Generator stator winding temperature field online simulation method

Abstract

The invention relates to the technical field of temperature field simulation application, in particular to a generator stator winding temperature field online simulation method, which comprises the following steps: establishing a simulation model, and obtaining an offline simulation calculation sample set; obtaining a historical normal operation data set, and correcting the obtained sample set to obtain a corrected sample set; performing order reduction processing on the simulation model to obtain an order reduction model; constructing a proxy model by using the reduced-order model; training a proxy model by using the corrected sample set; and on the basis of real-time working condition parameters, in combination with state monitoring data, multi-physics field online simulation analysis is completed by utilizing the trained agent model, and a simulation analysis result is obtained. By means of the simulation method, the problem that existing simulation precision is low can be effectively solved.

Description

Technical Field

[0001] The invention relates to the technical field of temperature field simulation application, and in particular to an online simulation method for the temperature field of a generator stator winding. Background Art

[0002] Clean and efficient coal-fired power will continue to play an important role in ensuring energy and power security for a long time. The safe and stable operation of large clean and efficient coal-fired power generating units plays an important role in ensuring the safety of the energy system. The flexible operation of coal-fired power units is a key means to improve the regulation capacity of new power systems. As a typical multi-parameter coupled nonlinear system, the high parameter development trend of the generating unit makes the operating characteristics increasingly complex.

[0003] Online simulation of multi-physics fields is an important foundation for the development of digital twins. The current multi-physics field simulation based on the fusion of artificial intelligence models can achieve high-efficiency, high-precision, and high-generalization performance online real-time evaluation and prediction of three-dimensional physical fields. Based on three-dimensional virtual shapes and combined with the dynamic properties of multi-physics fields, a three-dimensional digital prototype of equipment that integrates physical state information can be constructed to achieve comprehensive perception, evaluation and early warning of equipment status.

[0004] However, the state parameters of the current generator sets in flexible operation are complex and changeable, but the number of monitoring sensors is limited, and there are many monitoring blind spots inside the units. The existing multi-physics field simulation has the problems of slow speed and inability to simulate online.

[0005] For example, in the prior art, a Chinese invention patent document with publication number CN118657092A and publication date September 17, 2024 is proposed. The technical solution disclosed in the patent document is as follows: a multi-physics field calculation method and system for online simulation of digital twins, which establishes a multi-physics field coupling simulation calculation model of the simulated object, simplifies the multi-physics field coupling simulation calculation model, reduces the order of the temperature field simulation model, obtains a low-precision data set through temperature field analysis, and constructs a basic data-driven model through the low-precision data set and its corresponding sample space, thereby outputting the temperature field distribution result.

[0006] In the above technical solution, the initial multi-physics field model is not modified, and the sample space is preset, resulting in low simulation accuracy. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes an online simulation method for the temperature field of a generator stator winding, which can effectively solve the problem of low existing simulation accuracy.

[0008] The present invention is achieved by adopting the following technical solutions: A method for online simulation of a temperature field of a generator stator winding comprises the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set; Step S2. Obtain a historical normal operation data set, correct the sample set in step S1, and obtain a corrected sample set; reduce the order of the simulation model in step S1 to obtain a reduced-order model; and construct a proxy model using the reduced-order model; Step S3. Training the proxy model using the modified sample set; Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0009] The step S1 specifically refers to: establishing a simulation model based on the target unit, integrating electromagnetic, ventilation multi-physical fields and multi-operating condition boundary conditions to carry out offline finite element calculations, and obtaining a sample set.

[0010] When performing finite element calculations, the step size used for linearly related data is longer than the step size used for nonlinearly related data.

[0011] The method for obtaining the corrected sample set in step S2 is: obtaining a historical normal operating data set of the target unit, adjusting and adapting offline multi-physics field finite element calculation parameters according to the deviation between the historical normal operating data set and the sample set of the same operating condition, and calculating and obtaining the corrected sample set.

