Extra-high voltage transformer bushing transient temperature rise rapid prediction method

Abstract

The invention discloses an extra-high voltage transformer bushing transient temperature rise rapid prediction method, which comprises the steps of constructing a simulation model of an extra-high voltage transformer bushing, and simulating and extracting global node temperatures and corresponding working condition parameters of a bushing profile at multiple moments to form a data set; selecting global node temperature to perform orthogonal decomposition, and selecting first r-order modals to construct a modal matrix according to cumulative energy proportion truncation; projecting global node temperature in the data set to a modal space according to a modal matrix to obtain a modal coefficient vector as an output sequence, and performing sliding window on working condition parameters in the data set to form an input sequence; constructing a long-short-term memory network model, optimizing network hyper-parameters of the long-short-term memory network model, and training the model by adopting an input sequence and an output sequence to obtain a time sequence mapping model; and acquiring working condition parameters of the extra-high voltage transformer bushing in a to-be-measured time period, inputting the working condition parameters into the time sequence mapping model, predicting a modal vector at a future moment, and reconstructing transient temperature field distribution of the extra-high voltage transformer bushing in combination with inversion of a modal matrix.

Description

Technical Field

[0001] This invention relates to the field of transformer safety operation technology, specifically to a method for rapid prediction of transient temperature rise in bushings of ultra-high voltage transformers. Background Technology

[0002] UHV transformer bushings are crucial components in UHV power transmission and transformation projects, undertaking the core task of introducing high voltage and large current into the transformer housing. The integrity and thermal stability of their insulation structure directly affect the safe and stable operation of the transformer and even the entire UHV power grid. Therefore, real-time and accurate sensing and reconstruction of the internal temperature field of the bushing under complex operating conditions is a key link in preventing thermal insulation breakdown and realizing predictive maintenance of equipment.

[0003] Currently, the thermal condition assessment of bushings mainly relies on thermal path model calculations recommended by IEC standards, surface infrared thermography, or periodic oil contamination tests. These traditional methods have significant technical bottlenecks: on the one hand, thermal path models are based on steady-state assumptions and empirical formulas, simplifying complex physical field coupling mechanisms and making it difficult to accurately reflect transient temperature rise characteristics under drastic load fluctuations; on the other hand, infrared thermography can only obtain surface temperature information of the equipment and cannot penetrate to its internal core areas (such as conductive rods and capacitor cores), resulting in blind spots in the monitoring of internal "hot spots," which can easily lead to the omission of early overheating hazards.

[0004] In recent years, although numerical simulation methods based on finite element analysis (FEM) have been widely used in the electro-thermal-fluid multiphysics simulation of bushings to improve the accuracy of temperature field calculations, their engineering applications are still severely limited. Firstly, establishing a high-fidelity three-dimensional multiphysics coupled model involves massive degrees of freedom calculations, making the solution process extremely complex and time-consuming (usually requiring hours or even days). This high computational cost limits the method to offline post-hoc analysis, completely failing to meet the timeliness requirements of real-time online monitoring of equipment status in power grid dispatching. Furthermore, while purely data-driven methods (such as traditional neural networks) are computationally fast, they often lack constraints from physical mechanisms. They not only require massive amounts of training data, but also struggle to guarantee generalization ability and prediction reliability when faced with sudden operating conditions outside the training set. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, the rapid prediction method for transient temperature rise of UHV transformer bushings provided by this invention solves the problem that existing methods are unable to accurately reflect the distribution of transient hot spots and the coupling characteristics of multiple physical fields inside the bushing.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0007] A method for rapid prediction of transient temperature rise in bushings of ultra-high voltage transformers is provided, comprising the following steps:

[0008] S1. Construct an electro-thermal-fluid coupled transient simulation model of UHV transformer bushing, perform multi-condition transient simulation, extract global node temperature and corresponding operating condition parameters at multiple discrete moments of the bushing profile throughout the entire process, and form a dataset.

[0009] S2. Select global node temperatures from multiple time points in the dataset for orthogonal decomposition, and truncate the data based on the cumulative energy percentage. Construct the mode matrix for each selected mode and calculate the energy weight vector for each selected mode;

[0010] S3. Based on the modal matrix, project the global node temperature in the dataset to the modal space to obtain the modal coefficient vector as the output sequence, and perform sliding windowing on the operating parameters in the dataset to form the input sequence.

[0011] S4. Construct a long short-term memory network model, using the energy weighted mean square error constructed from the energy weight vector as the objective function, and use a multi-objective heat exchange optimization algorithm to optimize its network hyperparameters. Use the input sequence and output sequence to form a training set to train the model and obtain the time-series mapping model.

[0012] S5. Collect the operating parameters of the UHV transformer bushing during the test period and input them into the time-series mapping model to predict the mode vectors at future moments. Combine the mode matrix inversion to reconstruct the transient temperature field distribution of the UHV transformer bushing.

