Transformer state prediction and adaptive control method based on digital twinning

Abstract

The invention relates to the technical field of transformers, in particular to a transformer state prediction and adaptive control method based on digital twinning, which realizes real-time monitoring, fault prediction and adaptive control of a transformer by constructing a high-precision digital twinning model and a deep reinforcement learning algorithm. According to the method, finite element analysis software is used for modeling the physical structure of the transformer, a high-precision three-dimensional geometric model is generated, circuit simulation software is used for modeling electrical characteristics, a comprehensive digital twinborn model is formed, and the actual operation state of the transformer is comprehensively reflected. A fault module is added in the digital twin model, potential fault modes such as coil short circuit and iron core oversaturation are simulated, and faults are recognized and prevented in advance.

Description

Technical Field

[0001] The present invention relates to the technical field of transformers, and in particular to a transformer state prediction and adaptive control method based on digital twins. Background Art

[0002] At present, transformers are key equipment in power systems, and their safe and stable operation is crucial to the reliability of the entire power grid. Traditional transformer monitoring and management methods mainly rely on regular inspections and maintenance, which makes it difficult to grasp the operating status of transformers in real time and cannot predict and prevent potential faults in a timely manner. With the continuous expansion of the scale and increase in complexity of power systems, higher requirements are placed on the intelligent management of transformers. In recent years, digital twin technology has been widely used in the industrial field. Digital twin technology can achieve real-time monitoring, state prediction and optimization control of physical equipment by constructing a virtual model of physical entities. Applying digital twin technology to transformer management can build a digital twin model of the transformer, including detailed simulation of the physical structure of the transformer and its electrical characteristics. By collecting operating data in real time through sensors installed on the transformer and transmitting these data to the digital twin model, real-time monitoring of the transformer's operating status can be achieved.

[0003] However, it is not enough to rely solely on digital twin models for condition monitoring. It is also necessary to combine advanced artificial intelligence algorithms to achieve intelligent condition prediction and adaptive control. As a cutting-edge technology in the field of artificial intelligence, deep reinforcement learning algorithms have the ability to autonomously learn and optimize decisions, and can be used to generate adaptive control strategies for transformers. Through interaction with the environment, deep reinforcement learning algorithms can continuously optimize control strategies and improve the operating efficiency and reliability of transformers.

[0004] In terms of fault prediction, traditional methods mainly rely on expert experience and simple threshold judgments, which makes it difficult to accurately predict complex fault conditions. Using digital twin models for fault prediction can more comprehensively consider various operating parameters and environmental factors of the transformer, and improve the accuracy of fault prediction. When there is a significant deviation between the model prediction value and the actual measurement value, the system can issue an early warning signal in time to provide decision support for operation and maintenance personnel. Traditional transformer control methods often adopt fixed control strategies, which are difficult to adapt to complex and changeable operating environments. The adaptive control method based on digital twins and deep reinforcement learning can automatically adjust the operating parameters of the transformer, such as cooling settings, excitation current, etc., according to the fault prediction results, to avoid the occurrence of potential faults and improve the operating reliability and life of the transformer.

[0005] However, the existing transformer state prediction and control methods still have some shortcomings. First, the accuracy and real-time performance of the digital twin model need to be improved, requiring more efficient modeling methods and faster data processing capabilities. Second, the application of deep reinforcement learning algorithms in complex transformer control environments is not mature enough, and more actual operating data is needed to train and optimize the algorithms. Furthermore, the accuracy and robustness of the fault prediction model need to be further improved, especially for the prediction of some rare faults. Finally, the real-time and reliability of the adaptive control strategy need to be verified and optimized in practical applications.

[0006] In view of the above problems, the existing technology needs to be improved urgently. Summary of the invention

[0007] The present invention provides a transformer state prediction and adaptive control method based on digital twins. The present invention realizes real-time monitoring, fault prediction and adaptive control of transformers by constructing a high-precision digital twin model and a deep reinforcement learning algorithm. The present invention uses finite element analysis software to model the physical structure of the transformer, generates a high-precision three-dimensional geometric model, and uses circuit simulation software to model the electrical characteristics to form a comprehensive digital twin model that fully reflects the actual operating state of the transformer. A fault module is added to the digital twin model to simulate potential fault modes, such as coil short circuit, core oversaturation, etc., to identify and prevent faults in advance.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a transformer state prediction and adaptive control method based on digital twin technology, comprising the following steps: constructing a digital twin model of the transformer, which includes a detailed simulation of the physical structure of the transformer and its electrical characteristics; collecting the operating data of the transformer in real time through sensors installed on the transformer, and transmitting the operating data of the transformer to the digital twin model; generating an adaptive control strategy using a deep reinforcement learning algorithm, optimizing the control strategy through interaction with the environment, monitoring the operating status of the transformer in real time, and using the digital twin model to predict faults. When there is a significant deviation between the model prediction value and the actual measurement value, an early warning signal is issued; according to the fault prediction results, the operating parameters of the transformer are automatically adjusted to avoid the occurrence of potential faults.

