General control method of transformer and reactor comprehensive test system

Abstract

The present invention provides a general control method for a transformer and reactor comprehensive test system, which relates to the field of electrical control technology, including deploying a multi-dimensional sensor array to obtain power evolution data; establishing a multi-scale digital twin model to generate a power feature mapping matrix; using a graph neural network to realize cross-domain feature extraction, establishing a self-calibration optimization algorithm to generate an evolutionary power digital twin model; and constructing a hybrid power-classical fault diagnosis system to realize fault prediction and intelligent diagnosis. The present invention improves the accuracy and efficiency of transformer and reactor testing, and realizes the intelligence of multi-domain coupling analysis and fault prediction.

Description

Technical Field

[0001] The invention relates to an electrical appliance control technology, and in particular to a general control method for a transformer and reactor comprehensive test system. Background Art

[0002] Transformers and reactors are key equipment in the power system, and their safe and stable operation is directly related to the reliability of the entire power grid. With the continuous advancement of smart grid construction, higher requirements are placed on the condition monitoring and fault diagnosis of transformers and reactors. Traditional condition monitoring methods mainly rely on a single sensor to collect data and perform fault diagnosis through classic data analysis methods. In recent years, the development of digital twin technology has provided new ideas for equipment condition monitoring. By establishing a virtual mapping model of the equipment, comprehensive perception and prediction of the equipment's operating status can be achieved. At the same time, the rapid development of power computing technology has provided new solutions for the modeling and optimization of complex systems.

[0003] At present, the test systems for transformers and reactors have the following major problems in practical applications: Existing monitoring systems often separate physical, electrical, thermal and other multi-domain characteristics for separate analysis, making it difficult to effectively capture the dynamic coupling relationship between the various components of the equipment, resulting in insufficient accuracy and comprehensiveness in fault diagnosis. Traditional digital twin models lack adaptive optimization capabilities and cannot dynamically correct the model based on real-time operating data, causing the accuracy of virtual-to-real mapping to gradually decrease over time. Existing fault diagnosis methods are mainly based on classical algorithms, which have low computational efficiency and are difficult to handle high-dimensional, strongly coupled complex fault characteristics, affecting the timeliness and accuracy of fault warnings.

[0004] In order to solve the above problems, it is urgent to develop a new test system that can comprehensively consider the multi-domain coupling effect, has adaptive optimization capabilities, and integrates the advantages of power computing. The system should be able to achieve all-round status monitoring of transformers and reactors, high-precision fault diagnosis and intelligent early warning, thereby improving the operating reliability and maintenance efficiency of power equipment. Summary of the invention

[0005] The embodiment of the present invention provides a general control method for a transformer and reactor comprehensive test system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a general control method for a comprehensive test system for transformers and reactors, including:

[0008] Based on the deployed multi-dimensional sensor array, the power evolution data of transformers and reactors are acquired in real time. Based on the power evolution data, the correlation between the physical, electrical, and thermal characteristics of transformers and reactors is mapped, and the power calculation acceleration algorithm is used to calculate the correlation in real time to generate a power feature mapping matrix. The initial power digital twin model is constructed based on the power feature mapping matrix.

[0009] Graph neural networks are deployed using edge computing nodes, and the initial power digital twin model is used as the node feature input of the graph network to build a dynamic graph topology based on power evolution data. The dynamic coupling effect between different components of the transformer and reactor is captured through the graph attention mechanism. The dynamic graph topology is compared with the initial power digital twin model for power state error correction, the power state deviation is calculated, and a self-calibration optimization algorithm is established based on the power state deviation to dynamically correct the initial power digital twin model and generate an evolutionary power digital twin model with adaptive optimization capabilities.

[0010] A hybrid power-classical fault diagnosis system is constructed based on the evolutionary power digital twin model. The real-time collected power state evolution data is input into the hybrid power-classical neural network. The power characteristics are extracted through the power lines, and the classical neural network completes the fault mode recognition. When the abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state prediction of the fault development trajectory is carried out based on the power reinforcement learning algorithm. At the same time, the prediction results are fed back to the self-calibration optimization algorithm to realize the dynamic optimization of the power state of the evolutionary power digital twin model.

[0011] Based on the power evolution data mapping, the correlation between the physical characteristics, electrical characteristics, and thermal characteristics of the transformer and the reactor is carried out, and the power calculation acceleration algorithm is used to calculate the correlation in real time to generate the power feature mapping matrix; the initial power digital twin model is constructed according to the power feature mapping matrix, including:

[0012] The power evolution data including magnetic field distribution data, temperature field distribution data and current density distribution data are acquired synchronously at a sampling rate of 100MHz;

[0013] The magnetic field distribution data is encoded into a first power state to characterize physical properties, the temperature field distribution data is encoded into a second power state to characterize thermal properties, and the current density distribution data is encoded into a third power state to characterize electrical properties, thereby obtaining three sets of encoded characteristic power states;

[0014] Applying power gate operation to the three sets of characteristic power states generates power entanglement. Power entanglement couples physical properties, thermal properties, and electrical properties into a unified multi-body correlation state. The power state tomography technique is used to characterize the internal structural characteristics of the multi-body correlation state, and the power correlation matrix that characterizes the multi-physical field coupling law of the transformer and the reactor is obtained.

[0015] The power calculation acceleration algorithm is used to perform real-time evolution calculation of the power correlation matrix, and the power correlation information is mapped to the feature space through the power Fourier transform. The power error correction mechanism is introduced in the feature extraction process to ensure the calculation accuracy. The tensor network algorithm is used to compress and reorganize the feature information, and the power feature mapping matrix that characterizes the multi-physical field coupling law of the transformer and reactor is generated. The initial power digital twin model is constructed based on the power feature mapping matrix.

[0016] The dynamic graph topology is compared with the initial power digital twin model for power state error correction, the power state deviation is calculated, a self-calibration optimization algorithm is established based on the power state deviation, the initial power digital twin model is dynamically corrected, and an evolutionary power digital twin model with adaptive optimization capabilities is generated, including:

[0017] The dynamic graph topology is converted into the target power state through power state encoding; the predicted power state in the initial power digital twin model is compared with the target power state input into the power error correction circuit, which includes a power entanglement measurement gate and a power phase correction gate. The density matrix difference between the two power states is obtained through power state tomography technology, and the power state deviation is calculated;

[0018] A self-calibration optimization algorithm is constructed based on the power state deviation, the power state deviation is decomposed into a feature vector group, a variational power circuit is constructed based on the feature vector group, and the parameters of the variational power circuit are optimized by the power natural gradient method to obtain the optimal calibration parameters;

[0019] The initial power digital twin model is dynamically corrected using the optimal calibration parameters, and the power annealing strategy is used to map the optimal calibration parameters to the model parameter space to obtain the corrected model parameters;

[0020] The initial power digital twin model is updated based on the corrected model parameters, and the model prediction accuracy is continuously optimized through the Bayesian reasoning method, eventually generating an evolutionary power digital twin model with adaptive optimization capabilities.

[0021] The power state deviation is decomposed into a set of eigenvectors, and a variational power circuit is constructed based on the eigenvector group. The parameters of the variational power circuit are optimized by the power natural gradient method, and the optimal calibration parameters are obtained, including:

[0022] The power state deviation is converted into an eigenvector group through singular value decomposition, the deviation density matrix of the power state deviation is calculated, the deviation density matrix is ​​diagonalized to obtain the eigenvalue spectrum and the corresponding eigenvector group, and the eigenvector group is ranked according to the size of the eigenvalue spectrum;

[0023] Selecting a first eigenvector from the sorted eigenvector group, and when the cumulative contribution rate of the first eigenvector reaches a preset contribution threshold, constructing a variational power circuit based on the first eigenvector, wherein the variational power circuit is composed of a single power bit rotation gate layer and a multi-power bit entanglement gate layer alternately, and the rotation angle of each power gate is used as an adjustable parameter;

[0024] An optimization objective function is constructed based on a variational power circuit. The optimization objective function includes a distance measurement term between the power state output and the target state and a parameter regularization term. The adjustable parameters are optimized by the power natural gradient method, and the power Fisher information matrix corresponding to the adjustable parameters is calculated.

