Power equipment simulation analysis system based on machine learning
Through electrical-thermal coupled modeling, component-based step-down and deep learning optimization mechanisms, combined with GL-MLP-Trans and low-rank-Bi-LSTM models, the problem of insufficient computing efficiency and fault recognition capabilities in power equipment simulation analysis is solved, rapid modeling and efficient diagnosis are achieved, and simulation efficiency and fault recognition accuracy are improved.
Patent Information
- Application Number
- CN202511086492.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-09-02
- Estimated Expiration
- 2045-08-05
AI Technical Summary
The existing machine learning-based simulation analysis methods for power equipment have shortcomings in computing efficiency, modeling accuracy and fault identification capabilities, and it is difficult to meet the application scenarios with strong operating conditions and strong real-time requirements in large-scale power systems. In addition, traditional methods lack unified optimization and regulation of simulation accuracy and model stability, and huge redundant parameter scales.
Electric-thermal coupled physical modeling, component-based down-order modeling technology (ROM), deep learning optimization mechanism and low-rank fault analysis method are adopted, combined with GL-MLP-Trans model and low-rank-Bi-LSTM neural network, feature representation capabilities are enhanced through the multi-channel ghost node mechanism, combined with the Lyapunov exponential diagnostic strategy with feature spectrum direction weighted to achieve dynamic stability control, and a sub-module function approximate projection algorithm is used to filter key singular values to improve the robustness of the model and fault recognition accuracy.
It realizes rapid modeling and dynamic response of power equipment in complex environments, improves simulation efficiency and fault diagnosis capabilities, solves the problems of high computing complexity, rigid model and redundancy in traditional methods, and meets the needs of rapid response and lightweight under multiple operating conditions.
Smart Images

Figure CN120579470A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power equipment simulation, and in particular to a power equipment simulation analysis system based on machine learning. Background Art
[0002] With the continuous expansion of the scale of power systems and the increasing complexity of equipment operating environments, real-time perception of the operating status of power equipment and fault prediction have become the key to ensuring the safe and reliable operation of the system. In order to achieve a comprehensive understanding and efficient management of the operating behavior of power equipment, power equipment simulation technology has gradually become an important means of research and engineering practice. At present, traditional power equipment simulation methods mainly rely on sophisticated finite element modeling and numerical solution processes. For example, by establishing an electric-thermal coupling model, physical quantities such as equipment temperature rise and current density are accurately simulated. Although this type of method has certain advantages in terms of accuracy, its computational complexity is high and the simulation time is long, making it difficult to adapt to application scenarios with multiple working conditions and strong real-time requirements in large-scale power systems. In recent years, artificial intelligence, especially machine learning technology, has The rapid development of technology has provided a new path for the rapid simulation and intelligent fault analysis of power equipment. By introducing methods such as data-driven modeling, neural network structures, and reduced-order modeling, the operating status of power equipment can be efficiently predicted and evaluated while ensuring accuracy. However, existing simulation and analysis methods based on machine learning still have many shortcomings: First, traditional machine learning methods have limited ability to integrate operating condition characteristics with equipment structure information, making it difficult to accurately model the dynamic behavior between complex components. Second, in the process of assembling reduced-order models, there is a lack of a unified optimization and control mechanism between simulation accuracy and model stability. Third, existing fault diagnosis models generally have problems with large parameter sizes and significant redundancy, making it difficult to balance the generalization ability of modeling and the computational efficiency of operation. Summary of the Invention
[0003] The present invention aims to overcome the shortcomings of existing power equipment simulation and analysis methods in terms of computational efficiency, modeling accuracy and fault identification capabilities, and provides a power equipment simulation and analysis system based on machine learning. The system integrates electro-thermal coupling physical modeling, component-based reduced-order modeling technology (ROM), deep learning optimization mechanism and low-rank fault analysis method to achieve rapid simulation and intelligent fault diagnosis of power equipment; Among them, the power equipment simulation module proposes a rapid simulation method that integrates electro-thermal coupling modeling, orthogonal basis reduced-order analysis (POD) and component-based ROM assembly, and introduces the GL-MLP-Trans machine learning model to predict and optimize the simulation scheme; the model enhances the feature representation capability through the multi-channel ghost node mechanism, and combines the Lyap weighted feature spectrum direction The unov exponential diagnostic strategy realizes dynamic stability control, thereby improving the assembly accuracy and prediction robustness of ROM simulation; in terms of fault analysis, the present invention constructs a low-rank-Bi-LSTM neural network model, and effectively realizes the structural compression and performance improvement of the fault model by introducing low-rank perturbations and spectral projection gain mechanisms into the LSTM network weights; in particular, the submodular function approximation projection algorithm is used to screen key singular values to improve the fault identification accuracy and stability of the model in a limited parameter space; in general, the present invention organically combines physical modeling with machine learning to construct a power equipment simulation and analysis system that can respond quickly, is reconfigurable in structure, and has adaptive diagnostic capabilities, which can be widely used in scenarios such as smart grids, power operation and maintenance, and industrial equipment status assessment.
