Operation state identification method and device of transformer, medium and equipment

Abstract

The invention discloses an operation state identification method and device of a transformer, a medium and equipment. According to the method, the real-time operation data, the historical data and the physical attribute parameters of the converter transformer are acquired, and the data are combined with the preset multi-physical-field high-fidelity simulation model to carry out multi-physical-field solution, so that a plurality of predicted values are obtained. Meanwhile, a time sequence neural network model based on an LSTM-Transform network obtained by training real-time and historical data is utilized to predict a temperature residual value. And dynamically adjusting the fusion weight of the first temperature prediction value and the actual monitoring value by comparing the residual error of the first temperature prediction value and the actual monitoring value of the simulation model, and generating a mixed temperature prediction value. And finally, the mixed temperature prediction value is input into a knowledge layer decision-making system, the system uses a preset threshold value judgment and similarity matching algorithm to output an operation state identification result of the transformer, and the problem that the operation state of the transformer cannot be accurately and efficiently identified in the prior art is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring, and in particular to a method, device, medium, and equipment for identifying the operating status of a transformer. Background Technology

[0002] As power systems develop towards intelligence and efficiency, high-voltage direct current (HVDC) transmission technology is widely used due to its advantages in long-distance, large-capacity power transmission. As the core equipment in an HVDC system, the stability and reliability of the converter transformer's operation directly affect the safe operation of the entire power system.

[0003] Therefore, real-time monitoring and accurate analysis of the operating status of converter transformers are particularly important. Existing monitoring methods mainly include physical model-based monitoring methods and data-driven monitoring methods. Physical model-based methods rely on physical theories such as heat conduction, fluid mechanics, and electromagnetic fields, and can provide theoretically accurate and interpretable predictions. However, in practical applications, the predictive accuracy and adaptability of these models are limited due to the complexity of the operating environment of converter transformers. Data-driven methods analyze a large amount of historical operating data and use machine learning algorithms to build data models. Although they can learn the mapping relationships of historical data, they are unstable when dealing with extreme operating conditions or the early stages of abnormal evolution, and lack a deep understanding of the physical processes. These shortcomings prevent existing technologies from accurately and efficiently identifying the operating status of transformers. Summary of the Invention

[0004] This invention provides a method, apparatus, medium, and equipment for identifying the operating status of a transformer, in order to solve the problem that the existing technology cannot accurately and efficiently identify the operating status of a transformer.

[0005] Firstly, this application provides a method for identifying the operating status of a transformer, including: Acquire real-time operating data, historical operating data, and physical attribute parameters of the converter transformer; Based on the physical property parameters, and combined with a preset multiphysics high-fidelity simulation model, the multiphysics problem is solved by coupling the preset equations to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model. Based on real-time and historical operating data, and combined with a preset time-series neural network model, a predicted temperature residual value is obtained; wherein, the time-series neural network model is obtained by training the residual between the first temperature predicted value output by the simulation model and the actual monitored value, as well as historical operating data, using an LSTM-Transformer network. Based on the residual between the first temperature prediction value and the actual monitored value, the fusion weight of the first temperature prediction value and the temperature residual prediction value is adjusted, and a mixed temperature prediction value is generated by combining the preset weighted fusion algorithm. The mixed temperature prediction value is input into a preset knowledge layer decision system, so that the knowledge layer decision system outputs the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm.

[0006] This application achieves high-precision prediction and intelligent identification of the operating status of converter transformers by constructing a multi-physics high-fidelity simulation model and combining real-time and historical operating data. First, by coupling multiple pre-set equations, the simulation model can output multiple predicted values, providing physical priors for subsequent analysis. Second, a time-series neural network model based on LSTM-Transformer network dynamically corrects the prediction error of the simulation model, further improving prediction accuracy. By monitoring the residual between predicted and actual monitored values ​​in real time and dynamically adjusting the fusion weights, this application can adaptively optimize the prediction results, enhancing the model's adaptability and stability under complex operating conditions. Finally, the mixed temperature prediction values ​​are input into a knowledge-layer decision-making system, combined with an expert rule base and case reasoning module, to achieve rapid and accurate identification of the transformer's operating status. This process not only improves the accuracy and real-time performance of predictions but also enhances the model's interpretability and engineering adaptability, providing strong support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems and effectively solving the problem of inaccurate and inefficient identification of transformer operating status in existing technologies.

[0007] Furthermore, based on the physical property parameters and combined with a preset multiphysics high-fidelity simulation model, the multiphysics problem is solved by coupling the preset equations to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model, specifically as follows: The preset equations include preset fluid dynamics equations, preset electrostatic field equations, and preset heat conduction equations. Based on the preset fluid dynamics equations, and according to the preset physical property parameters of the oil and the preset volume force information, the velocity field distribution of the oil is calculated. Based on the preset electrostatic field equations, and according to the preset dielectric constant of the insulation system and the preset free charge density, the potential distribution and electric field strength are calculated. The Joule heat source distribution is calculated based on the described potential distribution and electric field strength. Based on the preset heat conduction equation, the velocity vector of the oil is input into the convection term of the heat conduction equation, and the Joule heat source distribution is input into the heat source term of the heat conduction equation. Combined with the preset oil density, preset specific heat capacity and preset thermal conductivity, the oil temperature distribution and hot spot temperature are obtained by solving.

[0008] This invention enables high-precision modeling and prediction of the operating state of converter transformers by coupling pre-defined fluid dynamics equations, electrostatic field equations, and heat conduction equations. First, using the fluid dynamics equations, the oil velocity field distribution is calculated based on the physical properties and volume force information of the oil. Next, using the electrostatic field equations, the potential distribution and electric field strength are calculated based on the dielectric constant and free charge density of the insulation system, thereby deriving the Joule heat source distribution. Finally, the oil velocity field is used as a convection term, and the Joule heat source distribution as a heat source term, input into the heat conduction equations. Combined with the oil's density, specific heat capacity, and thermal conductivity, the oil temperature distribution and hot spot temperatures are solved. This process not only ensures the physical consistency of the model but also improves prediction accuracy through the coupling of multiple physics fields, providing reliable theoretical support for the intelligent operation and maintenance of converter transformers.

[0009] Furthermore, the temporal neural network model is obtained by training an LSTM-Transformer network on the residual between the first temperature prediction value and the actual monitored value output by the simulation model, as well as historical operating data. Specifically: Calculate the residual between the first predicted temperature value and the actual monitored value, and define the residual as a supervision label; By acquiring real-time operational data and combining it with historical operational data, a time-series feature sequence is constructed. The temporal feature sequence is input into the LSTM-Transformer network for training, so that the LSTM-Transformer network extracts the long-short-term dependencies of the temporal feature sequence through the LSTM module, allocates global feature weights through the self-attention mechanism of the Transformer module, and optimizes the loss function using the supervision label during training until the value of the loss function no longer changes within a preset time period, at which point training stops and a trained temporal neural network model is obtained.

[0010] This application significantly improves the accuracy and adaptability of converter transformer operating status prediction by constructing a time-series neural network model based on an LSTM-Transformer network. First, the residual between the first temperature prediction value output by the multiphysics high-fidelity simulation model and the actual monitored value is calculated and defined as a supervision label to guide network training. Next, real-time operating data is acquired and combined with historical operating data to construct a time-series feature sequence, providing rich input information for the network. Then, the time-series feature sequence is input into the LSTM-Transformer network for training. During training, the LSTM module is responsible for extracting long-short-term dependencies in the time-series feature sequence, while the Transformer module dynamically allocates global feature weights through a self-attention mechanism, thereby better capturing important information in the data. By optimizing the loss function using supervision labels, the network can continuously adjust its parameters until the value of the loss function no longer changes within a preset time period, at which point training stops, resulting in a well-trained time-series neural network model. This process not only makes full use of historical and real-time data, but also dynamically corrects the prediction errors of the physical model through deep learning technology, significantly improving the model's prediction accuracy and adaptability to complex operating conditions, and providing strong support for the intelligent operation and maintenance of converter transformers.

[0011] Furthermore, the step of adjusting the fusion weights of the first temperature prediction value and the temperature residual prediction value based on the residual between the first temperature prediction value and the actual monitored value, and generating a hybrid temperature prediction value by combining a preset weighted fusion algorithm, specifically involves: Based on the residual, combined with the preset load fluctuation rate and the preset cooling state change rate, the fusion weight of the first temperature prediction value and the temperature residual prediction value is assigned. Based on the fusion weights, the first temperature prediction value, and the temperature residual prediction value, a mixed temperature prediction value is generated by combining a preset weight allocation threshold and a preset weighted fusion method.

[0012] This application effectively improves the accuracy and stability of converter transformer temperature prediction by dynamically adjusting fusion weights and combining them with a weighted fusion algorithm. Specifically, firstly, based on the residual between the first temperature prediction value and the actual monitored value, and combined with preset load fluctuation rate and cooling state change rate, the fusion weights of the first temperature prediction value and the temperature residual prediction value are dynamically allocated. This dynamic weight allocation mechanism can adaptively adjust the contribution ratio of the physical model and the data model according to changes in the current operating state. Subsequently, based on the allocated fusion weights, the first temperature prediction value, and the temperature residual prediction value, combined with preset weight allocation thresholds and a weighted fusion method, the final mixed temperature prediction value is generated. This fusion method not only fully utilizes the high fidelity of the physical model and the high adaptability of the data model, but also enhances the stability and responsiveness of the model under complex operating conditions through dynamic weight adjustment, thereby providing more accurate and reliable temperature prediction results for the intelligent operation and maintenance of converter transformers.