[0009] Furthermore, the construction of the digital twin model of the transformer includes the following steps: using finite element analysis software to model the physical structure of the transformer to generate a high-precision three-dimensional geometric model; using circuit simulation software to model the electrical characteristics of the transformer to generate a detailed electrical simulation model; coupling the physical model and the electrical model to form a comprehensive digital twin model that fully reflects the actual operating status of the transformer; adding a fault module to the digital twin model to simulate potential fault modes, and the fault module includes: coil short circuit, core oversaturation, increased winding loss, and flux imbalance.

[0010] Furthermore, the training of the deep reinforcement learning algorithm includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; using historical data to initially train the DRL algorithm so that the model can understand the normal behavior patterns of the transformer under different operating conditions; during actual operation, continuously updating the model parameters through the online learning algorithm to adapt to the actual operating environment of the transformer; the online learning algorithm uses the DDPG algorithm, which includes five stages: building a network structure; network training; parameter optimization; and evaluating network performance.

[0011] Furthermore, the network structure construction stage is as follows: building an actor-critic network architecture, the actor network generates control actions, and the critic network evaluates the value of each action; setting the parameters of the target network as a backup of the actor network to avoid overfitting; using normal distribution to initialize network parameters; and using the experience replay method to increase the diversity of training samples.

[0012] Furthermore, the interaction with the environment includes the following steps: binding the digital twin model with the actual transformer device to establish a communication link based on synchronized time; uploading the real-time collected operation data to the digital twin model to update the model status; generating a control strategy based on the current model status using a deep reinforcement learning algorithm; and sending the control strategy to the actual transformer device to enable it to operate according to the control strategy.

[0013] Furthermore, the fault prediction includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; training a fault prediction model, and using machine learning technology to establish a mapping relationship between fault characteristics and operating data; in the actual fault prediction stage, inputting operating data into the trained fault prediction model to generate fault prediction results; and processing potential faults according to the fault prediction results.

[0014] Furthermore, the training fault prediction model includes the following steps: building a machine learning model and selecting XGBoost or a neural network as a prediction model; setting the input of the fault prediction model as operating data and the output as the probability of fault; dividing the training set and the test set according to historical data, and adjusting the model parameters in a cross-validation manner to select an appropriate parameter combination; using the training set data to train the fault prediction model, and using the test set to evaluate the performance of the model.

[0015] Furthermore, the automatic adjustment of the transformer's operating parameters includes the following steps: selecting appropriate parameters as the target of adaptive control according to the transformer operating environment; generating a set of candidate control strategies when the fault prediction module detects a potential fault problem; performing multi-objective optimization according to the candidate strategies and their corresponding prediction results; and executing feedforward control to achieve rapid response, while monitoring the fault condition in real time and adjusting the control strategy.

[0016] Furthermore, the multi-objective optimization adopts the Pareto optimization method, including the following steps: defining optimization objectives, including reducing the temperature peak, reducing the maximum temperature of the winding, and optimizing the transformer power spectrum; defining multiple performance indicators based on these objectives; setting constraints, including the operating status of the transformer and the selection of control strategies; optimizing multiple objectives using genetic algorithms, particle swarm algorithms, or evolutionary algorithms to generate a set of Pareto solution sets; and selecting a set of optimal solutions that meet the constraints.

[0017] The advantages of the present invention are:

[0018] 1. The present invention provides a transformer state prediction and adaptive control method based on digital twin technology, which realizes real-time monitoring, fault prediction and adaptive control of the transformer by constructing a high-precision digital twin model and a deep reinforcement learning algorithm. The present invention uses finite element analysis software to model the physical structure of the transformer, generates a high-precision three-dimensional geometric model, and uses circuit simulation software to model the electrical characteristics to form a comprehensive digital twin model that fully reflects the actual operating state of the transformer. A fault module is added to the digital twin model to simulate potential fault modes, such as coil short circuit, core oversaturation, etc., to identify and prevent faults in advance.