[0013] Furthermore, methods for optimizing model network hyperparameters using multi-objective heat exchange optimization algorithms include:

[0014] S411. Initialize the population for the multi-objective heat exchange optimization algorithm. , For population size, , and These are individuals 1, 2, and M, respectively;

[0015] S412, Based on the initial temperature Maximum number of iterations and nonlinear cooling factor Construct a virtual temperature function:

[0016]

[0017] in, For the first Virtual temperature at the next iteration; The variable represents the current iteration number, with values ​​ranging from 1, 2, ... ;

[0018] S413, When virtual temperature > Threshold At that time, the hyperbolic tangent thermal expansion operator is used to update the global position of the individual:

[0019]

[0020] in, This represents the individual's current location. The globally optimal position; It is the hyperbolic tangent function; This is a periodic oscillation disturbance; Thermal noise coefficient; for Updated location; A random number within the interval [0,1];

[0021] S414, When virtual temperature ≤threshold At that time, the viscous damped contraction operator is used to update the global position of the individual:

[0022]

[0023] in, To produce a viscous damping effect that increases exponentially with decreasing temperature; The annealing cosine factor; A preset constant is used to prevent the denominator from being zero; It is the viscous damping coefficient;

[0024] S415, Update the location of the individual Substitute the values ​​into the model and pre-train the model using the training set, then calculate the fitness function value. ,like Then adopt Replace the original individual Otherwise, retain the original individual;

[0025] S416. Update the global optimal solution. ,when Less than the maximum number of iterations If the condition is met, return to step S412; otherwise, use the best individual in the current population as the optimal hyperparameter.

[0026] The beneficial effects of the above technical solution are: the optimization strategy explicitly constructs a virtual temperature field that decays nonlinearly with the number of iterations. and through threshold The optimization process of hyperparameters is divided into two physical phases: "thermal expansion" and "viscous damping." Its working mechanism is as follows:

[0027] During the high temperature stage ( The algorithm employs the hyperbolic tangent thermal expansion operator, where... The nonlinear saturation characteristics of the function and The periodic oscillations are coupled together, and a thermal noise term positively correlated with temperature is numerically superimposed. This mechanism simulates the intense thermal motion and expansion effects of particles in a high-energy state in the solution space, giving the population a strong global exploration capability and random jump ability, enabling the algorithm to effectively escape local extremum traps and avoid the model falling into premature convergence.

[0028] Subsequently, in the low-temperature phase ( The algorithm switches to the viscous damped contraction operator, where the denominator is... With exponential function This constitutes a viscous damping mechanism that increases exponentially with decreasing virtual temperature, in conjunction with The periodic decay of the annealing cosine factor enables adaptive fine-tuning of the individual update step size. This dual search mechanism based on thermodynamic phase transition allows the generator to maintain wide-area randomness in the early stage to fully cover the hyperparameter solution space, while automatically and smoothly converging to a high-precision global optimum in the final stage. This significantly improves the generalization prediction accuracy and robustness of the long short-term memory network model when facing complex, nonlinear transient temperature rise conditions.

[0029] Furthermore, step S1 further includes:

[0030] S11. Construct a transient simulation model of the electric-thermal-current coupling of an ultra-high voltage transformer bushing, and perform finite element mesh generation on the simulation model to obtain the results containing... The computational domain of each grid node;

[0031] S12. Set the electro-thermal-fluid multiphysics coupling control equations and material property parameters for the simulation model, and set the time step. Select load current Operating voltage and transformer top oil temperature Used as time-varying operating condition parameters for transient solution;

[0032] S13. Select the longitudinal section where the casing axis is located as the research object and extract the entire process. Cross-sectional global node temperature data at discrete time points and corresponding operating parameters, and Let N be the global node temperature at time 1 and time N;

[0033] The operating parameters include load current. Operating voltage and transformer top oil temperature ; and The load currents at times 1 and N are... and The load currents at times 1 and N are... and The transformer top oil temperature is at time 1 and time N.

[0034] Furthermore, step S2 further includes:

[0035] S21. Select data centers using preset conditions. The temperature data of the profile nodes at each time point, wherein the preset condition is an equal time interval sampling strategy or a non-uniform sampling strategy based on the temperature change rate.

[0036] S22. Flatten the temperature data of the profile nodes at each selected time point into column vectors, and arrange the column vectors of all time points to construct a temperature field snapshot matrix;

[0037] S23. Perform singular value decomposition on the temperature field snapshot matrix to obtain the left singular matrix containing the modes and the diagonal matrix of singular values ​​arranged in descending order.

[0038] S24. Calculate the i-th element in the left singular matrix. Eigenvalues ​​corresponding to the first mode , Let c be the c-th singular value in the singular value diagonal matrix, and calculate the first c. Accumulated energy percentage of first mode :

[0039]

[0040] in, For the front The j-th mode in the first-order mode;

[0041] S25, Adopt The smallest integer is used as the truncation order r, and a is a preset energy proportion less than 1. Before extraction... constructing the mode matrix of the first mode , , and These are the 1st, 2nd, and rth order modes;

[0042] S26. Based on the extracted previous... Calculate the normalized energy weighting factor for each mode by taking the eigenvalues ​​corresponding to the first-order modes. An energy weight vector is constructed using all energy weight factors. The calculation expression is:

[0043]

[0044] in, For the front The first mode in the first mode The eigenvalues ​​corresponding to each mode.

