Transformer digital twin management system and method based on multi-source data fusion

Abstract

The application discloses a transformer digital twin management system and method based on multi-source data fusion, relates to the technical field of data analysis, and constructs an operation parameter coupling network by multi-source monitoring data in the operation process of the transformer, determines the dynamic coupling strength between various operation parameters, identifies an abnormal node, determines an abnormal disturbance migration direction and migration strength in combination with the dynamic coupling strength and a time health gradient, updates a health evolution field, constructs an abnormal propagation prediction model according to a health potential value, the time health gradient and the dynamic coupling strength, inputs real-time operation data into the abnormal propagation prediction model, predicts a next time window health potential value, determines an early warning node, predicts an abnormal propagation path in combination with the dynamic coupling strength and the abnormal disturbance migration direction, generates transformer abnormal early warning information, improves transformer operation reliability, and reduces a fault outage risk.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a transformer digital twin management system and method based on multi-source data fusion. Background Technology

[0002] As a crucial piece of equipment in the power system, transformers directly impact the safety and stability of power system operation. With the development of intelligent monitoring and digital operation and maintenance technologies, more and more transformers are equipped with online monitoring devices. By collecting multi-source status data during equipment operation, the operating status, health level, and abnormal risks of the equipment are analyzed, providing a data foundation for transformer condition-based maintenance and fault early warning. At the same time, the development of digital twin technology makes it possible to realize equipment status mapping, operation process monitoring, and status analysis by constructing digital models corresponding to the actual equipment, providing a new technical approach to improve the level of intelligent operation and maintenance of transformers. However, existing transformer condition monitoring and digital twin analysis methods typically focus on assessing the current state based on real-time monitoring data. They mainly judge the health status of equipment through preset indicators, fixed models, or historical experience. They are unable to fully consider the dynamic changes caused by various factors such as load changes, environmental factors, and internal performance degradation during equipment operation. They are also unable to accurately reflect the development trend of equipment health status over time, which limits the ability to identify abnormal risks and predict faults. There is also a certain deviation between the digital twin model and the actual equipment operating status. Therefore, there is a need to provide an intelligent early warning method that can improve the accuracy of transformer operating status analysis and the reliability of fault early warning. Summary of the Invention

[0003] The purpose of this invention is to provide a transformer digital twin management system and method based on multi-source data fusion to solve the problems raised in the prior art.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a transformer digital twin management method based on multi-source data fusion, the method comprising: Step S100: Collect multi-source monitoring data during transformer operation, and standardize the multi-source monitoring data to obtain a unified operating parameter sequence; Step S200: Construct a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determine the dynamic coupling strength between each operating parameter; Step S300: Construct a healthy evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtain the health potential value and time health gradient corresponding to each operating parameter; Step S400: Identify abnormal nodes based on the health potential value, and determine the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and update the health evolution field. Step S500: Based on the health potential value, temporal health gradient and dynamic coupling strength of the abnormal node, construct an anomaly propagation prediction model; Step S600: Input the real-time running data into the trained anomaly propagation prediction model, predict the health potential value of each anomaly node in the next time window, determine the warning node based on the predicted health potential value, and predict the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, and generate transformer anomaly warning information.

[0005] Furthermore, step S100 includes: Collect multi-source operating data of the transformer, including electrical quantity data, mechanical quantity data, environmental quantity data, and chemical quantity data; The electrical data includes winding hot spot temperature, top oil temperature, load current, three-phase voltage, core grounding current, and partial discharge intensity; the mechanical data includes vibration signals; the environmental data includes ambient temperature and ambient humidity; and the chemical data includes dissolved gas content in the oil. The collected multi-source operational data are synchronized according to a unified sampling period to obtain a time-aligned data sequence. The time-aligned data sequence is filled with missing values ​​and corrected for outliers. The missing value filling process is to fill in the missing values ​​by interpolation based on the data of adjacent sampling points. The outlier correction process is to construct a sliding window centered on the current sampling point and include several sampling points before and after it, count all sampling values ​​in the window and calculate the median of the sampling values. If the absolute difference between the current sampling value and the median exceeds a preset threshold, the current sampling value is determined to be an outlier and the current sampling value is replaced by the median. The corrected operational data are normalized to obtain a unified sequence of operational parameters. By collecting multi-source operational data, including transformer electrical, mechanical, environmental, and chemical data, multi-dimensional and comprehensive monitoring of transformer operating status is achieved. Compared to a single data source, this provides a more comprehensive reflection of equipment operating characteristics and improves status awareness. Time synchronization of various data types according to a unified sampling period eliminates time deviations caused by inconsistent sampling frequencies from different data sources, providing a unified data foundation for subsequent multi-source data fusion analysis. Normalization of the corrected operational data eliminates differences in dimensions and numerical ranges between different parameters, giving all operational parameters a unified data scale. This facilitates feature fusion and analysis in subsequent status assessment, fault diagnosis, and prediction models, thereby improving the accuracy and stability of analysis results and providing reliable data support for intelligent transformer operation and maintenance.

[0006] Furthermore, step S200 includes: Each operating parameter in the unified operating parameter sequence is used as a parameter node; Divide the continuous sliding time window according to the preset time length, and obtain the data sequence corresponding to any two running parameters in the current time window for each time window. Based on the incremental changes of the two operating parameters between adjacent sampling points within the current time window, construct the change sequences corresponding to the two operating parameters respectively. Based on whether the signs of the change increments of corresponding sampling points in the two change sequences are consistent, we determine whether the change directions of each corresponding sampling point are the same, count the number of corresponding sampling points with the same change direction, and divide the number of corresponding sampling points by the total number of all corresponding sampling points to calculate the directional correlation coefficient between the two operating parameters. Based on the change increments of corresponding sampling points in two change sequences, the absolute difference between the change increments of corresponding sampling points is calculated. The absolute difference is used as the amplitude difference degree of corresponding sampling points. The amplitude difference degree of each corresponding sampling point is averaged to obtain the average amplitude difference degree. The amplitude correlation coefficient between the two operating parameters is calculated based on the average amplitude difference degree, A=1 / (1+B), where B represents the average amplitude difference degree. The directional correlation coefficient and the amplitude correlation coefficient are weighted and summed according to preset weights to obtain the initial correlation coefficient between the two operating parameters. The duration during which the change direction of two operating parameters remains consistent within the current time window is statistically analyzed. The amplitude difference degree corresponding to each sampling point of the two operating parameters within the current time window is collected. The absolute difference between the amplitude difference degree and the average amplitude difference degree is calculated. The absolute differences corresponding to all sampling points are summed to obtain the total amplitude deviation. The duration and the total amplitude deviation are normalized respectively. The normalized duration and the total amplitude deviation are weighted and summed according to a preset correction weight to obtain the continuous change coefficient between the two operating parameters. The continuous change coefficient is multiplied by the initial correlation coefficient to obtain the dynamic coupling strength between the two operating parameters. Each running parameter is used as a network node, and the dynamic coupling strength between any two running parameters is used as the weight of the connection edge between the corresponding nodes to construct the parameter coupling network for the current time window. Using each operating parameter in a unified operating parameter sequence as a network node, the dynamic relationship between these parameters is analyzed through a continuous sliding time window. This accurately reflects the coupling characteristics of the transformer's operating state over time, avoiding the problem of traditional static correlation analysis's inability to characterize time-varying relationships. By constructing separate operating parameter change sequences and calculating the initial correlation coefficient by combining the consistency of change direction and the similarity of change amplitude, the coordinated change trend between different operating parameters can be identified, while also considering the similarity of change degree, thus improving the accuracy of parameter correlation analysis. Introducing a continuous change coefficient comprehensively corrects the duration of continuous consistency in change direction and the stability of amplitude deviation, ensuring that the calculated dynamic coupling strength reflects whether the parameters are correlated. This effectively reduces the impact of instantaneous fluctuations, random noise, and accidental anomalies on the correlation results, realizing a networked expression of complex correlations between multiple transformer operating parameters. This more intuitively reveals the information transmission path and mutual influence mechanism between various operating parameters, providing a more accurate and reliable data foundation for key parameter identification, anomaly propagation analysis, health status assessment, and fault early warning.