[0004] The present invention provides a power equipment simulation and analysis system based on machine learning, which includes a data acquisition module, a data processing module, a power equipment simulation module and a fault analysis module;
[0005] The data acquisition module collects heterogeneous signal data in real time through electrical transformers, thermal sensors, vibration detectors, and intelligent switch quantity acquisition units deployed on the power equipment side. This heterogeneous signal data is parsed and standardized according to the IEC-61850 protocol, and then clock synchronization is performed to achieve millisecond-level time alignment. Each piece of data is associated with a unique device ID, acquisition node code, and channel number, and the processed data is encapsulated into a unified structure format to obtain a real-time data set. Historical operating data and equipment model parameters are collected and combined with the real-time data set to construct power equipment operating data.
[0006] The data processing module filters, reduces noise, and selects features for power equipment operation data to generate standardized power equipment data;
[0007] The power equipment simulation module uses a rapid power equipment simulation method to process standardized power equipment data and obtain a set of simulation results. The rapid power equipment simulation method is constructed by integrating electro-thermal coupling modeling, orthogonal basis reduction (POD) analysis, component-based ROM model construction, and machine learning optimization mechanisms.
[0008] The fault analysis module builds a low-rank Bi-LSTM model, processes the simulation result set through the low-rank Bi-LSTM model, performs fault analysis, and generates fault analysis results.
[0009] Furthermore, the process of processing the standardized power equipment data by the power equipment rapid simulation method to obtain a set of simulation results specifically includes the following steps:
[0010] Step S1: Extract the geometric structure, electrical boundaries, and physical parameters of standardized power equipment data, establish an electro-thermal coupling simulation model, and perform multi-operating condition simulations in combination with the Navier–Stokes equations and the energy conservation equation to obtain the time-series distributed temperature field and current density physical quantities to form simulation snapshot data;
[0011] Step S2: Expand the simulation snapshot data according to the spatial dimension, construct a snapshot matrix, and perform singular value decomposition to extract dynamic features and construct a low-dimensional POD (orthogonal basis reduction decomposition) basis function space. Based on the POD basis function space, combined with the structural topology partitioning information in the standardized power equipment data, the power equipment is decomposed into regions and functional modules to generate a componentized ROM model library. The componentized ROM model library outputs ROM component metadata and reduced-order description parameters.
[0012] Step S3: Set the target simulation working condition parameters and introduce the GL-MLP-Trans model. The GL-MLP-Trans model optimizes the multi-task MLP-Transformer model by introducing a multi-channel ghost node mechanism and a characteristic spectrum direction weighted-Lyapunov exponent diagnosis mechanism. The target simulation working condition parameters, ROM component metadata, and reduced-order description parameters are jointly constructed into a working condition feature vector set, which is input into the GL-MLP-Trans model for training. The optimal ROM combination scheme and POD truncation dimension are predicted to generate an assembled ROM simulation model. The GL-MLP-Trans model includes a Transformer model and a multi-task MLP model.