[0045] The beneficial effects of the above technical solution are as follows: This technical solution clearly constructs a physical field dimensionality reduction and feature attention allocation mechanism based on orthogonal decomposition (POD) energy spectrum. Its working mechanism is as follows: First, it utilizes the accumulated energy proportion... As an adaptive truncation criterion, the first r-order modes characterizing heat flow conduction inside the casing are accurately extracted from massive finite element mesh node data, achieving order-of-magnitude compression of data dimensions while ensuring the integrity of physical information.

[0046] Secondly, the scheme introduces a normalized energy weighting factor. In the formula, the molecule This directly maps the energy contribution amplitude of the k-th mode during the spatiotemporal evolution of the transient temperature rise. This weighting factor acts as a physical importance modulation coefficient in the subsequent training of the time-series prediction model. (The denominator is used for...) The normalization process constitutes a non-uniform penalty mechanism for prediction errors of different orders of modes. This allows the neural network to automatically focus on high-energy master modes that represent the core temperature rise trend during training, prioritizing the prediction accuracy of the main temperature distribution, while adaptively reducing the model's sensitivity to low-energy modes containing numerical noise or small thermal perturbations. As a result, the fidelity and noise robustness of the core hotspot region of the temperature field are significantly improved during reconstruction and inversion.

[0047] Furthermore, step S3 further includes:

[0048] S31. Flatten the global node temperatures of all discrete time points in the dataset into column vectors, and project them onto the modal space to obtain the full-time modal coefficient time series. As the output sequence:

[0049] , ,

[0050] in, , and These are the modal coefficient vectors at times 1, i, and N, respectively; Here is the modal matrix; T is the transpose. Let i be the global node temperature at time i; , and Let r be the vector corresponding to modes 1, 2, and r at time i;

[0051] S32. Set the sliding window length to... Based on all the operating condition parameters in the dataset, construct several input sequences:

[0052] ,

[0053] in, Let i be the input sequence at time i. , They are time i and i respectively Operating parameters For time i, the load current, operating voltage, and transformer top oil temperature are given.

[0054] S33, Using the input sequence and its corresponding A training sample is formed, and all training samples are used to form the training set.

[0055] Furthermore, the expression for the objective function is:

[0056]

[0057] in, The objective function is... The total number of training samples in the training set; Let be the vector corresponding to mode k at time h; This is the energy weight vector; For network hyperparameters The network prediction values ​​below, Network weights The L2 norm regularization term, which is the sum of squares of all elements in the network weight matrix, is used to constrain model complexity to prevent overfitting. is the L2 regularization coefficient, and is a hyperparameter used to adjust the model's fitting accuracy and generalization ability weights.

[0058] The beneficial effects of the above technical solution are as follows: the objective function expression explicitly constructs a dual constraint mechanism of physical energy weight guidance and structural risk regularization. Its working mechanism lies in the energy weight vector introduced in the first term of the formula. It acts as a physical characteristic sensitivity modulator, because The value is directly determined by the modal eigenvalues ​​after orthogonal decomposition. This mechanism forces the neural network to automatically focus its optimization on the high-energy-order modes (i.e., the main framework of the sleeve temperature rise) that carry the main heat flow information during backpropagation and gradient descent. This ensures that the model prioritizes the reconstruction accuracy of large-scale temperature rise trends (such as the heating of the conductive rod and the core of the capacitor) under limited training resources.

[0059] Formula 2 This leads to the introduction of L2 norm regularization constraints. This term penalizes the complexity of the network weight matrix, generating a "weight decay" effect during training, which effectively suppresses the model's tendency to overfit to high-frequency numerical noise in finite element simulation data. This design, which couples the "physical fidelity term" with the "mathematical regularization term," enables the final time-series mapping model to have both high fidelity in reproducing the complex thermal field inside the casing and strong generalization ability to cope with unseen operating conditions.

[0060] Furthermore, in step S4, the number of input layer nodes in the long short-term memory network model is the dimension of the operating condition parameters, and the number of output layer nodes corresponds to the modal truncation order. The network hyperparameter vector , This represents the number of neurons in the hidden layer. The initial learning rate, The L2 regularization coefficient is... This is the momentum decay factor for the optimizer.

[0061] Furthermore, the training set is input into the model, and the Adam optimizer is used to compute the objective function based on the backpropagation algorithm. Regarding the gradient of network weights, the network weight parameters are iteratively updated until the model converges, thus obtaining a time-series mapping model from the defined operating parameters to the modal coefficients.

