used in deep learning. Its application in lifespan prediction requires that lifespan-related data be transformed into a feature input format that CNN can process. The core logic is to map multi-dimensional monitoring data into a feature matrix and then extract the correlation patterns between aging features and lifespan through CNN.

[0041] like Figure 3 As shown, the original CNN convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer; The input layer maps the dissolved gas concentration, loading rate, operating temperature, and lifetime data in the oil into a feature matrix that matches the feature input format of the convolutional layer; and normalizes the input data by scaling it to 0-1 or standardizing it to mean 0 and variance 1 to avoid the difference in numerical range affecting training.

[0042] The feature matrix is ​​input into the convolutional layer. Based on the adaptive properties of the convolution kernel, the local features of the power transformer are mined through sliding calculation between the convolution kernel and the feature matrix.

[0043] A convolutional layer is a layer that applies filters to a feature matrix to extract or detect its features. The filters are applied multiple times to the feature matrix and create a feature that helps classify the feature matrix. (3) in, x Given the input feature matrix, y These are local features of the output. The weights of the convolution kernel, b For bias vectors, i ,j The pixel coordinates represent the local features; that is, the current output is the pixel coordinate of the local feature. i line, number j Column elements; m , n This represents the internal coordinates of the convolution kernel, i.e., the coordinates of the first iteration of the convolution kernel. m line, number n Column weights; k This represents the size parameter of the convolution kernel. When the convolution kernel is... k × k When it is big, then m , n The range of values ​​for is [0, ...]. k -1].

[0044] Furthermore, by using an activation function, a nonlinear transformation is performed on the local spatial characteristics of the power transformer to obtain the feature matrix after the nonlinear transformation.

[0045] To enhance the expressive power of the model, nonlinearity is introduced, enabling CNNs to fit complex data distributions; otherwise, multi-layer convolution is equivalent to a single-layer linear transformation.

[0046] Commonly used activation functions (Rectified Linear Unit, ReLU) are as follows: (4) Pooling layers downsample the feature matrix after nonlinear transformation to obtain a dimensionality-reduced feature matrix, reducing the number of data parameters, improving feature invariance, and preserving key feature information to obtain the dimensionality-reduced features.

[0047] Common pooling methods include Lp pooling, random pooling, and hybrid pooling, among which: Lp pooling Lp pooling is a type of operation that aggregates regions based on statistical metrics. Its general form is as follows: (5) in, d The channel vector represents the current processing channel number. d One feature channel; K The size parameter of the pooling window; A Input feature map; s For pooling stride; p Let Lp be the order of pooling; when p When = 1, it is mean pooling. p As we approach infinity, we use maximum pooling. p When the value is 2, it is L2 pooling.

[0048] Random pooling and hybrid pooling Random pooling involves selecting values ​​from a region with probability, retaining non-maximum values ​​for activation; Hybrid pooling is a linear combination of maximum pooling and mean pooling: (6) when When =1, it degenerates into pure maximum pooling; When = 0, it degenerates into pure mean pooling; 0 < When <1, it represents a linear fusion of the two pooling methods.

[0049] The feature sequence reshaping module arranges the dimensionality-reduced feature matrix according to the time dimension through a sliding window to obtain a temporal feature sequence, which serves as the input to the LSTM temporal modeling module.

[0050] LSTM is a time-recurrent neural network, a deep learning model commonly used to process sequential data. It introduces three gates: the input gate, the forget gate, and the output gate. It also proposes the concept of a cell based on RNN, using cells to replace the hidden neurons in the classic RNN.

[0051] like Figure 4 As shown, the LSTM time series modeling module includes multiple LSTM units arranged in time sequence. Each LSTM unit learns the long-short time series correlation in the time series feature vector through the synergistic effect of the forget gate, input gate, and output gate, so as to capture the dynamic time series features of transformer operation data and generate multiple time series corresponding to the real-time health feature vector of the transformer.

[0052] The fully connected layer aggregates the real-time health feature vectors of transformers from multiple time series into a one-dimensional vector; and uses the Softmax activation function to classify the one-dimensional vector, mapping it to the probability distribution of different lifespan intervals. The output layer selects the lifetime interval with the highest probability value based on the probability distribution of different lifetime intervals and a confidence threshold, and outputs the final lifetime prediction result.