[0012] The method of order reduction is physical coordinate order reduction method based on Krylov subspace method.

[0013] The agent model adopts a rolling condition optimization algorithm framework to achieve online continuous training of the agent model.

[0014] The method also includes step S5, evaluating the status of the unit and risk warning.

[0015] The step S5 specifically refers to: completing the unit status assessment and risk warning based on the simulation analysis results obtained in step S4 through deep integration of sensor status data, mechanism information and expert knowledge.

[0016] When conducting unit status assessment and risk warning, it also includes: using multi-dimensional state parameter cloud map real-time reconstruction technology to complete the three-dimensional visualization of the unit that integrates physical state information.

[0017] It also includes the use of transfer learning technology to apply the trained proxy model to simulation tasks other than the target unit.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In the present invention, after the initial simulation model is established and offline calculations are performed, the calculation parameters are adjusted according to the normal operation data set of the unit to obtain a corrected sample set, which can improve the simulation accuracy. Furthermore, the present invention reduces the amount of calculation and the calculation cycle without reducing the model accuracy by introducing multi-physics coupling model reduction and proxy model technology, and can achieve high-efficiency, high-precision, and high-generalization online real-time evaluation and prediction of three-dimensional physical fields.

[0019] The above method can solve the problems of limited internal sensors of the generator set resulting in many monitoring blind areas, slow existing multi-physics field simulation speed and inability to simulate online.

[0020] 2. In the present invention, by targeting the simulation process, the mechanism model plus dynamic data-driven fusion is adopted to realize dynamic tuning of the proxy model parameters. When the online simulation analysis is finally performed, it is essentially a fusion of the reduced-order model, the trained proxy model and the model of real-time data input. It has a high degree of approximation to the key working conditions and has stronger state estimation and prediction capabilities than traditional simulation. On the other hand, it also provides a reliable model input and result verification tool for iterative optimization of working conditions.

[0021] 3. In the present invention, different step sizes are used for different data to form multiple groups of finite element calculation inputs, which can optimize the overall calculation process.

[0022] 4. The present invention performs order reduction processing through the Krylov subspace method, which can significantly reduce the computing time and resource consumption while retaining the characteristics and features of the original model.

[0023] 5. The proxy model adopts a rolling condition optimization algorithm framework, which enables the model to adapt to new data and environmental conditions by updating the model parameters at each time step. In multi-physics field coupling simulations, such as the magnetic-thermal coupling of generator stator bars, the framework can handle the interaction between different physical fields. The physical field prediction proxy model construction method combined with the prior knowledge of the reduced-order model can improve the prediction accuracy and interpretability of the model under sparse observation data. This method uses the powerful fitting ability of deep neural networks to output reference physical fields as additional observations, and jointly constrains the optimization problem solving with sensor data, thereby improving the problem solving accuracy.

[0024] 6. The present invention deeply integrates sensor status data, mechanism information and expert knowledge, completes unit status assessment and risk warning based on simulation analysis results, realizes multi-dimensional fusion assessment, and the assessment results are more accurate.

[0025] 7. Through 3D visualization, real-time monitoring and analysis of the current status and performance of the product can be achieved to schedule preventive and predictive maintenance activities.

[0026] 8. In the present invention, the migration technology can also be used to apply the trained proxy model to simulation tasks other than the target unit, making the present invention more universally applicable. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, wherein: Figure 1 It is a schematic diagram of dynamic data driven simulation in the present invention. DETAILED DESCRIPTION

[0028] Example 1 As a basic embodiment of the present invention, the present invention includes an online simulation method for the temperature field of a generator stator winding, comprising the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set.

[0029] Step S2. Obtain a historical normal operation data set, correct the sample set in step S1, and obtain a corrected sample set. Perform order reduction processing on the simulation model in step S1 to obtain a reduced-order model. Use the reduced-order model to construct a proxy model.

[0030] Step S3: Train the proxy model using the revised sample set.