[0025] The product of the inverse matrix of the power Fisher information matrix and the objective function gradient is used as the parameter update direction, the Barzilai-Borwein method is used to adaptively adjust the optimization step size, and the momentum term is introduced in the parameter update process to accelerate convergence;

[0026] Multiple convergence criteria including gradient norm threshold, parameter update amount threshold and objective function improvement rate threshold are set. When multiple convergence criteria are met at the same time, the parameter optimization iteration is stopped and the optimal calibration parameters are output.

[0027] The initial power digital twin model is updated based on the corrected model parameters. The prediction accuracy of the model is continuously optimized through the power Bayesian reasoning method. The model structure is continuously adjusted in combination with the online learning mechanism. Finally, an evolutionary power digital twin model with adaptive optimization capabilities is generated, including:

[0028] The optimal calibration parameters are mapped to the variational power circuit of the initial power digital twin model, and the parameters of the initial power digital twin model are updated by adjusting the power gate rotation angle and entanglement strength of the variational power circuit to obtain a revised power digital twin model.

[0029] The power state of the revised power digital twin model is expressed in the form of a density operator. The power system observation data is obtained based on power state tomography. The power Dirichlet distribution is used to build a priori probability model. The posterior probability distribution is calculated by the power Bayesian theorem. The prediction accuracy of the revised power digital twin model is continuously optimized according to the expected value of the posterior probability distribution.

[0030] Based on the predicted longitude of the revised power digital twin model, the similarity index between the power state purity and the von Neumann entropy is calculated, the structural configuration of the variational power circuit is dynamically adjusted according to the similarity index, and the number of layers and power gate combinations of the variational power circuit are optimized using a progressive learning strategy to obtain the optimization result;

[0031] The optimization results of the variational power circuit are fed back into the reasoning process of the power Bayesian theorem, the prior probability model is updated, and the power Bayesian reasoning and structural optimization are repeated until the model prediction accuracy meets the preset accuracy threshold requirement, and finally an evolutionary power digital twin model with adaptive optimization capability is generated.

[0032] When the abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state is predicted based on the fault development trajectory. At the same time, the prediction results are fed back to the self-calibration optimization algorithm to achieve the dynamic optimization of the power state of the evolutionary power digital twin model, including:

[0033] Calculate the power fidelity, power trace distance and power relative entropy between the real-time power state and the standard power state, construct a multi-dimensional abnormal feature vector based on the power fidelity, power trace distance and power relative entropy, and identify the abnormal power state evolution characteristics when the deviation value of the multi-dimensional abnormal feature vector exceeds the preset abnormal threshold;

[0034] The abnormal density matrix of the identified abnormal power state evolution characteristics is input into the reinforcement learning strategy network constructed by the variational power circuit, and the first fault development trajectory is generated based on the abnormal density matrix of the abnormal power state evolution characteristics;

[0035] The abnormal power state evolution characteristics are input into the power reinforcement learning algorithm, a second fault development trajectory is generated based on the power reinforcement learning algorithm, and a prediction error is generated by comparing the first fault development trajectory with the second fault development trajectory;

[0036] Based on the prediction error, a power loss function is constructed, and the power gradient of the power loss function to the parameters of the evolutionary power digital twin model is calculated. The power stochastic gradient descent method is used to update the parameters of the evolutionary power digital twin model to complete the power state dynamic optimization of the evolutionary power digital twin model.

[0037] The method further comprises:

[0038] Generate intelligent diagnosis and early warning reports based on power state prediction results. The intelligent diagnosis and early warning reports include fault power feature description, multi-domain coupling analysis, development trend prediction, and optimal intervention strategy. The diagnosis results are used to continuously optimize the power state mapping accuracy of the evolutionary power digital twin model.

[0039] The power state prediction results are reconstructed through power state tomography technology to obtain a prediction density matrix, the expected value of the power observation quantity of the prediction density matrix is ​​calculated to construct a fault feature vector, and a fault power feature description is generated based on the fault feature vector;

[0040] The data of different physical domains in the fault power feature description are mapped into a power state space to construct a multi-body power state, the power mutual information and power entanglement of the multi-body power state are calculated, and the multi-domain coupling analysis results that characterize the fault propagation characteristics are generated;

[0041] According to the results of multi-domain coupling analysis, a power Markov chain model is constructed, and the transition probability of the fault state is calculated based on the power Markov chain model to generate the prediction results of the fault development trend;

[0042] Based on the prediction results of fault development trend, the controllability criterion of power system is established, and the optimal control equation is solved to obtain the optimal intervention strategy based on the controllability criterion of power system;

[0043] Integrate fault power feature description, multi-domain coupling analysis results, fault development trend prediction results and optimal intervention strategy to form an intelligent diagnosis and early warning report;

[0044] Construct a power likelihood function to calculate the prediction accuracy of the diagnosis results in the intelligent diagnosis and early warning report, and evolve the diagnosis results into an initial power digital twin model;

[0045] Based on the comparison between the output results of the evolutionary power digital twin model and the actual system state, the power state mapping deviation is continuously calculated, and the power variational algorithm is used to iteratively optimize the power state mapping parameters of the evolutionary power digital twin model to achieve continuous improvement in the power state mapping accuracy.

[0046] According to a second aspect of the embodiments of the present invention,

[0047] An electronic device is provided, comprising:

[0048] processor;

[0049] a memory for storing processor-executable instructions;

[0050] The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0053] The beneficial effects of this application are as follows:

[0054] By deploying a multi-dimensional sensor array to collect data in real time and establish a power digital twin model, combined with a power computing acceleration algorithm, it is possible to accurately map the correlation between the physical, electrical, and thermal characteristics of transformers and reactors, thereby improving the accuracy and real-time performance of system modeling and providing a reliable data foundation for subsequent fault diagnosis.

[0055] The graph neural network and graph attention mechanism are used to realize cross-domain feature extraction in the physical, electrical and thermal domains. Through the power state error correction comparison and self-calibration optimization algorithm, the digital twin model has the ability of adaptive evolution, which significantly improves the model's ability to characterize the dynamic characteristics of the equipment and the prediction accuracy.

[0056] Based on the hybrid power classical fault diagnosis system and the power reinforcement learning algorithm, the power state prediction of the fault development trajectory is realized. By continuously optimizing the power state mapping accuracy of the model, a closed-loop optimization mechanism is formed, which greatly improves the system's fault diagnosis accuracy and prediction ability, and provides a scientific basis for equipment maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a schematic diagram of the overall control of the transformer and reactor comprehensive test system according to an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of multi-dimensional evaluation of power state feature decomposition performance according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the multi-dimensional performance evaluation of the power digital twin model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. 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.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a schematic diagram of the overall control of the transformer and reactor comprehensive test system according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] Based on the deployed multi-dimensional sensor array, the power evolution data of transformers and reactors are acquired in real time. Based on the power evolution data, the correlation between the physical, electrical, and thermal characteristics of transformers and reactors is mapped, and the power calculation acceleration algorithm is used to calculate the correlation in real time to generate a power feature mapping matrix. The initial power digital twin model is constructed based on the power feature mapping matrix.

[0064] Graph neural networks are deployed using edge computing nodes, and the initial power digital twin model is used as the node feature input of the graph network to build a dynamic graph topology based on power evolution data. The dynamic coupling effect between different components of the transformer and reactor is captured through the graph attention mechanism. The dynamic graph topology is compared with the initial power digital twin model for power state error correction, the power state deviation is calculated, and a self-calibration optimization algorithm is established based on the power state deviation to dynamically correct the initial power digital twin model and generate an evolutionary power digital twin model with adaptive optimization capabilities.

[0065] A hybrid power-classical fault diagnosis system is constructed based on the evolutionary power digital twin model. The real-time collected power state evolution data is input into the hybrid power-classical neural network. The power characteristics are extracted through the power lines, and the classical neural network completes the fault mode recognition. When the abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state prediction of the fault development trajectory is carried out based on the power reinforcement learning algorithm. At the same time, the prediction results are fed back to the self-calibration optimization algorithm to realize the dynamic optimization of the power state of the evolutionary power digital twin model.

