A transformer bushing degradation trend prediction and health assessment method based on a time sequence neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN DERUILONG OPTOELECTRONICS TECHNOLOGY CO LTD
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-19
Smart Images
Figure CN122242863A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment condition monitoring and intelligent operation and maintenance technology, and in particular to a method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network. Background Technology
[0002] The function of transformer bushings is to achieve insulation of high-voltage conductors to ground and energy transmission. They directly determine whether the transformer and even the entire power grid can operate safely and stably. In actual operation, due to the long-term effects of electric field force, thermal stress, and external environment, they are subjected to long-term pressure, temperature, and external environmental damage. Prolonged operation can easily lead to insulation aging, excessive partial discharge, and increased dielectric loss. Once the transformer bushing is damaged, serious faults can easily occur, causing large-scale power outages and resulting in significant economic losses.
[0003] While current power systems have implemented online monitoring methods for bushings, including dielectric loss factor monitoring, leakage current monitoring, partial discharge detection, and dissolved gas analysis in oil, which can provide a basic understanding of the bushing's condition, there is still a significant gap between these methods and the actual goal.
[0004] Existing methods often rely on a single indicator threshold or empirical rules for judgment, which cannot fully reflect the evolutionary characteristics of the entire casing deterioration process. It is difficult to identify the gradual and hidden deterioration of the casing. Current methods are mostly instantaneous or short-term state analyses, without considering past operating conditions, changes in working conditions, changes in external factors, and other influencing factors. Therefore, the prediction of health status and remaining life is not very accurate.
[0005] With the continuous expansion of the power grid and the increase in the service life of equipment, traditional periodic maintenance or post-maintenance can no longer meet the power system's requirements for safety, economy and reliability. There is an urgent need to analyze and predict the trend of bushing deterioration based on long-term data, and use this as a quantitative basis for maintenance decisions.
[0006] In recent years, time series analysis and prediction technology based on deep learning has developed rapidly. However, in a large number of field applications, it is mostly used for equipment fault identification or condition classification. Its long-term time series prediction function has not been fully utilized to explore the continuous evolution law of the bushing deterioration process. There is a lack of technical means to incorporate working environment and load factors, making it difficult to meet the needs of health management and maintenance decision-making in actual engineering.
[0007] To address this, a method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network is proposed. Summary of the Invention
[0008] To overcome the shortcomings of existing technologies, this invention provides a method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network.
[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network, comprising: acquiring multi-source monitoring data of the transformer bushings; preprocessing the multi-source monitoring data and constructing a time-series input sequence to characterize the state evolution process of the transformer bushings; inputting the time-series input sequence into a time-series neural network model to obtain the deterioration trend prediction result of the transformer bushings; and determining the health assessment result of the transformer bushings based on the deterioration trend prediction result.
[0010] As a preferred technical solution of the present invention, the multi-source monitoring data includes one or more of online monitoring data, operating condition data, environmental data, and historical status data; the online monitoring data includes one or more of dielectric loss, leakage current, partial discharge characteristic quantity, and dissolved gas in oil; the operating condition data includes one or more of load current, operating temperature, and voltage; the environmental data includes one or more of ambient temperature and ambient humidity; and the historical status data includes one or more of historical health status and maintenance records.
[0011] As a preferred technical solution of the present invention, the multi-source monitoring data is preprocessed, including one or more of the following: time alignment, outlier handling, missing value filling, and normalization processing of data from different sources; the construction of the time series input sequence includes: organizing the preprocessed multi-source monitoring data into multi-dimensional time series samples with operating condition labels according to a preset time granularity; the operating condition labels include one or more of the following: load level, operating temperature, ambient temperature, and ambient humidity.
[0012] As a preferred technical solution of the present invention, the temporal neural network model includes one of a temporal convolutional network, a long short-term memory network, and a fusion model thereof; the deterioration trend prediction result includes the trend of changes in the state parameters of the transformer bushing within a future preset time range.
[0013] As a preferred embodiment of the present invention, the determination of the health assessment result includes: calculating a health index based on the deterioration trend prediction result; determining the health status of the transformer bushing based on the health index or the trend of health index changes; the calculation of the health index includes weighted fusion of multiple features characterizing the transformer bushing status; the health assessment result also includes the remaining life assessment result; the remaining life assessment result is estimated by performing trend analysis on the health index sequence and combining it with a preset failure threshold; after the deterioration trend prediction and health assessment, one or more of the following are output: health index, deterioration trend prediction result, remaining life assessment result, early warning information, and maintenance suggestions.