[0013] Furthermore, the step of inputting the mixed temperature prediction value into a preset knowledge layer decision-making system, so that the knowledge layer decision-making system outputs the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm, specifically: The mixed temperature prediction value is input into a preset knowledge layer decision system, so that the knowledge layer decision system, based on a preset expert rule base, determines whether the mixed temperature prediction value exceeds the preset hot spot temperature upper limit threshold and the temperature rise rate threshold, and obtains the threshold judgment result. By acquiring historical transformer cases and combining them with a preset similarity matching algorithm, a similarity search is performed on the current mixed temperature prediction value to obtain the similarity matching results. The threshold judgment result and the similarity matching result are fused together, and the judgment weight is set by a preset priority scheduling function to output the transformer operation status identification result.

[0014] This application achieves accurate identification and intelligent decision-making regarding the operating status of converter transformers by inputting the predicted mixed temperature value into a knowledge-layer decision-making system. First, utilizing a pre-defined expert rule base, the knowledge-layer decision-making system determines whether the predicted mixed temperature value exceeds preset hotspot temperature upper limit thresholds and temperature rise rate thresholds, thus obtaining a threshold judgment result. Next, combining historical transformer cases, a pre-defined similarity matching algorithm is used to perform a similarity search on the current predicted mixed temperature value, obtaining a similarity matching result. Finally, the threshold judgment result and the similarity matching result are fused, and a pre-defined priority scheduling function is used to set the judgment weights, outputting the final transformer operating status identification result. This process not only improves the accuracy and reliability of the prediction results but also enhances the model's interpretability and engineering adaptability, providing strong technical support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems.

[0015] Furthermore, the acquisition of real-time operating data and physical attribute parameters of the converter transformer specifically includes: The real-time operating data includes load current, oil flow rate, oil temperature, and ambient temperature and humidity. The physical property parameters include the density, specific heat capacity, thermal conductivity, viscosity, flow velocity field, internal structural parameters of the converter transformer, boundary conditions, and electrical excitation source of the cooling oil medium.

[0016] This application provides a comprehensive and accurate data foundation for high-precision modeling and intelligent analysis by acquiring real-time operating data and physical property parameters of the converter transformer. Specifically, the real-time operating data covers key operating indicators such as load current, oil flow rate, oil temperature, and ambient temperature and humidity, reflecting the transformer's operating status in real time. The physical property parameters include the density, specific heat capacity, thermal conductivity, viscosity, and flow velocity field of the cooling oil medium, as well as the internal structural parameters, boundary conditions, and electrical excitation sources of the converter transformer. These parameters accurately characterize the transformer's physical properties. By combining this data, the model can more accurately simulate the transformer's dynamic behavior, thereby achieving accurate prediction and intelligent decision-making under complex operating conditions, effectively improving operation and maintenance efficiency and system reliability.

[0017] Secondly, this application provides a transformer operating status identification device. The transformer operating status identification device includes: The acquisition module is used to acquire real-time operating data, historical operating data, and physical attribute parameters of the converter transformer; The simulation module is used to solve the multiphysics problem by coupling the preset equations based on the physical property parameters and a preset multiphysics high-fidelity simulation model, so as to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model. The prediction module is used to obtain the temperature residual prediction value based on real-time operation data and historical operation data, combined with a preset time-series neural network model; wherein, the time-series neural network model is obtained by training the residual between the first temperature prediction value output by the simulation model and the actual monitored value and historical operation data based on the LSTM-Transformer network. The fusion module is used to adjust the fusion weight of the first temperature prediction value and the temperature residual prediction value based on the residual between the first temperature prediction value and the actual monitoring value, and generate a mixed temperature prediction value by combining the preset weighted fusion algorithm. The identification module is used to input the mixed temperature prediction value into a preset knowledge layer decision system, so that the knowledge layer decision system can output the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm.

[0018] The transformer operating status identification device of this application integrates multiple functional modules to achieve high-precision prediction and intelligent identification of the operating status of converter transformers. The acquisition module collects real-time operating data, historical operating data, and physical attribute parameters, providing comprehensive data support for subsequent analysis. The simulation module utilizes these parameters to perform coupled solution through a multi-physics high-fidelity simulation model, outputting multiple predicted values ​​to ensure physical consistency of the predictions. The prediction module further combines a time-series neural network model to dynamically correct the prediction errors of the simulation model, improving prediction accuracy. The fusion module dynamically adjusts the fusion weights based on the residuals, optimizing the prediction results and enhancing the model's adaptability. Finally, the identification module inputs the mixed temperature prediction values ​​into a knowledge-layer decision system, combining threshold judgment and similarity matching algorithms to output the operating status identification result, achieving rapid and accurate diagnosis of the transformer's operating status. This device not only improves the accuracy and real-time performance of predictions but also enhances the model's interpretability and engineering adaptability, providing strong technical support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems.

[0019] Furthermore, the temporal neural network model is obtained by training an LSTM-Transformer network on the residual between the first temperature prediction value and the actual monitored value output by the simulation model, as well as historical operating data. Specifically: Calculate the residual between the first predicted temperature value and the actual monitored value, and define the residual as a supervision label; By acquiring real-time operational data and combining it with historical operational data, a time-series feature sequence is constructed. The temporal feature sequence is input into the LSTM-Transformer network for training, so that the LSTM-Transformer network extracts the long-short-term dependencies of the temporal feature sequence through the LSTM module, allocates global feature weights through the self-attention mechanism of the Transformer module, and optimizes the loss function using the supervision label during training until the value of the loss function no longer changes within a preset time period, at which point training stops and a trained temporal neural network model is obtained.

[0020] This application significantly improves the accuracy and adaptability of converter transformer operating status prediction by constructing a time-series neural network model based on an LSTM-Transformer network. First, the residual between the first temperature prediction value output by the multiphysics high-fidelity simulation model and the actual monitored value is calculated and defined as a supervision label to guide network training. Next, real-time operating data is acquired and combined with historical operating data to construct a time-series feature sequence, providing rich input information for the network. Then, the time-series feature sequence is input into the LSTM-Transformer network for training. During training, the LSTM module is responsible for extracting long-short-term dependencies in the time-series feature sequence, while the Transformer module dynamically allocates global feature weights through a self-attention mechanism, thereby better capturing important information in the data. By optimizing the loss function using supervision labels, the network can continuously adjust its parameters until the value of the loss function no longer changes within a preset time period, at which point training stops, resulting in a well-trained time-series neural network model. This process not only makes full use of historical and real-time data, but also dynamically corrects the prediction errors of the physical model through deep learning technology, significantly improving the model's prediction accuracy and adaptability to complex operating conditions, and providing strong support for the intelligent operation and maintenance of converter transformers.

[0021] Thirdly, this application provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the transformer operating status identification method as described above. Its beneficial effects are the same as those of the transformer operating status identification method provided in the first aspect of this application.

[0022] Fourthly, this application provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement any of the transformer operation status identification methods described in the first aspect. Attached Figure Description

[0023] Figure 1 : A schematic flowchart of an embodiment of the transformer operating status identification method provided in this application; Figure 2 : A schematic diagram of an embodiment of the transformer oil temperature analysis results provided in this application; Figure 3 : A schematic diagram of an embodiment of the transformer operation status identification device provided in this application. Detailed Implementation

[0024] 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.

[0025] Example 1 Please refer to Figure 1 In order to solve the problem that existing technologies cannot accurately and efficiently identify the operating status of transformers, this invention provides a method for identifying the operating status of transformers.

[0026] In this embodiment, the process of the transformer operation status identification method of this application is described in detail through steps S01-S05.

[0027] S01: Obtain real-time operating data, historical operating data, and physical attribute parameters of the converter transformer.

[0028] In a preferred embodiment of this invention, the acquisition of real-time operating data, historical operating data, and physical attribute parameters of the converter transformer specifically includes: The operating data of the converter transformer is collected in real time through a sensor network, including multi-dimensional characteristic information such as load current, oil flow rate, oil temperature, and ambient temperature and humidity.

[0029] Historical operating data of converter transformers are extracted from historical databases, including past load changes, oil temperature changes, and environmental conditions.

[0030] Obtain the physical property parameters of the converter transformer, including the density, specific heat capacity, thermal conductivity, viscosity, and flow velocity field of the cooling oil medium, the internal structural parameters of the converter transformer, boundary conditions, and electrical excitation sources (voltage and current).

[0031] S02: Based on the physical property parameters and combined with the preset multiphysics high-fidelity simulation model, the multiphysics is solved by coupling the preset equations to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model.

[0032] In a preferred embodiment of this invention, the step of solving the multiphysics problem by coupling preset equations based on the physical property parameters and a preset multiphysics high-fidelity simulation model to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model is as follows: The oil flow temperature rise coupling equation is used to perform thermal-flow field coupling modeling of the converter transformer winding region, and the oil temperature and hot spot temperature are calculated.