[0031] Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0032] Example 2 As a preferred embodiment of the present invention, the present invention includes an online simulation method for the temperature field of a generator stator winding, comprising the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set. Specifically, a simulation model can be established based on the target unit, and offline finite element calculations can be performed by integrating electromagnetic and ventilation multi-physics fields and multi-operating condition boundary conditions to obtain a sample set.

[0033] Step S2. The simulation model in step S1 is reduced in order to obtain a reduced-order model; and a proxy model is constructed using the reduced-order model. Specifically, the simulation model in step S1 can be reduced in order using a conventional reduction-order processing method in the art, and a fast input-output mapping can be constructed for the reduced-order model, and the proxy model can be obtained by selecting an artificial neural network (ANN), a Gaussian process (Kriging), or the like.

[0034] Obtain a historical normal operation data set, and correct the sample set in step S1 to obtain a corrected sample set. The method for obtaining the corrected sample set may specifically be: obtain a historical normal operation data set of the target unit, adjust and adapt the offline multi-physics field finite element calculation parameters according to the deviation between the historical normal operation data set and the sample set of the same operating condition, and calculate and obtain the corrected sample set.

[0035] Step S3: Train the proxy model using the revised sample set.

[0036] Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0037] Example 3 As another preferred embodiment of the present invention, the present invention includes an online simulation method for the temperature field of a generator stator winding, comprising the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set.

[0038] Step S2. Obtain a historical normal operation data set, correct the sample set in step S1, and obtain a corrected sample set. Perform order reduction processing on the simulation model in step S1 to obtain a reduced-order model. Use the reduced-order model to construct a proxy model.

[0039] Step S3: Using the modified sample set to train the proxy model. The proxy model uses a rolling condition optimization algorithm framework to achieve online continuous training of the proxy model.

[0040] Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0041] Step S5. Evaluate the unit status and risk warning. Specifically, through the deep integration of sensor status data, mechanism information and expert knowledge, based on the simulation analysis results obtained in step S4, complete the unit status evaluation and risk warning.

[0042] Example 4 As the best embodiment of the present invention, the present invention includes an online simulation method for the temperature field of a generator stator winding, comprising the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set. Specifically, a simulation model is established based on the target unit, and offline finite element calculations are carried out by integrating electromagnetic and ventilation multi-physical fields and multi-condition boundary conditions to obtain a sample set. A training sample set is constructed through a certain number of offline simulations, and different operating conditions are analyzed. Among them, the input simulation parameters are the characteristic operating points proposed above, including input parameters (such as current, voltage, cooling condition temperature, coordinate position, etc.) and output responses (such as winding temperature, electromagnetic loss, etc.), thereby forming a sample set covering a wide range of operating conditions. It helps to fully understand the performance of the generator under different cooling conditions and provide data support for the design and training of subsequent proxy models. When building an offline sample database, three aspects need to be clarified: ① Physical modeling: Determine the specific loss types to be included in the calculation, consider the impact of different heat dissipation forms on the system, and finally clarify the temperature value of the spatial location that needs to be calculated.

[0043] ② Calculation granularity: According to the requirements of the front-end display and volume restrictions, determine how many spatial regions the stator bar is divided into for temperature value calculation. Comprehensively consider the workload and time consumption of finite element calculation to balance accuracy and efficiency.

[0044] ③ Calculation step length: The choice of step length directly affects the accuracy of the final proxy model calculation. It should be moderate, neither too long nor too short. Combined with the actual input measurement points on site, a longer step length can be used for linearly related data, while a shorter step length is required for nonlinear data. Finally, multiple sets of finite element calculation inputs are formed in this way to optimize the overall calculation process.

[0045] Step S2. This step may specifically include step S 21 and step S 22 Among them, step S 21 and step S 22 There is no order requirement, step S 21 and step S 22 They can be performed one after the other or simultaneously.