[0053] Each step of the computation in each LSTM unit revolves around updating the cell state (long-term degradation trend) and generating the hidden state (current health characteristics), and there is a one-to-one correspondence between the cell state and the transformer's time-series characteristics (dissolved gases in oil, partial discharge, and time-series fluctuations in temperature / voltage / current). The updating and generation of the hidden state unfolds as follows: assuming the... t The time-series feature sequence output by the LSTM layer input feature sequence reshaping module at each time step corresponds to the transformer. t Local features at time t, then the input representation The hidden state at time (t-1) corresponds to the real-time health characteristics of the transformer at the previous time step. The hidden state is represented as... The core formula is as follows:

[0054] The forget gate determines which information is "forgotten" from the cell state. It is activated by a sigmoid function with a value between 0 and 1 (1 represents complete retention, 0 represents complete forgetting). This determines which information from the previous cell state is forgotten. Partial information, namely the cell state at time (t-1), corresponds to the long-term insulation degradation trend of the transformer:

[0055] (7) in, The retention value determines the long-term insulation degradation trend of the transformer. Retention ratio; during early signs of failure Discard irrelevant normal fluctuations; when a fault occurs Focusing on the current fault characteristics. This is the weight matrix. For the corresponding bias vector, The input from the previous moment, This is the input at the current moment. During normal operation, No abnormalities. ≈1, fully retaining the long-term, gradual aging trend; the early signs of failure (such as a slow increase in gas concentration). Discarding some irrelevant fluctuations leaves room for subsequent updates of abnormal features.

[0056] The input gate is responsible for filtering and updating new information to the cell state. It consists of two steps: first, the sigmoid function is used to determine which new information is "admitted", then the tanh function is used to generate candidate information to be updated (range -1-1), and finally the two are combined and updated to the cell state proportionally.

[0057] Screening and admission information: (8) in, This adds a degree value to the current state; the closer it is to 1, the more important the current feature is. This is the weight matrix. This is the corresponding bias vector. When When a sudden change in partial discharge occurs or the C2H2 concentration in the oil exceeds the standard, This abnormal feature is incorporated into the cell state; when When it is a normal fluctuation, To avoid redundant interference;

[0058] Generate candidate information: (9) in, The intermediate state, i.e., the candidate feature vector between [-1-1], corresponds to the nonlinear fault trend of the transformer at time t, during the normal operation phase. Gradual value selection; during the later stages of insulation aging The appearance of significant peak values ​​corresponds to fault characteristics such as a sudden increase in partial discharge and nonlinear growth in gas concentration. This is the weight matrix. For the corresponding bias vector, It is a hyperbolic tangent function. In the later stages of transformer insulation aging, The partial discharge level increased sharply from 100 pC to 1000 pC. Corresponding to the nonlinear characteristics of the mutation, and the smoothness of the normal phase Significant differences were formed.

[0059] Update cell status: (10) in, The updated cell state dynamically integrates the long-term degradation trend of the transformer with the current fault characteristics during normal operation. Preserve the long-term aging trend; when a failure occurs The update shows a fault-driven degradation trend; post-fault recovery... It integrates the remaining aging trend with the post-failure degradation characteristics. This refers to the memory state of the previous moment. This represents element-wise multiplication, retaining the parts to be forgotten, and adding new candidate information for admission. Normal operation phase: , , Preserve the long-term aging trend; when a failure occurs: (Discarding some normal trends) (Strengthen current fault characteristics) Updated to "Fault-driven degradation trend"; After a fault occurs: , , Preserve the decay or recovery trend after the failure.

[0060] The output gate controls which information from the cell state is output to the current hidden state (for subsequent calculations or predictions). It filters partial information from the cell state using the Sigmoid function, then compresses the range using the tanh function before outputting the final information. This will produce an output.

[0061] Filtered information: (11) in, The output level value, represented as an output coefficient between 0 and 1, determines the output proportion of features in the cell state during normal operation. Outputs stable and healthy characteristics; when a fault occurs Output the core characteristics of the fault. This is the weight matrix. This is the corresponding bias vector.