[0066] In an optional implementation, the correlation between the physical characteristics, electrical characteristics, and thermal characteristics of the transformer and the reactor is mapped based on the power evolution data, and the correlation is calculated in real time using a power calculation acceleration algorithm to generate a power feature mapping matrix; constructing an initial power digital twin model according to the power feature mapping matrix includes:

[0067] The power evolution data including magnetic field distribution data, temperature field distribution data and current density distribution data are acquired synchronously at a sampling rate of 100MHz;

[0068] The magnetic field distribution data is encoded into a first power state to characterize physical properties, the temperature field distribution data is encoded into a second power state to characterize thermal properties, and the current density distribution data is encoded into a third power state to characterize electrical properties, thereby obtaining three sets of encoded characteristic power states;

[0069] Applying power gate operation to the three sets of characteristic power states generates power entanglement. Power entanglement couples physical properties, thermal properties, and electrical properties into a unified multi-body correlation state. The power state tomography technique is used to characterize the internal structural characteristics of the multi-body correlation state, and the power correlation matrix that characterizes the multi-physical field coupling law of the transformer and the reactor is obtained.

[0070] The power calculation acceleration algorithm is used to perform real-time evolution calculation of the power correlation matrix, and the power correlation information is mapped to the feature space through the power Fourier transform. The power error correction mechanism is introduced in the feature extraction process to ensure the calculation accuracy. The tensor network algorithm is used to compress and reorganize the feature information, and the power feature mapping matrix that characterizes the multi-physical field coupling law of the transformer and reactor is generated. The initial power digital twin model is constructed based on the power feature mapping matrix.

[0071] A high-precision sensor array system is deployed at key locations of transformers and reactors. Hall effect sensors are used for magnetic field measurement, and 32 measuring points are arranged on the surface of the transformer core and windings. PT100 thermal resistors are used for temperature monitoring, and 24 measuring points are arranged to monitor the winding and core temperature. Rogowski coils are used for current density detection, and 16 measuring points are arranged to monitor the winding current distribution. All sensors use a 100MHz sampling rate for synchronous acquisition, and data parallel acquisition is achieved through a high-speed data acquisition card.

[0072] The physical characteristics of the magnetic field distribution data are encoded. Taking a certain measuring point of a 220kV transformer as an example, the magnetic field intensity is measured to be 1.2 Tesla, which is normalized to 0.8. The measuring point coordinates are (12cm, 18cm, 25cm). The magnetic flux density is 1.1 Weber / square meter, which is normalized to 0.73, forming the eigenvector (0.8, 0.73, 12, 18, 25). The data of 32 measuring points are encoded to form the first power state matrix. The temperature field distribution data is encoded as thermal characteristics. The temperature of a certain measuring point is 75℃, which is normalized to 0.625. The temperature gradient is 3℃ / cm, which is normalized to 0.6. The coordinates are (15cm, 20cm, 28cm), forming the eigenvector (0.625, 0.6, 15, 20, 28). The 24 temperature measuring points are encoded to obtain the second power state matrix. The current density distribution data is encoded as electrical characteristics. The current density of the measuring point is 4A / mm² normalized to 0.8, the leakage flux is 0.3 Weber normalized to 0.6, the coordinates are (16cm, 22cm, 30cm), and the characteristic vector is (0.8, 0.6, 16, 22, 30). The 16 current measuring points are encoded to form the third power state matrix.

[0073] First, the internal correlation of physical characteristics is calculated, and the correlation degree is calculated for each of the 32 magnetic field measurement points. The magnetic field strengths of adjacent measurement points A and B are 1.2 and 1.1 Tesla, respectively, and the distance is 5cm. The magnetic field gradient is calculated to be 0.02 Tesla / cm, and the normalization result is a correlation degree of 0.8, which constructs a 32×32 physical characteristic internal correlation submatrix. Similarly, the internal correlation of thermal characteristics of 24 temperature measurement points is calculated. The temperatures of measurement points C and D are 75℃ and 70℃, respectively, with a spacing of 4cm, a temperature gradient of 1.25℃ / cm, and a normalized correlation degree of 0.6, forming a 24×24 thermal characteristic internal correlation submatrix. The internal correlation of electrical characteristics of 16 current density measurement points is calculated. The current densities of measurement points E and F are 4.0 and 3.8A / mm², respectively, with a spacing of 3cm. The normalized correlation degree is 0.7, and a 16×16 electrical characteristic internal correlation submatrix is ​​obtained.

[0074] Continue to calculate the cross-correlation characteristics: the relationship between the physical quantity changes of the magnetic field measuring point G (1.2 Tesla) and the temperature measuring point H (80℃), the cross-correlation degree is 0.65, and a 32×24 physical-thermal characteristic cross-correlation submatrix is ​​constructed; the correlation analysis between the temperature measuring point I (75℃) and the current measuring point J (3.9A / mm²) obtains a correlation degree of 0.55, forming a 24×16 thermal-electric characteristic cross-correlation submatrix; the coupling analysis between the magnetic field measuring point K (1.1 Tesla) and the current measuring point L (4.1A / mm²) obtains a correlation degree of 0.75, and a 32×16 physical-electric characteristic cross-correlation submatrix is ​​constructed.

[0075] The above six sub-matrices are combined and arranged in the order of physical characteristics, thermal characteristics, and electrical characteristics to construct a complete power correlation matrix of 72×72. The matrix data is updated every 10μs, and the change rate of the correlation strength at adjacent time points is calculated. For example, if an element changes from 0.8 to 0.75, the change rate is 0.005 / μs.

[0076] A parallel computing architecture is used to process the power correlation matrix, and the 72×72 matrix is ​​divided into 9 8×8 sub-blocks for parallel operation. The correlation information is mapped to the feature space through the power Fourier transform, and the main eigenvalues ​​are extracted. Parity check and cyclic redundancy check are introduced to ensure the calculation accuracy, and the tensor network algorithm is used to compress the 72×72 matrix into a 24×24 feature mapping matrix.

[0077] Build a digital twin model based on the power feature mapping matrix. Establish a transformer geometry model including core, winding and other structures, and establish a physical model, thermal model and electrical model based on the feature mapping matrix to achieve real-time mapping and state prediction of physical, thermal and electrical characteristics.

[0078] An innovative power state encoding method is proposed to quantize the physical field data and realize the unified characterization of multiple physical quantities. Through high sampling rate synchronous acquisition and encoding conversion, data integrity and spatiotemporal consistency are guaranteed, laying a solid foundation for subsequent analysis. The power gate operation and state tomography technology are used to reveal the multi-physical field coupling law of transformers and reactors. By establishing a power correlation matrix, the quantitative characterization of the correlation between physical properties, thermal properties, and electrical properties is achieved, and the accuracy of feature extraction is improved. Combined with the power calculation acceleration algorithm and tensor network compression technology, the computing efficiency is significantly improved. By driving the digital twin model through the feature mapping matrix, real-time monitoring and prediction of the operating status of transformers and reactors are realized, providing important support for equipment status assessment and fault diagnosis.

[0079] In an optional implementation, the dynamic graph topology is compared with the initial power digital twin model for power state error correction, the power state deviation is calculated, a self-calibration optimization algorithm is established based on the power state deviation, the initial power digital twin model is dynamically corrected, and an evolutionary power digital twin model with adaptive optimization capability is generated, including:

[0080] The dynamic graph topology is converted into the target power state through power state encoding; the predicted power state in the initial power digital twin model is compared with the target power state input into the power error correction circuit, which includes a power entanglement measurement gate and a power phase correction gate. The density matrix difference between the two power states is obtained through power state tomography technology, and the power state deviation is calculated;

[0081] A self-calibration optimization algorithm is constructed based on the power state deviation, the power state deviation is decomposed into a feature vector group, a variational power circuit is constructed based on the feature vector group, and the parameters of the variational power circuit are optimized by the power natural gradient method to obtain the optimal calibration parameters;

[0082] The initial power digital twin model is dynamically corrected using the optimal calibration parameters, and the power annealing strategy is used to map the optimal calibration parameters to the model parameter space to obtain the corrected model parameters;

[0083] The initial power digital twin model is updated based on the corrected model parameters, and the model prediction accuracy is continuously optimized through the Bayesian reasoning method, eventually generating an evolutionary power digital twin model with adaptive optimization capabilities.