[0013] Step S4: Combined with the assembled ROM simulation model, perform reduced-order simulation, and use the Galerkin projection method to perform time-series evolution on the assembled ROM simulation model, quickly output dynamic response quantities, including temperature rise distribution, current density, and electrothermal coupling stress, to obtain a simulation response sequence; extract the temperature rise peak, maximum current density, voltage fluctuation amplitude, and energy consumption ratio of the simulation response sequence; structure these indicators according to the working condition dimension and module area division to form a multi-dimensional response result set, and output the simulation result set.
[0014] Furthermore, the process of predicting the optimal ROM combination solution and POD truncation dimension specifically includes the following steps:
[0015] Step S31: Input the working condition feature vector set into the Transformer model for training. During the training process, a multi-channel ghost node mechanism is introduced to generate a working condition-structure joint feature vector.
[0016] The multi-channel ghost node mechanism specifically involves adding multi-channel ghost node vectors to the intermediate representation output by the Transformer model, performing dimensionality expansion and dimensionality perturbation, and using a regularization penalty term to control the intensity of dimensional perturbation to prevent perturbation overfitting during training.
[0017] Step S32: The working condition-structure joint feature vector is input into the multi-task MLP model for training, and two branches are output. Branch 1 predicts whether each ROM component participates in assembly, and branch 2 predicts the target POD truncation dimension. During the training process, the characteristic spectrum direction weighted-Lyapunov exponent diagnosis mechanism is introduced to dynamically evaluate the stability and stage-by-stage regulation of the training state to generate the optimal ROM combination scheme and POD truncation dimension.
[0018] Furthermore, the process of generating the fault analysis result by the fault analysis module specifically includes the following steps:
[0019] Step E1: Establish a Bi-LSTM model, inject low-rank weight parameters into the weight matrix of the Bi-LSTM model, perform low-rank perturbation compensation on the main weight, achieve efficient parameter expression and gradient compression, construct a low-rank Bi-LSTM model, input the simulation result set into the low-rank Bi-LSTM model, capture the forward and backward evolution characteristics of the power equipment state, and obtain the initial low-rank weight matrix;
[0020] Step E2: Perform singular value decomposition on the initial low-rank weight matrix, extract the singular value set, calculate the first-order gradient and second-order gradient corresponding to the singular value set, obtain the sensitivity and local curvature of each singular value to the performance of the low-rank-Bi-LSTM model, and form a singular value importance score;
[0021] Step E3: Based on the singular value importance score, define the set quadratic performance loss function, perform submodular function approximation projection on the set quadratic performance loss function, and construct a spectral projection gain function. The spectral projection gain function has submodularity and monotonicity. Using the spectral projection gain function as the search space basis, introduce a global rank budget constraint, and use a greedy search algorithm to traverse the singular value set layer by layer. Iteratively select the singular value component with the maximum marginal performance gain to generate an optimized low-rank weight matrix.
[0022] Step E4: Re-inject the optimized low-rank weight matrix into the low-rank-Bi-LSTM model to complete the fault feature modeling and classification tasks and generate fault analysis results.
[0023] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:
[0024] The present invention constructs a rapid simulation method for power equipment that integrates electro-thermal coupling modeling, orthogonal basis order reduction analysis and component-based ROM assembly mechanism, thereby realizing rapid modeling and dynamic response prediction of the power equipment state under complex operating environments, effectively improving the simulation efficiency and adaptability of the system under multiple working conditions and multiple load conditions; compared with the traditional method that relies on high-precision finite element simulation, this method can significantly reduce the computational complexity and realize flexible assembly and rapid response of equipment-level models; in particular, the GL-MLP-Trans model introduced in the present invention improves the robustness and generalization ability of the simulation model under structural changes and working condition uncertainties through a multi-channel disturbance enhancement mechanism and a characteristic spectrum direction weighted stability control strategy, solving the problems of model rigidity and error instability in traditional simulation methods; this mechanism enables the simulation module to automatically adjust the model structure and order reduction parameters according to real-time working conditions, realizing reconfigurable and rapidly deployable engineering simulation capabilities.
