Oil-immersed transformer temperature field real-time calculation method

Abstract

The application discloses a kind of oil-immersed transformer temperature field real-time calculation method, it includes in the online operation stage of oil-immersed transformer, real-time acquisition current time's external dynamic working condition parameter constitutes real-time working condition parameter vector and the real temperature of all monitoring points in transformer inside constitutes real-time measurement temperature vector;According to the modal submatrix of all monitoring points in transformer inside and real-time measurement temperature vector, the inversion channel modal coefficient of monitoring point is calculated;Real-time working condition parameter vector is input into the modal coefficient prediction model that has been trained, and the working condition prediction channel modal coefficient of monitoring point is output;Dynamic fusion weight is generated using dynamic gating network, and inversion channel modal coefficient and working condition prediction channel modal coefficient are fused using dynamic fusion weight, to obtain fusion modal coefficient;According to the reduced modal matrix of oil-immersed transformer and fusion modal coefficient, the current time inside the transformer global three-dimensional temperature field reconstruction vector is calculated.

Description

Technical Field

[0001] This invention relates to the field of transformer fault monitoring technology, specifically to a method for real-time calculation of the temperature field of an oil-immersed transformer. Background Technology

[0002] Oil-immersed transformers are core hub equipment in ultra-high voltage and smart grids at all levels. The thermal aging rate and thermal breakdown risk of their internal insulation materials directly depend on the operating temperature. Therefore, real-time and accurate monitoring of the three-dimensional transient temperature field distribution inside the transformer, especially the precise capture and location of the hot spot temperature, is of extremely important engineering value for ensuring the safe and stable operation of the equipment, guiding dynamic capacity expansion of the transformer, and implementing predictive condition-based maintenance.

[0003] Currently, existing technologies for obtaining the internal temperature field of transformers mainly face the following technical bottlenecks: First, physical monitoring has blind spots. The most direct monitoring method is to embed temperature sensors such as fiber optic gratings in the transformer windings or oil channels for actual measurement. However, the internal structure of a transformer is extremely compact. Due to stringent high-voltage insulation distances, mechanical structural strength constraints, and space limitations in the insulating oil flow channels, the physical locations and number of sensors that can be effectively installed are extremely limited. This sparse monitoring method, which "uses points to represent the surface," cannot cover the complex thermal flow field inside the transformer, easily missing key local heating areas and creating significant monitoring blind spots, failing to present a panoramic temperature distribution. Second, traditional numerical simulation cannot meet real-time requirements. While numerical simulation using computational fluid dynamics (CFD) and the finite element method (FEM) to construct a transformer electro-thermal-fluid multiphysics coupled model can obtain a high-fidelity global temperature field, its calculation involves a nonlinear iterative process with a massive number of grid nodes. Simulation of a single transient condition often takes hours or even days, which cannot meet the millisecond-level real-time requirements of power grid dispatching and maintenance for online equipment status assessment, and can only be used for post-event offline analysis.

[0004] To address the slow computation speed of finite element method (FE) calculations, reduced-order model (ROM) techniques based on intrinsic orthogonal decomposition (POD) have been introduced in recent years. In practical applications, conventional POD-based reduced-order reconstruction has evolved into two independent technical routes: one is "POD + surrogate model," which achieves rapid prediction by establishing a mapping relationship between external operating conditions and modal coefficients; the other is "POD + finite measurement points," which directly infers modal coefficients using physical data from a small number of real-time monitoring points.

[0005] However, both of these single order reduction approaches have significant limitations in complex engineering applications: On the one hand, the simple data-driven surrogate model is highly dependent on the distribution of historical samples. When faced with sudden extreme working conditions outside the training set, early latent defects, or system aging drift, it is prone to serious generalization errors and lacks calibration based on the real physical state. On the other hand, although relying solely on a limited number of measurement points to inversely deduce modal coefficients provides a physical benchmark, its inversion accuracy is easily affected by sensor measurement noise, local working condition disturbances, and the location of measurement points, resulting in poor system robustness.

[0006] In the complex transient operating conditions of transformers, the two methods mentioned above often have their own advantages in different scenarios: for example, when the operating conditions are stable but the sensors are affected by electromagnetic noise, the surrogate model prediction is often more accurate; while when a sudden anomaly occurs or external operating parameters fail, the physical inversion of the measurement point can better reflect the true state.

[0007] Therefore, there is an urgent need for a real-time temperature field calculation method that can overcome the bottleneck of limited space measurement point layout and integrate the advantages of "surrogate model prediction" and "physical measurement point inversion" with dynamic credibility based on the reliability of the system's real-time operating status, so as to achieve a panoramic digital twin of transformers that takes into account high fidelity, high real-time performance and strong noise resistance. Summary of the Invention

[0008] To address the aforementioned shortcomings in existing technologies, this invention provides a real-time calculation method for the temperature field of an oil-immersed transformer. This method solves the problems of blind spots in internal temperature monitoring caused by the limited arrangement of traditional physical measuring points, and the time-consuming calculation of numerical simulation methods, which cannot meet the needs of online real-time evaluation.

[0009] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for real-time calculation of the temperature field of an oil-immersed transformer is provided, which includes the following steps: S1. During the online operation of the oil-immersed transformer, the external dynamic operating condition parameters at the current moment are collected in real time to form a real-time operating condition parameter vector, and the real temperature of all monitoring points inside the transformer is collected to form a real-time measured temperature vector. S2. Calculate the inversion channel mode coefficients of the monitoring points based on the modal submatrix of all monitoring points inside the transformer and the real-time measured temperature vector; S3. Input the real-time operating condition parameter vector into the trained modal coefficient prediction model and output the operating condition prediction channel modal coefficients of the monitoring points. S4. Dynamic fusion weights are generated using a dynamic gating network, and the dynamic fusion weights are used to fuse the modal coefficients of the inversion channel and the modal coefficients of the working condition prediction channel to obtain the fused modal coefficients. S5. Based on the reduced-order mode matrix of the oil-immersed transformer and fusion mode coefficients Calculate the reconstructed vector of the global three-dimensional temperature field inside the transformer at the current time t. .

[0010] In one possible implementation, the method for obtaining the reduced-order mode matrix of the oil-immersed transformer includes: A1. Construct an electro-thermal-current coupled transient simulation model for an oil-immersed transformer and perform dynamic multi-condition transient simulation to generate a dataset containing full-time-domain transient global temperature field data and dynamic operating condition parameters; A2. Select typical temperature field data from the dataset, construct a temperature field snapshot matrix and perform intrinsic orthogonal decomposition to obtain multiple temperature field spatial modes and their corresponding energy eigenvalues; A3. Select the previous step based on the cumulative energy percentage threshold. First-order modes, forming a reduced-order mode matrix The global temperature field at each time point is projected onto the reduced-order mode matrix to obtain the modal coefficients of the global temperature field at the corresponding time point.

[0011] In one possible implementation, the method for determining the arrangement positions of all monitoring points inside the transformer and the modal sub-matrix includes: S11. Select grid nodes that meet the preset conditions from all grid nodes in the coupled transient simulation model as optional measurement points, and construct a set of optional measurement point row numbers; S12. Based on the number of monitoring points to be arranged in the oil-immersed transformer, randomly select monitoring points from the set of available monitoring point row numbers to generate a monitoring point arrangement scheme; S13. For the measurement point layout scheme, select the reduced-order modes of the corresponding measurement points from the reduced-order mode matrix to form a mode submatrix; S14. Using modal submatrix, spatial distribution constraints of measurement points, and correlation of measurement points as evaluation indicators, construct an objective function for optimizing the layout of measurement points. S15. Based on the measurement point layout scheme, the measurement points not selected in the set of optional measurement point row numbers, and the measurement point layout optimization objective function, the measurement point layout scheme is iteratively optimized using the simulated annealing mechanism to obtain the optimal monitoring point layout scheme. S16. Based on the optimal monitoring point layout scheme, select the reduced-order modes of the corresponding monitoring points from the reduced-order mode matrix to form the optimal mode submatrix. The optimal monitoring point layout scheme and the optimal modal submatrix are adopted. This serves as the arrangement location and modal submatrix of all monitoring points inside the transformer.