[0062] Furthermore, step S5 further includes:

[0063] S51. Online data acquisition or setting the time to be measured. The operating parameters of ultra-high voltage transformer bushings are constructed, including current and past data. Input sequence of working condition characteristics at each time point , and They are respectively time and time Operating parameters;

[0064] S52, will Input a time-series mapping model, output the predicted mode coefficient vector at the time to be measured. , These are the vectors corresponding to the 1st, 2nd, and rth modes;

[0065] S53, Based on the modal matrix Global temperature field reconstruction is performed based on the inverse transformation principle of orthogonal decomposition:

[0066]

[0067] in, For the domain node temperature prediction vector;

[0068] S54. Based on steps S21 and S22, the predicted vector... The reverse reconstruction into a two-dimensional spatial temperature matrix, thanks to the linear superposition property of the reduced-order mode matrix, involves only basic matrix operations, achieving millisecond-level high-precision inversion of the global temperature field.

[0069] The beneficial effects of the above technical solution are as follows: This step constructs a fast inverse reconstruction mechanism based on linear mode superposition. Its working mechanism lies in: Formula The modality matrix was fully utilized. The orthogonal basis property. At the computational level, this relation transforms the extremely time-consuming iterative solution process of nonlinear partial differential equations (PDEs) in traditional finite element simulations into efficient matrix-vector multiplication operations.

[0070] At the physical level, this means that a complex transient temperature field at any given moment can be analytically converted into a linear weighted combination of several intrinsic modes. This mechanism avoids the need for individual thermal equilibrium calculations on tens of thousands of finite element mesh nodes, allowing the system to complete the inversion from low-dimensional modal coefficients to the global temperature field within a millisecond-level time window with minimal computational resources. This not only ensures that the reconstructed results have the same physical spatial resolution as offline simulations, but also enables real-time, panoramic visualization monitoring of transient hotspot distribution in the core areas inside the casing (such as conductive rods and capacitor cores).

[0071] Furthermore, the rapid prediction method for transient temperature rise of UHV transformer bushings also includes the use of... The coordinates and values ​​of the hottest spot inside the locking sleeve, when the maximum temperature... An alert is issued when the preset safety threshold is exceeded.

[0072] Compared with existing technologies, the rapid prediction method for transient temperature rise of bushings provided by this invention has the following advantages:

[0073] 1. A balance between computational efficiency and physical accuracy is achieved, with millisecond-level response capability. This invention employs an asymmetric computation strategy of "offline physical modeling + online ultra-fast inference." In the offline stage, a high-fidelity electro-thermal-fluid coupling simulation model is constructed and orthogonal decomposition (POD) is performed to pre-extract the reduced-order mode matrix of the bushing. In the online application stage, this mechanism transforms the time-consuming and lengthy iterative solution process of nonlinear partial differential equations (PDEs) in traditional numerical simulation into a computationally low-cost matrix linear weighted superposition operation. This enables the system to complete the inversion from operating parameters to the temperature field at a millisecond speed, preserving the physical spatial resolution of finite element simulation while meeting the timeliness requirements of power grid dispatch for real-time monitoring of equipment status.

[0074] 2. By integrating physical energy characteristics, the generalization ability and robustness of time-series prediction models are significantly improved. This invention innovatively introduces a normalized energy weight factor based on eigenvalues ​​when constructing the objective function of the Long Short-Term Memory (LSTM) network. This factor acts as a physical feature sensitivity regulator, forcing the neural network to automatically focus on high-energy modes (such as the core heating modes of conductive rods and capacitor cores) that carry the main heat flow information during training, prioritizing the prediction accuracy of the overall temperature rise trend. Simultaneously, this mechanism adaptively weakens the model's sensitivity to numerical noise or small perturbations contained in low-energy modes through energy spectrum analysis, effectively solving the problem of overfitting in purely data-driven models and significantly improving the model's noise resistance and prediction robustness when facing unseen transient fluctuation conditions.

[0075] 3. A Multi-Level Thermal Equilibrium Optimization Algorithm (MSTEO) is proposed to solve the problem of hyperparameter optimization in deep learning models. Addressing the difficulty in determining hyperparameters for LSTM networks, this invention proposes an optimization algorithm that simulates a thermodynamic phase transition mechanism. This algorithm constructs a virtual temperature field that decays nonlinearly with the number of iterations, dividing the optimization process into two stages: "thermal expansion" and "viscous damping." In the high-temperature stage, the hyperbolic tangent thermal expansion operator is used to endow the population with strong global exploration capabilities, effectively escaping local extremum traps. In the low-temperature stage, the algorithm switches to a viscous damping contraction operator to achieve adaptive and precise approximation of the optimal solution. This dual search mechanism avoids the subjectivity and blindness of manual empirical parameter tuning, ensuring that the final generated time-series mapping model is in a globally optimal parameter configuration state.

[0076] 4. Eliminates internal monitoring blind spots, achieving visualization and precise positioning of the entire temperature field. Unlike infrared thermometry, which only acquires the surface temperature of equipment, this invention, based on the principle of modal inverse reconstruction, can reconstruct the entire temperature field containing all grid nodes using a small number of predicted modal coefficients. This allows the monitoring system to "penetrate" the insulating jacket, directly observing the real-time temperature distribution of core components such as conductive rods, capacitor cores, and oil-paper insulation inside the bushing. Furthermore, by automatically locking onto the highest temperature point, it can accurately identify internal overheating hazards that are undetectable by traditional methods in the early stages of insulation breakdown or thermal runaway, thus achieving true panoramic proactive early warning. Attached Figure Description

[0077] Figure 1 This is a flowchart of a method for rapid prediction of transient temperature rise in bushings of ultra-high voltage transformers.