[0062] Generate the current hidden state for prediction or propagation to the next step: (12) in, The current output corresponds to the transformer's real-time health feature vector, which is directly input into the fully connected layer for fault type classification. During normal operation... Numerical stability corresponds to stable health characteristics; the early stage of a fault. The appearance of minute fluctuations corresponds to precursory characteristics; when a fault occurs, The appearance of significant peaks or abrupt changes corresponds to the core characteristics of the fault; after the fault, The decline gradually corresponds to the fault decay characteristic.

[0063] After the dynamic time-series features of the captured transformer operation data are aggregated into a one-dimensional vector in the fully connected layer and mapped to the target output space, the Softmax activation function is used to classify the one-dimensional vector and map it into the probability distribution of different lifespan intervals. The output layer outputs the final prediction result for each sample based on the probability distribution of different lifetime intervals.

[0064] During model training, this invention employs the Adam algorithm to backpropagate the training error, updating the model's parameters hierarchically, such as... Figure 5 As shown, the training curves include the training set accuracy iteration curve and the loss function curve. In the training set accuracy iteration curve, the accuracy gradually increases with the increase of the number of iterations, eventually approaching 100% and stabilizing. In the loss function curve, the loss gradually decreases and stabilizes with the increase of the number of iterations, reflecting that the model convergence tends to stabilize.

[0065] This invention also proposes a life prediction system for power transformers, comprising: The data acquisition module is used to acquire the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted. The life prediction module is used to input the insulation medium state characteristic parameters, operating load characteristic parameters and equipment thermal state characteristic parameters into the pre-trained prediction model, and output the life prediction results of the power transformer. The prediction model uses a CNN convolutional neural network as the base network, and adds a feature sequence reshaping module and an LSTM temporal modeling module after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer by time dimension through a sliding window to obtain the time-series feature sequence. The feature matrix is ​​obtained by mapping the insulation medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters. The LSTM time-series modeling module captures the dynamic time-series features of transformer operating data in the time-series feature sequence, generates multiple real-time health feature vectors of transformers corresponding to multiple time series, and aggregates them as input to the fully connected layer.

[0066] In this system, the data acquisition module mainly monitors the transformer's operating status in real time through sensors, collects data, and processes, cleans, and denoises the collected data to ensure data quality; the life prediction module uses deep learning algorithms to perform in-depth analysis of the extracted data features and integrates transformer operating data, environmental data, and historical fault information; the result output layer and decision layer realize dynamic prediction of transformer life and accurate assessment of remaining life (RUL).

[0067] The proposed power transformer life prediction method, through the construction of a prediction model, can take into account both the spatial correlation and temporal dynamics of multi-dimensional parameters, resulting in more comprehensive feature extraction. This method uses a CNN as the basic network, leveraging its powerful spatial feature extraction capabilities to fully explore the cross-correlation and spatial distribution patterns among the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters, thus solving the problem of weak spatial feature extraction in traditional LSTM models. Simultaneously, through the synergistic effect of the feature sequence reshaping module and the LSTM temporal modeling module, the static feature matrix output from the CNN pooling layer is transformed into a temporal feature sequence. Utilizing the LSTM's ability to capture long- and short-term temporal dependencies, it accurately characterizes the dynamic laws of transformer operating parameters evolving over time and the gradual cumulative effect of equipment aging, overcoming the shortcoming of traditional CNN models that ignore temporal characteristics, and achieving full-dimensional coverage of both spatial and temporal features.

[0068] By employing a sliding window processing method in the feature sequence reshaping module, the dimensionality-reduced feature matrix can be arranged in an orderly manner along the time dimension, ensuring the integrity and coherence of the temporal feature sequence. This lays the foundation for the LSTM module to efficiently capture dynamic temporal features. The multi-temporal real-time health feature vectors generated by the LSTM temporal modeling module are aggregated and input into the fully connected layer, integrating equipment health status information at different time points and accurately mapping the correspondence between equipment aging trends and lifespan intervals. Compared to existing single models, this method effectively avoids prediction bias caused by incomplete feature mining, significantly improving the accuracy and stability of lifespan prediction results. It can provide accurate data for transformer operation and maintenance planning, spare parts reserves, and maintenance arrangements, effectively reducing the risk of unplanned power outages and operation and maintenance costs.