[0012] The beneficial effects of the above technical solution are as follows: This solution clearly integrates the limiting conditions of sensor installation in engineering practice with the mathematical process of modal submatrix optimization. Specifically, it constructs a set of optional measurement point row numbers through physical masks such as insulation distance and structural avoidance, eliminates grid nodes that are not feasible for installation, and delineates the feasible domain boundary for safe sensor installation in engineering practice. Then, it iteratively optimizes the measurement point layout scheme by combining simulated annealing mechanism, which can quickly and accurately find the optimal measurement point position for installing temperature sensors.

[0013] In one possible implementation, the objective function for optimizing the measurement point layout is... The expression is:

[0014] ,

[0015]

[0016] in, The condition number of the modal submatrix; This refers to the spatial distribution constraint term of the measurement points; For the correlation term of the measurement points; for , and Weighting coefficients; Measurement point layout scheme The corresponding modal submatrix; and They are respectively The maximum and minimum singular values; It is a very small positive number; for The r-th order reduced mode; , and They are respectively The r-th reduced-order mode at the 1st, 2nd, and mth measurement points; This represents the distance between the spatial coordinates of two measuring points a and b. Let these be the row numbers for measurement points a and b. and These are the spatial coordinates of measurement points a and b, respectively. It is a quadratic norm; The preset minimum installation distance between measuring points; and These are the reduced-order modes corresponding to measurement points a and b, respectively. This is a transpose.

[0017] The beneficial effects of the above technical solution are as follows: This solution first calculates the condition number based on the maximum and minimum singular values ​​of the submatrix. This term numerically represents the error amplification factor of local measurement noise during the inversion solution of overdetermined equations; simultaneously and This quantifies both the physical congestion and the linear overlap of the modal responses. Therefore, the objective function, weighted by the aforementioned multiple indices, is... This constitutes a comprehensive constraint mechanism for high-fidelity screening and disturbance resistance of sparse measurement point networks. The resulting measurement point layout scheme not only strictly meets the stringent mechanical and insulation installation specifications of oil-immersed transformers, effectively avoiding infeasible solutions, but also guides measurement points to uniformly cover the most representative high-order mode sensitive areas of the system. Thus, while reasonably controlling the number and cost of sensors, it significantly reduces the sensitivity of the mode inversion process to local measurement noise.

[0018] In one possible implementation, step S15 further includes: S151. Determine the current measurement point layout scheme. Equal to the current optimal measurement point layout scheme And set the initial annealing temperature. Cooling coefficient Maximum number of iterations and the number of neighborhood searches at each annealing temperature ; S152, in the In the next annealing iteration, from the current measurement point layout scheme Randomly select a row number The set of measurable node row numbers is composed of the unselected measurable point row numbers from the set of optional measurable point row numbers. Select row number A new measurement point layout scheme was formed. ; S153. Calculate the change in the objective function. , and They are respectively and The calculated objective function value; S154, when season And proceed to step S155; when At that time, generate random numbers ,like Then let Proceed to step S155; otherwise, remain in the current state. The process remains unchanged, and proceeds to step S156. For the first Annealing temperature during the next annealing iteration; S155, Judgment Is it less than If so, then update. If the condition is met, proceed to step S156; otherwise, proceed directly to step S156. S156. Determine whether the annealing process has been completed at the current annealing temperature. If the next neighborhood search is successful, proceed to step S157; otherwise, return to step S152. S157. Determine whether the current iteration count has reached the maximum iteration count, whether the annealing temperature is lower than the preset termination temperature, or whether the change in the optimal objective function value in several consecutive iterations is less than a preset threshold. If yes, proceed to step S158; otherwise, update the next annealing temperature. and order Return to step S152; S158. The current optimal measurement point layout scheme is... As the optimal set of measurement points ,according to By determining the spatial coordinates corresponding to the row numbers of each measuring point, the optimal monitoring point layout scheme can be obtained. .

[0019] In one possible implementation, in the set Select row number The methods include: S1521, Read the reduced-order mode matrix The i-th order temperature field spatial mode in and their corresponding energy eigenvalues ,calculate Modal energy sensitivity at each measurement point: , in, line number Modal energy sensitivity at the corresponding measurement point; for In line number Modal values ​​at; For the first First-order mode energy weights; This represents the total order of the reduced-order modes; For the i-th order temperature field spatial mode Corresponding energy characteristic value; S1522. Calculate the weighted selection probability of the measurement point based on the modal energy sensitivity:

[0020] in, line number The weighted selection probability of the corresponding measurement point; L is the set The total number of Bank of China branches; It is a very small positive number; S1522, Based on weighted selection probability In the set Select row number .

[0021] The beneficial effects of the above technical solution are as follows: This solution is based on spatial modes With energy weight Calculation sensitivity This value represents the richness of thermodynamic information contained in a local region of the transformer, and is converted into a weighted selection probability. To guide the generation of new measurement point layout schemes During the iterative optimization process, new solutions generated within the neighborhood are analyzed. An evaluation is conducted for cases where the objective function deteriorates (i.e., The exponential form of the acceptance probability is given. This probability value, which is between zero and one, can be regarded as the dynamic weight of the algorithm tolerating inferior solutions at the current iteration temperature in order to explore a wider solution space.

[0022] Therefore, the probability term in the equation Combining sensitivity to select probability This constitutes an adaptive optimization mechanism that combines "physical feature-based heuristic guidance" with "random thermodynamic perturbation." This not only overcomes the drawbacks of conventional random initialization in generating the optimal monitoring point layout scheme, which easily falls into blind search and slow convergence in massive discrete grids, but also effectively addresses the problem of the highly non-convex objective function caused by physical constraints such as complex insulation and structural avoidance.

[0023] Specifically, the poor-quality solution tolerance mechanism in the high-temperature stage allows the search trajectory to flexibly overcome local minima, while the probabilistic convergence in the low-temperature stage ensures the refinement of the final spatial distribution. This dynamic balance between global exploration and local development not only stably acquires a globally competitive measurement point spatial topology within controllable iteration overhead, effectively avoiding engineering ineffective solutions such as measurement point clustering or falling into physically insensitive areas, but also maximizes the perception fidelity of finite physical sensors for the evolution of high-order modes across the entire transformer domain.

[0024] In one possible implementation, the modality coefficient prediction model is a BP neural network, whose loss function is... The expression is: , in, The modal coefficients represent the total number of offline training samples during model training. This represents the total order of the reduced-order modes; and The offline training set One sample in At this moment The actual modal coefficients and network predicted values ​​of the first-order modal coefficients; and These are the reduced-order mode matrices. Spatial modes of the i-th and j-th order temperature fields , Corresponding energy characteristic value.

[0025] The beneficial effects of the above technical solution are as follows: the offline training strategy of the modal coefficient prediction model in this solution explicitly incorporates energy feature values. The optimization objective is incorporated into the neural network, through the defined error weight formula. Differentiated penalties are applied to the prediction errors of modal coefficients at each order. Conventional neural network mean squared error tends to treat all dimensions of output equally, while in reality, the thermodynamic energies contained in different spatial modes differ by orders of magnitude. This strategy innovatively transforms the key physical prior knowledge of "modal energy distribution" into gradient-guided constraints at the algorithmic level, utilizing a logarithmic function. Its smooth adjustment characteristics enable the model to prioritize and accurately fit the high-energy modes that dominate the global macroscopic temperature distribution, while reasonably relaxing the penalty for low-energy minute thermal perturbation details.