[0025] In terms of fault analysis, the present invention proposes a Bi-LSTM fault diagnosis method based on low-rank perturbation and spectral projection gain mechanism optimization. By introducing singular value screening and structural compression mechanisms, the model parameter scale and training resource consumption are effectively controlled while ensuring modeling accuracy. This method can accurately capture the forward and backward evolution laws in the equipment operation state, and improve the system's sensitivity and recognition ability to potential fault characteristics. In particular, the spectral gain screening mechanism constructed in combination with the submodular function approximation projection algorithm can dynamically select key feature paths under a limited parameter budget, avoid redundant modeling and overfitting problems, and enhance the practicality and real-time response capability of the fault diagnosis model. Compared with traditional models based on rules or full features, this method greatly compresses the calculation and deployment overhead while improving the judgment accuracy, meeting the requirements of edge computing and on-site diagnosis for model lightweight and timeliness.
[0026] In summary, the present invention constructs an electric power equipment simulation and analysis system that integrates rapid modeling, intelligent prediction and efficient diagnosis. By synergistically integrating physical modeling and intelligent algorithms, it realizes rapid response, accurate evaluation and stable operation support of electric power equipment under actual complex working conditions. Compared with traditional technical means, the present invention has achieved essential breakthroughs in many aspects such as simulation efficiency, diagnostic accuracy, model flexibility and resource occupancy. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 This is a module diagram of a power equipment simulation and analysis system based on machine learning proposed by the present invention;
[0028] Figure 2 This is a graph of the weighted Lyapunov exponent convergence curve of the characteristic spectrum direction proposed in Example 4.
[0029] Figure 2 In the figure, the orange curve is the estimated value of the Lyapunov exponent, the blue dotted line is the sliding average, the light blue area is the confidence interval, and the red dot and the "peak" mark are the positions of the maximum Lyapunov values. The bottom shows the training steps (0-200), and the left side shows the Lyapunov exponent value. DETAILED DESCRIPTION
[0030] The technical solutions 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, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0031] Example 1, according to Figure 1 , the present invention provides a power equipment simulation and analysis system based on machine learning, the system includes a data acquisition module, a data processing module, a power equipment simulation module and a fault analysis module;
[0032] The data acquisition module collects heterogeneous signal data in real time through electrical transformers, thermal sensors, vibration detectors, and intelligent switch quantity acquisition units deployed on the power equipment side. This heterogeneous signal data is parsed and standardized according to the IEC-61850 protocol, and then clock synchronization is performed to achieve millisecond-level time alignment. Each piece of data is associated with a unique device ID, acquisition node code, and channel number, and the processed data is encapsulated into a unified structure format to obtain a real-time data set. Historical operating data and equipment model parameters are collected and combined with the real-time data set to construct power equipment operating data.
[0033] The data processing module filters, reduces noise, and selects features for power equipment operation data to generate standardized power equipment data;
[0034] The power equipment simulation module uses a rapid power equipment simulation method to process standardized power equipment data and obtain a set of simulation results. The rapid power equipment simulation method is constructed by integrating electro-thermal coupling modeling, orthogonal basis reduction (POD) analysis, component-based ROM model construction, and machine learning optimization mechanisms.
[0035] The fault analysis module builds a low-rank Bi-LSTM model, processes the simulation result set through the low-rank Bi-LSTM model, performs fault analysis, and generates fault analysis results.