[0046] Step S 21. Obtain a historical normal operation data set, and correct the sample set in step S1 to obtain a corrected sample set. Specifically, obtain a historical normal operation data set of the target unit, adjust and adapt the offline multi-physics field finite element calculation parameters according to the deviation between the historical normal operation data set and the sample set of the same operating condition, and calculate and obtain the corrected sample set.

[0047] Among them, the adjustment and adaptation of offline multi-physics field finite element calculation parameters can be based on data-driven parameter inversion optimization. The key is to establish a "deviation-parameter" mapping relationship and iteratively correct the physical parameters of the finite element model (such as material properties, boundary conditions, coupling coefficients, etc.) through the optimization algorithm to make the simulation results close to the real historical data.

[0048] This adjustment method can also adopt existing technology, and can be subsequently improved through proxy model acceleration + Bayesian optimization + physical regularization methods.

[0049] Step S 22. The simulation model of step S1 is subjected to order reduction processing to obtain a reduced-order model, and a proxy model is constructed using the reduced-order model.

[0050] Specifically, a suitable reduction processing method is selected, and the reduced order model is mainly obtained by reducing the stiffness matrix, damping matrix and mass matrix in the finite element. The reduced order model can reflect the main characteristics of the original simulation model. After converting the original simulation model into a reduced order model, it can not only maintain a high accuracy, but also efficiently obtain the approximate solution of large-scale geometric morphology and structural mechanical properties.

[0051] Existing physical order reduction methods are mainly divided into three types: physical coordinate order reduction, generalized coordinate order reduction and hybrid coordinate order reduction. Compared with the other two methods, the physical coordinate order reduction method has high computational efficiency, and its accuracy and convergence speed depend on the number and position of the selected master degrees of freedom. Commonly used physical coordinate order reduction methods include Guyan condensation method, dynamic condensation method, Krylov subspace method, etc. Based on the particularity of the generator stator winding, the present invention deliberately selects physical coordinate order reduction based on the Krylov subspace method.

[0052] Physical coordinate reduction based on Krylov subspace method is a widely used technique in many fields, especially when dealing with large sparse matrix linear equations. The core of this method is to find approximate solutions in Krylov subspace through an iterative process, thereby reducing the dimensionality and computational complexity of the problem.

[0053] set up A ∈ R n×n , r ∈ R n ,say ; Among them, the formula is a Krylov subspace generated by A and r, which has several basic properties: Krylov subspaces are nested, that is: K1∈K2∈ .... ∈K m ∈ .... , K m has a dimension not exceeding m， K m ( A , r ) = { x = p ( A ) r : p is a polynomial of degree not exceeding m - 1}.

[0054] In the Krylov subspace method, the Arnoldi process is an important orthogonalization technique used to generate an orthogonal basis for the Krylov subspace. The Arnoldi process is a method for calculating a set of bases for Km through the Gram - Schmidt orthogonalization process. The steps are as follows: First, normalize r, that is, ; Calculate the A norm ; Make a projection ; If , the calculation ends, otherwise normalize ; , loop back to the second step until j = m ends.

[0055] If at the k - th (k < m) step there is , the algorithm will terminate prematurely. At this time, must be linearly represented by v1, v2,..., v k . This process is crucial for ensuring the convergence and stability of the iterative method. The reduced - order model can significantly reduce the computational time and resource consumption while improving the computational efficiency, and can ensure that the relative error of the calculation results is within an acceptable range, while retaining the characteristics and features of the original model.

[0056] Construct a fast input - output mapping for the reduced - order model. Methods such as artificial neural network (ANN), Gaussian process (Kriging), etc. can be selected to obtain a surrogate model. The surrogate model can store a large number of parameters and rules, thus quickly responding in real - time simulation and reducing the burden of online calculation. The surrogate model can store a large number of parameters and rules, thus quickly responding in real - time simulation and reducing the burden of online calculation.

[0057] Step S3. Train the surrogate model using the corrected sample set.