[0078] Figure 2 This is a diagram of a transient simulation model of the electro-thermal-current coupling of an ultra-high voltage transformer bushing. Detailed Implementation

[0079] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0080] refer to Figure 1 , Figure 1 A flowchart of a rapid prediction method for transient temperature rise in bushings of ultra-high voltage transformers is shown; as follows: Figure 1 As shown, the method S includes steps S1 to S5.

[0081] In step S1, an electro-thermal-fluid coupled transient simulation model of the UHV transformer bushing is constructed, and multi-condition transient simulation is performed. The global node temperature and corresponding operating condition parameters at multiple discrete moments of the bushing profile throughout the entire process are extracted to form a dataset.

[0082] In one embodiment of the present invention, step S1 further includes:

[0083] S11. Construct a transient simulation model of the electric-thermal-current coupling of an ultra-high voltage transformer bushing, and perform finite element mesh generation on the simulation model to obtain the results containing... The computational domain consists of individual grid nodes; the simulation model covers components such as conductive rods, capacitor cores, insulating oil, insulating jackets, and flanges.

[0084] To facilitate understanding of the simulation model in this scheme, this scheme combines... Figure 2Detailed explanation:

[0085] like Figure 2 As shown, a three-dimensional finite element simulation model of the valve-side bushing of an ultra-high voltage converter transformer was constructed and thermal field analysis was performed. This simulation model was established based on the actual structure. The total length of the simulation model is 2000 mm, the central guide rod is 1242 mm long with a radius of 120 mm, the bushing has 20 layers of plates, the zero-layer plate is 8000 mm long with a conductor rod diameter of 146 mm, and the N-layer plates are 4400 mm long with a diameter of 380 mm. Simultaneously, the physical environment of the bushing operating across the medium was realistically reproduced: one end of the model was placed in the transformer oil tank, and the other end was placed in a closed pipe area filled with SF6 gas and containing basin-type insulators, to fully simulate the insulation environment within the transformer tank and GIS pipeline.

[0086] The simulation was conducted at an ambient temperature of 30°C and a transformer top oil temperature of 60°C. A rated load current of 4500A and an operating voltage of 800kV were applied to the central conductor as excitation. The thermal field distribution was analyzed using multiphysics coupling calculations. The simulation results show that the maximum steady-state temperature of the bushing is 119°C, mainly concentrated at the central conductor. The Joule heating of the bushing insulation is also concentrated near the central conductor. The heat energy at the central conductor is transferred to the connected medium through thermal radiation, conduction, and convection, and the temperature conduction in the solid medium exhibits a gradient.

[0087] S12. Set the electro-thermal-fluid multiphysics coupling control equations and material property parameters for the simulation model. The control equations include:

[0088] Electric field governing equations: , Represents the electric displacement vector. Indicates charge density; For Hamiltonian operators;

[0089] Magnetic field control equations: , Indicates magnetic field strength. Indicates the total current density;

[0090] Fluid field governing equations: , , Indicates the mass density of the fluid medium. Represents the fluid velocity vector. Indicates pressure, Indicates the dynamic viscosity coefficient of a fluid. Represents volume force;

[0091] Temperature field governing equations: , Indicates the density of the material. Indicates specific heat capacity. Indicates temperature. Indicates thermal conductivity, This indicates the heat source term.

[0092] During simulation, the time step is set. Select load current Operating voltage and transformer top oil temperature Used as time-varying operating condition parameters for transient solution.

[0093] S13. Select the longitudinal section containing the casing axis as the research object, and discretize this section into... The regular sampling grid is used to extract the entire process. Cross-sectional global node temperature data at discrete time points and corresponding operating parameters, and Let N be the global node temperature at time 1 and time N;

[0094] The operating parameters include load current. Operating voltage and transformer top oil temperature ; and The load currents at times 1 and N are... and The load currents at times 1 and N are... and The transformer top oil temperature is at time 1 and time N.

[0095] In step S2, the global node temperatures at multiple times in the dataset are selected for orthogonal decomposition, and the data is truncated according to the cumulative energy percentage. The modal matrix is ​​constructed by selecting each mode, and the energy weight vector of each selected mode is calculated. The detailed implementation steps are as follows:

[0096] S21. Select data centers using preset conditions. The temperature data of the profile nodes at each time point are selected, and the preset conditions are either an equal time interval sampling strategy or a non-uniform sampling strategy based on the temperature change rate, so as to ensure that the selected time points cover the transient drastic change stage, steady-state transition stage and dynamic equilibrium stage of the sleeve temperature rise.

[0097] S22. Flatten the temperature data of each selected profile node into a column vector, and arrange the column vectors of all times to construct a temperature field snapshot matrix. ,in and This is the global node temperature column vector after being flattened at times 1, 2, and m.