[0036] Example 2: This example is based on Example 1. In this example, the process of processing standardized power equipment data using the power equipment rapid simulation method to obtain a simulation result set specifically includes the following steps:
[0037] Step S1: Extract the geometric structure, electrical boundaries, and physical parameters of standardized power equipment data, establish an electro-thermal coupling simulation model, and perform multi-operating condition simulations in combination with the Navier–Stokes equations and the energy conservation equation to obtain the time-series distributed temperature field and current density physical quantities to form simulation snapshot data;
[0038] Step S2: Expand the simulation snapshot data according to the spatial dimension, construct a snapshot matrix, and perform singular value decomposition to extract dynamic features and construct a low-dimensional POD (orthogonal basis reduction decomposition) basis function space. Based on the POD basis function space, combined with the structural topology partitioning information in the standardized power equipment data, the power equipment is decomposed into regions and functional modules to generate a componentized ROM model library. The componentized ROM model library outputs ROM component metadata and reduced-order description parameters.
[0039] Step S3: Set the target simulation working condition parameters and introduce the GL-MLP-Trans model. The GL-MLP-Trans model optimizes the multi-task MLP-Transformer model by introducing a multi-channel ghost node mechanism and a characteristic spectrum direction weighted-Lyapunov exponent diagnosis mechanism. The target simulation working condition parameters, ROM component metadata, and reduced-order description parameters are jointly constructed into a working condition feature vector set, which is input into the GL-MLP-Trans model for training. The optimal ROM combination scheme and POD truncation dimension are predicted to generate an assembled ROM simulation model. The GL-MLP-Trans model includes a Transformer model and a multi-task MLP model.
[0040] The target simulation operating condition parameters include: environmental operating condition parameters, electrical load parameters, equipment operating condition settings, and thermal fluid parameters;
[0041] Step S4: Combined with the assembled ROM simulation model, perform reduced-order simulation, and use the Galerkin projection method to perform time-series evolution on the assembled ROM simulation model, quickly output dynamic response quantities, including temperature rise distribution, current density, and electrothermal coupling stress, to obtain a simulation response sequence; extract the temperature rise peak, maximum current density, voltage fluctuation amplitude, and energy consumption ratio of the simulation response sequence; structure these indicators according to the working condition dimension and module area division to form a multi-dimensional response result set, and output the simulation result set.
[0042] Embodiment 3: This embodiment is based on embodiment 1. In this embodiment, the process of processing standardized power equipment data to obtain a set of simulation results specifically includes the following steps:
[0043] Step Q1: Extract the geometric structure, electrical boundaries, and physical parameters of standardized power equipment data, establish an electro-thermal coupling simulation model, and perform multi-condition simulations in combination with the Navier–Stokes equations and energy conservation equations to obtain the time-series distribution of temperature fields and current density physical quantities, generating simulation snapshot data.
[0044] Step Q2: Expand the simulation snapshot data by spatial dimension, construct a snapshot matrix, and perform singular value decomposition (SVD) to extract dynamic features and construct a low-dimensional POD (orthogonal basis reduction) basis function space. Based on the POD basis function space and combined with the structural topology partitioning information in the standardized power equipment data, the power equipment is decomposed into regions and functional modules to generate a componentized ROM model library. The componentized ROM model library outputs ROM component metadata and reduced-order description parameters.
[0045] Step Q3: Introduce the Transformer model, collect the target simulation condition parameters, and construct the target simulation condition parameters together with the ROM component metadata and reduced-order description parameters into a condition feature vector set. Input it into the Transformer model for training, predict the optimal ROM combination scheme and POD truncation dimension, and generate an assembled ROM simulation model.
[0046] Step Q4: Combined with the assembled ROM simulation model, perform reduced-order simulation, and use the Galerkin projection method to perform time-series evolution on the assembled ROM simulation model, quickly output dynamic response quantities, including temperature rise distribution, current density, and electrothermal coupling stress, to obtain a simulation response sequence; extract the temperature rise peak, maximum current density, voltage fluctuation amplitude, and energy consumption ratio of the simulation response sequence; structure these indicators according to the operating condition dimension and module area division to form a multi-dimensional response result set, and output the simulation result set.