[0033] More specifically, a thermal-fluid field coupling model is performed on the winding region of the converter transformer using an oil flow temperature rise coupling equation to achieve quantitative prediction of oil temperature and hot spot temperature. The oil flow temperature rise coupling equation is as follows: (1) Where ρ is the oil density, c p Let Q be the specific heat capacity, k be the thermal conductivity, v be the flow velocity field, and Q be the flow velocity field. loss It represents the winding loss power density, reflecting the localized heating of components such as coils.

[0034] A cooling oil flow control model was established, and the incompressible Navier-Stokes momentum conservation equation was used to describe the flow characteristics of the oil in the winding channel and to calculate the velocity field distribution.

[0035] More specifically, to characterize the impact of oil flow characteristics in the complex channels inside the transformer on heat dissipation efficiency, a cooling oil flow control model is further established, using the incompressible Navier-Stokes momentum conservation equation for modeling. The calculation is as follows: (2) in, This represents the oil velocity vector, p represents the oil pressure, and μ represents the dynamic viscosity. This refers to volume forces, including gravity and possible electromagnetic disturbance forces.

[0036] An electric and magnetic field distribution model is introduced, and the internal heat source distribution is estimated using the static Maxwell equations, which are then used as the heat source term in the heat conduction model.

[0037] More specifically, to accurately estimate the internal heat source distribution and provide the heat source term Qloss to the heat conduction model, an electric field and magnetic field distribution model is introduced to simulate local induction and leakage magnetic flux, using the static Maxwell equations as follows: (3) in, Represents the electric potential distribution function. ρ represents the dielectric constant of an insulating system. e This represents the free charge density. Formulas (1), (2), and (3) represent the heat conduction, fluid dynamics, and electrostatic field distribution models, respectively, and there is a clear input-output relationship between the three: the flow velocity v calculated by formula (2) is used in the convection term in formula (1) and affects the temperature distribution; the electric potential calculated by formula (3) Electric field can be obtained Then calculate Joule heat Q, as a heat source lossInput into formula (1). Therefore, the temperature calculation depends on the flow rate and the electric heat source, which are provided by formulas (2) and (3) respectively, thereby realizing multi-physics coupling.

[0038] By integrating and solving the above three types of models, theoretical predictions of oil temperature distribution and hot spot locations of converter transformers in dynamic operating environments are obtained, providing physical priors and simulation benchmarks for data-driven models, thus realizing physical support for temperature prediction in hybrid drive structures.

[0039] This invention enables high-precision modeling and prediction of the operating state of converter transformers by coupling pre-defined fluid dynamics equations, electrostatic field equations, and heat conduction equations. First, using the fluid dynamics equations, the oil velocity field distribution is calculated based on the physical properties and volume force information of the oil. Next, using the electrostatic field equations, the potential distribution and electric field strength are calculated based on the dielectric constant and free charge density of the insulation system, thereby deriving the Joule heat source distribution. Finally, the oil velocity field is used as a convection term, and the Joule heat source distribution as a heat source term, input into the heat conduction equations. Combined with the oil's density, specific heat capacity, and thermal conductivity, the oil temperature distribution and hot spot temperatures are solved. This process not only ensures the physical consistency of the model but also improves prediction accuracy through the coupling of multiple physics fields, providing reliable theoretical support for the intelligent operation and maintenance of converter transformers.

[0040] S03: Based on real-time operating data and historical operating data, and combined with a preset time-series neural network model, a temperature residual prediction value is obtained; wherein, the time-series neural network model is obtained by training the residual between the first temperature prediction value output by the simulation model and the actual monitored value, as well as historical operating data, using an LSTM-Transformer network.

[0041] In a preferred embodiment of this invention, the step of obtaining the predicted temperature residual value based on real-time operating data and historical operating data, combined with a preset time-series neural network model, specifically involves: The temperature residual between the predicted values ​​and the actual monitored values ​​of the multiphysics high-fidelity simulation model is calculated, and this temperature residual is defined as the supervision label of the neural network.

[0042] The system acquires load current, oil flow rate, top oil temperature, and ambient temperature and humidity from real-time operating data, and combines this with historical operating data to construct a time-series feature sequence.

[0043] The time-series feature sequences are input into an LSTM-Transformer network for training. By optimizing the loss function and adjusting the network parameters, a trained time-series neural network model is obtained.

[0044] The LSTM-Transformer network is a hybrid neural network model combining Long Short-Term Memory (LSTM) and Transformer architectures, designed to leverage the strengths of both for processing sequential data. LSTM, a variant of Recurrent Neural Networks (RNNs), excels at capturing long-term dependencies in sequential data. Its unique gating mechanism effectively mitigates the vanishing or exploding gradient problems that traditional RNNs encounter when processing long sequences. The Transformer architecture, with its self-attention mechanism at its core, can process all elements in the sequence in parallel and dynamically assign different weights to each element, thus better capturing global dependencies and contextual information within the sequence. In the LSTM-Transformer network, the LSTM module typically performs preliminary processing on the sequential data to extract long-term dependency features. These features are then input into the Transformer module, where the self-attention mechanism further optimizes the feature representation, enhancing the model's ability to perceive global information. This combination enables LSTM-Transformer networks to achieve better performance than single network structures when dealing with complex sequence data tasks, such as natural language processing and time series prediction, while combining the ability of LSTM to model long-term dependencies with the ability of Transformer to capture global dependencies.

[0045] Using a trained temporal neural network model, predictions are made on real-time operating data to obtain predicted temperature residual values.

[0046] More specifically, to characterize the prediction errors of the physical model caused by boundary changes, parameter deviations, or operating condition disturbances during actual operation, a residual definition mechanism is introduced, and the calculation formula is as follows: (4) Among them, T phy (x,t) represents the predicted temperature output by the physical simulation model, T meas (x,t) represents the actual monitored temperature. This represents the temperature residual of the network prediction output. This residual term reflects the dynamic offset between the theoretical model and the actual operating state. As the core output of the neural network learning, it is used to fit external disturbances and model internal deficiencies of the system.

[0047] Secondly, multi-dimensional characteristic information of the converter transformer during operation is collected, including but not limited to: load current I(t), oil flow velocity v(t), and top oil temperature T. top The time series feature input vector consists of parameters such as (t), ambient temperature and humidity (RH(t)), etc. (5) Where d represents the feature dimension. This represents the combination of state features collected at time t, and this input sequence provides the necessary information for residual learning.

[0048] Based on the aforementioned feature inputs, a hybrid temporal neural network structure consisting of LSTM and Transformer (self-attention mechanism) modules is constructed. To fit and predict residuals The network prediction output can be represented as: (6) Wherein, ΔTpred(x,t) represents the temperature residual output by the network prediction, which means the difference between the predicted temperature Tphy(x,t) output by the physical simulation model and the actual monitored temperature Tmeas(x,t).

[0049] The residual mapping function for parameter θ is used to model long-term and short-term temporal dependencies through the LSTM module and extract global feature weights through the Transformer (self-attention mechanism), thereby effectively characterizing the impact of dynamic changes on temperature rise evolution.

[0050] This application significantly improves the accuracy and adaptability of converter transformer operating status prediction by constructing a time-series neural network model based on an LSTM-Transformer network. First, the residual between the first temperature prediction value output by the multiphysics high-fidelity simulation model and the actual monitored value is calculated and defined as a supervision label to guide network training. Next, real-time operating data is acquired and combined with historical operating data to construct a time-series feature sequence, providing rich input information for the network. Then, the time-series feature sequence is input into the LSTM-Transformer network for training. During training, the LSTM module is responsible for extracting long-short-term dependencies in the time-series feature sequence, while the Transformer module dynamically allocates global feature weights through a self-attention mechanism, thereby better capturing important information in the data. By optimizing the loss function using supervision labels, the network can continuously adjust its parameters until the value of the loss function no longer changes within a preset time period, at which point training stops, resulting in a well-trained time-series neural network model. This process not only makes full use of historical and real-time data, but also dynamically corrects the prediction errors of the physical model through deep learning technology, significantly improving the model's prediction accuracy and adaptability to complex operating conditions, and providing strong support for the intelligent operation and maintenance of converter transformers.

[0051] S04: Based on the residual between the first temperature prediction value and the actual monitored value, adjust the fusion weight of the first temperature prediction value and the temperature residual prediction value, and combine them with a preset weighted fusion algorithm to generate a mixed temperature prediction value.

[0052] In a preferred embodiment of this invention, the step of adjusting the fusion weights of the first temperature prediction value and the temperature residual prediction value based on the residual between the first temperature prediction value and the actual monitored value, and generating a mixed temperature prediction value by combining a preset weighted fusion algorithm, specifically involves: Real-time monitoring of the residual between the predicted values ​​and actual monitored values ​​of the multiphysics high-fidelity simulation model, the load fluctuation rate calculated based on the load current, and the real-time oil flow rate parameters.

[0053] The fusion weights of the physical model and the data model are dynamically allocated based on the absolute value of the residuals, the load volatility, and the degree to which the oil flow rate deviates from the rated value.

[0054] By using normalization constraints, we ensure that the sum of the weights of the physical model and the data model is 1, thus maintaining the rationality of the weight allocation.