[0007] Furthermore, step S300 includes: The running process is divided into multiple continuous sliding time windows according to a preset time length, and the parameter coupling network corresponding to each time window is obtained. Traverse the nodes corresponding to each running parameter in the current time window parameter coupling network, and obtain the dynamic coupling strength of the connection edge between the node corresponding to the current running parameter and its adjacent nodes. Based on the normal operating reference range corresponding to each operating parameter, it is determined whether the actual monitored value is within the normal operating reference range; when the actual monitored value is within the normal operating reference range, the deviation is set to zero; when the actual monitored value exceeds the normal operating reference range, the difference between the actual monitored value and the nearest boundary of the normal operating reference range is calculated as the deviation, and the parameter health loss value corresponding to each operating parameter is calculated based on the ratio between the deviation and the width of the normal operating reference range interval. The coupling influence of each adjacent node on the current operating parameter is obtained by multiplying the dynamic coupling strength between the current operating parameter and each adjacent node by the parameter health loss value of each adjacent node's operating parameter. All coupling influence values ​​are accumulated to obtain the cumulative coupling influence value corresponding to the current operating parameter. The health potential value corresponding to the current operating parameter is obtained by weighted summation of the parameter health loss value and the cumulative coupling influence value. The health potential value corresponding to each sliding time window is arranged in chronological order to form a health potential value time series corresponding to each operating parameter. The health potential value time series is interpolated and smoothed, and a health evolution field in continuous time dimension is constructed based on the health potential value time series corresponding to all operating parameters. Based on the rate of change of the health potential values ​​of each operating parameter with respect to time in the health evolution field, the time health gradient corresponding to each operating parameter is obtained; By dividing the operation process into multiple continuous sliding time windows and combining the parameter coupling network corresponding to each time window to conduct health status analysis, the temporal evolution of the transformer's operating status can be dynamically reflected, providing better continuity and timeliness compared to single-moment status evaluation. By calculating the parameter health loss value based on the normal operating reference range of each operating parameter, a quantitative characterization of the health status of each parameter is achieved, enabling different types of operating parameters to have a unified health evaluation scale. By combining the dynamic coupling strength between parameters with the parameter health loss value of adjacent nodes, the cumulative coupling impact value is calculated, fully considering the propagation effect and mutual influence of parameter anomalies in the coupling network, avoiding evaluation based solely on the status of a single parameter, and improving the comprehensiveness and accuracy of the health evaluation results.

[0008] Furthermore, step S400 includes: Based on the health potential value corresponding to each operating parameter, it is determined whether it is greater than the health abnormality threshold. When the health potential value is greater than the health abnormality threshold, the corresponding operating parameter is determined as the abnormal node corresponding to the current time window. The health abnormality threshold is calculated by statistically obtaining the mean and standard deviation of the historical health potential values, multiplying the preset health abnormality coefficient by the standard deviation, and adding it to the mean. For each abnormal node, the dynamic coupling strength corresponding to the connection edge between the abnormal node and each adjacent operating parameter is obtained. Adjacent operating parameters with dynamic coupling strength greater than the coupling strength threshold are selected as candidate migration nodes. The coupling strength threshold is calculated by statistically obtaining the average value and standard deviation of historical dynamic coupling strength based on historical dynamic coupling strength, multiplying the preset coupling strength coefficient by the standard deviation, and adding it to the mean value. The gradient change between the time health gradient corresponding to the abnormal node and the time health gradient corresponding to each candidate migration node is calculated respectively. Based on the sign relationship of the gradient change, the migration direction of the abnormal disturbance between the abnormal node and the candidate migration node is determined. For each candidate migration node, the absolute value of the time health gradient corresponding to the abnormal node is multiplied by the dynamic coupling strength between the abnormal node and the corresponding candidate migration node to obtain the migration strength corresponding to the abnormal disturbance migrating to each candidate migration node. Based on the migration direction and corresponding migration intensity of the abnormal disturbance to each candidate migration node, the proportion of the migration intensity of the abnormal node to each candidate migration node to the total migration intensity is calculated, and the proportion is multiplied by the health potential value corresponding to the abnormal node to calculate the abnormal migration amount of the abnormal node to each candidate migration node. For each abnormal node, the abnormal migration amount corresponding to each candidate migration node is accumulated to obtain the cumulative migration amount corresponding to the abnormal node. The health potential value corresponding to the abnormal node is then deducted and updated according to the cumulative migration amount to obtain the updated health potential value of the abnormal node. For each candidate migration node, the abnormal migration amounts from each abnormal node are accumulated to obtain the cumulative migration amount corresponding to the candidate migration node. The health potential value corresponding to the candidate migration node is then updated based on the cumulative migration amount to obtain the updated health potential value of the candidate migration node. Based on the updated health potential values ​​of abnormal nodes, the updated health potential values ​​of candidate migration nodes, and the health potential values ​​of nodes that did not participate in propagation, an updated health evolution field corresponding to the current time window is formed. By adaptively calculating health anomaly thresholds and coupling strength thresholds based on historical health potential values ​​and historical dynamic coupling strength, the identification of anomaly nodes and the screening of propagation paths can adapt to different operating conditions, improving the accuracy of anomaly detection and environmental adaptability. By calculating the anomaly migration amount, the health potential values ​​of anomaly nodes and candidate migration nodes are updated for migration out and migration in, respectively, realizing the dynamic transfer and redistribution of anomaly energy in the parameter coupling network. This allows the updated health evolution field to continuously reflect the overall health status changes after anomaly propagation, effectively revealing anomaly propagation paths, identifying anomaly propagation trends and potentially affected parameters, and providing a more accurate and reliable basis for fault source location, risk range prediction, anomaly propagation early warning, and maintenance decisions.