[0047] Example 4, according to Figure 2This embodiment is based on the second embodiment. In this embodiment, the process of predicting the optimal ROM combination scheme and POD truncation dimension specifically includes the following steps:
[0048] Step S31: Input the working condition feature vector set into the Transformer model for training. During the training process, a multi-channel ghost node mechanism is introduced to achieve convergence acceleration in the early stage of training and convergence stability control in the middle and late stages of training, and generate a working condition-structure joint feature vector;
[0049] The multi-channel ghost node mechanism specifically involves adding multi-channel ghost node vectors to the intermediate representation output by the Transformer model to perform dimensional expansion and dimensional perturbation. A regularized penalty term is used to control the intensity of the dimensional perturbation to prevent overfitting during training. This mechanism guides the Transformer model to break through local saddle points in the early stages of training, improving its ability to explore the input distribution and accelerating the convergence of contextual representations. In the middle and late stages of training, the intervention of ghost nodes is gradually weakened through a time decay strategy to achieve gradual convergence of the enhanced representation vector, improving the modeling robustness and generalization ability of the Transformer model for the set of working condition feature vectors.
[0050] Multi-channel ghost node mechanism structure and modeling method:
[0051] (1) Multi-channel disturbance coordination:
[0052] ;
[0053] in, represents the time step, represents the enhanced representation vector after adding ghost nodes, represents the original intermediate representation vector, Represents vector concatenation operation, represents the time decay factor, Represents an index variable, represents the total number of ghost node channels, Represents the dynamic aggregation weight calculated based on the context, Indicates the A ghost node vector is a learnable vector used to represent the perturbation dimension. It is initially random noise or perturbation and gradually participates in representation training.
[0054] (2) Regularized penalty control:
[0055] ;
[0056] in, represents the ghost node regularization loss, represents the regularization coefficient; Indicates the ghost node vectors The square of the L2 norm of ;
[0057] The multi-channel ghost node mechanism enhances the diversity and structural sensitivity of perturbation representation; dynamic aggregation makes perturbation representation scene-selective and semantically adaptable; time decay and regularization control work together to ensure rapid breakthrough of saddle points in the early stages of model training and gradual convergence in the middle and late stages.
[0058] Step S32: Input the working condition-structure joint feature vector into the multi-task MLP model for training, and output two branches. Branch 1 predicts whether each ROM component participates in the assembly, and branch 2 predicts the target POD truncation dimension. During the training process, the characteristic spectrum direction weighted-Lyapunov exponent diagnosis mechanism is introduced to dynamically evaluate the stability and stage-by-stage regulation of the training state to generate the optimal ROM combination scheme and POD truncation dimension. The formula used is as follows:
[0059] Characteristic spectrum direction weighted-Lyapunov exponent formula:
[0060] ;
[0061] in, Indicates the cumulative number of steps. Indicates the The weighted Lyapunov exponent estimate of the characteristic spectrum direction of the step, Represents the index of the current feature value during training, Represents the total number of eigenvalues of the Hessian matrix, that is, the parameter dimension of the multi-task MLP model; Indicates the Moment The weighting coefficients of the spectral directions; Indicates in The Hessian matrix of the step eigenvalues; represents the learning rate, Indicates in During the training, along the The logarithmic measure of the convergence of the perturbations in the characteristic directions;
[0062] In this embodiment, Figure 2 The figure shows the convergence curve of the weighted Lyapunov exponent of the characteristic spectrum direction during the training process of training steps 0-200;
[0063] Figure 2 In the middle, the orange curve is the estimated value of the Lyapunov exponent, the blue dashed line is the sliding average, the light blue area is the confidence interval, and the red dot and the "peak" mark are the positions of the maximum Lyapunov value.
[0064] The characteristic spectrum direction weighted-Lyapunov exponent diagnostic mechanism introduces a weighted expression of the Hessian characteristic spectrum structure on the basis of the classical Lyapunov exponent estimation method, thereby enhancing the ability to distinguish the convergence states of different disturbance directions and directional sensitivity.