[0026] Therefore, the energy weights in the formula By combining backpropagation updates with neural networks, an adaptive error adjustment mechanism guided by prior physical knowledge is constructed. This ensures that the parameter evolution direction of the data-driven model during training closely matches the real physical laws of transformer thermal field evolution, effectively avoiding the risk of macroscopic hotspot prediction deviations caused by overfitting low-energy secondary features. This not only improves the feature learning efficiency of the modal coefficient prediction model when facing complex transient conditions, but also ensures that the output predicted channel modal coefficients have stronger physical rationality and generalization ability.

[0027] In one possible implementation, the expression for generating the fused modal coefficients is: ,

[0028] in, For the current moment fusion mode coefficients; and For dynamic fusion weights, , and Both are variables, taking values ​​of 1 and 2. It is an exponential function; and All times are current. The confidence energy value of the monitoring point in the inversion channel or the working condition prediction channel; Temperature coefficient for adjusting the smoothness of weight distribution; and Each represents the current time. Inversion channel modal coefficients and predicted channel modal coefficients at monitoring points; This is the optimal modal submatrix; For the current moment Real-time temperature measurement vector of the transformer; This is a transpose.

[0029] In one possible implementation, the credibility energy value is calculated. and The methods include: The five-dimensional characteristic indexes for constructing a dynamic gating network are: the instantaneous residual of the monitoring point inversion channel. Instantaneous residuals of the operating condition prediction channel Historical reconstruction accumulated residuals Temperature fluctuation rate and mode coefficient volatility ; Based on the five-dimensional feature index, construct the feature vectors of the two channels in the dynamic gating network: , in, and These are the feature vectors of two channels in the dynamic gating network, namely the inversion channel and the working condition prediction channel. Extracting continuous The feature vectors at each time point form the historical reliability sequences of the two channels, respectively. and : , in, and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The inversion channel and the operating condition prediction channel are extracted using a GRU time series network through iterative loops. and Hidden state: , in, and These are the hidden states of the inversion channel and the operating condition prediction channel, respectively. For any time step; and These are nonlinear gated operation functions based on the network parameters of the inversion channel and the working condition prediction channel, respectively. and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The final hidden state and The inputs are fed into two independent fully connected layer branches, and the confidence energy value of the monitoring point inversion channel is calculated. Reliability energy value of the working condition prediction channel :

[0030]

[0031] in, and These are the learnable parameter sets that are not shared between the inversion channel and the operating condition prediction channel; and This is the weight matrix. and For bias terms, and This is the projection vector.

[0032] The beneficial effects of the above technical solution are as follows: This dual-channel dynamic credibility fusion strategy explicitly nests the multi-dimensional physical residual index with the temporal deep learning model, while also introducing the credibility energy value of the network output. With temperature coefficient The Softmax competition mechanism is used to dynamically allocate fusion weights. The constructed five-dimensional feature vector Numerically, the mapping accuracy of the inversion and prediction channels at the current moment and their thermodynamic stability within the historical time window were comprehensively quantified. Subsequently, the GRU time series network was used based on the historical reliability sequence. Extracting the spatiotemporal evolution patterns of errors and fluctuations allows for the understanding of hidden states. The output can take into account the historical trend of the system and effectively filter out misjudgments caused by single-point random mutations.

[0033] Therefore, the concatenation of multidimensional physical state assessment indices and deep gated networks constitutes an adaptive arbitration mechanism for "physical measurement point inversion" and "operating condition network prediction." This enables the computational system to effectively overcome the limitations of traditional single physical inversion, which is susceptible to local measurement noise contamination, and the generalization degradation of single data-driven models when faced with sudden changes in operating conditions without prior learning. The algorithm can dynamically adjust the confidence level of each channel based on the signal-to-noise ratio and operating condition familiarity of the real-time operating environment, and constrain the drastic oscillations of the weights. Thus, under complex dynamic operating conditions, it provides a low-dimensional state representation with strong anti-disturbance capability and reliability for high real-time reconstruction of the transformer's panoramic temperature field.

[0034] In one possible implementation, the expression for the instantaneous residual of the monitoring point inversion channel is: , in, For the current moment Real-time temperature reconstruction vector of the transformer; For the current moment The reconstructed temperature of the monitoring points in the inversion channel calculation; It is a quadratic norm; The expression for the instantaneous residual of the operating condition prediction channel is: , in, For the current moment The reconstructed temperature of the monitoring points calculated by the operating condition prediction channel; The expression for the cumulative residual of the historical reconstruction is: , in, for The real-time measured temperature vector at any given moment; for The historical fusion and reconstruction of moments evoke a sense of warmth. The modal coefficients at time 0; The expressions for the temperature change volatility and the modal coefficient volatility are as follows: , in, and They are respectively Real-time measured temperature vector and historical data fusion to reconstruct temperature; During the offline training phase, the dynamic gating network uses the modal coefficients from all time points in step A3 as supervision labels, and the mean squared error between the fused modal coefficients and the actual modal coefficients is used as the loss function.

[0035] in, Let be the modal coefficients at time t; N be the total number of time points. The actual modal coefficients at time t.

[0036] The beneficial effects of this invention are as follows: At the computational level, this solution effectively overcomes the problem of lengthy computation time in traditional numerical simulation methods. It reconstructs the vector through a three-dimensional temperature field reconstructed throughout the transformer's interior. The nonlinear multiphysics iterative process involving massive grid nodes in numerical simulation is abandoned and transformed into extremely low-dimensional linear matrix multiplication. This dimensionality reduction mapping greatly reduces the computational load of the system, enabling the system to complete the panoramic state update in a very short time, fully meeting the stringent requirements of millisecond-level real-time performance for online power grid assessment.

[0037] In terms of reconstruction accuracy, it ensures high fidelity and stability of physical field reconstruction. The fusion mode coefficients on which the reconstruction depends Through adaptive arbitration using a dynamic gating network, local measurement noise and generalization errors in the prediction model are effectively filtered out. Therefore, compared to single-channel reconstruction methods that rely on inversion from a single measurement point or prediction under a single operating condition, the three-dimensional global temperature field calculated by this method exhibits smaller numerical fluctuations, stronger anti-interference capabilities, and more stable and accurate temperature distribution results when facing complex transient operating conditions.

[0038] Through efficient dimensionality reduction and reconstruction calculations, the system can completely acquire the three-dimensional temperature field distribution across the entire transformer domain. This effectively covers hidden areas where sensors cannot be physically deployed due to insulation safety and structural constraints, such as winding hotspots and fault-prone locations, thus providing reliable digital twin early warning and defense. Global analysis of the high-real-time, high-precision 3D reconstruction results allows for accurate location of the real-time hotspot temperature. and its three-dimensional physical coordinates Furthermore, based on the reconstructed data across the entire domain, it can perform customized, multi-point continuous tracking and monitoring of the thermal state evolution of key areas of concern, such as critical insulation components. This ensures that the triggered overheating early warning signal is based on a solid dual physical and data benchmark, effectively reducing the risk of false alarms caused by monitoring delays or single-point distortions. It provides timely and reliable data support for transformer capacity expansion decisions and insulation safety state prediction when facing dynamic overloads. Attached Figure Description

[0039] Figure 1This is a flowchart illustrating the real-time calculation method for the temperature field of an oil-immersed transformer in this scheme. Figure 2 This is a 3D simulation model of an oil-immersed transformer. Figure 3 This is a 3D mesh partitioning diagram of an oil-immersed transformer. Figure 4 This is a diagram showing the magnetic flux density and magnetic field distribution of the core of an oil-immersed transformer. Figure 5 Diagram of winding losses for an oil-immersed transformer; Figure 6 A three-dimensional temperature field diagram of an oil-immersed transformer; Figure 7 This is a two-dimensional temperature-flow rate diagram for an oil-immersed transformer. Figure 8 For cumulative energy cutoff plot; Figure 9 Convergence diagram of the algorithm for optimizing the layout of modal sensitive measurement points; Figure 10 This is a temperature field error cloud map for an oil-immersed transformer. Figure 11 Box plot of the statistical distribution of global temperature field reconstruction error for dual-channel fusion, single prediction condition channel, and single measurement point inversion channel of dynamic gating network. Detailed Implementation

[0040] 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.