[0098] S23. Perform singular value decomposition on the temperature field snapshot matrix. This yields the left singular matrix containing the modes. and singular value diagonal matrices arranged in descending order ; They are the 1st, 2nd and 3rd respectively. First mode, These are the 1st, 2nd, and mth singular values, respectively. It's a diagonal matrix.

[0099] S24. Calculate the i-th element in the left singular matrix. Eigenvalues ​​corresponding to the first mode , Let c be the c-th singular value in the singular value diagonal matrix, and calculate the first c. Accumulated energy percentage of first mode :

[0100]

[0101] in, For the front The j-th mode in the first-order mode;

[0102] S25, Adopt The smallest integer is used as the truncation order r, and a is a preset energy proportion less than 1, preferably 99.5%, before extraction. constructing the mode matrix of the first mode , , and These are the 1st, 2nd, and rth order modes;

[0103] S26. Based on the extracted previous... Calculate the normalized energy weighting factor for each mode by taking the eigenvalues ​​corresponding to the first-order modes. An energy weight vector is constructed using all energy weight factors. The calculation expression is:

[0104]

[0105] in, For the front The first mode in the first mode The eigenvalues ​​corresponding to each mode.

[0106] In step S3, the global node temperatures in the dataset are projected onto the modal space based on the modal matrix to obtain the modal coefficient vector as the output sequence, and the operating parameters in the dataset are windowed to form the input sequence; step S3 further includes:

[0107] S31. Flatten the global node temperatures of all discrete time points in the dataset into column vectors, and project them onto the modal space to obtain the full-time modal coefficient time series. As the output sequence:

[0108] , ,

[0109] in, , and These are the modal coefficient vectors at times 1, i, and N, respectively; Here is the modal matrix; T is the transpose. Let i be the global node temperature at time i; , and Let r be the vector corresponding to modes 1, 2, and r at time i;

[0110] S32. Set the sliding window length to... Based on all the operating condition parameters in the dataset, construct several input sequences:

[0111] ,

[0112] in, Let i be the input sequence at time i. , They are time i and i respectively Operating parameters For time i, the load current, operating voltage, and transformer top oil temperature are given.

[0113] S33, Using the input sequence and its corresponding A training sample is formed, and all training samples are used to form the training set.

[0114] In step S4, a long short-term memory network model is constructed. The energy-weighted mean square error constructed from the energy weight vector is used as the objective function. The network hyperparameters are optimized using a multi-objective heat exchange optimization algorithm. The model is trained using the input sequence and the output sequence as the training set to obtain the time-series mapping model.

[0115] The number of input layer nodes in the long short-term memory network model built in step S4 corresponds to the dimension of the operating parameters, and the number of output layer nodes corresponds to the modal truncation order. The network hyperparameter vector , This represents the number of neurons in the hidden layer. The initial learning rate, The L2 regularization coefficient is... This is the momentum decay factor for the optimizer.

[0116] In one embodiment of the present invention, the method for optimizing the hyperparameters of a model network using a multi-objective heat exchange optimization algorithm includes:

[0117] S411. Initialize the population for the multi-objective heat exchange optimization algorithm. , For population size, , and These are individuals 1, 2, and M, respectively;

[0118] S412, Based on the initial temperature Maximum number of iterations and nonlinear cooling factor Construct a virtual temperature function:

[0119]

[0120] in, For the first Virtual temperature at the next iteration; The variable represents the current iteration number, with values ​​ranging from 1, 2, ... ;

[0121] S413, When virtual temperature > Threshold At that time, the hyperbolic tangent thermal expansion operator is used to update the global position of the individual:

[0122]

[0123] in, This represents the individual's current location. The globally optimal position; It is the hyperbolic tangent function; This is a periodic oscillation disturbance; Thermal noise coefficient; for Updated location; A random number within the interval [0,1];

[0124] S414, When virtual temperature ≤threshold At that time, the viscous damped contraction operator is used to update the global position of the individual:

[0125]

[0126] in, To produce a viscous damping effect that increases exponentially with decreasing temperature; The annealing cosine factor; A preset constant is used to prevent the denominator from being zero; It is the viscous damping coefficient;

[0127] S415, Update the location of the individual Substitute the values ​​into the model and pre-train the model using the training set, then calculate the fitness function value. ,like Then adopt Replace the original individual Otherwise, retain the original individual;

[0128] S416. Update the global optimal solution. ,when Less than the maximum number of iterations If the condition is met, return to step S412; otherwise, use the best individual in the current population as the optimal hyperparameter.

[0129] The objective function of the Long Short-Term Memory (LSTM) network model is expressed as follows:

[0130]

[0131] in, The objective function is... The total number of training samples in the training set; Let be the vector corresponding to mode k at time h; This is the energy weight vector; For network hyperparameters The network prediction values ​​below, Network weights The L2 norm regularization term, which is the sum of squares of all elements in the network weight matrix, is used to constrain model complexity to prevent overfitting. is the L2 regularization coefficient, and is a hyperparameter used to adjust the model's fitting accuracy and generalization ability weights.