[0055] Based on the adjusted fusion weights, and combining the predicted values ​​from the physical model with the temperature residual predicted values ​​from the data model, the final mixed temperature prediction value is calculated and generated.

[0056] More specifically, the residuals predicted by the neural network are weighted and fused with the predicted values ​​from the physical model to construct the final temperature prediction output expression: (7) in, This represents a hybrid predicted temperature, which combines the structured trends provided by the physical model with the nonlinear error compensation from neural network learning, effectively improving the overall accuracy and time-varying adaptability of the model.

[0057] The weighted fusion algorithm described is a method that generates a final output by assigning different weights to different inputs. In converter transformer operation status analysis, it combines the predicted values ​​from the physical model and the residual predicted values ​​from the data model, dynamically adjusting the weights based on the reliability of each model and the current operating status. For example, when the physical model performs better under certain operating conditions, it is given a higher weight; conversely, the weight of the data model is increased. This dynamic adjustment ensures the adaptability and accuracy of the model under different operating conditions, ultimately generating more reliable mixed temperature prediction values ​​for subsequent operation status identification.

[0058] Furthermore, this application introduces a dynamic weight adjustment mechanism, which consists of the following three sub-modules: State awareness module: Used for real-time monitoring of key system operating indicators, such as model prediction residuals (T). meas -T phy The system detects variations in operating status, such as load volatility and cooling status, and uses threshold judgment and trend recognition algorithms to identify areas of operational status variation, which serve as trigger signals for weight adjustments.

[0059] Weight Update Module: Based on the state-aware results, a fusion weight update function ω(t) is constructed to update the weight factors ω of the physical model and the temporal neural network model. phy (t) and ω data (t) Dynamic adjustments are made. These adjustments occur during the model prediction phase, rather than the model building or training phase, and aim to optimize the model fusion ratio in real time based on actual operating errors and system status, thereby improving overall prediction accuracy and adaptability.

[0060] Weight calculation satisfies: (8) Among them, T hybrid (t) represents the temperature prediction result of the fusion model at time t, where T is the temperature prediction result of the fusion model. phy (t) and T data (t) represent the independent predictions of the physical model and the data-driven model at time t, respectively; ω phy (t) and ω data (t) represents the corresponding dynamic fusion weights, reflecting the confidence or contribution of each model in the current state. The sum of the two satisfies the normalization constraint ω. phy (t)+ω data (t)=1. The system increases the contribution of the physical model when the data quality deteriorates, and enhances the data model compensation capability when the simulation scenario is simplified or the error is large, reflecting the time-varying flexibility and complementary advantages of the fusion strategy.

[0061] The self-feedback control module dynamically evaluates the error between the fused prediction output and the actual observations to achieve a closed-loop feedback mechanism for model performance. When the error is significant, it triggers a reassignment of weights and dynamic correction of network structure parameters, enabling the fusion system to achieve rapid convergence and adaptive repair capabilities.

[0062] This application effectively improves the accuracy and stability of converter transformer temperature prediction by dynamically adjusting fusion weights and combining them with a weighted fusion algorithm. Specifically, firstly, based on the residual between the first temperature prediction value and the actual monitored value, and combined with preset load fluctuation rate and cooling state change rate, the fusion weights of the first temperature prediction value and the temperature residual prediction value are dynamically allocated. This dynamic weight allocation mechanism can adaptively adjust the contribution ratio of the physical model and the data model according to changes in the current operating state. Subsequently, based on the allocated fusion weights, the first temperature prediction value, and the temperature residual prediction value, combined with preset weight allocation thresholds and a weighted fusion method, the final mixed temperature prediction value is generated. This fusion method not only fully utilizes the high fidelity of the physical model and the high adaptability of the data model, but also enhances the stability and responsiveness of the model under complex operating conditions through dynamic weight adjustment, thereby providing more accurate and reliable temperature prediction results for the intelligent operation and maintenance of converter transformers.

[0063] S05: Input the mixed temperature prediction value into the preset knowledge layer decision system, so that the knowledge layer decision system outputs the transformer operating status identification result based on the preset threshold judgment and the preset similarity matching algorithm.

[0064] In a preferred embodiment of this invention, the step of inputting the mixed temperature prediction value into a preset knowledge layer decision-making system, so that the knowledge layer decision-making system outputs the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm, specifically involves: The mixed temperature prediction values ​​are input into the preset knowledge layer decision system.

[0065] Based on a pre-defined expert rule base, threshold judgments are performed on the mixed temperature prediction values ​​to determine whether they exceed the pre-defined upper limit threshold for hot spot temperature and the temperature rise rate threshold.

[0066] The construction of the expert rule base is as follows: Based on long-term accumulated human experience and engineering knowledge in the operation and maintenance of converter transformers, an expert rule base was established. The rules cover standard expressions for temperature rise threshold judgment, load response threshold, and oil flow disturbance identification, and are in the following form: If Anomaly detection (9) Where, θ crit γ represents the upper limit threshold for hotspot temperature, and γ represents the temperature rise rate threshold. This rule-based model can output the temperature of the fusion model. It enables real-time monitoring and identification of abnormal trends, and has good interpretability and operability.

[0067] By acquiring historical transformer cases and combining them with a preset similarity matching algorithm, a similarity search is performed on the current mixed temperature prediction value to obtain the historical case most similar to the current state.

[0068] The preset similarity matching algorithm is a method for evaluating and comparing the similarity between data, determining the relationship between them by calculating the distance or similarity between the data. In converter transformer operating status identification, this algorithm finds the historical case most similar to the current state by comparing the current mixed temperature prediction value with the characteristics of historical transformer cases. For example, similarity can be calculated using metrics such as weighted Euclidean distance and Mahalanobis distance. The historical case with the highest similarity in the matching results will be used as a reference for the current state, and its corresponding response strategy can be used as an alternative output for the model's decision. This method can effectively utilize experience from historical data to enhance the model's decision-making ability in complex or ambiguous scenarios.

[0069] The process of obtaining historical transformer cases, combined with a preset similarity matching algorithm, involves performing a similarity search on the current mixed temperature prediction value, specifically as follows: To enhance the decision-making system's coverage in complex or ambiguous scenarios, a case-based reasoning (CBR) module driven by historical data is constructed. This is achieved by building a set of historical operational cases. Utilizing multi-dimensional working condition input Perform similarity retrieval to obtain the optimal match between the current state and past cases. Its formal definition is: (10) in, This represents a multidimensional similarity distance metric function, which can flexibly employ weighted Euclidean distance, Mahalanobis distance, or other custom metric systems. Matching result C * The associated response strategy will serve as an alternative output for the model's current decision.

[0070] The threshold judgment result and the similarity matching result are fused together, and the judgment weight is set by a preset priority scheduling function to realize the fusion judgment logic.

[0071] The final operating status identification result is output, which represents the operating category corresponding to the current operating status of the converter transformer and can serve as a direct basis for operation and maintenance scheduling, early warning control, or auxiliary diagnosis. This application uses a priority scheduling function. Set judgment weights to implement the fusion judgment logic: (11) Here, Decision represents the final operating condition identification or anomaly judgment result output by the knowledge layer decision-making system, used to characterize the operating category corresponding to the current operating status of the converter transformer. This result can serve as a direct basis for operation and maintenance scheduling, early warning control, or auxiliary diagnosis. As a fusion function, it supports multiple fusion strategies, including expert priority coverage, logical rule and case matching voting, and confidence weighting, to ensure that the output has consistency, stability and reasoning transparency.

[0072] This application achieves accurate identification and intelligent decision-making regarding the operating status of converter transformers by inputting the predicted mixed temperature value into a knowledge-layer decision-making system. First, utilizing a pre-defined expert rule base, the knowledge-layer decision-making system determines whether the predicted mixed temperature value exceeds preset hotspot temperature upper limit thresholds and temperature rise rate thresholds, thus obtaining a threshold judgment result. Next, combining historical transformer cases, a pre-defined similarity matching algorithm is used to perform a similarity search on the current predicted mixed temperature value, obtaining a similarity matching result. Finally, the threshold judgment result and the similarity matching result are fused, and a pre-defined priority scheduling function is used to set the judgment weights, outputting the final transformer operating status identification result. This process not only improves the accuracy and reliability of the prediction results but also enhances the model's interpretability and engineering adaptability, providing strong technical support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems.

[0073] As a preferred embodiment of this invention, in order to verify the effectiveness of the transformer operating status identification proposed in this application, a ±550kV converter transformer was selected as the experimental object. The performance of the first output result (PM) of the multiphysics high-fidelity simulation model, the second output result (DM) of the time sequence neural network model, and the third output result (HMM) which is a weighted combination of the output result of the multiphysics high-fidelity simulation model and the output result of the time sequence neural network model were compared and tested under normal, fault and unknown abnormal operating conditions.

[0074] The experimental data are divided into two categories: the first is multi-dimensional operating data collected by actual sensors, including winding hot spot temperature, oil temperature, oil flow rate, current and environmental parameters, covering the entire process of typical faults; the second is multi-physics simulation data generated based on the finite volume method (FVM), simulating non-ideal working conditions such as oil quality deterioration and heat dissipation degradation.