[0058] For the agent model, a rolling condition optimization algorithm framework is used to achieve online continuous training of the agent model. The rolling condition optimization algorithm framework enables the model to adapt to new data and environmental conditions by updating the model parameters at each time step. In multi-physics field coupling simulation, such as the magnetic-thermal coupling of the generator stator bars, the framework can handle the interaction between different physical fields. The physical field prediction agent model construction method combined with the prior knowledge of the reduced-order model can improve the prediction accuracy and interpretability of the model under sparse observation data. This method uses the powerful fitting ability of deep neural networks to output reference physical fields as additional observations, which are jointly constrained with sensor data to solve the optimization problem, thereby improving the problem solving accuracy. The agent model interacts with the environment through perception, decision-making and execution modules to achieve specific goals. The core concepts include state representation, goal setting and strategy selection. This framework allows the model to update its parameters at each time step to adapt to new data and environmental conditions.

[0059] In addition, the agent model has the ability to self-evolve and self-adapt. The agent model collects simulation data in the virtual environment through autonomous learning and search methods, and drives the model to optimize cognitive decisions based on historical data in the actual physical environment. Through continuous iterative updates, the difference between the digital twin and the physical entity is reduced, and the authenticity and accuracy of the simulation is improved. This ability enables the model to "learn" and improve over time, thereby providing more accurate predictions and diagnoses.

[0060] Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0061] The trained proxy model is a digital mapping of the physical object and is the basis for evaluation and diagnosis analysis. The accuracy of the modeling determines the accuracy, credibility and reliability of the analysis results. The proxy model is established using a modeling method that combines a mechanism model with data-driven fusion. On the one hand, it continuously improves the degree of approximation to key working conditions through online training, and on the other hand, it provides a reliable model input and result verification tool for iterative optimization of working conditions.

[0062] Among them, dynamic data-driven fusion simulation is a simulation method that "combines models and data". It continuously injects observations (data) of the real system into the simulation (model) and allows the data to dynamically correct the simulation (state, parameters), thereby improving the estimation and prediction capabilities based on simulation. Figure 1 As shown, since dynamic data-driven simulation integrates information from both model prediction and real-time observation, it can more accurately estimate the system state and predict the future evolution of the state. Neural networks have good nonlinear mapping capabilities and can provide more accurate solutions to nonlinear problems that are difficult to solve using conventional methods than general methods.

[0063] The core elements of dynamic data-driven fusion simulation are simulation model (also called system model), observation model, and data assimilation algorithm. In dynamic data-driven simulation, the data of the real system is continuously assimilated into the simulation, so that the simulation can dynamically adjust its own state to be closer to the state of the real system. If the model parameters are augmented into the model state, the model parameters can also be estimated together with the state as the data assimilation proceeds. By assimilating the data of the real system, a simulation can be obtained in which the simulation state dynamically approaches the state of the real system and the model parameters are dynamically tuned. Therefore, dynamic data-driven simulation has stronger state estimation and prediction capabilities than traditional simulation.

[0064] Based on the multi-physics field simulation reduced-order model library, neural network learning is used to construct an adaptive online dynamic simulation model that integrates mechanism and data, that is, a trained proxy model. It has the characteristics of small computational complexity, short computational cycle, and no reduction in model accuracy, thereby enabling online dynamic simulation.

[0065] Step S5. Evaluate the unit status and risk warning. Specifically, through the deep integration of sensor status data, mechanism information and expert knowledge, based on the simulation analysis results obtained in step S4, complete the unit status evaluation and risk warning. Utilize the real-time reconstruction technology of multi-dimensional state parameter cloud map to complete the three-dimensional digital prototype of the unit that integrates physical state information, and use advanced technologies such as big data storage and multi-level indexing to reduce the difficulty of multi-dimensional parameter reconstruction. By studying virtual-real mapping and real-time virtual visualization, the reconstruction of multi-dimensional state parameter cloud map under the real-time operating conditions of the generator is realized. It can be used by operation and maintenance personnel to monitor and analyze the current status and performance of the product in real time to schedule preventive and predictive maintenance activities.