[0069] This method is designed specifically for the operating scenarios and core parameters of building power transformers. It can accurately adapt to the working characteristics of transformer voltage regulation and power distribution in building power systems. The extracted feature parameters are highly consistent with the actual operating status of the equipment, and the prediction results can directly serve the intelligent operation and maintenance of building energy systems. At the same time, the model architecture has good scalability. The sliding window size, CNN convolution kernel parameters, and LSTM unit number can be flexibly adjusted according to different types of transformers and different operating conditions without requiring significant modifications to the core structure of the model. It has a wide range of applications and can meet the actual needs of the digital and sustainable upgrading of modern building power systems.

[0070] This method, through high-precision lifespan prediction, helps transformer operation and maintenance transform from a traditional passive maintenance model to a proactive predictive maintenance model. It can predict potential equipment failure risks and lifespan in advance, providing support for the optimal allocation of operation and maintenance resources and the efficient scheduling of energy systems. This not only effectively extends transformer lifespan and reduces operation and maintenance costs and resource consumption, but also ensures the continuous and stable operation of building power supply systems, improves the safety, reliability, and energy efficiency of building power systems, and provides core technological support for the intelligent upgrading of modern building power systems.

[0071] like Figure 6 As shown, this is a comparison chart of the predicted and true categories output by the CNN-LSTM deep learning-based power transformer life prediction model of this invention. The horizontal axis represents the sample number, with a total of 16 samples, as shown below. Figure 6 The numbers 1-16 are shown in the diagram; the vertical axis represents the category labels, including three categories: 1, 2, and 3; in the data points, blue represents the "true category" and red represents the "predicted category".

[0072] Category 1 (vertical axis = 1): The true category of all samples is completely consistent with the predicted category. That is, when the blue dot and the red dot coincide, the model's recognition accuracy for this category is 100%.

[0073] Category 2 (vertical axis = 2): The true category of all samples is exactly the same as the predicted category, and the model's recognition accuracy for this category is 100%.

[0074] Category 3 (vertical axis = 3) has a prediction error (samples with the true category 3 are predicted as category 2), indicating that the model's ability to distinguish category 3 is insufficient. It is necessary to increase the sample size of category 3, continuously iterate and optimize the model parameters, and further improve the model's prediction accuracy.

[0075] like Figure 7 As shown, Figure 6 The confusion matrix diagram of the corresponding test results shows that the model has a strong ability to identify categories 1 and 2, and a high "prediction accuracy" (precision) for category 3.

[0076] The prediction results show that the deep learning-based power transformer life prediction method proposed in this invention can effectively improve the accuracy of model prediction and verify that the algorithm has certain feasibility in power system applications.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0078] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for predicting the lifespan of a power transformer, characterized in that, Includes the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a temporal feature sequence; The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters; the LSTM time series modeling module captures the dynamic time series features of transformer operating data in the time series feature sequence, generates multiple time series corresponding to the real-time health feature vectors of the transformer, and aggregates them as the input of the fully connected layer.

2. The method for predicting the lifespan of a power transformer according to claim 1, characterized in that, The process of inputting insulating medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters into a pre-trained prediction model and outputting a power transformer life prediction result specifically includes: The prediction model uses a CNN convolutional neural network as the base network, and adds a feature sequence reshaping module and an LSTM temporal modeling module after the pooling layer of the original CNN convolutional neural network. The original CNN convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The LSTM timing modeling module includes multiple LSTM units arranged in a timing sequence; The input layer maps dissolved gas concentration, load rate, operating temperature, and lifespan data in the oil into a feature matrix that matches the feature input format of the convolutional layer. This feature matrix is ​​then input into the convolutional layer, where the local features of the power transformer are extracted by sliding the convolutional kernel against the feature matrix. An activation function is then used to perform a nonlinear transformation on the local spatial features of the power transformer to obtain a nonlinearly transformed feature matrix. Finally, the pooling layer downsamples the nonlinearly transformed feature matrix to obtain a dimensionally reduced feature matrix. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix according to the time dimension through a sliding window to obtain a time-series feature sequence. The time-series feature sequences are input into the LSTM unit respectively. Through the collaborative effect of the gating mechanism, each LSTM unit autonomously learns the long-short time-series correlation in the time-series feature sequences to capture the dynamic time-series features of transformer operation data and generate multiple time-series corresponding real-time health feature vectors of the transformer. The fully connected layer aggregates the real-time health feature vectors of multiple time-series transformers into a one-dimensional vector; and uses the Softmax activation function to classify the one-dimensional vector, mapping it to a probability distribution of different lifespan intervals. The output layer filters the lifetime interval with the highest probability value based on the probability distribution of different lifetime intervals and a confidence threshold, and outputs the final lifetime prediction result.