[0065] Embodiment 5: This embodiment is based on embodiment 4. In this embodiment, the process of generating the fault analysis result by the fault analysis module specifically includes the following steps:
[0066] Step E1: Establish a Bi-LSTM model, inject low-rank weight parameters into the weight matrix of the Bi-LSTM model, perform low-rank perturbation compensation on the main weight, achieve efficient parameter expression and gradient compression, construct a low-rank Bi-LSTM model, input the simulation result set into the low-rank Bi-LSTM model, capture the forward and backward evolution characteristics of the power equipment state, and obtain the initial low-rank weight matrix;
[0067] Step E2: Perform singular value decomposition on the initial low-rank weight matrix, extract the singular value set, calculate the first-order gradient and second-order gradient corresponding to the singular value set, obtain the sensitivity and local curvature of each singular value to the performance of the low-rank-Bi-LSTM model, and form a singular value importance score;
[0068] Step E3: Based on the singular value importance score, define the set quadratic performance loss function, perform submodular function approximation projection on the set quadratic performance loss function, and construct a spectral projection gain function. The spectral projection gain function has submodularity and monotonicity. Using the spectral projection gain function as the search space basis, introduce a global rank budget constraint, and use a greedy search algorithm to traverse the singular value set layer by layer. Iteratively select the singular value component with the maximum marginal performance gain to generate an optimized low-rank weight matrix.
[0069] Step E4: Re-inject the optimized low-rank weight matrix into the low-rank-Bi-LSTM model to complete the fault feature modeling and classification tasks and generate fault analysis results;
[0070] Submodular function approximation projection refers to the use of first-order gradient terms and second-order Hessian information to extract the local gain structure from the set quadratic performance loss function, constructing a function form that satisfies submodularity and monotonicity constraints, facilitating the use of greedy search for subset optimization. The formula used is as follows:
[0071] Define the set quadratic performance loss function:
[0072] ;
[0073] in, Indicates the range from rank The singular value vectors selected from , represents the set quadratic performance loss function, represents the Hessian matrix, represents the transpose operation, represents the gradient vector, represents the second-order term, represents the first-order term;
[0074] This function represents the overall performance loss, ranging from rank By selecting some singular value subsets (Other singular values are set to zero), construct the spectral projection gain function, and characterize the marginal contribution of different singular value combinations to the performance;
[0075] Spectral projection gain function formula:
[0076] ;
[0077] in, Represents a set of singular values The spectral projection gain function, represents the singular value vector Select index collection The corresponding subvector, Represents the set quadratic performance loss function About Singular Value Vectors The first-order gradient vector of express and The inner product of express The transpose of Represents the set quadratic performance loss function About Singular Value Subsets The local second-order approximation of Represents the set quadratic performance loss function About Singular Value Vectors The second-order partial derivative matrix of , that is, the Hessian submatrix, Represents the linear loss change term in the direction of gradient descent, represents the second-order approximation term.
[0078] Example 6. This example is based on Example 4. In this example, the fault analysis module generates a fault analysis result, which specifically includes the following steps: establishing a Bi-LSTM model, inputting the simulation result set into the Bi-LSTM model, and generating a fault analysis result.
[0079] Embodiment 7: This embodiment is based on embodiment 6. In this embodiment,
[0080] The power equipment simulation module introduces a rapid simulation method for power equipment. This method is constructed by integrating electro-thermal coupling modeling, orthogonal basis order reduction (POD) analysis, component-based ROM model construction, and machine learning optimization mechanisms. This method processes standardized power equipment data to obtain a set of simulation results.
[0081] Table 1 obtained in this embodiment is the optimal ROM combination scheme and POD truncation dimension:
[0082] Table 1
[0083] ;
[0084] Table 2 is a collection of simulation results:
[0085] Table 2
[0086] ;
[0087] The fault analysis module builds a low-rank Bi-LSTM model, processes the simulation result set through the low-rank Bi-LSTM model, performs fault analysis, and generates fault analysis results;
[0088] Table 3 shows the failure analysis results:
[0089] Table 3
[0090] .