[0019] 2. The present invention uses a deep reinforcement learning algorithm to generate an adaptive control strategy, and continuously optimizes the control strategy through interaction with the environment. The DRL algorithm can handle complex, nonlinear control tasks and adapt to the operating requirements of the transformer under different working conditions. The model parameters are continuously updated through the online learning algorithm to ensure that the control strategy can adapt to the actual operating environment, thereby improving real-time performance and accuracy. The sensors installed on the transformer collect operating data in real time and transmit it to the digital twin model to achieve real-time monitoring. The fault prediction model is trained using machine learning technology, and fault prediction results are generated by inputting real-time operating data, so that potential faults can be discovered in a timely manner and measures can be taken to avoid failures.

[0020] 3. In order to achieve rapid response, the present invention adopts a feedforward control strategy. When the fault prediction module detects a potential fault, a candidate control strategy is generated, and multi-objective optimization is performed, such as reducing the temperature peak, reducing the maximum temperature of the winding, etc. Through feedforward control, the system can quickly adjust the control strategy, monitor the fault situation in real time, and ensure stable operation. By automatically adjusting parameters such as load distribution, cooling settings, and excitation current, the operating state of the transformer is optimized, the operating efficiency and reliability are improved, overheating and overload problems are avoided, and the service life of the transformer is extended. In summary, the present invention significantly improves the operating efficiency and reliability of the transformer, avoids potential faults, and extends the service life. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Embodiment 1:

[0023] The present invention provides a transformer state prediction and adaptive control method based on digital twin technology proposed in this application, comprising the following steps: constructing a digital twin model of the transformer, which includes a detailed simulation of the physical structure of the transformer and its electrical characteristics; collecting the operating data of the transformer in real time through sensors installed on the transformer, and transmitting the operating data of the transformer to the digital twin model; generating an adaptive control strategy using a deep reinforcement learning algorithm, optimizing the control strategy through interaction with the environment, monitoring the operating state of the transformer in real time, and using the digital twin model to predict faults. When there is a significant deviation between the model prediction value and the actual measurement value, an early warning signal is issued; and automatically adjusting the operating parameters of the transformer according to the fault prediction results to avoid the occurrence of potential faults.

[0024] Among them, the construction of the digital twin model of the transformer can be done by using finite element analysis software to model the physical structure of the transformer and generate a high-precision three-dimensional geometric model; using circuit simulation software to model the electrical characteristics of the transformer and generate a detailed electrical simulation model; coupling the physical model and the electrical model to form a comprehensive digital twin model that fully reflects the actual operating status of the transformer. Furthermore, a fault module is added to the digital twin model to simulate potential fault modes, such as coil short circuit, core oversaturation, increased winding loss, and flux imbalance.

[0025] The training of deep reinforcement learning algorithm can be done by collecting historical operation data of transformer, including normal operation data and fault data; using historical data to initially train the DRL algorithm, so that the model can understand the normal behavior mode of transformer under different working conditions; in the actual operation process, the model parameters are continuously updated through online learning algorithm to adapt to the actual operating environment of transformer. The online learning algorithm can use DDPG algorithm, including five stages: building network structure, determining reward function, network training, parameter optimization and evaluating network performance.

[0026] Specifically, in the network structure construction stage, the actor-critic network architecture can be built, the actor network generates control actions, and the critic network evaluates the value of each action; the parameters of the target network are set as a backup of the actor network to avoid overfitting; the network parameters are initialized using normal distribution; and the experience replay method is used to increase the diversity of training samples. In the reward function determination stage, multiple indicators of transformer performance can be used as evaluation criteria, such as core loss, winding temperature, harmonic distortion, vibration frequency, etc.

[0027] The interaction with the environment includes binding the digital twin model with the actual transformer equipment to establish a communication link based on synchronized time; uploading the real-time collected operation data to the digital twin model and updating the model status; generating a control strategy based on the current model status using a deep reinforcement learning algorithm; and sending the control strategy to the actual transformer equipment so that it can operate according to the control strategy.

[0028] Fault prediction includes collecting historical operating data of the transformer, including normal operating data and fault data; training the fault prediction model and using machine learning technology to establish a mapping relationship between fault characteristics and operating data; in the actual fault prediction stage, inputting operating data into the trained fault prediction model to generate fault prediction results; and processing potential faults based on the fault prediction results.