[0075] Figure 2 The implementation process and effects of transformer operating status identification are described in detail. Figure 2 (a) presents the raw data collected by the temperature sensor, which includes multi-dimensional characteristic information such as load current, oil flow rate, oil temperature and ambient temperature and humidity. The fluctuations and noise of the curves show the complexity of the measurement. Figure 2 (b) shows the temperature data generated by the multiphysics high-fidelity simulation model (PM). The overall trend is basically consistent with the sensor data, but the fluctuations are smaller and the curves are smoother, which verifies that the simulation model has good predictive fitting ability. Figure 2Figure (c) shows the temperature residuals calculated by the temporal neural network model (DM). This model, based on an LSTM-Transformer network, was trained using the residuals between the first temperature prediction output from a multiphysics high-fidelity simulation model (PM) and the actual monitored values, as well as historical operating data. The temperature difference curve in Figure (c) reveals the change in the model's prediction error over time. A significant deviation can be observed in the early stages of operation, which then gradually converges to a more stable range of differences, demonstrating the model's adaptive correction capability.

[0076] Figure 2 (d) shows the operational status assessment results based on a multiphysics high-fidelity simulation model (PM). This model uses only simulation data to predict the operational status of the converter transformer, where 1 represents a fault state and 0 represents a normal state. As can be seen from the figure, the system frequently experiences fault states in the initial stage, and then the state stabilizes to normal, indicating that the simulation model can identify and correct abnormal operating conditions, demonstrating a certain early warning capability. Figure 2 Figure (f) shows the operational status assessment results based on a hybrid model (HMM). The hybrid model combines the outputs of a multiphysics high-fidelity simulation model and a temporal neural network model, generating a mixed temperature prediction value through a weighted fusion algorithm, and using this value to determine the operational status. In the figure, 1 represents a fault state and 0 represents a normal state. Compared to Figure (d), Figure (f) shows more stable or accurate operational status identification results, thanks to the hybrid model's effective suppression of measurement noise and instantaneous fluctuations while maintaining temperature trend characteristics. Figure 2 (e) illustrates the output of the hybrid temperature predictions, specifically generated by a hybrid model (HMM). This HMM combines the outputs of a multiphysics high-fidelity simulation model (PM) with the outputs of a temporal neural network model (DM). Through a weighted fusion algorithm, the first temperature prediction from the simulation model is integrated with the temperature residual predictions from the temporal neural network model, generating smoother and more continuous hybrid temperature predictions. These hybrid temperature predictions not only preserve the trend characteristics of temperature changes but also effectively suppress measurement noise and instantaneous fluctuations, thereby improving the accuracy and reliability of converter transformer operating status identification.

[0077] In terms of model structure, PM is based on heat-fluid theory for modeling, DM uses LSTM-Transformer to build a temporal network, and HMM integrates the two and introduces the exponential moving average (EMA) algorithm to achieve adaptive weight adjustment, combining the advantages of physical modeling with data-driven capabilities.

[0078] As shown in Table 1, the experimental results demonstrate that the Hidden Markov Model (HMM) significantly outperforms the control model in terms of average error, prediction stability, and temperature rise rate judgment. Especially in the early stages of a fault, it can promptly identify deviations in the temperature rise trend and dynamically adjust model weights, effectively improving response speed and accuracy. Furthermore, the HMM maintains good predictive performance under sudden disturbances and unknown anomalies. It can combine residual-triggered knowledge rule bases to achieve intelligent "emergency load reduction" alarms and match optimal historical case suggestions, exhibiting stronger generalization ability and intelligent adaptability.

[0079] Table 1. Temperature rise rate error and fault warning delay in the output results of different models. This application achieves high-precision prediction and intelligent identification of the operating status of converter transformers by constructing a multi-physics high-fidelity simulation model and combining real-time and historical operating data. First, by coupling multiple pre-set equations, the simulation model can output multiple predicted values, providing physical priors for subsequent analysis. Second, a time-series neural network model based on LSTM-Transformer network dynamically corrects the prediction error of the simulation model, further improving prediction accuracy. By monitoring the residual between predicted and actual monitored values ​​in real time and dynamically adjusting the fusion weights, this application can adaptively optimize the prediction results, enhancing the model's adaptability and stability under complex operating conditions. Finally, the mixed temperature prediction values ​​are input into a knowledge-layer decision-making system, combined with an expert rule base and case reasoning module, to achieve rapid and accurate identification of the transformer's operating status. This process not only improves the accuracy and real-time performance of predictions but also enhances the model's interpretability and engineering adaptability, providing strong support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems and effectively solving the problem of inaccurate and inefficient identification of transformer operating status in existing technologies.

[0080] Example 2 Please refer to Figure 3 This is a transformer operation status identification device provided in the embodiments of this application.

[0081] In this embodiment, the transformer operating status identification device includes an acquisition module 10, a simulation module 20, a prediction module 30, a fusion module 40, and an identification module 50.

[0082] The acquisition module 10 is used to acquire real-time operating data, historical operating data and physical attribute parameters of the converter transformer.

[0083] In a preferred embodiment of this invention, the acquisition of real-time operating data, historical operating data, and physical attribute parameters of the converter transformer specifically includes: The operating data of the converter transformer is collected in real time through a sensor network, including multi-dimensional characteristic information such as load current, oil flow rate, oil temperature, and ambient temperature and humidity.

[0084] Historical operating data of converter transformers are extracted from historical databases, including past load changes, oil temperature changes, and environmental conditions.

[0085] Obtain the physical property parameters of the converter transformer, including the density, specific heat capacity, thermal conductivity, viscosity, and flow velocity field of the cooling oil medium, the internal structural parameters of the converter transformer, boundary conditions, and electrical excitation sources (voltage and current).

[0086] The simulation module 20 is used to solve the multiphysics problem by coupling the preset equations based on the physical property parameters and a preset multiphysics high-fidelity simulation model, thereby obtaining multiple predicted values ​​output by the multiphysics high-fidelity simulation model.

[0087] In a preferred embodiment of this invention, the step of solving the multiphysics problem by coupling preset equations based on the physical property parameters and a preset multiphysics high-fidelity simulation model to obtain multiple predicted values ​​output by the multiphysics high-fidelity simulation model is as follows: The oil flow temperature rise coupling equation is used to perform thermal-flow field coupling modeling of the converter transformer winding region, and the oil temperature and hot spot temperature are calculated.

[0088] More specifically, a thermal-fluid field coupling model is performed on the winding region of the converter transformer using an oil flow temperature rise coupling equation to achieve quantitative prediction of oil temperature and hot spot temperature. The oil flow temperature rise coupling equation is as follows: (12) Where ρ is the oil density, c p Let Q be the specific heat capacity, k be the thermal conductivity, v be the flow velocity field, and Q be the flow velocity field. loss It represents the winding loss power density, reflecting the localized heating of components such as coils.

[0089] A cooling oil flow control model was established, and the incompressible Navier-Stokes momentum conservation equation was used to describe the flow characteristics of the oil in the winding channel and to calculate the velocity field distribution.

[0090] More specifically, to characterize the impact of oil flow characteristics in the complex channels inside the transformer on heat dissipation efficiency, a cooling oil flow control model is further established, using the incompressible Navier-Stokes momentum conservation equation for modeling. The calculation is as follows: (13) in, This represents the oil velocity vector, p represents the oil pressure, and μ represents the dynamic viscosity. This refers to volume forces, including gravity and possible electromagnetic disturbance forces.

[0091] An electric and magnetic field distribution model is introduced, and the internal heat source distribution is estimated using the static Maxwell equations, which are then used as the heat source term in the heat conduction model.

[0092] More specifically, to accurately estimate the internal heat source distribution and provide the heat source term Qloss to the heat conduction model, an electric field and magnetic field distribution model is introduced to simulate local induction and leakage magnetic flux, using the static Maxwell equations as follows: (14) in, Represents the electric potential distribution function. ρ represents the dielectric constant of an insulating system. e This represents the free charge density. Formulas (12), (13), and (14) represent the heat conduction, fluid dynamics, and electrostatic field distribution models, respectively, and there is a clear input-output relationship between the three: the flow velocity v calculated by formula (13) is used in the convection term in formula (12) and affects the temperature distribution; the electric potential calculated by formula (3) Electric field can be obtained Then calculate Joule heat Q, as a heat source loss Input into formula (12). Therefore, the temperature calculation depends on the flow rate and the electric heat source, which are provided by formulas (13) and (14) respectively, thereby achieving multiphysics coupling.

[0093] By integrating and solving the above three types of models, theoretical predictions of oil temperature distribution and hot spot locations of converter transformers in dynamic operating environments are obtained, providing physical priors and simulation benchmarks for data-driven models, thus realizing physical support for temperature prediction in hybrid drive structures.