[0066] Based on real-time operating parameters and combined with condition monitoring data, transfer learning technology can also be used to apply the trained proxy model to simulation tasks other than the target unit. The core of transfer learning is to use the learning results of existing models in similar tasks to reduce the need for a large amount of new data. Specifically, high-level features are extracted from the trained proxy model, which already contain rich physical information. Then, the trained proxy model is fine-tuned on an experimental data set other than the target unit, and the previously trained model parameters are adjusted through a small amount of experimental data to better adapt it to the new simulation task. In the process of fine-tuning the proxy model, a variety of techniques are also applied to improve the prediction accuracy and robustness of the model. For example, cross-validation techniques (such as K-fold cross-validation) are used to evaluate the performance of the model to ensure that the model performs consistently on different data sets and complete the online simulation analysis of multiple physical fields using the proxy model.

[0067] Example 5 As another preferred embodiment of the present invention, the present invention includes an online simulation method for the temperature field of a generator stator winding, comprising the following steps: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set. Carry out offline simulation analysis and calculation modeling based on the target unit, integrate electromagnetic and ventilation multi-physics fields and multi-condition boundary conditions to carry out offline finite element calculation and obtain a sample set.

[0068] Take a million-kilowatt generator as an example to calculate the multi-physics field offline sample library. First, the unit needs to be determined, such as the rated power, voltage, current, power factor, stator winding DC resistance, rotor winding DC resistance, rated excitation voltage, and rated excitation current.

[0069] Secondly, determine the calculation inputs, such as stator A, B, and C phase currents, stator voltage, cold air temperature, etc.

[0070] The above input parameters are required parameters, and other input conditions can be added according to actual calculation needs. The total amount of simulation condition data should be no less than 100; the simulation condition should at least cover the stator current change range of 20%~100% and the cold air temperature change of 20℃~50℃.

[0071] Finally, determine the calculation output parameters, such as the axial measuring point temperature of each layer of wire rods.

[0072] Among them, the axial measuring points are numbered 1 from the excitation end to n at the turning end; the value of n is required to be selected in the range of 100~500.

[0073] Step S2. Obtain a historical normal operation data set, correct the sample set in step S1, and obtain a corrected sample set. Based on the offline calculation sample library of step S1, taking a million-kilowatt generator set as an example, a total of 60 sets of actual operating condition simulation samples are completed to form a water-cooled unit sample database, which is used to correct the offline simulation sample set and obtain a corrected sample set.

[0074] The simulation model of step S1 is subjected to order reduction processing to obtain a reduced-order model, and a proxy model is constructed using the reduced-order model.

[0075] Step S3. Train the proxy model using the modified sample set. The proxy model uses a rolling condition optimization algorithm framework to achieve online continuous training of the proxy model. In order to improve the prediction accuracy and robustness of the model, a quasi-application error control strategy is applied during the training process.

[0076] First, the modified sample set is divided into a training set and a test set to ensure the representativeness, diversity, and randomness of the data. A cross-validation technique (such as K-fold cross-validation) is prepared to evaluate the performance of the model. By dividing the data set into K subsets, one of which is used for validation in turn, and the other subsets are used for training, the model's dependence on a specific data set can be effectively reduced, and its generalization ability can be improved.

[0077] Secondly, in terms of proxy model hyperparameter optimization, grid search and random search are used to optimize the model's hyperparameters. By systematically adjusting the parameters, the optimal parameter combination is found to improve the prediction accuracy of the model. Parameter optimization is one of the key steps to improve model performance. The model is trained using the training set under the optimized parameter settings. During the training process, the loss function and performance indicators of the model are monitored to ensure the stability and convergence of the model during the training process. Early stopping technology can also be used to prevent overfitting. After the model training is completed, the trained model is evaluated using the test set, and the prediction error and other performance indicators of the model (such as mean square error, R², etc.) are calculated to verify the prediction ability and robustness of the model. Through these evaluation indicators, the performance of the model in practical applications can be fully understood. The trained proxy model is applied to the winding performance simulation in a multi-physics environment.