[0009] Furthermore, step S500 includes: A series of consecutive days are selected as the training period. The health evolution field corresponding to multiple consecutive time windows within the training period is obtained. The abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes are extracted in chronological order for each time window. The health potential values, temporal health gradients, and dynamic coupling strengths are then normalized. The normalized sequence of abnormal node health potential values, the normalized sequence of time health gradients, and the normalized sequence of dynamic coupling strength corresponding to multiple consecutive time windows are arranged in chronological order to form training samples. The abnormal node health potential value corresponding to the future time window is used as the prediction target to construct the training dataset for the abnormal propagation prediction model. The training dataset is input into the anomaly propagation prediction model for training. The training process involves inputting the training samples in the training dataset into the long short-term memory network model in the order of time windows, using the health potential value of the abnormal node corresponding to the future time window as the supervision label, iteratively updating the network parameters through the backpropagation algorithm, and stopping the training when the prediction error is less than the preset error threshold or the training rounds reach the preset number of times, thus obtaining the trained anomaly propagation prediction model. A training dataset for an anomaly propagation prediction model is constructed based on the health evolution field corresponding to continuous time windows. This model integrates multi-dimensional temporal features such as the health potential value of anomaly nodes, temporal health gradient, and dynamic coupling strength, fully exploring the temporal correlation characteristics and parameter coupling patterns during anomaly propagation. Compared to methods using only a single state parameter, it exhibits higher prediction accuracy and stability. By normalizing the health potential value, temporal health gradient, and dynamic coupling strength, differences in the dimensions of different features are effectively eliminated. A long short-term memory network is used to learn from continuous time window samples, fully utilizing the influence of historical operating states on future anomaly evolution. This accurately captures long-term dependencies and dynamic trends during anomaly propagation. By using the health potential value of anomaly nodes in future time windows as a supervision label and employing a backpropagation algorithm to iteratively optimize network parameters, the model continuously improves its anomaly propagation prediction accuracy. The resulting anomaly propagation prediction model can predict the development trend of the health status of anomaly nodes and the degree of anomaly diffusion in advance, providing a forward-looking decision-making basis for equipment fault early warning, risk assessment, and maintenance plan formulation, thereby improving the intelligence level of transformer operating status prediction.

[0010] Furthermore, step S600 includes: The real-time transformer operation data is used to construct the health evolution field corresponding to the current time window according to the method of steps S300-S500. The health potential value, time health gradient and dynamic coupling strength between abnormal nodes are extracted and normalized. The normalized health potential value, time health gradient and dynamic coupling strength are then input into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window. The predicted health potential value corresponding to each abnormal node is compared with the preset warning threshold. When the predicted health potential value is greater than the preset warning threshold, the abnormal node is determined as a warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted warning node and the predicted propagation path, transformer abnormal warning information is generated and sent to the staff. This method enables advance prediction of abnormal health states in the next forecast time window. Compared to traditional threshold-based alarm methods based on the current state, it fully utilizes the temporal evolution of anomaly propagation, shifting from "post-event alarms" to "pre-event warnings," thus improving the timeliness and accuracy of anomaly warnings. By comparing the predicted health potential with preset warning thresholds, it accurately identifies warning nodes that may experience future anomalies. Combined with the dynamic coupling strength between warning nodes and the migration direction of anomaly disturbances, it predicts the future propagation path of anomalies, forecasting their development trend and spread range. This allows maintenance personnel to anticipate potential risks, rationally arrange maintenance and handling measures, effectively reduce the risk of equipment failure caused by anomaly propagation, improve the intelligence level and fault warning capability of transformer operation status monitoring, and ensure the safety, stability, and reliability of power equipment operation.

[0011] To better implement the above methods, a transformer digital twin management system based on multi-source data fusion is also proposed. The system includes a unified operating parameter sequence module, a dynamic coupling strength module, a health evolution field module, an updated health evolution field module, an anomaly propagation prediction model module, and an early warning information module. Unified operating parameter sequence module: Collects multi-source monitoring data during transformer operation, standardizes the multi-source monitoring data, and obtains a unified operating parameter sequence; Dynamic coupling strength module: Constructs a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determines the dynamic coupling strength between each operating parameter; Health Evolution Field Module: Constructs a health evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtains the health potential value and time health gradient corresponding to each operating parameter; Update the health evolution field module: Identify abnormal nodes based on the health potential value, and determine the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and update the health evolution field. Anomaly propagation prediction model module: Based on the health potential value, temporal health gradient and dynamic coupling strength of the anomaly node, an anomaly propagation prediction model is constructed; Early warning information module: Input real-time running data into the trained anomaly propagation prediction model, predict the health potential value of each anomaly node in the next time window, determine the early warning node based on the predicted health potential value, and predict the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, and generate transformer anomaly early warning information.

[0012] Furthermore, the anomaly propagation prediction model module includes a unit for constructing the training dataset and a unit for training the anomaly propagation prediction model: Constructing a training dataset unit: Select several consecutive days as the training period, obtain the health evolution field corresponding to multiple consecutive time windows within the training period, extract the abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes corresponding to each time window in chronological order, normalize the health potential values, temporal health gradients, and dynamic coupling strengths, and form training samples by combining the normalized abnormal node health potential value sequences, normalized temporal health gradient sequences, and normalized dynamic coupling strength sequences corresponding to multiple consecutive time windows in chronological order, and use the health potential values ​​of abnormal nodes corresponding to future time windows as the prediction target to construct the training dataset for the anomaly propagation prediction model. Training the anomaly propagation prediction model unit: The training dataset is input into the anomaly propagation prediction model for training. The training process involves inputting the training samples in the training dataset into the long short-term memory network model in the order of time windows, using the health potential value of the anomaly node corresponding to the future time window as the supervision label, and iteratively updating the network parameters through the backpropagation algorithm. When the prediction error is less than the preset error threshold or the training rounds reach the preset number of times, the training is stopped, and the trained anomaly propagation prediction model is obtained.

[0013] Furthermore, the early warning information module includes a health potential prediction unit and an early warning information generation unit: Predicting health potential value unit: Construct the health evolution field corresponding to the current time window using the real-time collected transformer operation data according to the method of steps S300-S500, extract the health potential value, time health gradient and dynamic coupling strength between abnormal nodes, and complete the normalization process. Then, input the normalized health potential value, time health gradient and dynamic coupling strength into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window. The unit generates early warning information by comparing the predicted health potential value corresponding to each abnormal node with the preset early warning threshold. When the predicted health potential value is greater than the preset early warning threshold, the abnormal node is identified as an early warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the early warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted early warning node and the predicted propagation path, transformer abnormal early warning information is generated and sent to the staff.