[0014] Compared with the prior art, the beneficial effects that this invention can achieve are: This application constructs a multi-dimensional time-series input sequence with operating condition labels by performing time alignment and unified processing on multi-source monitoring data such as dielectric loss, leakage current, partial discharge characteristics, operating load parameters, and ambient temperature and humidity. It then uses a time-series neural network to jointly model historical and current data, thereby enabling the prediction of future state changes in the bushing. As a result, this application can effectively identify the gradual and hidden degradation of the bushing, improve the ability to detect early risks, and realize the transformation from instantaneous state judgment to continuous degradation trend analysis.
[0015] In constructing the time-series input sequence, this application incorporates operating condition information such as load level, operating temperature, ambient temperature, and ambient humidity as tag information, inputting them along with monitoring features into the time-series neural network model. This enables the model to learn the differences in the casing degradation process under different operating conditions. In this way, this application can not only reflect the changes in the casing's physical parameters but also demonstrate the impact of the operating environment and load conditions on the casing degradation rate. Consequently, the obtained health index and remaining life level are more consistent with the actual field conditions, improving the reliability and stability of the prediction results in engineering applications.
[0016] After obtaining the future state prediction results, this application further calculates the health index based on the predicted state parameters, and classifies the remaining life of the bushing according to the change law of the health index and its correspondence with historical failure samples. By outputting the health index, deterioration trend, remaining life level and early warning information, this application can expand the equipment status from judging whether it is abnormal to the quantitative analysis level of how the health level deteriorates and how much remaining life is left. This provides a direct basis for the formulation of condition-based maintenance plans, adjustment of monitoring frequency, optimization of maintenance resource allocation and risk early warning, and significantly improves the engineering practicality of this application.
[0017] This application is mainly based on multi-source monitoring data already collected by existing integrated online monitoring systems, operation management systems, and environmental monitoring systems. It achieves in-depth analysis of bushing status through data processing, time-series neural network prediction, and health assessment algorithms without the need for additional complex hardware. At the same time, the time-series neural network model used in this application can also adopt a lightweight design, which reduces computational complexity while ensuring prediction accuracy. It is suitable for deployment at substations or dispatch centers. Therefore, this application not only improves the intelligence level of bushing status monitoring, but also has the advantages of low implementation cost, easy deployment, and easy promotion.
[0018] This application can predict the deterioration trend and assess the health of a single bushing, and can also perform parallel analysis on multiple phase bushings or multiple devices of the same transformer. By comparing the health index, deterioration rate and remaining life level of different objects, differentiated risk ranking and maintenance recommendations can be formed. As a result, preventive maintenance can be prioritized for high-risk equipment, while routine monitoring and routine maintenance can be maintained for equipment in better health, thereby improving the rationality of maintenance resource allocation and enhancing the precision and economy of operation and maintenance decisions. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a time-aligned diagram showing the input composition of the monitoring data for this invention. Figure 3 This is a schematic diagram of the casing deterioration trend prediction model of the present invention; Figure 4 This is a flowchart of the method for calculating the Health Index (HI) and assessing the remaining life expectancy level according to the present invention. Figure 5 This is a schematic diagram illustrating the casing health assessment results and maintenance decisions under actual operating conditions of the present invention. Detailed Implementation
[0020] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.
[0021] Example 1: As Figure 1-5 As shown, a method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network is completed collaboratively by three levels: multi-source data acquisition, data processing and modeling, and assessment output and application. This forms a closed-loop processing procedure from online status acquisition and time-series trend prediction to health assessment and maintenance recommendation output. In the multi-source data acquisition layer, the accessed data includes four categories: online monitoring data, operating condition data, environmental data, and historical data. Online monitoring data is used to characterize the insulation and discharge status of the bushing, and may include dielectric loss factor, leakage current, partial discharge characteristics, and dissolved gas information in the oil. Operating condition data is used to characterize the electrical and thermal loads borne by the bushing, and may include load current, voltage, and operating temperature. Environmental data is used to characterize changes in the external environment, and may include ambient temperature and humidity. Historical data is used to characterize the past evolution of the equipment, and may include historical health status and maintenance records. By unifying the access of data from the above different sources, a basic data set for the long-term deterioration and evolution process of the bushing can be established.