[0132] The training set is input into the model, and the Adam optimizer is used to calculate the objective function based on the backpropagation algorithm. Regarding the gradient of network weights, the network weight parameters are iteratively updated until the model converges, thus obtaining a time-series mapping model from the defined operating parameters to the modal coefficients.

[0133] In step S5, the operating parameters of the UHV transformer bushing during the test period are collected and input into the time-series mapping model to predict the mode vectors at future times. The transient temperature field distribution of the UHV transformer bushing is then reconstructed by combining the mode matrix inversion.

[0134] In one embodiment of the present invention, step S5 further includes:

[0135] S51. Online data acquisition or setting the time to be measured. The operating parameters of ultra-high voltage transformer bushings are constructed, including current and past data. Input sequence of working condition characteristics at each time point , and They are respectively time and time Operating parameters;

[0136] S52, will Input a time-series mapping model, output the predicted mode coefficient vector at the time to be measured. , These are the vectors corresponding to the 1st, 2nd, and rth modes;

[0137] S53, Based on the modal matrix Global temperature field reconstruction is performed based on the inverse transformation principle of orthogonal decomposition:

[0138]

[0139] in, For the domain node temperature prediction vector;

[0140] S54. Based on steps S21 and S22, the predicted vector... The reverse reconstruction into a two-dimensional spatial temperature matrix, thanks to the linear superposition property of the reduced-order mode matrix, involves only basic matrix operations, achieving millisecond-level high-precision inversion of the global temperature field.

[0141] In this scheme, the following is adopted: The system automatically locks the coordinates and values ​​of the hottest spot inside the bushing (including the conductive rod and capacitor core), thereby accurately identifying potential internal overheating hazards that are difficult to detect with traditional surface monitoring in the early stages of a fault. An early warning is issued immediately when the preset safety threshold is exceeded, effectively preventing serious accidents such as insulation breakdown and explosion.

[0142] In summary, the prediction method provided in this solution solves the problems mentioned in the background technology, such as the inability of traditional thermal circuit models and surface temperature measurement technology to accurately reflect the transient hot spot distribution and multi-physics coupling characteristics inside the bushing, and the excessive computation time of existing numerical simulation methods, which cannot meet the needs of online monitoring.

Claims

1. A method for rapid prediction of transient temperature rise in bushings of ultra-high voltage transformers, characterized in that, Including the following steps: S1. Construct an electro-thermal-fluid coupled transient simulation model of UHV transformer bushing, perform multi-condition transient simulation, extract global node temperature and corresponding operating condition parameters at multiple discrete moments of the bushing profile throughout the entire process, and form a dataset. S2. Select global node temperatures from multiple time points in the dataset for orthogonal decomposition, and truncate the data based on the cumulative energy percentage. Construct the mode matrix for each selected mode and calculate the energy weight vector for each selected mode; S3. Based on the modal matrix, project the global node temperature in the dataset to the modal space to obtain the modal coefficient vector as the output sequence, and perform sliding windowing on the operating parameters in the dataset to form the input sequence. S4. Construct a long short-term memory network model, using the energy weighted mean square error constructed from the energy weight vector as the objective function, and use a multi-objective heat exchange optimization algorithm to optimize its network hyperparameters. Use the input sequence and output sequence to form a training set to train the model and obtain the time-series mapping model. S5. Collect the operating parameters of the UHV transformer bushing during the test period and input them into the time-series mapping model to predict the mode vectors at future moments. Combine the mode matrix inversion to reconstruct the transient temperature field distribution of the UHV transformer bushing.

2. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 1, characterized in that, Methods for optimizing model network hyperparameters using multi-objective heat exchange optimization algorithms include: S411. Initialize the population for the multi-objective heat exchange optimization algorithm. , For population size, , and These are individuals 1, 2, and M, respectively; S412, Based on the initial temperature Maximum number of iterations and nonlinear cooling factor Construct a virtual temperature function: in, For the first Virtual temperature at the next iteration; The variable represents the current iteration number, with values ​​ranging from 1, 2, ... ; S413, When virtual temperature > Threshold At that time, the hyperbolic tangent thermal expansion operator is used to update the global position of the individual: in, This represents the individual's current location. The globally optimal position; It is the hyperbolic tangent function; This is a periodic oscillation disturbance; Thermal noise coefficient; for Updated location; A random number within the interval [0,1]; S414, When virtual temperature ≤threshold At that time, the viscous damped contraction operator is used to update the global position of the individual: in, To produce a viscous damping effect that increases exponentially with decreasing temperature; The annealing cosine factor; A preset constant is used to prevent the denominator from being zero; It is the viscous damping coefficient; S415, Update the location of the individual Substitute the values ​​into the model and pre-train the model using the training set, then calculate the fitness function value. ,like Then adopt Replace the original individual Otherwise, retain the original individual; S416. Update the global optimal solution. ,when Less than the maximum number of iterations If the condition is met, return to step S412; otherwise, use the best individual in the current population as the optimal hyperparameter.

3. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 1, characterized in that, Step S1 further includes: S11. Construct a transient simulation model of the electric-thermal-current coupling of an ultra-high voltage transformer bushing, and perform finite element mesh generation on the simulation model to obtain the results containing... The computational domain of each grid node; S12. Set the electro-thermal-fluid multiphysics coupling control equations and material property parameters for the simulation model, and set the time step. Select load current Operating voltage and transformer top oil temperature Used as time-varying operating condition parameters for transient solution; S13. Select the longitudinal section where the casing axis is located as the research object and extract the entire process. Cross-sectional global node temperature data at discrete time points and corresponding operating parameters, and Let N be the global node temperature at time 1 and time N; The operating parameters include load current. Operating voltage and transformer top oil temperature ; and The load currents at times 1 and N are... and The load currents at times 1 and N are... and The transformer top oil temperature is at time 1 and time N.

4. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 1, characterized in that, Step S2 further includes: S21. Select data centers using preset conditions. The temperature data of the profile nodes at each time point, wherein the preset condition is an equal time interval sampling strategy or a non-uniform sampling strategy based on the temperature change rate. S22. Flatten the temperature data of the profile nodes at each selected time point into column vectors, and arrange the column vectors of all time points to construct a temperature field snapshot matrix; S23. Perform singular value decomposition on the temperature field snapshot matrix to obtain the left singular matrix containing the modes and the diagonal matrix of singular values ​​arranged in descending order. S24. Calculate the i-th element in the left singular matrix. Eigenvalues ​​corresponding to the first mode , Let c be the c-th singular value in the singular value diagonal matrix, and calculate the first c. Accumulated energy percentage of first mode : in, For the front The j-th mode in the first-order mode; S25, Adopt The smallest integer is used as the truncation order r, and a is a preset energy proportion less than 1. Before extraction... constructing the mode matrix of the first mode , , and For the 1st, 2nd, and rth order modes; take The smallest integer b is used as the truncation order r, and a is a preset energy ratio less than 1; S26. Based on the extracted previous... Calculate the normalized energy weighting factor for each mode by taking the eigenvalues ​​corresponding to the first-order modes. An energy weight vector is constructed using all energy weight factors. The calculation expression is: in, For the front The first mode in the first mode The eigenvalues ​​corresponding to each mode.

5. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 1, characterized in that, Step S3 further includes: S31. Flatten the global node temperatures of all discrete time points in the dataset into column vectors, and project them onto the modal space to obtain the full-time modal coefficient time series. As the output sequence: ， ， in, , and These are the modal coefficient vectors at times 1, i, and N, respectively; Here is the modal matrix; T is the transpose. Let i be the global node temperature at time i; , and Let r be the vector corresponding to modes 1, 2, and r at time i; S32. Set the sliding window length to... Based on all the operating condition parameters in the dataset, construct several input sequences: ， in, Let i be the input sequence at time i. , They are time i and i respectively Operating parameters For time i, the load current, operating voltage, and transformer top oil temperature are given. S33, Using the input sequence and its corresponding A training sample is formed, and all training samples are used to form the training set.

6. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 2 or 4, characterized in that, The expression for the objective function is: in, The objective function is... The total number of training samples in the training set; Let be the vector corresponding to mode k at time h; This is the energy weight vector; For network hyperparameters The network prediction values ​​below, Network weights L2 norm regularization term; is the L2 regularization coefficient.

7. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 4, characterized in that, The number of input layer nodes in the long short-term memory network model built in step S4 corresponds to the dimension of the operating parameters, and the number of output layer nodes corresponds to the modal truncation order. The network hyperparameter vector , This represents the number of neurons in the hidden layer. The initial learning rate, The L2 regularization coefficient is... This is the momentum decay factor for the optimizer.

8. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 4, characterized in that, The training set is input into the model, and the Adam optimizer is used to calculate the objective function based on the backpropagation algorithm. Regarding the gradient of network weights, the network weight parameters are iteratively updated until the model converges, thus obtaining a time-series mapping model from the defined operating parameters to the modal coefficients.

9. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 4, characterized in that, Step S5 further includes: S51. Online data collection or setting the time to be measured. The operating parameters of ultra-high voltage transformer bushings are constructed, including current and past data. Input sequence of working condition characteristics at each time point , and They are respectively time and time Operating parameters; S52, will Input a time-series mapping model, output the predicted mode coefficient vector at the time to be measured. , These are the vectors corresponding to the 1st, 2nd, and rth modes; S53, Based on the modal matrix Global temperature field reconstruction is performed based on the inverse transformation principle of orthogonal decomposition: in, For the domain node temperature prediction vector; S54. Based on steps S21 and S22, the predicted vector... It is reverse-engineered into a two-dimensional spatial temperature matrix.

10. The method for rapid prediction of transient temperature rise in UHV transformer bushings according to claim 9, characterized in that, It also includes the use of The coordinates and values ​​of the hottest spot inside the locking sleeve, when the maximum temperature... An alert is issued when the preset safety threshold is exceeded.

Images