3. The method for predicting the lifespan of a power transformer according to claim 2, characterized in that, The process involves inputting the time-series feature sequences into the LSTM units, where each LSTM unit, through the collaborative effect of a gating mechanism, autonomously learns the long-short time-series correlations within the time-series feature sequences to capture the dynamic time-series features of transformer operation data, generating multiple time-series corresponding to the transformer's real-time health feature vectors. Specifically, this includes: In each LSTM unit, the input at the current time step in the temporal feature sequence is determined as... ; The hidden state at time t-1 is determined as follows This indicates the real-time health characteristics of the transformer at the previous moment; The output of the forget gate is: （7） in, The retention value output by the forget gate; This is the weight matrix. For the corresponding bias vector, This is the hidden state from the previous moment. The input for each LSTM unit at the current time; The cell state updated through the input gate is: （10） in, This represents the updated cell state. This represents the memory state of the previous moment, indicating the long-term insulation degradation trend of the transformer; The degree of importance added to the current state is indicated by a value closer to 1, which means the current feature is more important. This represents the intermediate state, indicating the nonlinear fault trend of the transformer at time t. This indicates element-wise multiplication; The hidden state at the current moment is generated through the output gate. This serves as the real-time health feature vector of the transformer at the current moment: in, The output level value of the output gate is represented as an output coefficient between 0 and 1; It is the hyperbolic tangent function; The real-time health feature vectors of the transformers corresponding to each time series in different time series feature sequences are obtained.

4. The method for predicting the lifespan of a power transformer according to claim 1, characterized in that, The acquisition of the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted specifically includes: The characteristic parameters of the insulating medium of a power transformer include the concentration of dissolved gases in the oil; Obtain operational load characteristic parameters, including load rate; The thermal characteristic parameters of the equipment include operating temperature; It also includes data preprocessing of the obtained dissolved gas concentration, load rate and operating temperature in the oil, and using the preprocessed data as input to the prediction model.

5. A life prediction system for power transformers, characterized in that, include: The data acquisition module is used to acquire the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted. The life prediction module is used to input the insulation medium state characteristic parameters, operating load characteristic parameters and equipment thermal state characteristic parameters into the pre-trained prediction model, and output the life prediction results of the power transformer. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a temporal feature sequence; The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters; the LSTM time series modeling module captures the dynamic time series features of transformer operating data in the time series feature sequence, generates multiple time series corresponding to the real-time health feature vectors of the transformer, and aggregates them as the input of the fully connected layer.

6. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a temporal feature sequence; The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters; the LSTM time series modeling module captures the dynamic time series features of transformer operating data in the time series feature sequence, generates multiple time series corresponding to the real-time health feature vectors of the transformer, and aggregates them as the input of the fully connected layer.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the following steps: Obtain the insulation medium state characteristic parameters, operating load characteristic parameters, and equipment thermal state characteristic parameters of the power transformer to be predicted; The insulation medium condition characteristics, operating load characteristics, and equipment thermal condition characteristics are input into the pre-trained prediction model, and the life prediction results of the power transformer are output. The prediction model is based on a CNN convolutional neural network, and a feature sequence reshaping module and an LSTM temporal modeling module are added after the pooling layer of the original CNN convolutional neural network. The feature sequence reshaping module arranges the dimensionality-reduced feature matrix output by the pooling layer according to the time dimension through a sliding window to obtain a temporal feature sequence; The feature matrix is ​​obtained by mapping the insulating medium state feature parameters, operating load feature parameters, and equipment thermal state feature parameters; the LSTM time series modeling module captures the dynamic time series features of transformer operating data in the time series feature sequence, generates multiple time series corresponding to the real-time health feature vectors of the transformer, and aggregates them as the input of the fully connected layer.

Images