[0094] This invention enables high-precision modeling and prediction of the operating state of converter transformers by coupling pre-defined fluid dynamics equations, electrostatic field equations, and heat conduction equations. First, using the fluid dynamics equations, the oil velocity field distribution is calculated based on the physical properties and volume force information of the oil. Next, using the electrostatic field equations, the potential distribution and electric field strength are calculated based on the dielectric constant and free charge density of the insulation system, thereby deriving the Joule heat source distribution. Finally, the oil velocity field is used as a convection term, and the Joule heat source distribution as a heat source term, input into the heat conduction equations. Combined with the oil's density, specific heat capacity, and thermal conductivity, the oil temperature distribution and hot spot temperatures are solved. This process not only ensures the physical consistency of the model but also improves prediction accuracy through the coupling of multiple physics fields, providing reliable theoretical support for the intelligent operation and maintenance of converter transformers.

[0095] The prediction module 30 is used to obtain the temperature residual prediction value based on real-time operation data and historical operation data, combined with a preset time-series neural network model; wherein, the time-series neural network model is obtained by training the residual between the first temperature prediction value output by the simulation model and the actual monitoring value and historical operation data based on the LSTM-Transformer network.

[0096] In a preferred embodiment of this invention, the step of obtaining the predicted temperature residual value based on real-time operating data and historical operating data, combined with a preset time-series neural network model, specifically involves: The temperature residual between the predicted values ​​and the actual monitored values ​​of the multiphysics high-fidelity simulation model is calculated, and this temperature residual is defined as the supervision label of the neural network.

[0097] The system acquires load current, oil flow rate, top oil temperature, and ambient temperature and humidity from real-time operating data, and combines this with historical operating data to construct a time-series feature sequence.

[0098] The time-series feature sequences are input into an LSTM-Transformer network for training. By optimizing the loss function and adjusting the network parameters, a trained time-series neural network model is obtained.

[0099] The LSTM-Transformer network is a hybrid neural network model combining Long Short-Term Memory (LSTM) and Transformer architectures, designed to leverage the strengths of both for processing sequential data. LSTM, a variant of Recurrent Neural Networks (RNNs), excels at capturing long-term dependencies in sequential data. Its unique gating mechanism effectively mitigates the vanishing or exploding gradient problems that traditional RNNs encounter when processing long sequences. The Transformer architecture, with its self-attention mechanism at its core, can process all elements in the sequence in parallel and dynamically assign different weights to each element, thus better capturing global dependencies and contextual information within the sequence. In the LSTM-Transformer network, the LSTM module typically performs preliminary processing on the sequential data to extract long-term dependency features. These features are then input into the Transformer module, where the self-attention mechanism further optimizes the feature representation, enhancing the model's ability to perceive global information. This combination enables LSTM-Transformer networks to achieve better performance than single network structures when dealing with complex sequence data tasks, such as natural language processing and time series prediction, while combining the ability of LSTM to model long-term dependencies with the ability of Transformer to capture global dependencies.

[0100] Using a trained temporal neural network model, predictions are made on real-time operating data to obtain predicted temperature residual values.

[0101] More specifically, to characterize the prediction errors of the physical model caused by boundary changes, parameter deviations, or operating condition disturbances during actual operation, a residual definition mechanism is introduced, and the calculation formula is as follows: (15) Among them, T phy (x,t) represents the predicted temperature output by the physical simulation model, T meas (x,t) represents the actual monitored temperature. This represents the temperature residual of the network prediction output. This residual term reflects the dynamic offset between the theoretical model and the actual operating state. As the core output of the neural network learning, it is used to fit external disturbances and model internal deficiencies of the system.

[0102] Secondly, multi-dimensional characteristic information of the converter transformer during operation is collected, including but not limited to: load current I(t), oil flow velocity v(t), and top oil temperature T. top The time series feature input vector consists of parameters such as (t), ambient temperature and humidity (RH(t)), etc. (16) Where d represents the feature dimension. This represents the combination of state features collected at time t, and this input sequence provides the necessary information for residual learning.

[0103] Based on the aforementioned feature inputs, a hybrid temporal neural network structure consisting of LSTM and Transformer (self-attention mechanism) modules is constructed. To fit and predict residuals The network prediction output can be represented as: (17) Wherein, ΔTpred(x,t) represents the temperature residual output by the network prediction, which means the difference between the predicted temperature Tphy(x,t) output by the physical simulation model and the actual monitored temperature Tmeas(x,t).

[0104] The residual mapping function for parameter θ is used to model long-term and short-term temporal dependencies through the LSTM module and extract global feature weights through the Transformer (self-attention mechanism), thereby effectively characterizing the impact of dynamic changes on temperature rise evolution.

[0105] This application significantly improves the accuracy and adaptability of converter transformer operating status prediction by constructing a time-series neural network model based on an LSTM-Transformer network. First, the residual between the first temperature prediction value output by the multiphysics high-fidelity simulation model and the actual monitored value is calculated and defined as a supervision label to guide network training. Next, real-time operating data is acquired and combined with historical operating data to construct a time-series feature sequence, providing rich input information for the network. Then, the time-series feature sequence is input into the LSTM-Transformer network for training. During training, the LSTM module is responsible for extracting long-short-term dependencies in the time-series feature sequence, while the Transformer module dynamically allocates global feature weights through a self-attention mechanism, thereby better capturing important information in the data. By optimizing the loss function using supervision labels, the network can continuously adjust its parameters until the value of the loss function no longer changes within a preset time period, at which point training stops, resulting in a well-trained time-series neural network model. This process not only makes full use of historical and real-time data, but also dynamically corrects the prediction errors of the physical model through deep learning technology, significantly improving the model's prediction accuracy and adaptability to complex operating conditions, and providing strong support for the intelligent operation and maintenance of converter transformers.

[0106] The fusion module 40 is used to adjust the fusion weight of the first temperature prediction value and the temperature residual prediction value according to the residual between the first temperature prediction value and the actual monitoring value, and generate a mixed temperature prediction value by combining the preset weighted fusion algorithm.

[0107] In a preferred embodiment of this invention, the step of adjusting the fusion weights of the first temperature prediction value and the temperature residual prediction value based on the residual between the first temperature prediction value and the actual monitored value, and generating a mixed temperature prediction value by combining a preset weighted fusion algorithm, specifically involves: Real-time monitoring of the residual between the predicted values ​​and actual monitored values ​​of the multiphysics high-fidelity simulation model, the load fluctuation rate calculated based on the load current, and the real-time oil flow rate parameters.

[0108] The fusion weights of the physical model and the data model are dynamically allocated based on the absolute value of the residuals, the load volatility, and the degree to which the oil flow rate deviates from the rated value.

[0109] By using normalization constraints, we ensure that the sum of the weights of the physical model and the data model is 1, thus maintaining the rationality of the weight allocation.

[0110] Based on the adjusted fusion weights, and combining the predicted values ​​from the physical model with the temperature residual predicted values ​​from the data model, the final mixed temperature prediction value is calculated and generated.

[0111] More specifically, the residuals predicted by the neural network are weighted and fused with the predicted values ​​from the physical model to construct the final temperature prediction output expression: (18) in, This represents a hybrid predicted temperature, which combines the structured trends provided by the physical model with the nonlinear error compensation from neural network learning, effectively improving the overall accuracy and time-varying adaptability of the model.

[0112] The weighted fusion algorithm described is a method that generates a final output by assigning different weights to different inputs. In converter transformer operation status analysis, it combines the predicted values ​​from the physical model and the residual predicted values ​​from the data model, dynamically adjusting the weights based on the reliability of each model and the current operating status. For example, when the physical model performs better under certain operating conditions, it is given a higher weight; conversely, the weight of the data model is increased. This dynamic adjustment ensures the adaptability and accuracy of the model under different operating conditions, ultimately generating more reliable mixed temperature prediction values ​​for subsequent operation status identification.

[0113] Furthermore, this application introduces a dynamic weight adjustment mechanism, which consists of the following three sub-modules: State awareness module: Used for real-time monitoring of key system operating indicators, such as model prediction residuals (T). meas -T phy The system detects variations in operating status, such as load volatility and cooling status, and uses threshold judgment and trend recognition algorithms to identify areas of operational status variation, which serve as trigger signals for weight adjustments.

[0114] Weight Update Module: Based on the state-aware results, a fusion weight update function ω(t) is constructed to update the weight factors ω of the physical model and the temporal neural network model. phy (t) and ω data (t) Dynamic adjustments are made. These adjustments occur during the model prediction phase, rather than the model building or training phase, and aim to optimize the model fusion ratio in real time based on actual operating errors and system status, thereby improving overall prediction accuracy and adaptability.

[0115] Weight calculation satisfies: (19) Among them, T hybrid (t) represents the temperature prediction result of the fusion model at time t, where T is the temperature prediction result of the fusion model. phy (t) and T data (t) represent the independent predictions of the physical model and the data-driven model at time t, respectively; ω phy (t) and ω data(t) represents the corresponding dynamic fusion weights, reflecting the confidence or contribution of each model in the current state. The sum of the two satisfies the normalization constraint ω. phy (t)+ω data (t)=1. The system increases the contribution of the physical model when the data quality deteriorates, and enhances the data model compensation capability when the simulation scenario is simplified or the error is large, reflecting the time-varying flexibility and complementary advantages of the fusion strategy.

[0116] The self-feedback control module dynamically evaluates the error between the fused prediction output and the actual observations to achieve a closed-loop feedback mechanism for model performance. When the error is significant, it triggers a reassignment of weights and dynamic correction of network structure parameters, enabling the fusion system to achieve rapid convergence and adaptive repair capabilities.