[0041] refer to Figure 1 , Figure 1 A flowchart illustrating the real-time calculation method for the temperature field of an oil-immersed transformer is shown; for example... Figure 1 As shown, the method S includes steps S1 to S5.

[0042] In step S1, during the online operation phase of the oil-immersed transformer, the external dynamic operating condition parameters at the current moment are collected in real time to form a real-time operating condition parameter vector, and the actual temperatures of all monitoring points inside the transformer are collected to form a real-time measured temperature vector.

[0043] In one embodiment of the present invention, the method for determining the arrangement positions of all monitoring points inside the transformer and the mode sub-matrix includes: S11. Select grid nodes that meet the preset conditions from all grid nodes in the coupled transient simulation model as optional measurement points, and construct a set of optional measurement point row numbers; the preset conditions are measurable nodes that meet the conditions of insulation distance, mechanical installation space, oil flow channel avoidance, structural component avoidance and wiring.

[0044] S12. Based on the number of monitoring points to be arranged in the oil-immersed transformer, randomly select monitoring points from the set of available monitoring point row numbers to generate a monitoring point arrangement scheme; S13. For the measurement point layout scheme, select the reduced-order modes of the corresponding measurement points from the reduced-order mode matrix to form a mode submatrix; S14. Using the modal submatrix, spatial distribution constraints of measurement points, and correlation of measurement points as evaluation indicators, construct an objective function for optimizing the measurement point layout; the objective function for optimizing the measurement point layout... The expression is:

[0045] ,

[0046]

[0047] in, The condition number of the modal submatrix; This refers to the spatial distribution constraint term of the measurement points; For the correlation term of the measurement points; for , and Weighting coefficients; Measurement point layout scheme The corresponding modal submatrix; and They are respectively The maximum and minimum singular values; It is a very small positive number; for The r-th order reduced mode; , and They are respectively The r-th reduced-order mode at the 1st, 2nd, and mth measurement points; This represents the distance between the spatial coordinates of two measuring points a and b. Let these be the row numbers for measurement points a and b. and These are the spatial coordinates of measurement points a and b, respectively. It is a quadratic norm; The preset minimum installation distance between measuring points; and These are the reduced-order modes corresponding to measurement points a and b, respectively. This is a transpose.

[0048] S15. Based on the measurement point layout scheme, the measurement points not selected in the set of optional measurement point row numbers, and the measurement point layout optimization objective function, the measurement point layout scheme is iteratively optimized using the simulated annealing mechanism to obtain the optimal monitoring point layout scheme. S16. Based on the optimal monitoring point layout scheme, select the reduced-order modes of the corresponding monitoring points from the reduced-order mode matrix to form the optimal mode submatrix. :

[0049] In a preferred example, the number of monitoring points is set. The value is 8, at the initial annealing temperature. Cooling coefficient Iterative optimization is performed under the given parameters. For example... Figure 9 As shown, the objective function tends to stabilize and converge after approximately 150 iterations. The optimal physical space coordinates for the arrangement of the eight fiber optic grating measurement points were obtained. These eight optimal monitoring points correspond to the most sensitive critical areas inside the transformer, specifically including: the interphase oil passage at the upper end of the high-voltage winding (3 measurement points), the side of the insulating pad at the upper end of the low-voltage winding (2 measurement points), the side wall of the yoke clamp on the iron core (1 measurement point), and the oil outlet pipe inlet area at the top of the oil tank (2 measurement points).

[0050] The optimal monitoring point layout scheme and the optimal modal submatrix are adopted. This serves as the arrangement location and modal submatrix of all monitoring points inside the transformer.

[0051] In one embodiment of the present invention, step S15 further includes: S151. Determine the current measurement point layout scheme. Equal to the current optimal measurement point layout scheme And set the initial annealing temperature. Cooling coefficient Maximum number of iterations and the number of neighborhood searches at each annealing temperature ; S152, in the In the next annealing iteration, from the current measurement point layout scheme Randomly select a row number The set of measurable node row numbers is composed of the unselected measurable point row numbers from the set of optional measurable point row numbers. Select row number A new measurement point layout scheme was formed. ; In implementation, this scheme is preferably implemented in a collection. Select row number The methods include: S1521, Read the reduced-order mode matrix The i-th order temperature field spatial mode in and their corresponding energy eigenvalues ,calculate Modal energy sensitivity at each measurement point: , in, line number Modal energy sensitivity at the corresponding measurement point; for In line number Modal values ​​at; For the first First-order mode energy weights; This represents the total order of the reduced-order modes; For the i-th order temperature field spatial mode Corresponding energy characteristic value; S1522. Calculate the weighted selection probability of the measurement point based on the modal energy sensitivity:

[0052] in, line number The weighted selection probability of the corresponding measurement point; L is the set The total number of Bank of China branches; It is a very small positive number; S1522, Based on weighted selection probability In the set Select row number .

[0053] S153. Calculate the change in the objective function. , and They are respectively and The calculated objective function value; S154, when season And proceed to step S155; when At that time, generate random numbers ,like Then let Proceed to step S155; otherwise, remain in the current state. The process remains unchanged, and proceeds to step S156. For the first Annealing temperature during the next annealing iteration; S155, Judgment Is it less than If so, then update. If the condition is met, proceed to step S156; otherwise, proceed directly to step S156. S156. Determine whether the annealing process has been completed at the current annealing temperature. If the next neighborhood search is successful, proceed to step S157; otherwise, return to step S152. S157. Determine whether the current iteration count has reached the maximum iteration count, whether the annealing temperature is lower than the preset termination temperature, or whether the change in the optimal objective function value in several consecutive iterations is less than a preset threshold. If yes, proceed to step S158; otherwise, update the next annealing temperature. and order Return to step S152; S158. The current optimal measurement point layout scheme is... As the optimal set of measurement points ,according to By determining the spatial coordinates corresponding to the row numbers of each measuring point, the optimal monitoring point layout scheme can be obtained. .

[0054] In implementation, this scheme preferably uses the following methods to obtain the reduced-order mode matrix of the oil-immersed transformer: A1. Construct an electro-thermal-current coupled transient simulation model for an oil-immersed transformer and perform dynamic multi-condition transient simulation to generate a dataset containing full-time-domain transient global temperature field data and dynamic operating condition parameters; The transient simulation model of an oil-immersed transformer with electro-thermal-current coupling includes the core, windings, clamps, insulating oil, oil tank, radiator, and oil duct structure. Based on the actual spatial location and heat transfer relationship of each component, the three-dimensional geometric model is meshed using finite element methods to obtain a model containing... The computational domain of each grid node.

[0055] like Figure 2 and Figure 3 As shown, a simulation model was established based on the actual structural parameters of an oil-immersed transformer, model ODFPS-1000000 / 1000. This oil-immersed transformer adopts a single-phase, double-column core structure. The oil tank includes the core, windings, clamps, magnetic shielding, and insulating oil area. Each column has multiple coaxial winding units, arranged radially along the core. Axial and radial oil channels are provided inside the windings, forming the insulating oil flow path; the radial oil channels refer to the oil channels along the radial direction of the transformer windings, or the direction of the winding width.