[0022] In the data processing and modeling layer, the incoming data is first preprocessed. Preprocessing includes time alignment, anomaly handling, missing value imputation, and normalization. Time alignment addresses inconsistent sampling frequencies from different data sources; anomaly handling removes abrupt and invalid values; missing value imputation ensures sequence continuity; and normalization eliminates dimensional differences. After preprocessing, a multidimensional time-series sequence with operating condition labels is constructed from the multi-source data to serve as input to the time-series neural network model. In the modeling section, a time-series neural network is used to model the casing degradation trend. The network used can be a temporal convolutional network, a lightweight long short-term memory network, or a prediction model formed by combining both. During model training and optimization, loss function, hyperparameter tuning, and cross-validation are combined to make the model suitable for on-site deployment and long-term operation.
[0023] In the evaluation output and application layer, the health index is first calculated based on the prediction results. The health index is obtained by fusing multiple indicators and uses dynamic weight allocation to achieve a common representation of the comprehensive health status by different monitoring quantities. The resulting health index HI value is between 0 and 1. Then, the degradation trend is output based on the future state prediction results, and the state changes within a certain number of future time steps are visualized. Next, the remaining life is assessed based on the relationship between the health index change pattern and the failure threshold, and the assessment results are divided into four levels of early warning status: healthy, alert, abnormal, and severe. Based on the above results, condition-based maintenance decisions, operation and maintenance plan optimization suggestions, risk assessment reports, and early warning information push content are further generated and can be displayed in the visualization interface in the form of a health status dashboard and trend change curves. A real-time data update loop can also be formed between the data acquisition layer and the data processing modeling layer. The operation results in the evaluation output and application layer can also have a reverse effect on the model training and optimization process, thereby forming a feedback loop of continuous updating and continuous optimization.
[0024] Example 2: As Figure 1-5As shown, this embodiment illustrates the input structure, time alignment method, and time-series input sample generation process for multi-source monitoring data, including: Multi-source monitoring data is categorized into high-frequency, medium-frequency, low-frequency, and operating condition data based on sampling frequency. High-frequency data may include partial discharge and leakage current, with a sampling interval of 1 minute. Medium-frequency data may include dielectric loss factor and operating temperature, with a sampling interval of 1 hour. Low-frequency data may include dissolved gas analysis results in oil and ambient temperature and humidity, with a sampling interval of 1 day. Operating condition data may include event-type or low-frequency change information such as load current. Since the above data forms independent asynchronous time series in their original state, they cannot be directly spliced and modeled; therefore, unified time alignment processing is required first. In one specific implementation, a resampling strategy is first used to map data of different frequencies to a unified time base. For data unification from high frequency to low frequency, the average value can be used to obtain statistical results corresponding to the unified time step. For data unification from low frequency to high frequency, forward padding can be used to retain the most recent valid value. The unified time base is selected as 1 hour. Subsequently, outlier handling and normalization are performed on the aligned data. Outlier handling can use the 3-principle to remove outliers, and linear interpolation can be used to fill in the gaps after outlier removal. Normalization can use the Min-Max method to map each feature to the interval between 0 and 1. After the above processing, the original asynchronous multi-source data is converted into multi-dimensional time series samples on a unified time axis.
[0025] During the sample construction phase, 30 consecutive days of data are used as an input window, and time-series samples are constructed with a step size of 1 hour, resulting in an input sequence with 720 time steps. Each time step corresponds to a multi-dimensional input vector, and each input vector contains both monitoring features and operating condition labels. In one specific implementation, the input vector has 12 dimensions, of which 8 dimensions are monitoring features and 4 dimensions are operating condition labels. The input sequence can be denoted as X = [X_(Tk), X_(T-k+1),, X_(T0)], where each Xt contains the monitoring features and operating condition labels at that time.