[0078] Through rapid simulation, the performance of the winding under different physical field conditions is evaluated to provide reliable prediction results. This step not only verifies the practicality of the model, but also provides an important reference for actual engineering applications. Through this systematic training and evaluation method, the prediction accuracy and robustness of the model will be effectively improved, and its reliability and effectiveness in practical applications will be ensured. In this way, the most suitable proxy model for electromagnetic-thermal coupling characteristics can be selected, laying a solid foundation for subsequent research and application of proxy model correction based on actual operating data.

[0079] Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

[0080] Step S5. Evaluate the unit status and risk warning. Specifically, through the deep integration of sensor status data, mechanism information and expert knowledge, based on the simulation analysis results obtained in step S4, complete the unit status evaluation and risk warning.

[0081] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical scheme and technical concept of the present invention without creative mental labor, which all fall within the scope of protection of the present invention.

Claims

1. A method for online simulation of the temperature field of a generator stator winding, characterized in that: The following steps are involved: Step S1. Establish a simulation model and obtain an offline simulation calculation sample set; Step S2. Obtain a historical normal operation data set, correct the sample set in step S1, and obtain a corrected sample set; reduce the order of the simulation model in step S1 to obtain a reduced-order model; and construct a proxy model using the reduced-order model; Step S3. Training the proxy model using the modified sample set; Step S4. Based on the real-time operating parameters and in combination with the state monitoring data, the trained proxy model is used to complete the multi-physics field online simulation analysis to obtain the simulation analysis results.

2. The method for online simulation of the temperature field of a generator stator winding according to claim 1, characterized in that: The step S1 specifically refers to: establishing a simulation model based on the target unit, integrating electromagnetic, ventilation multi-physical fields and multi-operating condition boundary conditions to carry out offline finite element calculations, and obtaining a sample set.

3. The method for online simulation of the temperature field of a generator stator winding according to claim 2, characterized in that: When performing finite element calculations, the step size used for linearly related data is longer than the step size used for nonlinearly related data.

4. The method for online simulation of the temperature field of a generator stator winding according to claim 1, characterized in that: The method for obtaining the corrected sample set in step S2 is: obtaining a historical normal operating data set of the target unit, adjusting and adapting offline multi-physics field finite element calculation parameters according to the deviation between the historical normal operating data set and the sample set of the same operating condition, and calculating and obtaining the corrected sample set.

5. The method for online simulation of the temperature field of a generator stator winding according to claim 1, characterized in that: The method of order reduction is physical coordinate order reduction method based on Krylov subspace method.

6. The method for online simulation of the temperature field of a generator stator winding according to claim 1, characterized in that: The agent model adopts a rolling condition optimization algorithm framework to achieve online continuous training of the agent model.

7. The method for online simulation of the temperature field of a generator stator winding according to claim 1, characterized in that: The method also includes step S5, evaluating the status of the unit and risk warning.

8. The method for online simulation of the temperature field of a generator stator winding according to claim 7, characterized in that: The step S5 specifically refers to: completing the unit status assessment and risk warning based on the simulation analysis results obtained in step S4 through deep integration of sensor status data, mechanism information and expert knowledge.

9. The method for online simulation of the temperature field of a generator stator winding according to claim 7, characterized in that: When conducting unit status assessment and risk warning, it also includes: using multi-dimensional state parameter cloud map real-time reconstruction technology to complete the three-dimensional visualization of the unit that integrates physical state information.

10. The method for online simulation of the temperature field of a generator stator winding according to any one of claims 1 to 9, characterized in that: It also includes the use of transfer learning technology to apply the trained proxy model to simulation tasks other than the target unit.

Images