[0084] Collect real-time operation data of 220kV transformers, including data from 32 magnetic field measurement points, 24 temperature measurement points, and 16 current density measurement points. Normalize the data of each measurement point, such as the magnetic field measurement point A value of 1.2 Tesla is normalized to 0.8, the temperature measurement point B value of 75℃ is normalized to 0.625, and the current density measurement point C value of 4A / mm² is normalized to 0.8. Encode the feature vectors of all measurement point data, and generate a 5-dimensional feature vector containing physical value and spatial position for each measurement point. Combine the feature vectors of all measurement points to form the target power state matrix.

[0085] The initial digital twin model predicts the current state of the equipment based on historical data and generates a predicted power state matrix. The predicted power state and the target power state are input into the power error correction circuit. The power entanglement measurement gate measures the correlation between the two states. For example, the correlation between the predicted value 0.85 and the actual value 0.8 at a certain measuring point is 0.94. The power phase correction gate calculates the phase difference between the two states and obtains a phase deviation value of 0.06. The density matrix of the two states is reconstructed through power state tomography, and the difference in the density matrix is ​​calculated to obtain the power state deviation. Taking a certain measuring point as an example, the state deviation value is 0.05, indicating that there is a 5% deviation between the predicted value and the actual value.

[0086] The power state deviation is decomposed into a set of eigenvectors. For the 32 magnetic field measurement points, the deviation value of each measurement point forms an eigenvector; the 24 temperature measurement points and 16 current density measurement points are processed in the same way. A variational power circuit is constructed based on the eigenvector group, and the circuit contains adjustable weight parameters. The weight parameters are iteratively optimized using the power natural gradient method. When the parameter change between two adjacent iterations is less than 0.001, the optimization is stopped to obtain the optimal calibration parameter group.

[0087] The power annealing strategy is used to map the optimal calibration parameters to the model parameter space. The initial temperature is set to 1.0, and the temperature attenuation coefficient for each iteration is 0.95. At each temperature, the optimal mapping relationship is determined by local search. When the temperature drops to 0.01, the annealing process is completed and the corrected model parameters are obtained. Taking the parameters of a magnetic field measurement point as an example, the original model parameter 0.85 is changed to 0.82 after correction.

[0088] Update the initial digital twin model with the corrected parameters. Continuously optimize the model prediction accuracy through Bayesian reasoning method. Set the prior probability distribution, calculate the posterior probability based on the new data, and update the model parameter distribution. After 100 iterations of optimization, the model prediction accuracy increased from the original 92% to 96%, forming an evolutionary digital twin model with adaptive optimization capabilities.

[0089] The innovative use of power state coding and power error correction technology enables accurate comparison between the dynamic graph topology and the digital twin model. Through power state tomography reconstruction and density matrix difference analysis, the model prediction deviation is accurately calculated, providing a reliable basis for subsequent optimization. A self-calibration optimization method based on power state deviation is proposed, which combines the variational power circuit and the power natural gradient algorithm to achieve adaptive optimization of model parameters. The power annealing strategy is used for parameter mapping to ensure the convergence and stability of the optimization process. An evolutionary digital twin model with adaptive optimization capabilities is constructed. The model prediction accuracy is continuously optimized through the Bayesian reasoning method, which significantly improves the prediction accuracy of the model. The model can automatically adjust parameters as the operating status of the equipment changes, maintaining a high prediction performance.

[0090] In an optional implementation, the power state deviation is decomposed into a set of characteristic vectors, a variational power circuit is constructed based on the set of characteristic vectors, and the parameters of the variational power circuit are optimized by the power natural gradient method to obtain the optimal calibration parameters, including:

[0091] The power state deviation is converted into an eigenvector group through singular value decomposition, the deviation density matrix of the power state deviation is calculated, the deviation density matrix is ​​diagonalized to obtain the eigenvalue spectrum and the corresponding eigenvector group, and the eigenvector group is ranked according to the size of the eigenvalue spectrum;

[0092] Selecting a first eigenvector from the sorted eigenvector group, and when the cumulative contribution rate of the first eigenvector reaches a preset contribution threshold, constructing a variational power circuit based on the first eigenvector, wherein the variational power circuit is composed of a single power bit rotation gate layer and a multi-power bit entanglement gate layer alternately, and the rotation angle of each power gate is used as an adjustable parameter;

[0093] An optimization objective function is constructed based on a variational power circuit. The optimization objective function includes a distance measurement term between the power state output and the target state and a parameter regularization term. The adjustable parameters are optimized by the power natural gradient method, and the power Fisher information matrix corresponding to the adjustable parameters is calculated.

[0094] The product of the inverse matrix of the power Fisher information matrix and the objective function gradient is used as the parameter update direction, the Barzilai-Borwein method is used to adaptively adjust the optimization step size, and the momentum term is introduced in the parameter update process to accelerate convergence;

[0095] Multiple convergence criteria including gradient norm threshold, parameter update amount threshold and objective function improvement rate threshold are set. When multiple convergence criteria are met at the same time, the parameter optimization iteration is stopped and the optimal calibration parameters are output.

[0096] A multidimensional sensor array is used to collect the operating state data of power equipment in real time, and the deviation between the current power state and the ideal power state is calculated based on the collected power state data. The power state deviation data is decomposed by the singular value decomposition method to obtain a series of eigenvectors and their corresponding eigenvalues. The obtained eigenvalues ​​are sorted from large to small, and the contribution rate corresponding to each eigenvector is calculated. The contribution rate of the eigenvector is measured by the ratio of the eigenvalue corresponding to the eigenvector to the sum of all eigenvalues. The eigenvector with the largest contribution rate is selected as the first eigenvector. When the cumulative contribution rate of the first eigenvector reaches the preset 90% threshold, it can be determined to use the eigenvector to construct the variational power circuit.

[0097] The variational power circuit adopts a structure in which a single power bit rotary gate layer and a multi-power bit entangled gate layer are alternately stacked. The single power bit rotary gate layer contains a number of rotary gates with independently adjustable rotation angles, which are used to adjust the phase of a single power bit. The multi-power bit entangled gate layer contains an entangled gate between two power bits, which is used to build an association between power bits. The rotation angle of each power gate is used as an adjustable parameter. For example, for a system with 4 power bits, 4 layers of single power bit rotary gate layers and 3 layers of multi-power bit entangled gate layers can be set to be alternately stacked, containing a total of 28 adjustable rotation angle parameters.

[0098] When constructing the optimization objective function, the fidelity between the power state output and the target state is used as the distance measurement term. At the same time, the parameter regularization term is introduced to suppress the overfitting phenomenon caused by excessive parameters. In the parameter optimization process, the power Fisher information matrix element corresponding to each parameter is first calculated. For any two parameters, the Fisher information between them is obtained by calculating the inner product of the partial derivatives of the circuit output with respect to the two parameters. The Fisher information corresponding to all parameters constitutes the power Fisher information matrix.

[0099] The power natural gradient method is used for parameter optimization. The inverse matrix of the power Fisher information matrix is ​​multiplied by the gradient vector of the objective function to the parameter to obtain the direction of parameter update. The parameter update step size is adaptively adjusted using the Barzilai-Borwein method, and the optimal step size is estimated by comparing the parameter difference and gradient difference between two adjacent iterations. The momentum term is introduced in the parameter update process, and the momentum coefficient is set to 0.9, which can speed up the optimization convergence speed.

[0100] Set multiple convergence criteria: gradient norm is less than 0.001, parameter update amount is less than 0.0001, and the objective function improvement rate is less than 0.01% for three consecutive iterations. When these convergence criteria are met at the same time, the iteration is stopped and the optimal calibration parameters are output. According to experimental data, in a system with 4 power bits, it can converge to the optimal solution after about 200 iterations, and the fidelity between the final circuit output and the target state can reach more than 99.9%.