[0026] To further illustrate the composition of the input samples, the following input content can be constructed at continuous time slices T-3, T-2, T-1, and the current time T0: At time T-3, the monitoring features may include tan=0.005, leakage current 2.3mA, and partial discharge 15pC, corresponding to operating condition labels of 80% load and 35℃; at time T-2, the monitoring features may include tan=0.006, leakage current 2.5mA, and partial discharge 18pC, corresponding to operating condition labels of 85% load and 38℃; at time T-1... The monitoring characteristics may include tan=0.007, leakage current 2.8mA, and partial discharge 22pC, corresponding to the operating condition labels of 90% load and 40℃ temperature. At the current time T0, the monitoring characteristics may include tan=0.008, leakage current 3.2mA, and partial discharge 25pC, corresponding to the operating condition labels of 95% load and 42℃ temperature. As can be seen from the above time series slices, this method does not only perform instantaneous analysis on a single monitoring value, but uses the monitoring characteristics and operating condition labels in a continuous time window to jointly describe the state evolution process of the bushing.
[0027] Example 3: As Figure 1-5 As shown, this embodiment is used to further illustrate the structure of the time-series neural network model and the casing degradation trend prediction process based on the model, including: The model input is a multidimensional temporal tensor, denoted by (L, D), where L represents the time step and D represents the feature dimension. After the input sequence enters the temporal neural network, the temporal features are first extracted by the temporal convolutional network module, then the long-term dependencies are extracted by the lightweight LSTM module, and finally the output layer generates the state prediction results within a preset future time range, denoted by (N, M), where N represents the prediction step and M represents the state dimension.
[0028] The temporal convolutional network module consists of multiple cascaded convolutional blocks, each with dilated convolutions and residual connections. In one specific implementation, the temporal convolutional network includes three convolutional blocks with dilation coefficients of 1, 2, and 4, each employing a dilated convolutional structure with a kernel size of 3. By setting different dilation coefficients, the receptive field can be expanded without significantly increasing the parameter size, thereby extracting sequence dependencies over a longer time range. Residual connections are used to alleviate gradient degradation during deep network training, making the network more stable. After processing by the three convolutional blocks, high-level temporal features are output to characterize the local patterns and long-term trends in the sleeve state change process.
[0029] After advanced temporal features are input into a lightweight LSTM module, the temporal evolution relationship is further modeled. The LSTM module includes an input gate, a forget gate, an output gate, cell states, and hidden states. Through the gating mechanism, historical information is selectively retained and the current state is updated, thereby improving the ability to characterize long-term dependencies. In one specific implementation, the number of hidden layer units in the LSTM is set to 64, and the Dropout probability is set to 0.2 to reduce the risk of overfitting. After processing by the LSTM module, the obtained temporal features are input into a fully connected layer, and a prediction output is generated after passing through a Dropout layer. The output content is the state prediction result for several future time steps.
[0030] During model training, historical data is used as input, and the entire processing flow is completed through steps such as feature extraction, temporal modeling, state prediction, trend analysis, health assessment, and risk warning. In one specific implementation, historical data from 30 consecutive days is input into the model to predict state parameters for the next 7 days (168 hours). Mean squared error is used as the primary loss function during training, with an L2 regularization term added to improve generalization ability. The optimizer is Adam, with a learning rate of 0.001. The training run consists of 20 epochs with a batch size of 32. Five-fold cross-validation is used for validation. The fusion of TCN and LSTM structures balances temporal feature extraction efficiency with long-term dependency modeling capabilities, reducing model complexity while maintaining prediction accuracy, making it suitable for field engineering applications.
[0031] Example 4: Figure 1-5 As shown, this embodiment further illustrates the process of health index calculation, remaining life expectancy assessment, and decision output in engineering applications, including: First, health assessment features are constructed based on the future state parameters output by the model. The health assessment process can be divided into three stages: predicted state acquisition, feature weight allocation, and multi-indicator fusion. The predicted state acquisition stage outputs state quantities at several future time points. The feature weight allocation stage sets weight vectors W for different state quantities and determines the contribution of each state feature using methods such as entropy weighting. The multi-indicator fusion stage performs a weighted summation of each feature to obtain the health index HI. The health index can be calculated using a normalized and weighted fusion method, typically as HI(t) = 1 - i=1 n w i (x i (t)-x i _min) / (x i _max-x i The calculation is performed using _min), where x i (t) represents the value of the i-th feature at time t, x i _min and x i_max represents the minimum and maximum values of the feature, respectively, w i The value represents the corresponding weight. The HI obtained by this method is in the range of 0 to 1. The closer the value is to 1, the healthier the casing is, and the closer the value is to 0, the higher the risk of failure.