[0014] Compared with the prior art, the beneficial effects of the present invention are: the present invention comprehensively considers the consistency of the direction of change of operating parameters, the consistency of the change amplitude, and the continuous change characteristics, establishes dynamic coupling strength, and constructs a parameter coupling network, which can describe whether there is a correlation between various operating parameters, and at the same time reflects the dynamic evolution of the correlation relationship with the change of operating state. Compared with the traditional method of using fixed correlation coefficients or static correlation models, it can more accurately reveal the coupling mechanism between operating parameters inside the transformer. Based on the health loss of parameters and the dynamic coupling effect between parameters, a health potential value is constructed, and a health evolution field in the continuous time dimension is further formed. A time health gradient is introduced to describe the trend of health status change, so that the health status is extended from the traditional single index evaluation to a continuous dynamic evolution process, which can more accurately reflect the development trend of equipment health status and improve the ability of digital twin model to express the operating status. This invention identifies abnormal nodes based on health potential values ​​and calculates the migration direction, intensity, and amount of abnormal disturbances by combining dynamic coupling strength and time health gradient. It dynamically updates the health evolution field and simulates the propagation of abnormal disturbances in the parameter network. Compared with existing technologies that only locate a single abnormal parameter, this invention can analyze the abnormal diffusion process and propagation law, providing a more intuitive basis for fault mechanism analysis. Compared with the method of setting a fixed threshold based on the current monitoring value for alarm, this invention can predict the development of abnormalities in advance, improve the accuracy of early warning, and reduce the false alarm rate and missed alarm rate. By combining digital twin models with anomaly propagation prediction, current abnormal states can be identified, and future early warning nodes and anomaly propagation paths can be predicted. This helps maintenance personnel to formulate maintenance strategies in advance, shifting from post-fault handling to proactive prevention, improving transformer operational reliability, and reducing the risk of fault outages. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of the transformer digital twin management system based on multi-source data fusion according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a technical solution: a transformer digital twin management method based on multi-source data fusion, the method comprising: Step S100: Collect multi-source monitoring data during transformer operation, and standardize the multi-source monitoring data to obtain a unified operating parameter sequence; Step S100 includes: Collect multi-source operating data of the transformer, including electrical quantity data, mechanical quantity data, environmental quantity data, and chemical quantity data; The electrical data includes winding hot spot temperature, top oil temperature, load current, three-phase voltage, core grounding current, and partial discharge intensity; the mechanical data includes vibration signals; the environmental data includes ambient temperature and ambient humidity; and the chemical data includes dissolved gas content in the oil. The collected multi-source operational data are synchronized according to a unified sampling period to obtain a time-aligned data sequence. The time-aligned data sequence is filled with missing values ​​and corrected for outliers. The missing value filling process is to fill in the missing values ​​by interpolation based on the data of adjacent sampling points. The outlier correction process is to construct a sliding window centered on the current sampling point and include several sampling points before and after it, count all sampling values ​​in the window and calculate the median of the sampling values. If the absolute difference between the current sampling value and the median exceeds a preset threshold, the current sampling value is determined to be an outlier and the current sampling value is replaced by the median. The corrected operational data are normalized to obtain a unified sequence of operational parameters. For example, taking a 220kV oil-immersed power transformer as an example, its continuous operation status was monitored. A uniform sampling period of 5 minutes was set, and 24 hours of continuous operation data were collected, resulting in 288 sets of sampling data. The electrical data included winding hot spot temperature, top oil temperature, load current, three-phase voltage, core grounding current, and partial discharge intensity; the mechanical data included vibration signals; the environmental data included ambient temperature and humidity; and the chemical data included dissolved gas content in the oil. Taking a set of raw data collected at 08:00 as an example, the winding hot spot temperature is 86.5℃, the top oil temperature is 67.8℃, the load current is 512A, the three-phase voltages of A, B, and C are 220.8kV, 221.3kV, and 220.5kV, respectively, the core grounding current is 0.18A, the partial discharge intensity is 95pC, the effective vibration value is 0.62mm / s, the ambient temperature is 31.4℃, the ambient humidity is 68%, and the dissolved hydrogen content in the oil is 118μL / L; Because there are discrepancies in the sampling times of different sensors, for example, the sampling time for partial discharge monitoring data is 08:00:03, the sampling time for vibration monitoring data is 08:00:01, and the sampling time for other monitoring data is 08:00:00, therefore, according to a unified sampling period of 5 minutes, with 08:00:00 as the unified time node, all kinds of monitoring data are synchronized to the same time series to obtain a time-aligned running data series. Taking the winding hotspot temperature data as an example, at 08:35, due to a communication anomaly, data was missing. The adjacent sampled values ​​before and after it were 84.6℃ and 85.4℃, respectively. Therefore, linear interpolation was used to complete the data, and the completed data was (84.6+85.4) / 2=85.0℃. Taking the top oil temperature data at 08:55 as an example, its sampled value was 93.7℃. A sliding window containing 5 sampled values ​​was constructed with 2 sampled points before and after it. The top oil temperature data in the window were 68.1℃, 67.9℃, 93.7℃, 68.3℃ and 68.0℃, respectively. The median of the sampled values ​​in the window was calculated to be 68.1℃. The anomaly judgment threshold was set to 8℃. Since the absolute difference between the current sampled value and the median was 25.6℃, which was greater than the preset threshold, 93.7℃ was judged to be an anomaly value, and the median of 68.1℃ was used to replace the anomaly value. After completing missing value completion and outlier correction, various operating data were normalized. A minimum-maximum normalization method was used. The normal operating range for winding hot spot temperature was set to 40℃~120℃, so the normalized value for 86.5℃ was (86.5-40) / (120-40)=0.581; the normal operating range for top layer oil temperature was set to 30℃~95℃, so the normalized value for 67.8℃ was 0.582; the normal operating range for load current was set to 0~1000A, so the normalized value for 512A was 0.512; and the normal operating range for partial discharge intensity was set to 0~500. pC, then the normalized value of 95pC is 0.190; the normal range of vibration effective value is set to 0~2.0mm / s, then the normalized value of 0.62mm / s is 0.310; the normal range of ambient temperature is set to -20℃~50℃, then the normalized value is 0.734; the normal range of ambient humidity is set to 0%~100%, then the normalized value is 0.680; the normal range of dissolved hydrogen content in oil is set to 0~1000μL / L, then the normalized value is 0.118. The normalized results corresponding to each monitoring parameter are arranged in a unified time sequence to form a unified operating parameter sequence. The unified operating parameter sequence corresponding to 08:00 is [0.581, 0.582, 0.512, 0.736, 0.738, 0.735, 0.180, 0.190, 0.310, 0.734, 0.680, 0.118]. Each element corresponds to the winding hot spot temperature, top oil temperature, load current, A-phase voltage, B-phase voltage, C-phase voltage, core grounding current, partial discharge intensity, vibration effective value, ambient temperature, ambient humidity, and dissolved hydrogen content in the oil.