[0101] Figure 2 This is a schematic diagram of multi-dimensional evaluation of the power state feature decomposition performance according to an embodiment of the present invention:

[0102] This performance comparison chart shows the trend of the cumulative contribution rate of three different technical solutions as the number of feature vectors changes. Among them, this technical solution (square mark point) is the best performing solution in the figure. From the curve trend, it can be seen that when the number of feature vectors is small (about 10), a high cumulative contribution rate (75%) can be achieved, indicating that this solution has stronger feature extraction and data compression capabilities. When the number of feature vectors increases to 30, the cumulative contribution rate has reached 97%, and then tends to stabilize, and finally reaches a high contribution rate of 99.5% when the number of feature vectors is 50, indicating that this technical solution can express the most data information with the least feature vectors; the traditional SVD method (triangle mark point) is used as a traditional singular value decomposition method for comparison, and its performance is between this technical solution and the benchmark method. When the number of feature vectors is When the number of eigenvectors reaches 10, the cumulative contribution rate is about 55%, and the growth rate is relatively slow. More eigenvectors are needed to achieve a higher cumulative contribution rate. When the number of eigenvectors reaches 50, the final contribution rate is 90%, which shows that the traditional SVD method is not as efficient as this technical solution in data compression; the benchmark method (circular mark point) has the lowest performance among the three solutions, and its cumulative contribution rate grows the slowest. When the number of eigenvectors is 10, it only reaches a cumulative contribution rate of 40%, and more eigenvectors are needed to achieve better performance. Even when the number of eigenvectors reaches 50, its cumulative contribution rate is only 84%, which is much lower than the other two methods. This shows that the benchmark method is less efficient in data dimensionality reduction and information extraction. From the comparison of these three curves, it can be clearly seen that this technical solution can retain more original information under the same number of eigenvectors, has better dimensionality reduction effect and higher data compression efficiency, especially when the number of eigenvectors is small, the advantages of this technical solution are more obvious, which is of great significance for computing efficiency and storage space optimization in practical applications.

[0103] By decomposing the eigenvector of the power state deviation, the characteristic components that have the most significant impact on the system are identified, the number of degrees of freedom that need to be calibrated is reduced, and the optimization complexity is significantly reduced. The optimization strategy that combines the variational power circuit with the power natural gradient method fully utilizes the geometric characteristics of the power system and avoids the gradient vanishing problem of the traditional gradient descent method during the optimization process. The introduction of multiple convergence criteria and momentum acceleration mechanism not only ensures the reliability of the optimization results, but also improves the convergence speed, making the calibration process more efficient and stable.

[0104] In an optional implementation, the initial power digital twin model is updated based on the corrected model parameters, the model prediction accuracy is continuously optimized through the power Bayesian reasoning method, and the model structure is continuously adjusted in combination with the online learning mechanism, and finally an evolutionary power digital twin model with adaptive optimization capability is generated, including:

[0105] The optimal calibration parameters are mapped to the variational power circuit of the initial power digital twin model, and the parameters of the initial power digital twin model are updated by adjusting the power gate rotation angle and entanglement strength of the variational power circuit to obtain a revised power digital twin model.

[0106] The power state of the revised power digital twin model is expressed in the form of a density operator. The power system observation data is obtained based on power state tomography. The power Dirichlet distribution is used to build a priori probability model. The posterior probability distribution is calculated by the power Bayesian theorem. The prediction accuracy of the revised power digital twin model is continuously optimized according to the expected value of the posterior probability distribution.

[0107] Based on the predicted longitude of the revised power digital twin model, the similarity index between the power state purity and the von Neumann entropy is calculated, the structural configuration of the variational power circuit is dynamically adjusted according to the similarity index, and the number of layers and power gate combinations of the variational power circuit are optimized using a progressive learning strategy to obtain the optimization result;

[0108] The optimization results of the variational power circuit are fed back into the reasoning process of the power Bayesian theorem, the prior probability model is updated, and the power Bayesian reasoning and structural optimization are repeated until the model prediction accuracy meets the preset accuracy threshold requirement, and finally an evolutionary power digital twin model with adaptive optimization capability is generated.

[0109] First, the initial power digital twin model is updated based on the optimal calibration parameters. The optimal calibration parameters are converted into control parameters in the variational power circuit through mapping relationships, including the rotation angle of the single power bit rotation gate and the entanglement strength of the multi-power bit entanglement gate. For a system consisting of 4 power bits, the rotation angle is set to a range of 0 to 2π, and the entanglement strength is set to a range of 0 to 1. According to the mapping results of the optimal calibration parameters, the specific values ​​of the 28 control parameters in the variational power circuit are adjusted to complete the parameter update of the initial power digital twin model.

[0110] Next, the power Bayesian inference method is used to continuously optimize the prediction accuracy of the model. First, the power state of the modified power digital twin model is expressed as a density operator, and the observation data of the actual power system is obtained through power state tomography technology. Specifically, the projection data of the power state is collected under different measurement bases, and the number of measurement bases is not less than the square of the power state dimension. The power Dirichlet distribution is used to construct a priori probability model, in which the hyperparameters are determined based on historical operating data. The observation data is combined with the prior probability through the power Bayesian theorem to calculate the posterior probability distribution. According to the expected value of the posterior probability distribution, the prediction parameters of the power digital twin model are continuously adjusted to continuously improve the prediction accuracy.

[0111] Based on the prediction results of the revised model, the similarity index is calculated to evaluate the model performance. The similarity index consists of two parts: power state purity and von Neumann entropy. Power state purity reflects the closeness of the prediction result to the target state, and its value range is 0 to 1. Von Neumann entropy characterizes the uncertainty of the prediction result, and its value range is 0 to the maximum entropy value. When the prediction accuracy is low, the structural configuration of the variational power circuit is dynamically adjusted. A progressive learning strategy is adopted to gradually increase the number of circuit layers and optimize the combination of single power bit rotation gates and multi-power bit entanglement gates. For example, 1-2 layers can be added to the original 4-layer structure, and different gate type combinations can be tried until the optimal configuration is found.

[0112] The optimization results of the variational power circuit are used as feedback information to update the prior probability model in the power Bayesian reasoning. Specifically, according to the prediction effect after structural optimization, the hyperparameters of the power Dirichlet distribution are adjusted so that the prior probability reflects the system characteristics more accurately. The process of power Bayesian reasoning and structural optimization is repeated until the model prediction accuracy reaches the preset 99.9% accuracy threshold requirement. In practical applications, this process usually requires 3-5 rounds of iterations to converge, and finally generates an evolutionary power digital twin model with adaptive optimization capabilities.

[0113] Figure 3 This is a schematic diagram of multi-dimensional performance evaluation of the power digital twin model according to an embodiment of the present invention:

[0114] The figure shows the prediction accuracy comparison results of three different digital twin solutions. Among them, the "present technical solution" marked with a square represents the evolutionary power digital twin model based on power state error correction and self-calibration optimization. Its prediction accuracy increases rapidly from the initial 12.5% ​​to 37.5% at the 10th iteration, 57.5% at the 20th iteration, and 70% at the 30th iteration. After that, the growth trend gradually slowed down, and finally reached a stable prediction accuracy of 78% at the 50th iteration; the "traditional Bayesian method" marked with a triangle line represents the solution of model optimization using traditional Bayesian reasoning. Its prediction accuracy improves relatively slowly, gradually increasing from the initial 5%, reaching only 38.8% at the 30th iteration, and finally stabilizing at 42%; the "ordinary digital twin" marked with a diamond line represents a conventional digital twin model without an adaptive optimization mechanism. Its prediction accuracy improves the slowest, starting from the initial 3%, and even after 50 iterations, it can only reach a prediction accuracy of 29.8%. From the comparison of the curve trends, it can be seen that this technical solution not only has a faster convergence speed, but also can achieve higher prediction accuracy. At the 20th iteration, it has reached a prediction accuracy of 57.5%, far exceeding the other two methods. This shows that by introducing the power state error correction and self-calibration optimization mechanism, the learning ability and prediction performance of the model are significantly improved. At the same time, the smooth convergence of the curve also proves that the solution has good stability.