[0032] To illustrate the calculation method of the health index, in one specific embodiment, the input feature vector may include tan=0.008, leakage current=3.2mA, partial discharge=25pC, H2=85L / L, CH4=32L / L, C2H2=2L / L, oil temperature=65℃ and ambient temperature=28℃. After weighted fusion of the above features, a health index HI=0.82 or HI=0.78 can be obtained, corresponding to the status assessment results at different time points. The health index can also be used to form a historical evolution curve in a time series manner to show the trend of the casing health status from the past to the present.
[0033] In the remaining life assessment phase, a trend analysis is first performed on the health index sequence, and then the remaining life is estimated based on the degradation rate and failure threshold. Trend analysis can use linear regression or exponential fitting to obtain the current degradation rate k of the bushing. The remaining life estimate can be calculated based on the difference between the current health index HI_current and the failure threshold HI_threshold, typically expressed as RUL=(HI_current-HI_threshold) / k, where HI_threshold can be determined based on historical failure samples and expert experience. In one specific implementation, the failure threshold can be set to 0.3. After estimation, the equipment is classified according to the health index range and the remaining life range, and corresponding maintenance recommendations are formed. For example, for a healthy state, corresponding to HI 0.8 and a remaining life greater than 5 years, a routine inspection strategy can be adopted. For a state of concern, the monitoring frequency can be increased and preventative maintenance prepared when the health index drops to the 0.6 to 0.8 range. For abnormal and severe states, more proactive maintenance and handling measures can be taken based on a lower health index range combined with the remaining life results.
[0034] In engineering applications, the evaluation results can be directly presented in the transformer bushing intelligent health monitoring system and intelligent maintenance decision support system. Taking the C-phase high-voltage bushing of the T1 main transformer at the 500kV Dongshan substation as an example, the basic equipment information may include equipment number TB-500-001-C, commissioning date June 15, 2015, operating duration 8 years and 7 months, and rated voltage 500kV. Real-time monitoring data may include tan=0.008, leakage current=3.2mA, partial discharge=25pC, and oil temperature 65℃. The health index monitoring section displays the current HI=0 on the dashboard. 0.78 corresponds to a "Caution" status; the degradation trend prediction and remaining life assessment section uses a trend curve to show the future degradation direction and gives a remaining life prediction of approximately 2.5 years; the early warning information can provide a prompt that the equipment is expected to enter an abnormal level within 3 months, and suggests increasing the monitoring frequency to once a week, while preparing a maintenance plan; the maintenance decision support system can further provide maintenance strategy suggestions based on the health assessment results, and provide a comprehensive judgment of equipment status, operating status, and risk level. For example, in the current case, the equipment health level is "Caution," the operating status is "Normal," and the risk level is "Medium Risk."
[0035] Example 5: Figure 1-5 As shown, this embodiment is used to further illustrate the scenario of parallel evaluation of multiple devices and differentiated operation and maintenance decision-making, including: Synchronous monitoring is performed on the A-phase, B-phase, and C-phase bushings of the same transformer, and their respective multi-dimensional time-series input sequences are constructed. The input data of the three-phase bushings are aligned under a unified time reference, and each input sample contains both monitoring features and operating condition labels. To reflect the differences in different operating conditions, label information such as load level and ambient temperature is introduced into the input samples of the three-phase bushings. In one specific embodiment, the load levels of phases A, B, and C are set to 85%, 82%, and 90%, respectively, and the ambient temperature at their respective installation locations is recorded simultaneously.
[0036] After inputting the three-phase bushing samples into the same TCN-LSTM fusion model for training and testing, the degradation trend prediction results of each of the three-phase bushings can be obtained. Under a specific working condition, when the C-phase bushing is in a high-temperature heavy-load state, that is, when the ambient temperature is greater than 35℃ and the load rate is greater than 90%, the model can predict that the bushing of this phase will show an accelerated degradation trend. After further health assessment, the health indices of the three-phase bushings at the same time point are as follows: Phase A HI = 0.85, corresponding to a healthy state; Phase B HI = 0.82, corresponding to a state of concern; and Phase C HI = 0.78, corresponding to a state of concern. After trend fitting of the health index sequence, the degradation rates can be obtained respectively: Phase A is approximately -0.0015 / day, Phase B is approximately -0.0021 / day, and Phase C is approximately -0.0028 / day, with Phase C showing the fastest degradation rate. Based on this, it can be further estimated that the remaining life of the Phase C bushing is approximately 2.5 years, while the remaining life of the Phase A bushing is approximately 5.8 years, with Phase C only about 42% of that of Phase A.