[0018] Step S200: Construct a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determine the dynamic coupling strength between each operating parameter; Step S200 includes: Each operating parameter in the unified operating parameter sequence is used as a parameter node; Divide the continuous sliding time window according to the preset time length, and obtain the data sequence corresponding to any two running parameters in the current time window for each time window. Based on the incremental changes of the two operating parameters between adjacent sampling points within the current time window, construct the change sequences corresponding to the two operating parameters respectively. Based on whether the signs of the change increments of corresponding sampling points in the two change sequences are consistent, we determine whether the change directions of each corresponding sampling point are the same, count the number of corresponding sampling points with the same change direction, and divide the number of corresponding sampling points by the total number of all corresponding sampling points to calculate the directional correlation coefficient between the two operating parameters. Based on the change increments of corresponding sampling points in two change sequences, the absolute difference between the change increments of corresponding sampling points is calculated. The absolute difference is used as the amplitude difference degree of corresponding sampling points. The amplitude difference degree of each corresponding sampling point is averaged to obtain the average amplitude difference degree. The amplitude correlation coefficient between the two operating parameters is calculated based on the average amplitude difference degree, A=1 / (1+B), where B represents the average amplitude difference degree. The directional correlation coefficient and the amplitude correlation coefficient are weighted and summed according to preset weights to obtain the initial correlation coefficient between the two operating parameters. The duration during which the change direction of two operating parameters remains consistent within the current time window is statistically analyzed. The amplitude difference degree corresponding to each sampling point of the two operating parameters within the current time window is collected. The absolute difference between the amplitude difference degree and the average amplitude difference degree is calculated. The absolute differences corresponding to all sampling points are summed to obtain the total amplitude deviation. The duration and the total amplitude deviation are normalized respectively. The normalized duration and the total amplitude deviation are weighted and summed according to a preset correction weight to obtain the continuous change coefficient between the two operating parameters. The continuous change coefficient is multiplied by the initial correlation coefficient to obtain the dynamic coupling strength between the two operating parameters. Each running parameter is used as a network node, and the dynamic coupling strength between any two running parameters is used as the weight of the connection edge between the corresponding nodes to construct the parameter coupling network for the current time window. For example, taking 08:00 to 08:55 as the current analysis time window, with a uniform sampling period of 5 minutes, this time window contains a total of 12 sets of continuous sampling data. The calculation process of the dynamic coupling relationship between the winding hot spot temperature denoted as P1 and the top oil temperature denoted as P2 is taken as an example. Each operating parameter in the unified operating parameter sequence is taken as a parameter node, including winding hot spot temperature, top oil temperature, load current, A-phase voltage, B-phase voltage, C-phase voltage, core grounding current, partial discharge intensity, vibration effective value, ambient temperature, ambient humidity and dissolved hydrogen content in oil, forming a total of 12 parameter nodes. The continuous sliding window is divided according to the preset time window length of 60 minutes, and the parameter coupling network is updated step by step with a sliding step size of 5 minutes. Taking the normalized data corresponding to the winding hot spot temperature and the top oil temperature within the current time window of 08:00 to 08:55 as an example, the data sequences corresponding to the winding hot spot temperature are 0.581, 0.584, 0.588, 0.592, 0.596, 0.601, 0.605, 0.609, 0.612, 0.616, 0.620, and 0.624; the data sequences corresponding to the top oil temperature are 0.582, 0.585, 0.588, 0.590, 0.594, 0.597, 0.600, 0.603, 0.606, 0.608, 0.611, and 0.615. Based on the incremental changes between adjacent sampling points of the two operating parameters mentioned above, change sequences are constructed respectively. The change sequence of winding hot spot temperature is [+0.003, +0.004, +0.004, +0.004, +0.005, +0.004, +0.004, +0.003, +0.004, +0.004, +0.004]; the change sequence of top oil temperature is [+0.003, +0.003, +0.002, +0.004, +0.003, +0.003, +0.003, +0.003, +0.002, +0.003, +0.004]. Since the change increments of the above 11 groups are all positive, the change directions of the 11 sampling points are all consistent. The number of groups with consistent change directions is 11, and the total number of comparisons is 11. Therefore, the directional correlation coefficient is 11 / 11 = 1.000. The amplitude differences corresponding to each sampling point are 0, 0.001, 0.002, 0, 0.002, 0.001, 0.001, 0, 0.002, 0.001, and 0, respectively. Therefore, the average amplitude difference is (0+0.001+0.002+0+0.002+0.001+0.001+0+0.002+0.001+0) / 11=0.00091. According to the formula for calculating the amplitude correlation coefficient, the amplitude correlation coefficient is 1 / (1+0.00091)=0.9991. If the weight of the directional correlation coefficient is set to 0.6 and the weight of the amplitude correlation coefficient is set to 0.4, then the initial correlation coefficient between the two operating parameters is 0.6×1.000+0.4×0.9991=0.9996; Then, the duration during which the two operating parameters changed in the same direction within the current time window was calculated. Since the two parameters always increased synchronously within the entire 60-minute time window, the duration of consistency was 60 minutes. The normalization range of the duration was set to 0 to 60 minutes, and the normalization result of the duration was 60 / 60 = 1.000. Based on the amplitude difference degree corresponding to the aforementioned 11 sampling points, the absolute difference between each sampling point and the average amplitude difference degree of 0.00091 is calculated, which are 0.00091, 0.00009, 0.00109, 0.00091, 0.00109, 0.00009, 0.00009, 0.00009, 0.00091, 0.00109, 0.00009, and 0.00091. The sum of the above absolute differences yields a total amplitude deviation of 0.00728. The normalization range of the total amplitude deviation is set to 0 to 0.020, then the normalized total amplitude deviation is 0.00728 / 0.020 = 0.364. With the duration correction weight set to 0.7 and the stability correction weight set to 0.3, the calculated result of the duration variation coefficient is 0.7×1.000+0.3×0.364=0.8092. Multiplying the duration variation coefficient by the initial correlation coefficient, the dynamic coupling strength between the winding hot spot temperature and the top oil temperature is 0.9996×0.8092=0.8088.

[0019] Step S300: Construct a healthy evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtain the health potential value and time health gradient corresponding to each operating parameter; Step S300 includes: The running process is divided into multiple continuous sliding time windows according to a preset time length, and the parameter coupling network corresponding to each time window is obtained. Traverse the nodes corresponding to each running parameter in the current time window parameter coupling network, and obtain the dynamic coupling strength of the connection edge between the node corresponding to the current running parameter and its adjacent nodes. Based on the normal operating reference range corresponding to each operating parameter, it is determined whether the actual monitored value is within the normal operating reference range; when the actual monitored value is within the normal operating reference range, the deviation is set to zero; when the actual monitored value exceeds the normal operating reference range, the difference between the actual monitored value and the nearest boundary of the normal operating reference range is calculated as the deviation, and the parameter health loss value corresponding to each operating parameter is calculated based on the ratio between the deviation and the width of the normal operating reference range interval. The coupling influence of each adjacent node on the current operating parameter is obtained by multiplying the dynamic coupling strength between the current operating parameter and each adjacent node by the parameter health loss value of each adjacent node's operating parameter. All coupling influence values ​​are accumulated to obtain the cumulative coupling influence value corresponding to the current operating parameter. The health potential value corresponding to the current operating parameter is obtained by weighted summation of the parameter health loss value and the cumulative coupling influence value. The health potential value corresponding to each sliding time window is arranged in chronological order to form a health potential value time series corresponding to each operating parameter. The health potential value time series is interpolated and smoothed, and a health evolution field in continuous time dimension is constructed based on the health potential value time series corresponding to all operating parameters. Based on the rate of change of the health potential values ​​of each operating parameter with respect to time in the health evolution field, the time health gradient corresponding to each operating parameter is obtained.