[0117] This application effectively improves the accuracy and stability of converter transformer temperature prediction by dynamically adjusting fusion weights and combining them with a weighted fusion algorithm. Specifically, firstly, based on the residual between the first temperature prediction value and the actual monitored value, and combined with preset load fluctuation rate and cooling state change rate, the fusion weights of the first temperature prediction value and the temperature residual prediction value are dynamically allocated. This dynamic weight allocation mechanism can adaptively adjust the contribution ratio of the physical model and the data model according to changes in the current operating state. Subsequently, based on the allocated fusion weights, the first temperature prediction value, and the temperature residual prediction value, combined with preset weight allocation thresholds and a weighted fusion method, the final mixed temperature prediction value is generated. This fusion method not only fully utilizes the high fidelity of the physical model and the high adaptability of the data model, but also enhances the stability and responsiveness of the model under complex operating conditions through dynamic weight adjustment, thereby providing more accurate and reliable temperature prediction results for the intelligent operation and maintenance of converter transformers.

[0118] The identification module 50 is used to input the mixed temperature prediction value into a preset knowledge layer decision system, so that the knowledge layer decision system can output the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm.

[0119] In a preferred embodiment of this invention, the step of inputting the mixed temperature prediction value into a preset knowledge layer decision-making system, so that the knowledge layer decision-making system outputs the transformer operating status identification result based on a preset threshold judgment and a preset similarity matching algorithm, specifically involves: The mixed temperature prediction values ​​are input into the preset knowledge layer decision system.

[0120] Based on a pre-defined expert rule base, threshold judgments are performed on the mixed temperature prediction values ​​to determine whether they exceed the pre-defined upper limit threshold for hot spot temperature and the temperature rise rate threshold.

[0121] The construction of the expert rule base is as follows: Based on long-term accumulated human experience and engineering knowledge in the operation and maintenance of converter transformers, an expert rule base was established. The rules cover standard expressions for temperature rise threshold judgment, load response threshold, and oil flow disturbance identification, and are in the following form: If Anomaly detection (20) Where, θ crit γ represents the upper limit threshold for hotspot temperature, and γ represents the temperature rise rate threshold. This rule-based model can output the temperature of the fusion model. It enables real-time monitoring and identification of abnormal trends, and has good interpretability and operability.

[0122] By acquiring historical transformer cases and combining them with a preset similarity matching algorithm, a similarity search is performed on the current mixed temperature prediction value to obtain the historical case most similar to the current state.

[0123] The preset similarity matching algorithm is a method for evaluating and comparing the similarity between data, determining the relationship between them by calculating the distance or similarity between the data. In converter transformer operating status identification, this algorithm finds the historical case most similar to the current state by comparing the current mixed temperature prediction value with the characteristics of historical transformer cases. For example, similarity can be calculated using metrics such as weighted Euclidean distance and Mahalanobis distance. The historical case with the highest similarity in the matching results will be used as a reference for the current state, and its corresponding response strategy can be used as an alternative output for the model's decision. This method can effectively utilize experience from historical data to enhance the model's decision-making ability in complex or ambiguous scenarios.

[0124] The process of obtaining historical transformer cases, combined with a preset similarity matching algorithm, involves performing a similarity search on the current mixed temperature prediction value, specifically as follows: To enhance the decision-making system's coverage in complex or ambiguous scenarios, a case-based reasoning (CBR) module driven by historical data is constructed. This is achieved by building a set of historical operational cases. Utilizing multi-dimensional working condition input Perform similarity retrieval to obtain the optimal match between the current state and past cases. Its formal definition is: (twenty one) in, This represents a multidimensional similarity distance metric function, which can flexibly employ weighted Euclidean distance, Mahalanobis distance, or other custom metric systems. Matching result C * The associated response strategy will serve as an alternative output for the model's current decision.

[0125] The threshold judgment result and the similarity matching result are fused together, and the judgment weight is set by a preset priority scheduling function to realize the fusion judgment logic.

[0126] The final operating status identification result is output, which represents the operating category corresponding to the current operating status of the converter transformer and can serve as a direct basis for operation and maintenance scheduling, early warning control, or auxiliary diagnosis. This application uses a priority scheduling function. Set judgment weights to implement the fusion judgment logic: (twenty two) Here, Decision represents the final operating condition identification or anomaly judgment result output by the knowledge layer decision-making system, used to characterize the operating category corresponding to the current operating status of the converter transformer. This result can serve as a direct basis for operation and maintenance scheduling, early warning control, or auxiliary diagnosis. As a fusion function, it supports multiple fusion strategies, including expert priority coverage, logical rule and case matching voting, and confidence weighting, to ensure that the output has consistency, stability and reasoning transparency.

[0127] This application achieves accurate identification and intelligent decision-making regarding the operating status of converter transformers by inputting the predicted mixed temperature value into a knowledge-layer decision-making system. First, utilizing a pre-defined expert rule base, the knowledge-layer decision-making system determines whether the predicted mixed temperature value exceeds preset hotspot temperature upper limit thresholds and temperature rise rate thresholds, thus obtaining a threshold judgment result. Next, combining historical transformer cases, a pre-defined similarity matching algorithm is used to perform a similarity search on the current predicted mixed temperature value, obtaining a similarity matching result. Finally, the threshold judgment result and the similarity matching result are fused, and a pre-defined priority scheduling function is used to set the judgment weights, outputting the final transformer operating status identification result. This process not only improves the accuracy and reliability of the prediction results but also enhances the model's interpretability and engineering adaptability, providing strong technical support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems.

[0128] As a preferred embodiment of this invention, in order to verify the effectiveness of the transformer operating status identification proposed in this application, a ±550kV converter transformer was selected as the experimental object. The performance of the first output result (PM) of the multiphysics high-fidelity simulation model, the second output result (DM) of the time sequence neural network model, and the third output result (HMM) which is a weighted combination of the output result of the multiphysics high-fidelity simulation model and the output result of the time sequence neural network model were compared and tested under normal, fault and unknown abnormal operating conditions.

[0129] The experimental data are divided into two categories: the first is multi-dimensional operating data collected by actual sensors, including winding hot spot temperature, oil temperature, oil flow rate, current and environmental parameters, covering the entire process of typical faults; the second is multi-physics simulation data generated based on the finite volume method (FVM), simulating non-ideal working conditions such as oil quality deterioration and heat dissipation degradation.

[0130] Figure 2 The implementation process and effects of converter transformer operation status identification are described in detail. Figure 2 (a) presents the raw data collected by the temperature sensor, which includes multi-dimensional characteristic information such as load current, oil flow rate, oil temperature and ambient temperature and humidity. The fluctuations and noise of the curves show the complexity of the measurement. Figure 2 (b) shows the temperature data generated by the multiphysics high-fidelity simulation model (PM). The overall trend is basically consistent with the sensor data, but the fluctuations are smaller and the curves are smoother, which verifies that the simulation model has good predictive fitting ability. Figure 2 Figure (c) shows the temperature residuals calculated by the temporal neural network model (DM). This model, based on an LSTM-Transformer network, was trained using the residuals between the first temperature prediction output from a multiphysics high-fidelity simulation model (PM) and the actual monitored values, as well as historical operating data. The temperature difference curve in Figure (c) reveals the change in the model's prediction error over time. A significant deviation can be observed in the early stages of operation, which then gradually converges to a more stable range of differences, demonstrating the model's adaptive correction capability.

[0131] Figure 2 (d) shows the operational status assessment results based on a multiphysics high-fidelity simulation model (PM). This model uses only simulation data to predict the operational status of the converter transformer, where 1 represents a fault state and 0 represents a normal state. As can be seen from the figure, the system frequently experiences fault states in the initial stage, and then the state stabilizes to normal, indicating that the simulation model can identify and correct abnormal operating conditions, demonstrating a certain early warning capability. Figure 2 Figure (f) shows the operational status assessment results based on a hybrid model (HMM). The hybrid model combines the outputs of a multiphysics high-fidelity simulation model and a temporal neural network model, generating a mixed temperature prediction value through a weighted fusion algorithm, and using this value to determine the operational status. In the figure, 1 represents a fault state and 0 represents a normal state. Compared to Figure (d), Figure (f) shows more stable or accurate operational status identification results, thanks to the hybrid model's effective suppression of measurement noise and instantaneous fluctuations while maintaining temperature trend characteristics. Figure 2(e) illustrates the output of the hybrid temperature predictions, specifically generated by a hybrid model (HMM). This HMM combines the outputs of a multiphysics high-fidelity simulation model (PM) with the outputs of a temporal neural network model (DM). Through a weighted fusion algorithm, the first temperature prediction from the simulation model is integrated with the temperature residual predictions from the temporal neural network model, generating smoother and more continuous hybrid temperature predictions. These hybrid temperature predictions not only preserve the trend characteristics of temperature changes but also effectively suppress measurement noise and instantaneous fluctuations, thereby improving the accuracy and reliability of converter transformer operating status identification.

[0132] In terms of model structure, PM is based on heat-fluid theory for modeling, DM uses LSTM-Transformer to build a temporal network, and HMM integrates the two and introduces the exponential moving average (EMA) algorithm to achieve adaptive weight adjustment, combining the advantages of physical modeling with data-driven capabilities.