[0037] Based on the above differentiated assessment results, the intelligent maintenance decision support system can generate the optimal maintenance plan. For C-phase bushings, which deteriorate faster and have a shorter remaining lifespan, preventive maintenance is recommended first; for A-phase and B-phase bushings, which are in relatively good condition, routine monitoring and maintenance strategies are maintained. Compared with the original unplanned power outages and extensive maintenance methods, adopting the above differentiated operation and maintenance decision can save approximately 85% of power outage time, reduce operation and maintenance costs by approximately 42%, and increase equipment utilization by approximately 28%. This indicates that the method of the present invention is not only applicable to single bushing health assessment, but also to scenarios involving parallel assessment of multiple devices and optimal resource allocation.
[0038] Example 6: As Figure 1-5 As shown, this embodiment further illustrates the deployment method of the present invention at the substation side or dispatch center, including: deploying a lightweight model inference engine at the substation side and connecting it to the integrated online monitoring system, operation management system, and environmental monitoring system. The inference engine receives bushing online monitoring data, operating condition data, and environmental data in real time, and constructs multi-dimensional time-series samples at a uniform time granularity. Whenever new time window data is received, a trend prediction and health assessment calculation is triggered, automatically outputting the current health index, remaining life level, and early warning information; the results can also be used as a data source for continuous model optimization, thus forming a closed-loop operation mechanism of real-time data updates, rolling predictions, and continuous model optimization.
[0039] In addition to deployment at substations, the above method can also be implemented in dispatch centers or other platforms with data access and computing capabilities. As long as it can achieve unified access to multi-source monitoring data, time alignment, time series input construction, trend prediction, health assessment, and decision output, it can be regarded as a specific implementation of this method.
[0040] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. A method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network, characterized in that, include: Acquire multi-source monitoring data of transformer bushings; Multi-source monitoring data are preprocessed, and a time-series input sequence is constructed to characterize the state evolution of transformer bushings. The time-series input sequence is input into a time-series neural network model to obtain the prediction results of the deterioration trend of the transformer bushing; The health assessment results of the transformer bushings are determined based on the deterioration trend prediction results.
2. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 1, characterized in that, The multi-source monitoring data includes one or more of the following: online monitoring data, operating condition data, environmental data, and historical status data; The online monitoring data includes one or more of the following: dielectric loss, leakage current, partial discharge characteristic quantities, and dissolved gases in oil. The operating condition data includes one or more of load current, operating temperature, and voltage; The environmental data includes one or more of ambient temperature and ambient humidity; The historical status data includes one or more of the following: historical health status and maintenance records.
3. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 2, characterized in that, The multi-source monitoring data is preprocessed, including one or more of the following: time alignment, outlier handling, missing value imputation, and normalization.
4. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 3, characterized in that, The construction time-series input sequence includes: The preprocessed multi-source monitoring data is organized into multi-dimensional time-series samples with operating condition labels according to the preset time granularity.
5. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 4, characterized in that, The operating condition label includes one or more of the following: load level, operating temperature, ambient temperature, and ambient humidity.
6. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 5, characterized in that, The temporal neural network model includes one of the following: temporal convolutional network, long short-term memory network, and their fusion model; The degradation trend prediction results include the trend of changes in the state parameters of the transformer bushing within a future preset time range.
7. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 6, characterized in that, The determination of the health assessment results includes: The health index is calculated based on the predicted deterioration trend. The health status of transformer bushings is determined based on the health index or the trend of the health index.
8. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 7, characterized in that, The calculation of the health index involves weighted fusion of multiple features characterizing the condition of transformer bushings.
9. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 8, characterized in that, The health assessment results also include remaining life expectancy assessment results; The remaining life assessment result is obtained by performing trend analysis on the health index sequence and combining it with a preset failure threshold.
10. The method for predicting the deterioration trend and assessing the health of transformer bushings based on a time-series neural network according to claim 9, characterized in that, After predicting the degradation trend and conducting a health assessment, the system outputs one or more of the following: health index, degradation trend prediction results, remaining life assessment results, early warning information, and maintenance recommendations.