[0029] Automatically adjusting the operating parameters of the transformer includes selecting appropriate parameters as the target of adaptive control according to the transformer operating environment; generating a set of candidate control strategies when the fault prediction module detects a potential fault problem; performing multi-objective optimization based on the candidate strategies and their corresponding prediction results; and executing feedforward control to achieve rapid response, while monitoring the fault condition in real time and adjusting the control strategy.

[0030] As a preferred implementation, multi-objective optimization can adopt the Pareto optimization method, including defining optimization objectives, such as reducing the temperature peak, reducing the maximum temperature of the winding, and optimizing the transformer power spectrum; defining multiple performance indicators based on these objectives; setting constraints, such as the operating status of the transformer and the selection of control strategies; optimizing multiple objectives using genetic algorithms, particle swarm algorithms, or evolutionary algorithms to generate a set of Pareto solution sets; and selecting a set of optimal solutions that meet the constraints.

[0031] Therefore, the present application solves the technical problems of transformer state prediction and adaptive control by constructing a digital twin model, generating an adaptive control strategy using a deep reinforcement learning algorithm, monitoring the operating status of the transformer in real time and predicting faults, and automatically adjusting the operating parameters of the transformer according to the fault prediction results. This technical solution, through the combination of a digital twin model and a deep reinforcement learning algorithm, can monitor and predict the operating status of the transformer in real time, and automatically adjust the operating parameters when a fault occurs, thereby avoiding the occurrence of potential faults and improving the operating efficiency and safety of the transformer. Compared with the prior art, the technical solution of the present application has higher prediction accuracy and faster response speed, and can effectively reduce the failure rate and maintenance cost of the transformer.

[0032] Furthermore, the present application also proposes that constructing a digital twin model of a transformer includes the following steps: using finite element analysis software to model the physical structure of the transformer to generate a high-precision three-dimensional geometric model; using circuit simulation software to model the electrical characteristics of the transformer to generate a detailed electrical simulation model; coupling the physical model and the electrical model to form a comprehensive digital twin model that fully reflects the actual operating status of the transformer; adding a fault module to the digital twin model to simulate potential fault modes, and the fault module includes: coil short circuit, core oversaturation, increased winding loss, and flux imbalance.

[0033] Specifically, finite element analysis software such as ANSYS or COMSOL can be used to perform high-precision three-dimensional modeling of the physical structure of the transformer. Circuit simulation software such as PSpice or MATLAB / Simulink can be used to perform detailed simulation of the electrical characteristics of the transformer. When coupling the physical model and the electrical model, data interface technology can be used to ensure data synchronization and consistency between the two models. The simulation of the fault module can be achieved through preset fault scenarios and parameter settings. For example, the coil short circuit can be simulated by adjusting the resistance value of the coil, and the core oversaturation can be simulated by changing the magnetic permeability of the core.

[0034] Therefore, the present application can fully reflect the actual operating status of the transformer by constructing a high-precision digital twin model, and add a fault module to the model to simulate potential fault modes. This not only improves the simulation accuracy of the model, but also provides reliable data support for subsequent fault prediction and adaptive control. Compared with the prior art, the technical solution of the present application can more accurately predict the operating status of the transformer, detect potential faults in a timely manner, and make adjustments through adaptive control strategies to avoid the occurrence of faults and improve the operational reliability and safety of the transformer.

[0035] Furthermore, the present application also proposes that the training of the deep reinforcement learning algorithm includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; using historical data to perform initial training on the DRL algorithm so that the model can understand the normal behavior patterns of the transformer under different operating conditions; during actual operation, continuously updating the model parameters through an online learning algorithm to adapt to the actual operating environment of the transformer; the online learning algorithm uses the DDPG algorithm, which includes five stages: building a network structure; determining a reward function; network training; parameter optimization; and evaluating network performance.

[0036] Specifically, the network structure construction stage is: build the actor-critic network architecture, the actor network generates control actions, and the critic network evaluates the value of each action; set the parameters of the target network as a backup of the actor network to avoid overfitting; use normal distribution to initialize network parameters; use the experience replay method to increase the diversity of training samples. The reward function determination stage is: the reward function uses multiple indicators of transformer performance as evaluation criteria, including core loss, winding temperature, harmonic distortion, and vibration frequency; the reward function uses the following formula: where n is the total number of samples in a single training cycle, Vn is the transformer performance of the nth sample, Vset is the threshold for performance setting, P is the reward for a single sample, μ is the weight parameter, α is the bias parameter, j is the number of iterations, and αj is the weight parameter at the jth iteration.