[0115] The power Bayesian inference method is used to continuously optimize the model prediction accuracy, making full use of prior knowledge and observation data, significantly improving the accuracy and reliability of the prediction. Based on the similarity index and progressive learning strategy, the dynamic optimization of the variational power circuit structure is realized, so that the model has the ability to adaptively adjust and can better adapt to system changes. The iterative optimization mechanism is adopted to feed back the structural optimization results to the Bayesian inference process, forming a closed-loop optimization framework, ensuring the continuous improvement of the model performance while maintaining the high efficiency and stability of the optimization process.

[0116] In an optional implementation, when abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state is predicted for the fault development trajectory based on the power reinforcement learning algorithm, and the prediction results are fed back to the self-calibration optimization algorithm to achieve dynamic optimization of the power state of the evolutionary power digital twin model, including:

[0117] Calculate the power fidelity, power trace distance and power relative entropy between the real-time power state and the standard power state, construct a multi-dimensional abnormal feature vector based on the power fidelity, power trace distance and power relative entropy, and identify the abnormal power state evolution characteristics when the deviation value of the multi-dimensional abnormal feature vector exceeds the preset abnormal threshold;

[0118] The abnormal density matrix of the identified abnormal power state evolution characteristics is input into the reinforcement learning strategy network constructed by the variational power circuit, and the first fault development trajectory is generated based on the abnormal density matrix of the abnormal power state evolution characteristics;

[0119] The abnormal power state evolution characteristics are input into the power reinforcement learning algorithm, a second fault development trajectory is generated based on the power reinforcement learning algorithm, and a prediction error is generated by comparing the first fault development trajectory with the second fault development trajectory;

[0120] Based on the prediction error, a power loss function is constructed, and the power gradient of the power loss function to the parameters of the evolutionary power digital twin model is calculated. The power stochastic gradient descent method is used to update the parameters of the evolutionary power digital twin model to complete the power state dynamic optimization of the evolutionary power digital twin model.

[0121] The power operation state data is obtained through the real-time monitoring system, and the real-time power state is compared and analyzed with the pre-set standard power state. First, the power fidelity is calculated to characterize the degree of overlap between the two power states. Its value range is between 0 and 1. The closer the value is to 1, the higher the overlap. At the same time, the power trace distance is calculated to measure the geometric distance between the two power states, and the value is a non-negative real number. In addition, the power relative entropy is calculated to measure the information difference between the two power states, and the value is also a non-negative real number. These three indicators are combined to form a multi-dimensional abnormal feature vector. Set the abnormal threshold: when the power fidelity is lower than 0.95, the power trace distance is greater than 0.1, and the power relative entropy is greater than 0.05, it is judged that abnormal power state evolution characteristics have occurred.

[0122] When an abnormal situation is detected, the density matrix of the abnormal power state is input into the pre-built reinforcement learning strategy network. The network consists of multiple layers of variational power circuits, including encoding layer, feature extraction layer and prediction layer. The encoding layer converts the input density matrix into a feature representation suitable for processing, the feature extraction layer extracts fault feature information through the combination of multiple layers of power gates, and the prediction layer outputs the first fault development trajectory. For example, for a certain abnormal situation, a trajectory containing the evolution process of key state parameters is generated through the prediction of 10 time points.

[0123] At the same time, the power reinforcement learning algorithm is started. The algorithm adopts a deep Q network structure and contains three core components: state space, action space, and reward function. The state space contains all possible states of the power system, the action space contains all control actions that the system can take, and the reward function is used to evaluate the quality of the action. The algorithm searches for the optimal action sequence in the state space through an exploration-exploitation strategy to generate the second fault development trajectory. The two trajectories are compared, the difference in the predicted values ​​at each time point is calculated, and a prediction error sequence is generated.

[0124] Based on the prediction error sequence, the power loss function is constructed, which comprehensively considers the mean and variance of the prediction error. By calculating the gradient information of the loss function to the model parameters, the power stochastic gradient descent method is used to update the model parameters. In the specific update process, an adaptive learning rate strategy is adopted, and the initial learning rate is set to 0.01, and the learning rate is dynamically adjusted according to the change of loss during training. After each parameter update, the model prediction effect is re-evaluated until the prediction error converges below the preset threshold. Practice has shown that through 3-5 rounds of iterative optimization, the prediction error can be reduced to less than 20% of the original.

[0125] The multi-dimensional abnormal feature vector is used to accurately identify abnormal states. The combination of multiple evaluation indicators improves the reliability of anomaly detection and reduces the probability of false alarms and missed alarms. The dual-trajectory prediction mechanism is used to combine variational circuit prediction with reinforcement learning prediction, making full use of the advantages of the two methods and improving the accuracy of fault development prediction. Based on the dynamic optimization mechanism driven by prediction errors, the adaptive adjustment of model parameters is achieved, so that the digital twin model has the ability of continuous learning and performance improvement, ensuring the stability of prediction performance.

[0126] In an optional implementation, the method further includes:

[0127] Generate intelligent diagnosis and early warning reports based on power state prediction results. The intelligent diagnosis and early warning reports include fault power feature description, multi-domain coupling analysis, development trend prediction, and optimal intervention strategy. The diagnosis results are used to continuously optimize the power state mapping accuracy of the evolutionary power digital twin model.

[0128] The power state prediction results are reconstructed through power state tomography technology to obtain a prediction density matrix, the expected value of the power observation quantity of the prediction density matrix is ​​calculated to construct a fault feature vector, and a fault power feature description is generated based on the fault feature vector;

[0129] The data of different physical domains in the fault power feature description are mapped into a power state space to construct a multi-body power state, the power mutual information and power entanglement of the multi-body power state are calculated, and the multi-domain coupling analysis results that characterize the fault propagation characteristics are generated;

[0130] According to the results of multi-domain coupling analysis, a power Markov chain model is constructed, and the transition probability of the fault state is calculated based on the power Markov chain model to generate the prediction results of the fault development trend;

[0131] Based on the prediction results of fault development trend, the controllability criterion of power system is established, and the optimal control equation is solved to obtain the optimal intervention strategy based on the controllability criterion of power system;

[0132] Integrate fault power feature description, multi-domain coupling analysis results, fault development trend prediction results and optimal intervention strategy to form an intelligent diagnosis and early warning report;

[0133] Construct a power likelihood function to calculate the prediction accuracy of the diagnosis results in the intelligent diagnosis and early warning report, and evolve the diagnosis results into an initial power digital twin model;

[0134] Based on the comparison between the output results of the evolutionary power digital twin model and the actual system state, the power state mapping deviation is continuously calculated, and the power variational algorithm is used to iteratively optimize the power state mapping parameters of the evolutionary power digital twin model to achieve continuous improvement in the power state mapping accuracy.

[0135] First, the power state prediction results are reconstructed through power state tomography technology to obtain the prediction density matrix. Specifically, the maximum likelihood estimation method is used to reconstruct the power state. By measuring the power state projection probability under different bases, the likelihood function is constructed and maximized to obtain the optimal prediction density matrix. For example, for a two-level power system, multiple measurements can be performed under the three bases of X, Y, and Z to obtain the power state projection probability under each base, and then the maximum likelihood estimation is solved through the iterative optimization algorithm to obtain the prediction density matrix.

[0136] Next, the expected value of the power observation quantity of the prediction density matrix is ​​calculated to construct the fault feature vector. A set of complete power observation quantity operators, such as Pauli operators σx, σy, σz, is selected, and their expected values ​​under the prediction density matrix are calculated, and these expected values ​​are combined into the fault feature vector. For example, for a two-level system, the three expected values ​​of 〈σx〉, 〈σy〉, and 〈σz〉 can be calculated to form a three-dimensional fault feature vector.

[0137] Based on the fault feature vector, a fault power feature description is generated. The fault feature vector is mapped to the fault symptom space, and a fault classification model is trained through machine learning algorithms such as support vector machines or random forests. The new fault feature vector is classified using the model to obtain a description of the fault type and severity. For example, the fault can be divided into three levels: minor, moderate, and severe, and the specific faulty components and fault modes are given.