[0020] Step S400: Identify abnormal nodes based on the health potential value, and determine the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and update the health evolution field. Step S400 includes: Based on the health potential value corresponding to each operating parameter, it is determined whether it is greater than the health abnormality threshold. When the health potential value is greater than the health abnormality threshold, the corresponding operating parameter is determined as the abnormal node corresponding to the current time window. The health abnormality threshold is calculated by statistically obtaining the mean and standard deviation of the historical health potential values, multiplying the preset health abnormality coefficient by the standard deviation, and adding it to the mean. For each abnormal node, the dynamic coupling strength corresponding to the connection edge between the abnormal node and each adjacent operating parameter is obtained. Adjacent operating parameters with dynamic coupling strength greater than the coupling strength threshold are selected as candidate migration nodes. The coupling strength threshold is calculated by statistically obtaining the average value and standard deviation of historical dynamic coupling strength based on historical dynamic coupling strength, multiplying the preset coupling strength coefficient by the standard deviation, and adding it to the mean value. The gradient change between the time health gradient corresponding to the abnormal node and the time health gradient corresponding to each candidate migration node is calculated respectively. Based on the sign relationship of the gradient change, the migration direction of the abnormal disturbance between the abnormal node and the candidate migration node is determined. For each candidate migration node, the absolute value of the time health gradient corresponding to the abnormal node is multiplied by the dynamic coupling strength between the abnormal node and the corresponding candidate migration node to obtain the migration strength corresponding to the abnormal disturbance migrating to each candidate migration node. Based on the migration direction and corresponding migration intensity of the abnormal disturbance to each candidate migration node, the proportion of the migration intensity of the abnormal node to each candidate migration node to the total migration intensity is calculated, and the proportion is multiplied by the health potential value corresponding to the abnormal node to calculate the abnormal migration amount of the abnormal node to each candidate migration node. For each abnormal node, the abnormal migration amount corresponding to each candidate migration node is accumulated to obtain the cumulative migration amount corresponding to the abnormal node. The health potential value corresponding to the abnormal node is then deducted and updated according to the cumulative migration amount to obtain the updated health potential value of the abnormal node. For each candidate migration node, the abnormal migration amounts from each abnormal node are accumulated to obtain the cumulative migration amount corresponding to the candidate migration node. The health potential value corresponding to the candidate migration node is then updated based on the cumulative migration amount to obtain the updated health potential value of the candidate migration node. Based on the updated health potential values ​​of abnormal nodes, the updated health potential values ​​of candidate migration nodes, and the health potential values ​​of nodes that did not participate in propagation, an updated health evolution field corresponding to the current time window is formed.

[0021] Step S500: Based on the health potential value, temporal health gradient and dynamic coupling strength of the abnormal node, construct an anomaly propagation prediction model; Step S500 includes: A series of consecutive days are selected as the training period. The health evolution field corresponding to multiple consecutive time windows within the training period is obtained. The abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes are extracted in chronological order for each time window. The health potential values, temporal health gradients, and dynamic coupling strengths are then normalized. The normalized sequence of abnormal node health potential values, the normalized sequence of time health gradients, and the normalized sequence of dynamic coupling strength corresponding to multiple consecutive time windows are arranged in chronological order to form training samples. The abnormal node health potential value corresponding to the future time window is used as the prediction target to construct the training dataset for the abnormal propagation prediction model. The training dataset is input into the anomaly propagation prediction model for training. The training process involves inputting the training samples in the training dataset into the long short-term memory network model in the order of time windows, using the health potential value of the abnormal node corresponding to the future time window as the supervision label, and iteratively updating the network parameters through the backpropagation algorithm. When the prediction error is less than the preset error threshold or the training rounds reach the preset number, the training is stopped, and the trained anomaly propagation prediction model is obtained.

[0022] Step S600: Input the real-time running data into the trained anomaly propagation prediction model, predict the health potential value of each anomaly node in the next time window, determine the warning node based on the predicted health potential value, and predict the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, and generate transformer anomaly warning information. Step S600 includes: The real-time transformer operation data is used to construct the health evolution field corresponding to the current time window according to the method of steps S300-S500. The health potential value, time health gradient and dynamic coupling strength between abnormal nodes are extracted and normalized. The normalized health potential value, time health gradient and dynamic coupling strength are then input into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window. The predicted health potential value corresponding to each abnormal node is compared with the preset warning threshold. When the predicted health potential value is greater than the preset warning threshold, the abnormal node is determined as a warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted warning nodes and the predicted propagation path, transformer abnormal warning information is generated and sent to the staff.

[0023] To better implement the above methods, a transformer digital twin management system based on multi-source data fusion is also proposed. The system includes a unified operating parameter sequence module, a dynamic coupling strength module, a health evolution field module, an updated health evolution field module, an anomaly propagation prediction model module, and an early warning information module. Unified operating parameter sequence module: Collects multi-source monitoring data during transformer operation, standardizes the multi-source monitoring data, and obtains a unified operating parameter sequence; Dynamic coupling strength module: Constructs a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determines the dynamic coupling strength between each operating parameter; Health Evolution Field Module: Constructs a health evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtains the health potential value and time health gradient corresponding to each operating parameter; Update the health evolution field module: Identify abnormal nodes based on the health potential value, and determine the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and update the health evolution field. Anomaly propagation prediction model module: Based on the health potential value, temporal health gradient and dynamic coupling strength of the anomaly node, an anomaly propagation prediction model is constructed; The anomaly propagation prediction model module includes a unit for constructing a training dataset and a unit for training the anomaly propagation prediction model. Constructing a training dataset unit: Select several consecutive days as the training period, obtain the health evolution field corresponding to multiple consecutive time windows within the training period, extract the abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes corresponding to each time window in chronological order, normalize the health potential values, temporal health gradients, and dynamic coupling strengths, and form training samples by combining the normalized abnormal node health potential value sequences, normalized temporal health gradient sequences, and normalized dynamic coupling strength sequences corresponding to multiple consecutive time windows in chronological order, and use the health potential values ​​of abnormal nodes corresponding to future time windows as the prediction target to construct the training dataset for the anomaly propagation prediction model. Training the anomaly propagation prediction model unit: The training dataset is input into the anomaly propagation prediction model for training. The training process involves inputting the training samples in the training dataset into the long short-term memory network model in the order of time windows, using the health potential value of the anomaly node corresponding to the future time window as the supervision label, and iteratively updating the network parameters through the backpropagation algorithm. When the prediction error is less than the preset error threshold or the training rounds reach the preset number of times, the training is stopped, and the trained anomaly propagation prediction model is obtained.

[0024] Early warning information module: Input real-time running data into the trained anomaly propagation prediction model, predict the health potential value of each anomaly node in the next time window, determine the early warning node based on the predicted health potential value, and predict the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, and generate transformer anomaly early warning information. The early warning information module includes a health potential prediction unit and an early warning information generation unit. Predicting health potential value unit: Construct the health evolution field corresponding to the current time window using the real-time collected transformer operation data according to the method of steps S300-S500, extract the health potential value, time health gradient and dynamic coupling strength between abnormal nodes, and complete the normalization process. Then, input the normalized health potential value, time health gradient and dynamic coupling strength into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window. The unit generates early warning information by comparing the predicted health potential value corresponding to each abnormal node with the preset early warning threshold. When the predicted health potential value is greater than the preset early warning threshold, the abnormal node is identified as an early warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the early warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted early warning node and the predicted propagation path, transformer abnormal early warning information is generated and sent to the staff.