[0037] Therefore, this application can effectively improve the accuracy and efficiency of transformer state prediction and adaptive control through the training of deep reinforcement learning algorithms. Specifically, by collecting and analyzing historical operating data, the model can better understand the behavior patterns of the transformer under different working conditions, thereby improving the accuracy of the prediction. The application of online learning algorithms enables the model to update parameters in real time and adapt to the actual operating environment of the transformer, further improving the adaptability and effectiveness of the control strategy. In addition, by constructing an actor-critic network architecture and setting target network parameters, the problem of model overfitting is avoided and the stability and reliability of the model are improved. The setting of the reward function ensures that the control strategy can comprehensively consider multiple performance indicators of the transformer, thereby achieving a more comprehensive and optimized control effect.

[0038] Specific implementation method for claim 4

[0039] Furthermore, the present application also proposes that the network structure construction stage is as follows: building an actor-critic network architecture, the actor network generates control actions, and the critic network evaluates the value of each action; setting the parameters of the target network as a backup of the actor network to avoid overfitting; using normal distribution to initialize network parameters; and using the experience replay method to increase the diversity of training samples.

[0040] Specifically, the actor-critic network architecture is a commonly used deep reinforcement learning model, in which the actor network is responsible for generating control actions, while the critic network is responsible for evaluating the value of each action. In this way, the model is able to optimize the control strategy through trial and error. In order to avoid overfitting of the model during training, the parameters of the target network are set as a backup of the actor network, which can maintain a certain stability during training. In addition, the initialization of the network parameters adopts a normal distribution to ensure that the initial values ​​of the parameters have a certain degree of randomness, thereby increasing the generalization ability of the model. The experience replay method increases the diversity of training samples by storing and replaying past experiences, which helps the model better learn and adapt to different operating environments.

[0041] Therefore, this application effectively improves the performance of deep reinforcement learning algorithms in transformer state prediction and adaptive control by constructing an actor-critic network architecture and combining target network parameter backup, normal distribution initialization of network parameters, and experience replay method. Specifically, this technical solution solves the problem of model overfitting and insufficient diversity of training samples in complex operating environments, thereby improving the optimization effect of the control strategy and the adaptability of the model.

[0042] Furthermore, the present application also proposes that the reward function uses multiple indicators of transformer performance as evaluation criteria, including core loss, winding temperature, harmonic distortion, and vibration frequency; the reward function adopts the following formula: where n is the total number of samples in a single training cycle, Vn is the transformer performance of the nth sample, Vset is the threshold for performance setting, P is the reward for a single sample, μ is the weight parameter, α is the bias parameter, j is the number of iterations, and αj is the weight parameter at the jth iteration.

[0043] Among them, core loss refers to the energy loss caused by hysteresis and eddy current effects during the magnetization process of the transformer core; winding temperature refers to the heat generated by the current passing through the transformer winding during operation; harmonic distortion refers to the non-fundamental wave components contained in the transformer output voltage or current; vibration frequency refers to the vibration frequency generated by electromagnetic and mechanical forces during the operation of the transformer. These indicators are collected in real time by sensors and transmitted to the digital twin model to evaluate the operating status of the transformer.

[0044] Furthermore, the reward function can be implemented through software programming. Specifically, during the training process of the deep reinforcement learning algorithm, the above indicators can be used as input, the reward value of each sample can be obtained through the calculation formula, and the control strategy can be adjusted according to the reward value. As a preferred implementation, the Python programming language can be used in combination with a deep learning framework such as TensorFlow or PyTorch to realize the calculation of the reward function and the optimization of the control strategy.

[0045] Therefore, by using a multi-index reward function, the operating state of the transformer can be more comprehensively evaluated, thereby generating a better adaptive control strategy and improving the operating efficiency and reliability of the transformer. Compared with the prior art, the technical solution of the present application can more accurately predict and control the operating state of the transformer, reduce the possibility of failure, and extend the service life of the transformer.

[0046] Furthermore, the present application also proposes to bind the digital twin model with the actual transformer equipment and establish a communication link based on synchronized time; upload the real-time collected operation data to the digital twin model and update the model status; generate a control strategy based on the current model status using a deep reinforcement learning algorithm; and send the control strategy to the actual transformer equipment so that it operates according to the control strategy.