[0138] Then, the data of different physical domains in the fault power feature description are mapped into power state space to construct a multi-body power state. The data of 32 measuring points in the magnetic field domain (such as the magnetic field intensity of measuring point A is 1.2 Tesla) are normalized to 0.75; the data of 16 measuring points in the electric field domain (such as the current density of measuring point B is 4.0A / mm²) are normalized to 0.67; the data of 24 measuring points in the temperature domain (such as the temperature of measuring point C is 75℃) are normalized to 0.75. The normalized data of different physical domains are organized into feature vectors according to the spatial position information, and the correlation coefficients between physical domains are calculated (such as the magnetic field-electric field correlation coefficient is 0.92), and finally a multi-body power state vector (0.75, 0.67, 0.75) containing all physical quantity information is constructed.

[0139] The power mutual information and power entanglement of multi-body power states are calculated to generate multi-domain coupling analysis results that characterize the fault propagation characteristics. Power mutual information is used to measure the correlation between different physical domains, while power entanglement reflects the strength of the power correlation between them. By comparing the power mutual information and entanglement between different pairs of physical domains, the main paths and key coupling points of fault propagation can be identified. For example, a heat map can be drawn to show the power mutual information strength between different physical domains, and the highly coupled areas that exceed the threshold are marked.

[0140] According to the results of multi-domain coupling analysis, a power Markov chain model is constructed. The fault states of different physical domains are used as the states of the Markov chain, and the state transition probability is determined according to the strength of the power mutual information and entanglement. For example, the power mutual information value can be normalized as the basis for the state transition probability, and then appropriately adjusted according to the system characteristics.

[0141] Based on the power Markov chain model, the transition probability of the fault state is calculated to generate the prediction result of the fault development trend. Through multi-step iterative calculation, the probability distribution of each fault state at different time scales is obtained. You can select a specific time point, such as 1 hour, 1 day, 1 week later, to give the probability prediction of various possible fault states. For example, a pie chart can be used to show the probability distribution of different states that the system may be in, such as normal, minor fault, serious fault, etc. after 1 day.

[0142] Based on the prediction results of fault development trend, the controllability criterion of the power system is established. The power Lie algebra method is used to analyze the commutation relationship between the system Hamiltonian and the control Hamiltonian to determine whether the system is fully controllable. For incompletely controllable systems, the dimension and structure of the controllable subspace are determined. For example, the controllability index of the system and the orthogonal basis vectors of the controllable subspace can be given.

[0143] Based on the controllability criterion of the power system, the optimal control equation is solved to obtain the optimal intervention strategy. The variational power algorithm is used to express the control objective function as the expected value of the parameterized power circuit. The circuit parameters are iteratively updated through the classical optimization algorithm to finally obtain the optimal control sequence. For example, a series of power gate operations and their parameters can be given to describe how to control the system from the current fault state to the target state.

[0144] Integrate the fault power feature description, multi-domain coupling analysis results, fault development trend prediction results and optimal intervention strategy to form an intelligent diagnosis and early warning report. The report should include the power feature description of the fault, the identification results of key coupling domains, the fault probability prediction at different time scales, and detailed intervention operation suggestions. For example, a comprehensive report containing text descriptions, charts and operation guidelines can be generated.

[0145] Construct the power likelihood function and calculate the prediction accuracy of the diagnosis results in the intelligent diagnosis warning report. Compare the predicted power state with the actual measurement results and calculate the fidelity or trace distance between them as a measure of prediction accuracy. For example, a percentage value of the prediction accuracy can be given, as well as a trend chart of the accuracy over time.

[0146] The diagnostic results are input into the evolutionary power digital twin model. The predicted power state, multi-domain coupling analysis results and other information are used as inputs to the power digital twin model to update the model parameters and states.

[0147] Based on the comparison between the output of the evolutionary power digital twin model and the actual system state, the power state mapping deviation is continuously calculated. The power state of the actual system is reconstructed using power state tomography technology, and the fidelity or trace distance is calculated with the power state output by the model to obtain the power state mapping deviation. For example, a curve of the mapping deviation changing over time can be plotted to monitor model performance.

[0148] The power variational algorithm is used to iteratively optimize the power state mapping parameters of the evolutionary power digital twin model. The power state mapping deviation is used as the loss function, a parameterized power circuit is constructed to represent the mapping relationship, and the circuit parameters are updated through optimization algorithms such as gradient descent. For example, the power circuit structure and parameter values ​​before and after optimization, as well as the convergence curve of the loss function during the optimization process, can be given.

[0149] Through continuous iteration of the above steps, the accuracy of power state mapping can be continuously improved, thereby continuously improving the performance of the intelligent diagnosis and early warning system.

[0150] First, through power state prediction and multi-domain coupling analysis, this method can fully capture the power characteristics and complex interactions of the system, providing more accurate and comprehensive fault diagnosis and early warning capabilities than traditional methods. It can not only identify the fault of a single component, but also reveal the correlation between different physical domains, which helps to discover hidden fault propagation paths and root causes. Secondly, this method combines the power Markov chain model and power optimal control theory to accurately predict the development trend of faults and give the optimal intervention strategy. This forward-looking analysis and decision support can help operation and maintenance personnel take timely and effective preventive measures to significantly reduce the risk of system failures and maintenance costs. Finally, by introducing the evolutionary power digital twin model and power variational algorithm, this method realizes the adaptive optimization of the diagnostic and early warning system. The system can continuously learn and update, adapt to changes in equipment status and operating environment, and maintain long-term high-precision diagnostic capabilities. This continuous evolutionary feature makes this method particularly suitable for long-term health management of complex dynamic systems.

[0151] According to a second aspect of the embodiments of the present invention,

[0152] An electronic device is provided, comprising:

[0153] processor;

[0154] a memory for storing processor-executable instructions;

[0155] The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0156] According to a third aspect of the embodiments of the present invention,

[0157] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0158] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0159] 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 with 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 general control method for a transformer and reactor comprehensive test system, characterized in that: include: Acquire power evolution data of transformers and reactors in real time based on the deployed multi-dimensional sensor array; Based on the power evolution data, the correlation between the physical, electrical, and thermal characteristics of transformers and reactors is mapped, and the power calculation acceleration algorithm is used to calculate the correlation in real time to generate a power feature mapping matrix; the initial power digital twin model is constructed based on the power feature mapping matrix; Graph neural networks are deployed using edge computing nodes, and the initial power digital twin model is used as the node feature input of the graph network to build a dynamic graph topology based on power evolution data. The dynamic coupling effects between different components of transformers and reactors are captured through the graph attention mechanism. Compare the dynamic graph topology with the initial power digital twin model for power state error correction, calculate the power state deviation, establish a self-calibration optimization algorithm based on the power state deviation, dynamically correct the initial power digital twin model, and generate an evolutionary power digital twin model with adaptive optimization capabilities; A hybrid power-classical fault diagnosis system is built based on the evolutionary power digital twin model. The real-time collected power state evolution data is input into the hybrid power-classical neural network. The power characteristics are extracted through the power lines, and the classical neural network completes the fault mode recognition. When the abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state prediction of the fault development trajectory is carried out based on the power reinforcement learning algorithm. At the same time, the prediction results are fed back to the self-calibration optimization algorithm to realize the dynamic optimization of the power state of the evolutionary power digital twin model.

2. The method according to claim 1, characterized in that Based on the power evolution data, the correlation between the physical, electrical and thermal characteristics of transformers and reactors is mapped, and the power calculation acceleration algorithm is used to calculate the correlation in real time to generate a power feature mapping matrix; Building the initial power digital twin model based on the power feature mapping matrix includes: The power evolution data including magnetic field distribution data, temperature field distribution data and current density distribution data are acquired synchronously at a sampling rate of 100MHz; The magnetic field distribution data is encoded into a first power state to characterize physical properties, the temperature field distribution data is encoded into a second power state to characterize thermal properties, and the current density distribution data is encoded into a third power state to characterize electrical properties, thereby obtaining three sets of encoded characteristic power states; Applying power gate operation to the three sets of characteristic power states generates power entanglement. Power entanglement couples physical properties, thermal properties, and electrical properties into a unified multi-body correlation state. The power state tomography technique is used to characterize the internal structural characteristics of the multi-body correlation state, and the power correlation matrix that characterizes the multi-physical field coupling law of the transformer and the reactor is obtained. The power calculation acceleration algorithm is used to perform real-time evolution calculation of the power correlation matrix, and the power correlation information is mapped to the feature space through the power Fourier transform. The power error correction mechanism is introduced in the feature extraction process to ensure the calculation accuracy. The tensor network algorithm is used to compress and reorganize the feature information, and the power feature mapping matrix that characterizes the multi-physical field coupling law of the transformer and reactor is generated. The initial power digital twin model is constructed based on the power feature mapping matrix.