[0025] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A transformer digital twin management method based on multi-source data fusion, characterized in that, The methods include: Step S100: Collect multi-source monitoring data during transformer operation, and standardize the multi-source monitoring data to obtain a unified operating parameter sequence; Step S200: Construct a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determine the dynamic coupling strength between each operating parameter; Step S300: Construct a healthy evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtain the health potential value and time health gradient corresponding to each operating parameter; Step S400: Identify abnormal nodes based on the health potential value, and determine the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and update the health evolution field. Step S500: Based on the health potential value, temporal health gradient and dynamic coupling strength of the abnormal node, construct an anomaly propagation prediction model; Step S600: Input the real-time running data into the trained anomaly propagation prediction model, predict the health potential value of each anomaly node in the next time window, determine the warning node based on the predicted health potential value, and predict the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, and generate transformer anomaly warning information.

2. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S100 includes the following steps: Collect multi-source operating data of the transformer, including electrical quantity data, mechanical quantity data, environmental quantity data, and chemical quantity data; The electrical data includes winding hot spot temperature, top oil temperature, load current, three-phase voltage, core grounding current, and partial discharge intensity; the mechanical data includes vibration signals; the environmental data includes ambient temperature and ambient humidity; and the chemical data includes dissolved gas content in the oil. The collected multi-source operational data are synchronized according to a unified sampling period to obtain a time-aligned data sequence. The time-aligned data sequence is filled with missing values ​​and corrected for outliers. The missing value filling process is to fill in the missing values ​​by interpolation based on the data of adjacent sampling points. The outlier correction process is to construct a sliding window centered on the current sampling point and include several sampling points before and after it, count all sampling values ​​in the window and calculate the median of the sampling values. If the absolute difference between the current sampling value and the median exceeds a preset threshold, the current sampling value is determined to be an outlier and the current sampling value is replaced by the median. The corrected operational data are normalized to obtain a unified sequence of operational parameters.

3. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S200 includes the following steps: Each operating parameter in the unified operating parameter sequence is used as a parameter node; Divide the continuous sliding time window according to the preset time length, and obtain the data sequence corresponding to any two running parameters in the current time window for each time window. Based on the incremental changes of the two operating parameters between adjacent sampling points within the current time window, construct the change sequences corresponding to the two operating parameters respectively. Based on whether the signs of the change increments of corresponding sampling points in the two change sequences are consistent, we determine whether the change directions of each corresponding sampling point are the same, count the number of corresponding sampling points with the same change direction, and divide the number of corresponding sampling points by the total number of all corresponding sampling points to calculate the directional correlation coefficient between the two operating parameters. Based on the change increments of corresponding sampling points in two change sequences, the absolute difference between the change increments of corresponding sampling points is calculated. The absolute difference is used as the amplitude difference degree of corresponding sampling points. The amplitude difference degree of each corresponding sampling point is averaged to obtain the average amplitude difference degree. The amplitude correlation coefficient between the two operating parameters is calculated based on the average amplitude difference degree, A=1 / (1+B), where B represents the average amplitude difference degree. The directional correlation coefficient and the amplitude correlation coefficient are weighted and summed according to preset weights to obtain the initial correlation coefficient between the two operating parameters. The duration during which the change direction of two operating parameters remains consistent within the current time window is statistically analyzed. The amplitude difference degree corresponding to each sampling point of the two operating parameters within the current time window is collected. The absolute difference between the amplitude difference degree and the average amplitude difference degree is calculated. The absolute differences corresponding to all sampling points are summed to obtain the total amplitude deviation. The duration and the total amplitude deviation are normalized respectively. The normalized duration and the total amplitude deviation are weighted and summed according to a preset correction weight to obtain the continuous change coefficient between the two operating parameters. The continuous change coefficient is multiplied by the initial correlation coefficient to obtain the dynamic coupling strength between the two operating parameters. By treating each running parameter as a network node and using the dynamic coupling strength between any two running parameters as the weight of the connection edge between the corresponding nodes, a parameter coupling network for the current time window is constructed.

4. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S300 includes the following steps: The running process is divided into multiple continuous sliding time windows according to a preset time length, and the parameter coupling network corresponding to each time window is obtained. Traverse the nodes corresponding to each running parameter in the current time window parameter coupling network, and obtain the dynamic coupling strength of the connection edge between the node corresponding to the current running parameter and its adjacent nodes. Based on the normal operating reference range corresponding to each operating parameter, it is determined whether the actual monitored value is within the normal operating reference range; when the actual monitored value is within the normal operating reference range, the deviation is set to zero; when the actual monitored value exceeds the normal operating reference range, the difference between the actual monitored value and the nearest boundary of the normal operating reference range is calculated as the deviation, and the parameter health loss value corresponding to each operating parameter is calculated based on the ratio between the deviation and the width of the normal operating reference range interval. The coupling influence of each adjacent node on the current operating parameter is obtained by multiplying the dynamic coupling strength between the current operating parameter and each adjacent node by the parameter health loss value of each adjacent node's operating parameter. All coupling influence values ​​are accumulated to obtain the cumulative coupling influence value corresponding to the current operating parameter. The health potential value corresponding to the current operating parameter is obtained by weighted summation of the parameter health loss value and the cumulative coupling influence value. The health potential value corresponding to each sliding time window is arranged in chronological order to form a health potential value time series corresponding to each operating parameter. The health potential value time series is interpolated and smoothed, and a health evolution field in continuous time dimension is constructed based on the health potential value time series corresponding to all operating parameters. Based on the rate of change of the health potential values ​​of each operating parameter with respect to time in the health evolution field, the time health gradient corresponding to each operating parameter is obtained.

5. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S400 includes the following steps: Based on the health potential value corresponding to each operating parameter, it is determined whether it is greater than the health abnormality threshold. When the health potential value is greater than the health abnormality threshold, the corresponding operating parameter is determined as the abnormal node corresponding to the current time window. The health abnormality threshold is calculated by statistically obtaining the mean and standard deviation of the historical health potential values, multiplying the preset health abnormality coefficient by the standard deviation, and adding it to the mean. For each abnormal node, the dynamic coupling strength corresponding to the connection edge between the abnormal node and each adjacent operating parameter is obtained. Adjacent operating parameters with dynamic coupling strength greater than the coupling strength threshold are selected as candidate migration nodes. The coupling strength threshold is calculated by statistically obtaining the average value and standard deviation of historical dynamic coupling strength based on historical dynamic coupling strength, multiplying the preset coupling strength coefficient by the standard deviation, and adding it to the mean value. The gradient change between the time health gradient corresponding to the abnormal node and the time health gradient corresponding to each candidate migration node is calculated respectively. Based on the sign relationship of the gradient change, the migration direction of the abnormal disturbance between the abnormal node and the candidate migration node is determined. For each candidate migration node, the absolute value of the time health gradient corresponding to the abnormal node is multiplied by the dynamic coupling strength between the abnormal node and the corresponding candidate migration node to obtain the migration strength corresponding to the abnormal disturbance migrating to each candidate migration node. Based on the migration direction and corresponding migration intensity of the abnormal disturbance to each candidate migration node, the proportion of the migration intensity of the abnormal node to each candidate migration node to the total migration intensity is calculated, and the proportion is multiplied by the health potential value corresponding to the abnormal node to calculate the abnormal migration amount of the abnormal node to each candidate migration node. For each abnormal node, the abnormal migration amount corresponding to each candidate migration node is accumulated to obtain the cumulative migration amount corresponding to the abnormal node. The health potential value corresponding to the abnormal node is then deducted and updated according to the cumulative migration amount to obtain the updated health potential value of the abnormal node. For each candidate migration node, the abnormal migration amounts from each abnormal node are accumulated to obtain the cumulative migration amount corresponding to the candidate migration node. The health potential value corresponding to the candidate migration node is then updated based on the cumulative migration amount to obtain the updated health potential value of the candidate migration node. Based on the updated health potential values ​​of abnormal nodes, the updated health potential values ​​of candidate migration nodes, and the health potential values ​​of nodes that did not participate in propagation, an updated health evolution field corresponding to the current time window is formed.

6. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S500 includes the following steps: A series of consecutive days are selected as the training period. The health evolution field corresponding to multiple consecutive time windows within the training period is obtained. The abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes are extracted in chronological order for each time window. The health potential values, temporal health gradients, and dynamic coupling strengths are then normalized. The normalized sequence of abnormal node health potential values, the normalized sequence of time health gradients, and the normalized sequence of dynamic coupling strength corresponding to multiple consecutive time windows are arranged in chronological order to form training samples. The abnormal node health potential value corresponding to the future time window is used as the prediction target to construct the training dataset for the abnormal propagation prediction model. The training dataset is input into the anomaly propagation prediction model for training. The training process involves inputting the training samples in the training dataset into the long short-term memory network model in the order of time windows, using the health potential value of the abnormal node corresponding to the future time window as the supervision label, and iteratively updating the network parameters through the backpropagation algorithm. When the prediction error is less than the preset error threshold or the training rounds reach the preset number, the training is stopped, and the trained anomaly propagation prediction model is obtained.

7. The transformer digital twin management method based on multi-source data fusion according to claim 1, characterized in that, Step S600 includes the following steps: The real-time transformer operation data is used to construct the health evolution field corresponding to the current time window according to the method of steps S300-S500. The health potential value, time health gradient and dynamic coupling strength between abnormal nodes are extracted and normalized. The normalized health potential value, time health gradient and dynamic coupling strength are then input into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window. The predicted health potential value corresponding to each abnormal node is compared with the preset warning threshold. When the predicted health potential value is greater than the preset warning threshold, the abnormal node is determined as a warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted warning nodes and the predicted propagation path, transformer abnormal warning information is generated and sent to the staff.

8. A transformer digital twin management system based on multi-source data fusion, used to implement the transformer digital twin management method based on multi-source data fusion as described in any one of claims 1-7, characterized in that, The system includes a unified operating parameter sequence module, a dynamic coupling strength module, a health evolution field module, an updated health evolution field module, an anomaly propagation prediction model module, and an early warning information module. The unified operating parameter sequence module: collects multi-source monitoring data during transformer operation, performs standardized processing on the multi-source monitoring data, and obtains a unified operating parameter sequence; The dynamic coupling strength module: constructs a parameter coupling network between operating parameters based on the unified operating parameter sequence, and determines the dynamic coupling strength between each operating parameter; The health evolution field module: constructs a health evolution field based on the dynamic coupling strength and the operating status of each operating parameter, and obtains the health potential value and time health gradient corresponding to each operating parameter; The updated health evolution field module: identifies abnormal nodes based on the health potential value, and determines the migration direction and migration intensity of abnormal disturbances in the health evolution field by combining the dynamic coupling strength and time health gradient, and updates the health evolution field. The anomaly propagation prediction model module constructs an anomaly propagation prediction model based on the health potential value, temporal health gradient, and dynamic coupling strength corresponding to the anomaly node. The early warning information module inputs real-time operating data into the trained anomaly propagation prediction model, predicts the health potential value corresponding to each anomaly node in the next time window, determines the early warning node based on the predicted health potential value, and predicts the anomaly propagation path by combining the dynamic coupling strength between anomaly nodes and the migration direction of anomalies, thereby generating transformer anomaly early warning information.

9. The transformer digital twin management system based on multi-source data fusion according to claim 8, characterized in that, The anomaly propagation prediction model module includes a unit for constructing a training dataset and a unit for training the anomaly propagation prediction model: The unit for constructing the training dataset involves selecting several consecutive days as the training period, obtaining the health evolution field corresponding to multiple consecutive time windows within the training period, extracting the abnormal nodes, health potential values, temporal health gradients, and dynamic coupling strengths between abnormal nodes corresponding to each time window in chronological order, normalizing the health potential values, temporal health gradients, and dynamic coupling strengths, and combining the normalized abnormal node health potential value sequence, normalized temporal health gradient sequence, and normalized dynamic coupling strength sequence corresponding to multiple consecutive time windows in chronological order to form training samples, and using the health potential values ​​of abnormal nodes corresponding to future time windows as the prediction target to construct the training dataset for the anomaly propagation prediction model. The training anomaly propagation prediction model unit: The training dataset is input into the anomaly propagation prediction model for training. The training process is to input the training samples in the training dataset into the long short-term memory network model in the order of time windows, use the health potential value of the abnormal node corresponding to the future time window as the supervision label, and iteratively update the network parameters through the backpropagation algorithm. When the prediction error is less than the preset error threshold or the training rounds reach the preset number, the training is stopped, and the trained anomaly propagation prediction model is obtained.

10. The transformer digital twin management system based on multi-source data fusion according to claim 8, characterized in that, The early warning information module includes a health potential prediction unit and an early warning information generation unit: The predicted health potential unit: constructs the health evolution field corresponding to the current time window according to the method of steps S300-S500 based on the real-time collected transformer operation data, extracts the health potential value, time health gradient and dynamic coupling strength between abnormal nodes, completes the normalization process, and inputs the normalized health potential value, time health gradient and dynamic coupling strength into the trained abnormal propagation prediction model to obtain the predicted health potential value of each abnormal node corresponding to the next prediction time window; The unit that generates early warning information compares the predicted health potential value corresponding to each abnormal node with a preset early warning threshold. When the predicted health potential value is greater than the preset early warning threshold, the abnormal node is identified as an early warning node. Based on the dynamic coupling strength and abnormal disturbance migration direction between the early warning nodes, the future propagation path of the abnormal disturbance is predicted. Based on the predicted early warning node and the predicted propagation path, transformer abnormal early warning information is generated and sent to the staff.

Images