[0047] Specifically, the binding process can be achieved through the Internet of Things technology to ensure data synchronization between the digital twin model and the actual transformer equipment. The real-time collected operating data includes but is not limited to parameters such as temperature, pressure, and current. After these data are collected by sensors, they are transmitted to the digital twin model through wireless or wired networks. After receiving the data, the model updates the status and generates the corresponding control strategy using the pre-trained deep reinforcement learning algorithm. The generated control strategy is sent to the actual transformer equipment through the communication link, and the equipment makes corresponding operational adjustments based on the received strategy.

[0048] As a preferred implementation, the deep reinforcement learning algorithm can use the DDPG (Deep Deterministic Policy Gradient) algorithm, which generates control actions through an actor-critic network architecture and evaluates the value of each action. In the process of updating the model state, the experience replay method can be used to increase the diversity of training samples, thereby improving the generalization ability of the model.

[0049] Therefore, through the above technical means, this application can realize real-time monitoring and adaptive control of the transformer status, effectively avoid the occurrence of potential faults, and improve the operating efficiency and safety of the transformer. Compared with the prior art, this application realizes more accurate and real-time control strategy generation and execution through the close integration of digital twin models and actual equipment, significantly improving the response speed and control accuracy of the system.

[0050] Furthermore, the present application also proposes that fault prediction includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; training a fault prediction model, and using machine learning technology to establish a mapping relationship between fault characteristics and operating data; in the actual fault prediction stage, inputting operating data into the trained fault prediction model to generate fault prediction results; and processing potential faults based on the fault prediction results.

[0051] Among them, training the fault prediction model includes the following steps: building a machine learning model and selecting XGBoost or neural network as the prediction model; setting the input of the fault prediction model as operating data and the output as the probability of failure; dividing the training set and test set according to historical data, and adjusting the model parameters by cross-validation to select the most suitable parameter combination; using the training set data to train the fault prediction model, and using the test set to evaluate the performance of the model.

[0052] Specifically, the machine learning model can choose XGBoost or neural network, both of which are excellent in processing nonlinear relationships and high-dimensional data. The input data includes the operating data of the transformer, such as current, voltage, temperature, etc., and the output is the probability of failure, which helps to evaluate the status of the transformer in real time. By adjusting the model parameters through cross-validation, the generalization ability of the model on different data sets can be ensured, thereby improving the accuracy of the prediction.

[0053] Furthermore, the training set data is used to train the model, while the test set data is used to evaluate the performance of the model. Evaluation indicators can include accuracy, recall, F1 score, etc. These indicators can fully reflect the prediction ability of the model. In this way, the reliability and effectiveness of the fault prediction model in practical applications can be ensured.

[0054] Therefore, the fault prediction method of the present application can effectively improve the operational safety of the transformer by collecting historical data, training prediction models, and predicting and processing potential faults in real time. Compared with the prior art, the method of the present application can not only monitor the status of the transformer in real time, but also predict potential faults in advance through machine learning technology, thereby avoiding the occurrence of faults and improving the operating efficiency and reliability of the transformer.

[0055] Furthermore, the present application also proposes that training a fault prediction model includes the following steps: building a machine learning model and selecting XGBoost or a neural network as a prediction model; setting the input of the fault prediction model as operating data and the output as the probability of failure; dividing the training set and the test set according to historical data, and adjusting the model parameters by cross-validation to select the most appropriate parameter combination; using the training set data to train the fault prediction model, and using the test set to evaluate the performance of the model.

[0056] Among them, XGBoost is an algorithm based on gradient boosting decision trees, which is suitable for processing high-dimensional data and large-scale data sets, and has high prediction accuracy and speed. Neural network is a computational model that simulates the structure of human brain neurons. It realizes complex nonlinear mapping through the connection and activation function of multiple layers of neurons, and is suitable for processing complex pattern recognition and prediction problems. In practical applications, the appropriate prediction model can be selected according to the operating data characteristics and computing resources of the transformer.

[0057] Furthermore, cross-validation is a model evaluation method that divides the data set into multiple subsets, uses one of the subsets as the test set and the remaining subsets as the training set in turn, to reduce the risk of model overfitting and improve the generalization ability of the model. During the model parameter adjustment process, different parameter combinations can be systematically traversed through methods such as grid search or random search to find the optimal parameter configuration.