3. The method according to claim 1, characterized in that The dynamic graph topology is compared with the initial power digital twin model for power state error correction, the power state deviation is calculated, a self-calibration optimization algorithm is established based on the power state deviation, the initial power digital twin model is dynamically corrected, and an evolutionary power digital twin model with adaptive optimization capabilities is generated, including: The dynamic graph topology is converted into the target power state through power state encoding; the predicted power state in the initial power digital twin model is compared with the target power state input into the power error correction circuit, which includes a power entanglement measurement gate and a power phase correction gate. The density matrix difference between the two power states is obtained through power state tomography technology, and the power state deviation is calculated; A self-calibration optimization algorithm is constructed based on the power state deviation, the power state deviation is decomposed into a feature vector group, a variational power circuit is constructed based on the feature vector group, and the parameters of the variational power circuit are optimized by the power natural gradient method to obtain the optimal calibration parameters; The initial power digital twin model is dynamically corrected using the optimal calibration parameters, and the power annealing strategy is used to map the optimal calibration parameters to the model parameter space to obtain the corrected model parameters; The initial power digital twin model is updated based on the corrected model parameters, and the model prediction accuracy is continuously optimized through the Bayesian reasoning method, eventually generating an evolutionary power digital twin model with adaptive optimization capabilities.

4. The method according to claim 3, characterized in that The power state deviation is decomposed into a set of eigenvectors, and a variational power circuit is constructed based on the eigenvector group. The parameters of the variational power circuit are optimized by the power natural gradient method, and the optimal calibration parameters are obtained, including: The power state deviation is converted into an eigenvector group through singular value decomposition, the deviation density matrix of the power state deviation is calculated, the deviation density matrix is ​​diagonalized to obtain the eigenvalue spectrum and the corresponding eigenvector group, and the eigenvector group is ranked according to the size of the eigenvalue spectrum; Selecting a first eigenvector from the sorted eigenvector group, and when the cumulative contribution rate of the first eigenvector reaches a preset contribution threshold, constructing a variational power circuit based on the first eigenvector, wherein the variational power circuit is composed of a single power bit rotation gate layer and a multi-power bit entanglement gate layer alternately, and the rotation angle of each power gate is used as an adjustable parameter; An optimization objective function is constructed based on a variational power circuit. The optimization objective function includes a distance measurement term between the power state output and the target state and a parameter regularization term. The adjustable parameters are optimized by the power natural gradient method, and the power Fisher information matrix corresponding to the adjustable parameters is calculated. The product of the inverse matrix of the power Fisher information matrix and the objective function gradient is used as the parameter update direction, the Barzilai-Borwein method is used to adaptively adjust the optimization step size, and the momentum term is introduced in the parameter update process to accelerate convergence; Multiple convergence criteria including gradient norm threshold, parameter update amount threshold and objective function improvement rate threshold are set. When multiple convergence criteria are met at the same time, the parameter optimization iteration is stopped and the optimal calibration parameters are output.

5. The method according to claim 3, characterized in that The initial power digital twin model is updated based on the corrected model parameters. The prediction accuracy of the model is continuously optimized through the power Bayesian reasoning method. The model structure is continuously adjusted in combination with the online learning mechanism. Finally, an evolutionary power digital twin model with adaptive optimization capabilities is generated, including: The optimal calibration parameters are mapped to the variational power circuit of the initial power digital twin model, and the parameters of the initial power digital twin model are updated by adjusting the power gate rotation angle and entanglement strength of the variational power circuit to obtain a revised power digital twin model. The power state of the revised power digital twin model is expressed in the form of a density operator. The power system observation data is obtained based on power state tomography. The power Dirichlet distribution is used to build a priori probability model. The posterior probability distribution is calculated by the power Bayesian theorem. The prediction accuracy of the revised power digital twin model is continuously optimized according to the expected value of the posterior probability distribution. Based on the predicted longitude of the revised power digital twin model, the similarity index between the power state purity and the von Neumann entropy is calculated, the structural configuration of the variational power circuit is dynamically adjusted according to the similarity index, and the number of layers and power gate combinations of the variational power circuit are optimized using a progressive learning strategy to obtain the optimization result; The optimization results of the variational power circuit are fed back into the reasoning process of the power Bayesian theorem, the prior probability model is updated, and the power Bayesian reasoning and structural optimization are repeated until the model prediction accuracy meets the preset accuracy threshold requirement, and finally an evolutionary power digital twin model with adaptive optimization capability is generated.

6. The method according to claim 1, characterized in that When the abnormal power state evolution characteristics are identified, the power reinforcement learning algorithm is started, and the power state is predicted based on the fault development trajectory. At the same time, the prediction results are fed back to the self-calibration optimization algorithm to achieve the dynamic optimization of the power state of the evolutionary power digital twin model, including: Calculate the power fidelity, power trace distance and power relative entropy between the real-time power state and the standard power state, construct a multi-dimensional abnormal feature vector based on the power fidelity, power trace distance and power relative entropy, and identify the abnormal power state evolution characteristics when the deviation value of the multi-dimensional abnormal feature vector exceeds the preset abnormal threshold; The abnormal density matrix of the identified abnormal power state evolution characteristics is input into the reinforcement learning strategy network constructed by the variational power circuit, and the first fault development trajectory is generated based on the abnormal density matrix of the abnormal power state evolution characteristics; The abnormal power state evolution characteristics are input into the power reinforcement learning algorithm, a second fault development trajectory is generated based on the power reinforcement learning algorithm, and a prediction error is generated by comparing the first fault development trajectory with the second fault development trajectory; Based on the prediction error, a power loss function is constructed, and the power gradient of the power loss function to the parameters of the evolutionary power digital twin model is calculated. The power stochastic gradient descent method is used to update the parameters of the evolutionary power digital twin model to complete the power state dynamic optimization of the evolutionary power digital twin model.

7. The method according to claim 1, characterized in that The method also includes: Generate intelligent diagnosis and early warning reports based on power state prediction results. The intelligent diagnosis and early warning reports include fault power feature description, multi-domain coupling analysis, development trend prediction, and optimal intervention strategy. The diagnosis results are used to continuously optimize the power state mapping accuracy of the evolutionary power digital twin model. The power state prediction results are reconstructed through power state tomography technology to obtain a prediction density matrix, the expected value of the power observation quantity of the prediction density matrix is ​​calculated to construct a fault feature vector, and a fault power feature description is generated based on the fault feature vector; The data of different physical domains in the fault power feature description are mapped into a power state space to construct a multi-body power state, the power mutual information and power entanglement of the multi-body power state are calculated, and the multi-domain coupling analysis results that characterize the fault propagation characteristics are generated; According to the results of multi-domain coupling analysis, a power Markov chain model is constructed, and the transition probability of the fault state is calculated based on the power Markov chain model to generate the prediction results of the fault development trend; Based on the prediction results of fault development trend, the controllability criterion of power system is established, and the optimal control equation is solved to obtain the optimal intervention strategy based on the controllability criterion of power system; Integrate fault power feature description, multi-domain coupling analysis results, fault development trend prediction results and optimal intervention strategy to form an intelligent diagnosis and early warning report; Construct a power likelihood function to calculate the prediction accuracy of the diagnosis results in the intelligent diagnosis and early warning report, and evolve the diagnosis results into an initial power digital twin model; Based on the comparison between the output results of the evolutionary power digital twin model and the actual system state, the power state mapping deviation is continuously calculated, and the power variational algorithm is used to iteratively optimize the power state mapping parameters of the evolutionary power digital twin model to achieve continuous improvement in the power state mapping accuracy.

Images