[0091] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.
Claims
1. A power equipment simulation and analysis system based on machine learning, comprising a data processing module, wherein the data processing module collects power equipment operation data, processes the power equipment operation data, and generates standardized power equipment data; characterized in that: The system also includes a power equipment simulation module and a fault analysis module; The power equipment simulation module processes standardized power equipment data using a fast power equipment simulation method to obtain a set of simulation results; The fault analysis module constructs a low-rank-Bi-LSTM model, processes the device simulation results through the low-rank-Bi-LSTM model to perform fault analysis, and generates a fault analysis result.
2. The power equipment simulation and analysis system based on machine learning according to claim 1, characterized in that: The process of obtaining a set of simulation results using the power equipment rapid simulation method specifically includes the following steps: Step S1: Processing standardized power equipment data to generate simulation snapshot data; Step S2: generating a componentized ROM model library based on the simulation snapshot data, wherein the componentized ROM model library outputs ROM component meta-information and reduced-order description parameters; Step S3: Set the target simulation condition parameters, introduce the GL-MLP-Trans model, and construct the target simulation condition parameters together with the ROM component metadata and reduced-order description parameters into a condition feature vector set. The condition feature vector set is input into the GL-MLP-Trans model for training, and the optimal ROM combination scheme and POD truncation dimension are predicted to generate an assembled ROM simulation model. Step S4: Combine the assembled ROM simulation model and output a simulation result set.
3. The power equipment simulation and analysis system based on machine learning according to claim 2, characterized in that: The GL-MLP-Trans model includes the Transformer model and the multi-task MLP model.
4. The power equipment simulation and analysis system based on machine learning according to claim 3, characterized in that: The process of predicting the optimal ROM combination solution and POD truncation dimension specifically includes the following steps: Step S31: inputting the working condition feature vector set into the Transformer model for training to generate a working condition-structure joint feature vector; Step S32: The working condition-structure joint feature vector is input into the multi-task MLP model for training. During the training process, the characteristic spectrum direction weighted-Lyapunov exponent diagnosis mechanism is introduced to dynamically evaluate the stability and stage-by-stage regulation of the training state to generate the optimal ROM combination scheme and POD truncation dimension.
5. The power equipment simulation and analysis system based on machine learning according to claim 4, characterized in that: In step S31, a multi-channel ghost node mechanism is introduced during the training process. Multi-channel ghost node vectors are added to the intermediate representation output by the Transformer model to perform dimensional perturbation, and a regular penalty term is used to control the intensity of the dimensional perturbation.
6. The power equipment simulation and analysis system based on machine learning according to claim 1, characterized in that: The fault analysis module generates fault analysis results, which specifically includes the following steps: Step E1: Establish a Bi-LSTM model, inject low-rank weight parameters into the weight matrix of the Bi-LSTM model, construct a low-rank-Bi-LSTM model, input the simulation result set into the low-rank-Bi-LSTM model, and obtain the initial low-rank weight matrix; Step E2: Perform singular value decomposition on the initial low-rank weight matrix to obtain a set of singular values to form a singular value importance score; Step E3: Based on the singular value importance score, define the set quadratic performance loss function, perform submodular function approximation projection on the set quadratic performance loss function, and construct a spectral projection gain function; using the spectral projection gain function as the search space basis, introduce a global rank budget constraint, and use a greedy search algorithm to traverse the singular value set to generate an optimized low-rank weight matrix; Step E4: Re-inject the optimized low-rank weight matrix into the low-rank-Bi-LSTM model to generate fault analysis results.
Citation Information
Patent Citations
Simulation prediction method and device based on big data, equipment and storage medium
CN119066617A
Power grid communication equipment fault identification method and system based on multi-source heterogeneous data representation learning
CN119397360A
Pump turbine runner analysis method and system
CN119885967A
Electrical equipment fault diagnosis and prediction analysis system
CN120372505A
Nuclear power plant equipment fault diagnosis method
CN120386327A