[0056] The oil tank height of the ODFPS-1000000 / 1000 oil-immersed transformer is 8.6m, the length is 13.2m, and the width is 5.4m; the core height is 5.8m, and the center distance between core columns is 3.2m; the axial height of the winding is 4.6m; the inner diameter of the low-voltage winding is 1.60m and the outer diameter is 2.10m; the inner diameter of the high-voltage winding is 2.30m and the outer diameter is 3.05m; the width of the main oil passage is 60mm to 120mm, and the width of the radial oil passage is 6mm to 12mm. The basic operating parameters of the transformer in this embodiment are shown in Table 1.

[0057] Table 1 Parameters of ODFPS-1000000 / 1000 Transformer

[0058] When performing dynamic multi-condition transient simulations, it is necessary to set the electro-thermal-fluid multiphysics coupling control equations and material property parameters: The simulation model is configured with coupled control equations for electromagnetic field, temperature field and fluid field, and material property parameters are assigned to the iron core, winding, insulating oil, oil tank and insulating structure. The material property parameters include one or more of the following: electrical conductivity, magnetic permeability, density, specific heat capacity, thermal conductivity, dynamic viscosity and thermal property parameters of insulating oil.

[0059] The electromagnetic field is used to calculate the core loss and winding loss, which serve as the heat source term for solving the temperature field; the temperature field is used to calculate the transient temperature of each node inside the transformer; and the fluid field is used to describe the flow and heat transfer process of the insulating oil in the oil channels and heat dissipation structure.

[0060] Electric field governing equations: , Represents the electric displacement vector. Indicates charge density; For Hamiltonian operators; Magnetic field control equations: , Indicates magnetic field strength. Indicates the total current density; 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; Temperature field governing equations: , Indicates the density of the material. Indicates specific heat capacity. Indicates temperature. Indicates thermal conductivity, This represents the heat source term obtained from electromagnetic losses.

[0061] Set dynamic operating parameters and transient boundary conditions: Set dynamic operating parameters, including insulating oil flow rate. Load factor Ambient temperature Operating frequency One or more of the following; set the heat source boundary, oil flow boundary, and heat dissipation boundary for transient simulation based on the dynamic operating condition parameters. Load factor Used to correct winding copper losses, operating frequency Used to correct core loss and insulating oil flow rate Used to correct fluid heat transfer conditions, ambient temperature. Used to set the convective heat transfer boundary on the outer surface of the oil tank or radiator.

[0062] After setting the relevant parameters of the simulation model, perform dynamic multi-condition transient simulation: Set the simulation time step and total simulation time The electro-thermal-current coupled transient simulation model was solved under multiple dynamic operating condition combinations to obtain the transient response results of the internal temperature field of the oil-immersed transformer as a function of time under different operating conditions. The discrete time can be expressed as: ,in, This represents the total number of discrete moments during the transient simulation.

[0063] A coupled electromagnetic field-fluid field-temperature field calculation method based on the finite element method for oil-immersed transformers is adopted. First, the dynamic operating parameters at the current time step are read; the leakage magnetic field distribution of the transformer winding at the current time step is calculated to obtain the winding loss and core loss, and the loss is converted into heat source density; then, using the flow field and temperature field of the previous time step as initial conditions, the mass conservation equation and momentum conservation equation of the insulating oil flow process are solved to obtain the flow velocity distribution in the oil channel and oil tank at the current time step; then, the heat source density is applied as a load to the coupled fluid field-temperature field calculation model to establish the whole-field fluid-structure interaction calculation equation, and the whole-field oil flow distribution and temperature distribution at the current time step are solved; during the solution process, the influence of the insulating oil physical properties and heat source density changes with temperature is considered, and the converged flow field and temperature field distribution results are obtained through multiple iterations within the current time step; the converged result of the current time step is used as the initial state of the next time step, the time step is advanced, and the above coupled calculation process is repeated until the dynamic simulation of the whole time domain is completed.

[0064] Figure 4This is a diagram showing the magnetic flux density and magnetic field distribution of the core of the oil-immersed transformer. Figure 5 This is a diagram showing the winding losses of the oil-immersed transformer. Figure 6 This is a three-dimensional temperature field diagram of the oil-immersed transformer. Figure 7 This is a two-dimensional temperature-flow rate diagram of the oil-immersed transformer. Figure 6 and Figure 7 It can be seen that the overall internal temperature of the transformer increases with the height, and a hot spot is formed at the upper end of the winding, with a temperature of 327K. The temperature in the phase-to-phase area between the high-voltage winding and the low-voltage winding is relatively high. This is because the space between adjacent windings is small, the oil flow resistance is large, and the local oil flow velocity is low, which leads to a decrease in the convective heat transfer efficiency in this area.

[0065] This scheme employs the aforementioned multiphysics coupling iterative calculation process to explicitly couple the electromagnetic loss heating within the transformer with the flow heat transfer process of the insulating oil. It also incorporates temperature-varying insulating oil properties and heat source density into multiple iterative cycles. The working mechanism involves first obtaining an accurate non-uniform heat source density through electromagnetic field solving; then, obtaining the true local oil flow velocity by solving the mass and momentum conservation equations, which directly determines the convective heat transfer efficiency at the fluid-solid boundary; and finally, the design of nonlinearly updating the insulating oil properties with temperature accurately maps the true physical evolution of the thermally buoyant-driven oil flow in numerical terms.

[0066] Therefore, the operations of passing data from the previous time step, iterating within the current time step, and advancing to the next time step constitute a high-fidelity nonlinear feedback mechanism for global thermo-fluid dynamics. This enables the transient simulation model to fully capture and reconstruct the physical details of local oil flow stagnation and hot spot aggregation within narrow interphase oil channels when facing complex load fluctuations. This avoids the significant truncation errors caused by traditional steady-state thermal circuit models or empirical formulas, thus providing a high-quality full-time-domain dataset with rigorous physical benchmarks for subsequent intrinsic orthogonal decomposition order reduction and dynamic reliability gating networks.

[0067] Extract the global node temperatures and corresponding operating parameters across the entire time domain to form a dataset: At each discrete time Extract the global grid node temperatures of the oil-immersed transformer and form a global node temperature column vector: .in, Indicates the first Each grid node at time... Temperature value, This indicates the total number of grid nodes in the entire domain.

[0068] Simultaneously, extract the dynamic operating condition parameter vector corresponding to that moment: .in, for Dynamic operating conditions at any given moment.

[0069] The whole process The global node temperature vectors and their corresponding dynamic operating condition parameter vectors at each discrete time point are associated and stored to construct a full-time-domain transient temperature field dataset: and corresponding operating condition parameter dataset: .

[0070] A2. Select typical temperature field data from the dataset, construct a temperature field snapshot matrix, and perform intrinsic orthogonal decomposition to obtain multiple temperature field spatial modes and their corresponding energy eigenvalues; the detailed implementation process is as follows: Typical temperature field data are selected from the dataset, including temperature field data during the stable operating phase and temperature field data during the fluctuating operating phase. The stable operating phase is used to characterize the temperature distribution characteristics of the oil-immersed transformer under stable operating conditions, while the fluctuating operating phase is used to characterize the transient temperature response characteristics when the load factor, insulating oil flow rate, ambient temperature, or operating frequency changes.

[0071] Extract the column vectors of each typical temperature field and arrange them in chronological order to construct a temperature field snapshot matrix: .in, Indicates the first The global node temperature vector at a typical moment. This indicates the number of typical temperature field data sets selected, and .

[0072] This typical data selection and snapshot matrix construction strategy explicitly arranges the high-dimensional global temperature column vector into a time series and specifically introduces two highly representative types of temperature field data—one from the stable operating phase and the other from the fluctuating operating phase—to jointly construct the snapshot matrix. Its working mechanism lies in the fact that the data from the stable phase numerically and rigorously maps the fundamental thermodynamic distribution benchmark of the oil-immersed transformer under stable operating conditions; while the data from the fluctuating phase captures with high fidelity the transient temperature response characteristics and nonlinear evolution boundaries excited within the system when the load factor, insulating oil flow rate, ambient temperature, or operating frequency undergoes drastic changes.