[0058] Therefore, through the above steps, an efficient and accurate fault prediction model can be constructed, which can monitor the operating status of the transformer in real time, predict potential faults in time, and provide reliable data support for the adaptive control of the transformer.

[0059] Specifically, this technical solution achieves in-depth analysis and fault prediction of transformer operation data by introducing machine learning technology, solving the shortcomings of traditional methods in fault prediction accuracy and real-time performance. Compared with existing technologies, this solution not only improves the accuracy of fault prediction, but also enhances the robustness and adaptability of the model through cross-validation and model parameter adjustment, providing strong support for intelligent operation and maintenance of transformers.

[0060] Furthermore, the present application also proposes that automatically adjusting the operating parameters of the transformer includes the following steps: selecting appropriate parameters as the target of adaptive control according to the operating environment of the transformer; generating a set of candidate control strategies when the fault prediction module detects a potential fault problem; performing multi-objective optimization according to the candidate strategies and their corresponding prediction results; and executing feedforward control to achieve rapid response, while monitoring the fault condition in real time and adjusting the control strategy.

[0061] Specifically, the parameters include cooling settings and excitation current. Among them, the cooling settings include controlling the temperature and wind speed of the cooling channel to optimize the heat dissipation efficiency of the transformer; the excitation current quickly reduces or increases the amplitude of the excitation current to enable the transformer to quickly enter a light load or heavy load state. Multi-objective optimization adopts the Pareto optimization method, which includes the following steps: define optimization goals, including reducing the temperature peak, reducing the maximum temperature of the winding, and optimizing the transformer power spectrum; define multiple performance indicators based on these goals; set constraints, including the operating status of the transformer and the choice of control strategy; use genetic algorithms, particle swarm algorithms or evolutionary algorithms to optimize multiple objectives and generate a set of Pareto solution sets; select a set of optimal solutions to meet the constraints.

[0062] Therefore, the stage of automatically adjusting the working parameters of the transformer is as follows: obtaining fault prediction results, including potential fault types, time and location of fault occurrence; automatically generating multiple control strategies based on the fault prediction results, which are designed to mitigate the impact of the fault or avoid the occurrence of the fault; starting the feedforward control mechanism to quickly adjust the operating parameters of the transformer according to the generated control strategy, including load distribution, cooling settings, and excitation current; by adjusting the control strategy, the operating state of the transformer can be restored to normal within the expected time; when the fault prediction module detects a new fault, the optimization algorithm will generate a new control strategy based on the new fault information.

[0063] As a preferred implementation, the present application can effectively avoid the occurrence of potential faults and improve the operational stability and reliability of the transformer by automatically adjusting the operating parameters of the transformer. Specifically, by real-time monitoring of the operating status of the transformer and automatically adjusting the operating parameters according to the fault prediction results, it is possible to quickly respond to and handle potential fault problems, thereby reducing the impact of the fault on the operation of the transformer. Compared with the prior art, the technical solution of the present application has a higher response speed and stronger adaptive ability, and can better ensure the safe operation of the transformer.

[0064] Furthermore, the present application also proposes that automatically adjusting the operating parameters of the transformer includes the following steps: selecting appropriate parameters as the target of adaptive control according to the operating environment of the transformer; generating a set of candidate control strategies when the fault prediction module detects a potential fault problem; performing multi-objective optimization according to the candidate strategies and their corresponding prediction results; and executing feedforward control to achieve rapid response, while monitoring the fault condition in real time and adjusting the control strategy.

[0065] Specifically, the parameters include cooling settings and excitation current. Among them, the cooling settings include controlling the temperature and wind speed of the cooling channel to optimize the heat dissipation efficiency of the transformer. The excitation current quickly reduces or increases the amplitude of the excitation current, so that the transformer quickly enters a light load or heavy load state. As a preferred embodiment, the multi-objective optimization adopts the Pareto optimization method, including the following steps: defining optimization goals, including reducing the temperature peak, reducing the maximum temperature of the winding, and optimizing the transformer power spectrum; defining multiple performance indicators based on these goals; setting constraints, including the operating status of the transformer and the choice of control strategy; using genetic algorithms, particle swarm algorithms or evolutionary algorithms to optimize multiple objectives and generate a set of Pareto solution sets; selecting a set of optimal solutions to meet the constraints.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer state prediction and adaptive control method based on digital twin technology, characterized in that: The following steps are involved: Build a digital twin model of the transformer, which includes a detailed simulation of the transformer’s physical structure and its electrical characteristics; The operating data of the transformer is collected in real time through sensors installed on the transformer, and the operating data of the transformer is transmitted to the digital twin model; Use deep reinforcement learning algorithms to generate adaptive control strategies, optimize control strategies through interaction with the environment, monitor the operating status of the transformer in real time, and use digital twin models to predict faults. When there is a significant deviation between the model prediction value and the actual measurement value, an early warning signal is issued; based on the fault prediction results, the operating parameters of the transformer are automatically adjusted to avoid potential faults.