[0133] As shown in Table 2, the experimental results demonstrate that the Hidden Markov Model (HMM) significantly outperforms the control model in terms of average error, prediction stability, and temperature rise rate judgment. Especially in the early stages of a fault, it can promptly identify deviations in the temperature rise trend and dynamically adjust model weights, effectively improving response speed and accuracy. Furthermore, the HMM maintains good predictive performance under sudden disturbances and unknown anomalies. It can combine residual-triggered knowledge rule bases to achieve intelligent "emergency load reduction" alarms and match optimal historical case suggestions, exhibiting stronger generalization ability and intelligent adaptability.

[0134] Table 2. Temperature rise rate error and fault warning delay in the output results of different models. The transformer operating status identification device of this application integrates multiple functional modules to achieve high-precision prediction and intelligent identification of the operating status of converter transformers. The acquisition module collects real-time operating data, historical operating data, and physical attribute parameters, providing comprehensive data support for subsequent analysis. The simulation module utilizes these parameters to perform coupled solution through a multi-physics high-fidelity simulation model, outputting multiple predicted values ​​to ensure physical consistency of the predictions. The prediction module further combines a time-series neural network model to dynamically correct the prediction errors of the simulation model, improving prediction accuracy. The fusion module dynamically adjusts the fusion weights based on the residuals, optimizing the prediction results and enhancing the model's adaptability. Finally, the identification module inputs the mixed temperature prediction values ​​into a knowledge-layer decision system, combining threshold judgment and similarity matching algorithms to output the operating status identification result, achieving rapid and accurate diagnosis of the transformer's operating status. This device not only improves the accuracy and real-time performance of predictions but also enhances the model's interpretability and engineering adaptability, providing strong technical support for the intelligent operation and maintenance of converter transformers in high-voltage direct current transmission systems.

[0135] Example 3: This application provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the aforementioned method for identifying the operating status of a transformer. The transformer operation status identification method, if implemented as a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0136] Example 4 This application provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the transformer operation status identification methods as described in Embodiment 1.

[0137] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method of identifying an operating state of a transformer, characterized by, The method comprises the following steps: obtaining real-time operation data, historical operation data and physical property parameters of a converter transformer; based on the physical property parameters, combining a preset multi-physics high-fidelity simulation model, and solving the multi-physics by coupling preset equations to obtain multiple predicted values output by the multi-physics high-fidelity simulation model; based on the real-time operation data and the historical operation data, combining a preset time series neural network model to obtain a temperature residual prediction value; wherein the time series neural network model is obtained by training the residual between a first temperature prediction value output by the simulation model and an actual monitoring value and historical operation data based on an LSTM-Transformer network; based on the residual between the first temperature prediction value and the actual monitoring value, adjusting the fusion weight of the first temperature prediction value and the temperature residual prediction value, combining a preset weighted fusion algorithm to generate a hybrid temperature prediction value; inputting the hybrid temperature prediction value into a preset knowledge layer decision system to make the knowledge layer decision system output a transformer operation state recognition result based on a preset threshold value judgment and a preset similarity matching algorithm.

2. The operating state recognition method of a transformer according to claim 1, characterized in that, The method comprises the following steps: The preset equations include a preset fluid mechanics equation, a preset electrostatic field equation and a preset heat conduction equation; based on the preset fluid mechanics equation, the flow velocity field distribution of the oil is calculated based on the preset physical property parameters of the oil and the preset volume force information; based on the preset electrostatic field equation, the potential distribution and the electric field intensity are calculated based on the preset dielectric constant of the insulation system and the preset free charge density; the Joule heat source distribution is calculated based on the potential distribution and the electric field intensity; based on the preset heat conduction equation, the velocity vector of the oil is input into the convection term of the heat conduction equation, the Joule heat source distribution is input into the heat source term of the heat conduction equation, and the oil temperature distribution and the hot spot temperature are solved based on the preset density of the oil, the preset specific heat capacity and the preset thermal conductivity.

3. The method of claim 1, wherein The time series neural network model is obtained by training the residual between the first temperature prediction value output by the simulation model and the actual monitoring value and the historical operation data based on an LSTM-Transformer network, and the method comprises the following steps: calculating the residual between the first temperature prediction value and the actual monitoring value, and defining the residual as a supervision label; obtaining real-time operation data, combining historical operation data to construct a time series feature sequence; The time sequence feature sequence is input into an LSTM-Transformer network for training, so that the LSTM-Transformer network extracts long and short term dependencies of the time sequence feature sequence through an LSTM module, distributes global feature weights through a self-attention mechanism of a Transformer module, and optimizes a loss function using the supervision label during the training process until the value of the loss function no longer changes within a preset time period, the training is stopped, and a trained time sequence neural network model is obtained.

4. The operating state recognition method of a transformer according to claim 1, characterized in that, The fusion weight of the first temperature prediction value and the temperature residual prediction value is adjusted according to the residual of the first temperature prediction value and the actual monitoring value, and a hybrid temperature prediction value is generated by combining a preset weighted fusion algorithm, specifically: According to the residual, the fusion weight of the first temperature prediction value and the temperature residual prediction value is allocated in combination with a preset load fluctuation rate and a preset cooling state change rate; According to the fusion weight, the first temperature prediction value and the temperature residual prediction value, a hybrid temperature prediction value is generated by combining a preset weight distribution threshold and a preset weighted fusion method.

5. The operating state recognition method of a transformer according to claim 1, characterized in that, The hybrid temperature prediction value is input into a preset knowledge layer decision system, so that the knowledge layer decision system outputs a transformer operating state recognition result based on a preset threshold judgment and a preset similarity matching algorithm, specifically: The hybrid temperature prediction value is input into a preset knowledge layer decision system, so that the knowledge layer decision system judges whether the hybrid temperature prediction value exceeds a preset hotspot temperature upper threshold and a temperature rise rate threshold based on a preset expert rule base, to obtain a threshold judgment result; A historical transformer case is obtained, and a similarity search is performed on the current hybrid temperature prediction value in combination with a preset similarity matching algorithm, to obtain a similarity matching result; The threshold judgment result and the similarity matching result are fused and processed, a judgment weight is set through a preset priority scheduling function, and a transformer operating state recognition result is output.

6. The operating state recognition method of a transformer according to claim 1, characterized in that, The real-time operating data and physical property parameters of the converter transformer are obtained, specifically: The real-time operating data includes load current, oil flow rate, oil temperature and environmental temperature and humidity; The physical property parameters include cooling oil medium density, specific heat capacity, thermal conductivity, viscosity, flow rate field, converter transformer internal structure parameters, boundary conditions and electrical excitation source.

7. An operating state recognition device for a transformer, characterized by It includes: An acquisition module is configured to acquire real-time operating data, historical operating data and physical property parameters of a converter transformer; A simulation module is configured to obtain a plurality of prediction values output by a multi-physical field high-fidelity simulation model by coupling a plurality of preset equations for multi-physical field solving in combination with the physical property parameters and a preset multi-physical field high-fidelity simulation model; A prediction module is configured to obtain a temperature residual prediction value in combination with a preset time sequence neural network model based on real-time operating data and historical operating data; wherein the time sequence neural network model is obtained by training an LSTM-Transformer network based on a residual of a first temperature prediction value output by the simulation model and actual monitoring values and historical operating data. The fusion module is configured to adjust a fusion weight of the first temperature prediction value and the temperature residual prediction value according to a residual of the first temperature prediction value and an actual monitoring value, combine a preset weighted fusion algorithm, and generate a mixed temperature prediction value; The identification module is configured to input the mixed temperature prediction value into a preset knowledge layer decision system, so that the knowledge layer decision system outputs a transformer operation state identification result based on a preset threshold judgment and a preset similarity matching algorithm.

8. The operating state recognition device of a transformer according to claim 7, characterized in that The time sequence neural network model is obtained by training a residual of the first temperature prediction value and an actual monitoring value and historical operation data output by the simulation model based on an LSTM-Transformer network, specifically as follows: A residual between the first temperature prediction value and the actual monitoring value is calculated, and the residual is defined as a supervision label. Real-time operation data is acquired, and historical operation data is combined to construct a time sequence feature sequence. The time sequence feature sequence is input into the LSTM-Transformer network for training, so that the LSTM-Transformer network extracts a long-short term dependency relationship of the time sequence feature sequence through an LSTM module, assigns a global feature weight through a self-attention mechanism of a Transformer module, and optimizes a loss function using the supervision label in a training process until a value of the loss function no longer changes within a preset time period, and the training is stopped to obtain a trained time sequence neural network model.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium includes a stored computer program, wherein the computer program controls a device where the computer readable storage medium is located to execute the transformer operation state identification method according to any one of claims 1 to 6 when the computer program is running.

10. A terminal device, comprising: The computer readable storage medium includes a stored computer program, wherein the computer program controls a device where the computer readable storage medium is located to execute the transformer operation state identification method according to any one of claims 1 to 6 when the computer program is running. The computer readable storage medium includes a stored computer program, wherein the computer program controls a device where the computer readable storage medium is located to execute the transformer operation state identification method according to any one of claims 1 to 6 when the computer program is running.

Images