[0073] Therefore, the aforementioned joint sampling, encompassing both normal baselines and transient mutations, constitutes a panoramic feature extraction mechanism for complex time-varying thermodynamic behavior. This enables the constructed snapshot matrix to significantly eliminate redundant time-series data across the entire time domain while fully preserving the spatiotemporal evolution details of the system under extremely variable environments within the matrix column space. This ensures that the subsequently extracted dominant spatial modes not only accurately reconstruct the physical field under normal operating conditions but also possess extremely strong spatiotemporal representation robustness and wide-area generalization ability when facing sudden, unknown operating conditions that have not been learned.

[0074] snapshot matrix of the temperature field Perform singular value decomposition to obtain ,in, It is a left singular matrix. Indicates the first Spatial modes of the temperature field; It is a singular value diagonal matrix, and , It is a right singular matrix.

[0075] Calculate the energy eigenvalues ​​corresponding to each order of the temperature field spatial mode based on singular values. ,in, Indicates the first Energy eigenvalues ​​corresponding to the spatial modes of the temperature field.

[0076] A3. Select the previous step based on the cumulative energy percentage threshold. First-order modes, forming a reduced-order mode matrix The global temperature field at each time point is then projected onto the reduced-order mode matrix to obtain the modal coefficients of the global temperature field at the corresponding time point. The detailed implementation process of A3 is as follows: Before calculation Accumulated energy percentage of spatial modes in the first-order temperature field According to the preset cumulative energy percentage threshold Select the smallest integer that satisfies the following formula : ; the preceding The spatial mode composition of the first-order temperature field is a reduced-order mode matrix. .

[0077] In a preferred example, 100 sets of typical temperature field data are selected from the dataset to construct a temperature field snapshot matrix, and intrinsic orthogonal decomposition and cumulative energy truncation are performed using the method described above. The cumulative energy percentage threshold is... Setting it to 99.99%, the first six temperature field spatial modes were ultimately selected as the reduced-order modes, such as... Figure 8 As shown.

[0078] This reduced-order mode truncation strategy explicitly transforms eigenvalues ​​into cumulative energy percentages and introduces a stringent preset threshold as a boundary to dynamically determine the minimum integer value. And construct the reduced-order mode matrix Its working mechanism lies in the energy eigenvalues ​​of each order generated by the intrinsic orthogonal decomposition. Numerically, it directly represents the variance of the thermodynamic system (i.e., the amount of physical information) contained in the corresponding spatial mode; The dominant mode gathers most of the macroscopic temperature distribution and transient evolution laws inside the transformer, while the small eigenvalues ​​that are forcibly discarded at the tail end often correspond to local irregular thermal disturbances or redundant numerical noise.

[0079] Therefore, the aforementioned adaptive truncation criterion based on a high cumulative energy ratio constitutes an energy-dominant mathematical dimensionality reduction and high-frequency noise filtering mechanism for high-dimensional complex temperature fields. This enables the system to achieve high-fidelity physical feature mapping of the entire thermal field of a transformer with massive grid nodes while compressing dimensions. This not only breaks through the high-dimensional computational barrier of traditional numerical simulation and avoids the negative impact of high-frequency micro-interferences on subsequent measurement point inversion, but also provides the necessary core dimensionality reduction features to achieve high-precision millisecond-level panoramic temperature field reconstruction in the online stage.

[0080] The whole process Projecting the global nodal temperature vectors at discrete time points onto the reduced-order mode matrix The modal coefficients at each time point are obtained. ,in, express The modal coefficient vector at each time step. By iterating through all discrete time steps, a modal coefficient dataset covering the entire time domain is generated. .

[0081] In step S2, the inversion channel mode coefficients of the monitoring points are calculated based on the mode sub-matrices of all monitoring points inside the transformer and the real-time measured temperature vector. .

[0082] In step S3, the real-time operating condition parameter vector is input into the trained modal coefficient prediction model, and the operating condition prediction channel modal coefficients of the monitoring points are output. In practice, this scheme preferably uses a BP neural network as the modal coefficient prediction model, and its loss function is... The expression is: , in, The modal coefficients represent the total number of offline training samples during model training. This represents the total order of the reduced-order modes; and The offline training set One sample in At this moment The actual modal coefficients and network predicted values ​​of the first-order modal coefficients; and These are the reduced-order mode matrices. Spatial modes of the i-th and j-th order temperature fields , Corresponding energy characteristic value.

[0083] In step S4, a dynamic gating network is used to generate dynamic fusion weights, and the dynamic fusion weights are used to fuse the modal coefficients of the inversion channel and the modal coefficients of the operating condition prediction channel to obtain fused modal coefficients.

[0084] In one embodiment of the present invention, the expression for generating the fused mode coefficients is: , in, For the current moment fusion mode coefficients; and For dynamic fusion weights, , and Both are variables, taking values ​​of 1 and 2. It is an exponential function; and All times are current. The confidence energy value of the monitoring point in the inversion channel or the working condition prediction channel; Temperature coefficient for adjusting the smoothness of weight distribution; and Each represents the current time. Inversion channel modal coefficients and predicted channel modal coefficients at monitoring points; This is the optimal modal submatrix; For the current moment Real-time temperature measurement vector of the transformer; This is a transpose.

[0085] During implementation, this scheme preferably calculates the credibility energy value. and The methods include: The five-dimensional characteristic indexes for constructing a dynamic gating network are: the instantaneous residual of the monitoring point inversion channel. Instantaneous residuals of the operating condition prediction channel Historical reconstruction accumulated residuals Temperature fluctuation rate and mode coefficient volatility ; Based on the five-dimensional feature index, construct the feature vectors of the two channels in the dynamic gating network: , in, and These are the feature vectors of two channels in the dynamic gating network, namely the inversion channel and the working condition prediction channel. Extracting continuous The feature vectors at each time point form the historical reliability sequences of the two channels, respectively. and : , in, and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The inversion channel and the operating condition prediction channel are extracted using a GRU time series network through iterative loops. and Hidden state: , in, and These are the hidden states of the inversion channel and the operating condition prediction channel, respectively. For any time step; and These are nonlinear gated operation functions based on the network parameters of the inversion channel and the working condition prediction channel, respectively. and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The final hidden state and The inputs are fed into two independent fully connected layer branches, and the confidence energy value of the monitoring point inversion channel is calculated. Reliability energy value of the working condition prediction channel :

[0086]

[0087] in, and These are learnable parameter sets that are not shared between the inversion channel and the operating condition prediction channel. The specific mechanism by which these parameters function in the network is as follows: and This is the weight matrix, used to assign weights to the hidden states output by the GRU network. and The high-dimensional linear affine transformation of the feature space is essentially to map the temporal state features to a dedicated credibility feature space. The contribution weights of historical dynamic features of different dimensions to the final result are reflected by the internal values ​​of the matrix, thereby learning and extracting the key feature combination relationships that affect the credibility of the current channel. and The bias term is used to provide a baseline translation offset for the above affine transformation. Its function is to introduce a basic threshold for network activation to correct the overall offset of the feature mapping results, thereby effectively improving the adaptability and fitting flexibility of the reliability evaluation model to different complex operating states and fluctuation data distributions of transformers. and This is the projection vector. Since the hidden state remains in the high-dimensional feature space after activation by the tanh function, this projection vector is obtained through an inner product operation (i.e., transpose multiplication). This process involves dimensionality reduction, compression, and aggregation of high-dimensional nonlinear features. Ultimately, it maps these features to a single scalar value (i.e., the quantifiable channel confidence energy value). and This allows for a direct quantitative evaluation of the current reliability of each channel.