2. According to claim 1, a transformer state prediction and adaptive control method based on digital twin technology is characterized in that: The construction of the digital twin model of the transformer includes the following steps: The physical structure of the transformer is modeled using finite element analysis software to generate a high-precision three-dimensional geometric model. The electrical characteristics of the transformer are modeled using circuit simulation software to generate a detailed electrical simulation model. The physical model and the electrical model are coupled to form a comprehensive digital twin model that fully reflects the actual operating status of the transformer. A fault module is added to the digital twin model to simulate potential failure modes, including: coil short circuit, core oversaturation, increased winding loss, and magnetic flux imbalance.

3. The transformer state prediction and adaptive control method based on digital twin technology according to claim 1 is characterized in that: The training of the deep reinforcement learning algorithm includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; using historical data to initially train the DRL algorithm so that the model can understand the normal behavior patterns of the transformer under different operating conditions; in the actual operation process, continuously updating the model parameters through the online learning algorithm to adapt to the actual operating environment of the transformer; the online learning algorithm uses the DDPG algorithm, which includes five stages: building a network structure; network training; parameter optimization; and evaluating network performance.

4. The transformer state prediction and adaptive control method based on digital twin technology according to claim 3 is characterized in that: The network structure construction stage is as follows: building an actor-critic network architecture, the actor network generates control actions, and the critic network evaluates the value of each action; setting the parameters of the target network as a backup of the actor network to avoid overfitting; initializing the network parameters using a normal distribution; and using an experience replay method to increase the diversity of training samples.

5. The transformer state prediction and adaptive control method based on digital twin technology according to claim 1 is characterized in that: The interaction with the environment includes the following steps: binding the digital twin model with the actual transformer device to establish a communication link based on synchronized time; uploading the real-time collected operation data to the digital twin model to update the model status; generating a control strategy based on the current model status using a deep reinforcement learning algorithm; and sending the control strategy to the actual transformer device to enable it to operate according to the control strategy.

6. The transformer state prediction and adaptive control method based on digital twin technology according to claim 1 is characterized in that: The fault prediction includes the following steps: collecting historical operating data of the transformer, including normal operating data and fault data; training a fault prediction model, and using machine learning technology to establish a mapping relationship between fault characteristics and operating data; in the actual fault prediction stage, inputting operating data into the trained fault prediction model to generate fault prediction results; and processing potential faults according to the fault prediction results.

7. The transformer state prediction and adaptive control method based on digital twin technology according to claim 6 is characterized in that: The training fault prediction model includes the following steps: building a machine learning model and selecting XGBoost or a neural network as a prediction model; setting the input of the fault prediction model as operating data and the output as the probability of fault; dividing the training set and the test set according to historical data, and adjusting the model parameters in a cross-validation manner to select an appropriate parameter combination; using the training set data to train the fault prediction model, and using the test set to evaluate the performance of the model.

8. The transformer state prediction and adaptive control method based on digital twin technology according to claim 1 is characterized in that: The automatic adjustment of the transformer's operating parameters includes the following steps: selecting appropriate parameters as the target of adaptive control according to the transformer's operating environment; generating a set of candidate control strategies when a fault prediction module detects a potential fault problem; performing multi-objective optimization according to the candidate strategies and their corresponding prediction results; and executing feedforward control to achieve rapid response, while monitoring the fault situation in real time and adjusting the control strategy.

9. The transformer state prediction and adaptive control method based on digital twin technology according to claim 8 is characterized in that: The multi-objective optimization adopts the Pareto optimization method, including the following steps: defining optimization objectives, including reducing the temperature peak, reducing the maximum temperature of the winding, and optimizing the transformer power spectrum; defining multiple performance indicators based on these objectives; setting constraints, including the operating state of the transformer and the selection of control strategies; optimizing the multiple objectives using a genetic algorithm, a particle swarm algorithm, or an evolutionary algorithm to generate a set of Pareto solution sets; and selecting a set of optimal solutions to meet the constraints.