[0088] The expression for the instantaneous residual of the monitoring point inversion channel is as follows: , in, For the current moment Real-time temperature reconstruction vector of the transformer; For the current moment The reconstructed temperature of the monitoring points in the inversion channel calculation; It is a quadratic norm; The expression for the instantaneous residual of the operating condition prediction channel is: , in, For the current moment The reconstructed temperature of the monitoring points calculated by the operating condition prediction channel; The expression for the cumulative residual of the historical reconstruction is: , in, for The real-time measured temperature vector at any given moment; for The historical fusion and reconstruction of moments evoke a sense of warmth. The modal coefficients at time 0; The expressions for the temperature change volatility and the modal coefficient volatility are as follows: , in, and They are respectively Real-time measured temperature vector and historical data fusion to reconstruct temperature; During the offline training phase, the dynamic gating network uses the modal coefficients from all time points in step A3 as supervision labels, and the mean squared error between the fused modal coefficients and the actual modal coefficients is used as the loss function.

[0089] in, Let be the modal coefficients at time t; N be the total number of time points. The actual modal coefficients at time t.

[0090] In step S5, based on the reduced-order mode matrix of the oil-immersed transformer... and fusion mode coefficients Calculate the reconstructed vector of the global three-dimensional temperature field inside the transformer at the current time t. .

[0091] During implementation, this plan also includes The data is reverse-engineered into a three-dimensional spatial temperature distribution map, and the global maximum temperature is extracted as the real-time hotspot temperature at the current moment. Locate the three-dimensional physical space coordinates corresponding to the hottest spot. When the hottest temperature An overheat warning signal is triggered when the preset safe insulation temperature threshold is exceeded, or when the temperature rise rate exceeds the safe temperature rise threshold.

[0092] In a preferred embodiment, by applying this scheme to the online monitoring of oil-immersed transformers, the results show that the original 10.2-hour calculation time required for three-dimensional transient multiphysics finite element analysis was transformed into matrix multiplication and neural network forward inference in a low-dimensional space using this method. Specifically, the calculation time for reconstructing the single-step global temperature field (covering 2,438,457 degrees of freedom nodes) was significantly reduced to 25 milliseconds, achieving a leapfrog improvement in computational efficiency and meeting the millisecond-level real-time requirements for online assessment of power grid equipment status.

[0093] At the same time, such as Figure 10 The reconstructed temperature error cloud map shown indicates that, under complex transient conditions with drastic fluctuations in load and flow rate, the global temperature field reconstructed by this scheme is highly consistent with the actual three-dimensional simulation temperature field. The maximum absolute error of the global reconstruction is only 3.42K, which meets the error tolerance requirements for transformer hot spot temperature rise assessment and engineering monitoring.

[0094] To more comprehensively evaluate the accuracy and robustness of the dual-channel fusion reconstruction method of this invention, mean absolute error (MAE) and mean relative error (MRE) are introduced as global evaluation indicators. MAE measures the overall level of absolute deviation between the reconstructed temperature and the true temperature, while MRE characterizes the percentage deviation of the overall error relative to the true temperature.

[0095] Figure 11 This paper presents a box plot illustrating the statistical distribution of the global temperature field reconstruction error between the dual-channel fusion model of this invention and the single prediction condition channel and single measurement point inversion channel. The upper and lower edges of the box represent the 20% and 80% quantiles, respectively; the thick horizontal line represents the mean; and the upper and lower whisker lines represent the maximum and minimum values. Figure 11 It can be intuitively seen that the error polarization phenomenon of the measurement point inversion channel is the most serious, with the highest mean values ​​of MAE and MRE and the largest fluctuation range (mean MRE of about 4.5%, extreme error approaching 7%); the prediction condition channel is the second worst; while the dual-channel model proposed in this invention not only has a greatly compressed mean error (mean MAE of about 1.2K, mean MRE of about 1.5%), but also has the most concentrated error distribution range and the smallest upper and lower limit range.

[0096] The underlying physical and algorithmic working mechanism is as follows: a single measurement point inversion channel is highly dependent on the real-time measurement accuracy of a limited number of sensors. Once a local measurement point is subjected to complex electromagnetic noise interference or measurement drift occurs, its error will be amplified many times through the inversion matrix, resulting in drastic numerical fluctuations in the reconstruction results. As a pure data-driven model, a single prediction condition channel often suffers from generalization decay when faced with sudden transient fluctuations in operating conditions that deviate from the historical training set due to the lack of real-time physical state calibration.

[0097] In contrast, the dual-channel model of this invention benefits from the adaptive arbitration mechanism of the Dynamically Gated Network (DCGN), which can dynamically adjust the fusion weights based on the signal-to-noise ratio and familiarity with the operating conditions in the real-time operating environment. It leverages the smoothing characteristics of the prediction model in macroscopic trends to effectively suppress random high-frequency noise at physical measurement points, while also utilizing data from measured physical anchor points to correct the generalization bias of the surrogate model in real time. This deep complementary mechanism of physical benchmark and data-driven approach significantly enhances the system's fault tolerance when facing complex power grid conditions and measurement disturbances, thereby ensuring high-fidelity accuracy and strong robustness against disturbances in the temperature field digital twin results at the macroscopic statistical level.

[0098] In summary, the real-time temperature field calculation method proposed in this scheme overcomes the computational bottleneck of traditional numerical simulation, which is time-consuming, while ensuring high-fidelity physical accuracy. It also effectively eliminates the monitoring blind spots caused by limited physical measurement points, and realizes high real-time, panoramic digital twin and accurate online early warning under complex transient conditions of oil-immersed transformers.

Claims

1. A method for real-time calculation of the temperature field of an oil-immersed transformer, characterized in that, Including the following steps: S1. During the online operation of the oil-immersed transformer, the external dynamic operating condition parameters at the current moment are collected in real time to form a real-time operating condition parameter vector, and the real temperature of all monitoring points inside the transformer is collected to form a real-time measured temperature vector. S2. Calculate the inversion channel mode coefficients of the monitoring points based on the modal submatrix of all monitoring points inside the transformer and the real-time measured temperature vector; S3. Input the real-time operating condition parameter vector into the trained modal coefficient prediction model and output the operating condition prediction channel modal coefficients of the monitoring points. S4. Dynamic fusion weights are generated using a dynamic gating network, and the dynamic fusion weights are used to fuse the modal coefficients of the inversion channel and the modal coefficients of the working condition prediction channel to obtain the fused modal coefficients. S5. Based on the reduced-order mode matrix of the oil-immersed transformer and fusion mode coefficients Calculate the reconstructed vector of the global three-dimensional temperature field inside the transformer at the current time t. .

2. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 1, characterized in that, The method for obtaining the reduced-order mode matrix of the oil-immersed transformer includes: A1. Construct an electro-thermal-current coupled transient simulation model for an oil-immersed transformer and perform dynamic multi-condition transient simulation to generate a dataset containing full-time-domain transient global temperature field data and dynamic operating condition parameters; A2. Select typical temperature field data from the dataset, construct a temperature field snapshot matrix and perform intrinsic orthogonal decomposition to obtain multiple temperature field spatial modes and their corresponding energy eigenvalues; A3. Select the previous one based on the cumulative energy percentage threshold. First-order modes, forming a reduced-order mode matrix The global temperature field at each time point is projected onto the reduced-order mode matrix to obtain the modal coefficients of the global temperature field at the corresponding time point.

3. The method for real-time calculation of the temperature field of an oil-immersed transformer according to claim 2, characterized in that, The method for determining the layout of all monitoring points inside the transformer and the modal sub-matrix includes: S11. Select grid nodes that meet the preset conditions from all grid nodes in the coupled transient simulation model as optional measurement points, and construct a set of optional measurement point row numbers; S12. Based on the number of monitoring points to be arranged in the oil-immersed transformer, randomly select monitoring points from the set of available monitoring point row numbers to generate a monitoring point arrangement scheme; S13. For the measurement point layout scheme, select the reduced-order modes of the corresponding measurement points from the reduced-order mode matrix to form a mode submatrix; S14. Using modal submatrix, spatial distribution constraints of measurement points, and correlation of measurement points as evaluation indicators, construct an objective function for optimizing the layout of measurement points. S15. Based on the measurement point layout scheme, the measurement points that were not selected in the set of available measurement point row numbers, and the objective function for measurement point layout optimization, the measurement point layout scheme is iteratively optimized using a simulated annealing mechanism to obtain the optimal monitoring point layout scheme. S16. Based on the optimal monitoring point layout scheme, select the reduced-order modes of the corresponding monitoring points from the reduced-order mode matrix to form the optimal mode submatrix. The optimal monitoring point layout scheme and the optimal modal submatrix are adopted. This serves as the arrangement location and modal submatrix of all monitoring points inside the transformer.

4. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 3, characterized in that, The objective function for optimizing the layout of measuring points The expression is: ， in, The condition number of the modal submatrix; This refers to the spatial distribution constraint term of the measurement points; For the correlation term of the measurement points; for , and Weighting coefficients; Measurement point layout scheme The corresponding modal submatrix; and They are respectively The maximum and minimum singular values; It is a very small positive number; for The r-th order reduced mode; , and They are respectively The r-th reduced-order mode at the 1st, 2nd, and mth measurement points; This represents the distance between the spatial coordinates of two measuring points a and b. Let these be the row numbers for measurement points a and b. and These are the spatial coordinates of measurement points a and b, respectively. It is a quadratic norm; The preset minimum installation distance between measuring points; and These are the reduced-order modes corresponding to measurement points a and b, respectively. This is a transpose.

5. The method for real-time calculation of the temperature field of an oil-immersed transformer according to claim 3, characterized in that, Step S15 further includes: S151. Determine the current measurement point layout scheme. Equal to the current optimal measurement point layout scheme And set the initial annealing temperature. Cooling coefficient Maximum number of iterations and the number of neighborhood searches at each annealing temperature ; S152, in the In the next annealing iteration, from the current measurement point layout scheme Randomly select a row number The set of measurable node row numbers is composed of the unselected measurable point row numbers from the set of optional measurable point row numbers. Select row number A new measurement point layout scheme was formed. ; S153. Calculate the change in the objective function. , and They are respectively and The calculated objective function value; S154, when season And proceed to step S155; when At that time, generate random numbers ,like Then let Proceed to step S155; otherwise, remain in the current state. The process remains unchanged, and proceeds to step S156. For the first Annealing temperature during the next annealing iteration; S155, Judgment Is it less than If so, then update. If the condition is met, proceed to step S156; otherwise, proceed directly to step S156. S156. Determine whether the annealing process has been completed at the current annealing temperature. If the next neighborhood search is successful, proceed to step S157; otherwise, return to step S152. S157. Determine whether the current iteration count has reached the maximum iteration count, whether the annealing temperature is lower than the preset termination temperature, or whether the change in the optimal objective function value in several consecutive iterations is less than a preset threshold. If yes, proceed to step S158; otherwise, update the next annealing temperature. and order Return to step S152; S158. The current optimal measurement point layout scheme is... As the optimal set of measurement points ,according to By determining the spatial coordinates corresponding to the row numbers of each measuring point, the optimal monitoring point layout scheme can be obtained. .

6. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 5, characterized in that, In the set Select row number The methods include: S1521, Read the reduced-order mode matrix The i-th order temperature field spatial mode in and their corresponding energy eigenvalues ,calculate Modal energy sensitivity at each measurement point: ， in, line number Modal energy sensitivity at the corresponding measurement point; for In line number Modal values ​​at; For the first First-order mode energy weights; This represents the total order of the reduced-order modes; For the i-th order temperature field spatial mode Corresponding energy characteristic value; S1522. Calculate the weighted selection probability of the measurement point based on the modal energy sensitivity: in, line number The weighted selection probability of the corresponding measurement point; L is the set The total number of Bank of China codes; It is a very small positive number; S1522, Based on weighted selection probability In the set Select row number .

7. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 1, characterized in that, The modality coefficient prediction model is a BP neural network, and its loss function is... The expression is: ， in, The modal coefficients represent the total number of offline training samples during model training. This represents the total order of the reduced-order modes; and The offline training set One sample in At this moment The actual modal coefficients and network predicted values ​​of the first-order modal coefficients; and These are the reduced-order mode matrices. Spatial modes of the i-th and j-th order temperature fields , Corresponding energy characteristic value.

8. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 3, characterized in that, The expression for generating the fused modal coefficients is: ， in, For the current moment fusion mode coefficients; and For dynamic weight fusion, , and All are variables, taking values ​​of 1 and 2. It is an exponential function; and All times are current. The confidence energy value of the monitoring point in the inversion channel or the working condition prediction channel; Temperature coefficient for adjusting the smoothness of weight distribution; and Each represents the current time. Inversion channel modal coefficients and predicted channel modal coefficients at monitoring points; This is the optimal modal submatrix; For the current moment Real-time temperature measurement vector of the transformer; This is a transpose.

9. The method for real-time calculation of temperature field of oil-immersed transformer according to claim 8, characterized in that, Calculate the credibility energy value and The methods include: The five-dimensional characteristic indexes for constructing a dynamic gating network are: the instantaneous residual of the monitoring point inversion channel. Instantaneous residuals of the operating condition prediction channel Historical reconstruction accumulated residuals Temperature fluctuation rate and mode coefficient volatility ; Based on the five-dimensional feature index, construct the feature vectors of the two channels in the dynamic gating network: ， in, and These are the feature vectors of two channels in the dynamic gating network, namely the inversion channel and the working condition prediction channel. Extracting continuous The feature vectors at each time point form the historical reliability sequences of the two channels, respectively. and : ， in, and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The inversion channel and the operating condition prediction channel are extracted using a GRU time series network through iterative loops. and Hidden state: ， in, and These are the hidden states of the inversion channel and the operating condition prediction channel, respectively. For any time step; and These are nonlinear gated operation functions based on the network parameters of the inversion channel and the working condition prediction channel, respectively. and These are the inversion channel and the operating condition prediction channel, respectively. The feature vector at time step; The final hidden state and The inputs are fed into two independent fully connected layer branches, and the confidence energy value of the monitoring point inversion channel is calculated. Reliability energy value of the working condition prediction channel : in, and These are the learnable parameter sets that are not shared between the inversion channel and the operating condition prediction channel; and This is the weight matrix. and For bias terms, and This is the projection vector.

10. The method for real-time calculation of the temperature field of an oil-immersed transformer according to claim 9, characterized in that, The expression for the instantaneous residual of the monitoring point inversion channel is: ， in, For the current moment Real-time temperature reconstruction vector of the transformer; For the current moment The reconstructed temperature of the monitoring points calculated in the inversion channel; It is a quadratic norm; The expression for the instantaneous residual of the operating condition prediction channel is: ， in, For the current moment The reconstructed temperature of the monitoring points calculated by the operating condition prediction channel; The expression for the cumulative residual of the historical reconstruction is: ， in, for The real-time measured temperature vector at any given moment; for The historical fusion and reconstruction of moments evoke a sense of warmth. The modal coefficients at time 0; The expressions for the temperature change volatility and the modal coefficient volatility are as follows: ， in, and They are respectively Real-time measured temperature vector and historical data fusion to reconstruct temperature; During the offline training phase, the dynamic gating network uses the modal coefficients from all time points in step A3 as supervision labels, and the mean squared error between the fused modal coefficients and the actual modal coefficients is used as the loss function. in, Let be the modal coefficients at time t; N be the total number of time points. The actual modal coefficients at time t.

Images