Real-time monitoring method and system for hot-spot temperature of oil-immersed transformer

Abstract

The invention discloses an oil-immersed transformer hot-spot temperature real-time monitoring method and system, and relates to the technical field of temperature monitoring, and the method comprises the steps of multi-mode sensing fusion and oil flow field reconstruction, adaptive Kalman filtering-wavelet packet combined denoising, and physical information neural network dynamic compensation. By deploying a multi-parameter embedded sensor array, winding temperature field distribution is captured in real time, and the hot spot tracking capability is optimized; temperature signal distortion caused by oil flow disturbance and electromagnetic interference is eliminated in combination with a time-frequency domain combined denoising algorithm; and a time-varying thermal parameter and harmonic loss model is embedded, so that high-precision temperature compensation is realized, and complex working conditions are adapted. According to the invention, the problem of monitoring misalignment of a traditional monitoring method under a dynamic working condition can be solved, and high-precision data support is provided for intelligent operation and maintenance of the transformer.

Description

Technical Field

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

[0002] As a core component of power systems, the internal hot spot temperature of oil-immersed transformers directly affects insulation life and operational safety. Traditional monitoring methods rely on temperature sensors located at fixed positions. However, in actual operation, factors such as oil flow disturbances, sudden load changes, and equipment aging cause dynamic changes in the temperature field distribution. Traditional methods struggle to accurately capture the migration trajectory of hot spots, resulting in monitoring lag and error accumulation, which severely restricts the reliability of overload capacity assessment and fault early warning.

[0003] The current technology mainly faces three major bottlenecks: Rigid sensor layout: Fixed sensor arrays have insufficient spatial resolution, making it impossible to track the dynamic migration of hotspots in real time, and the positioning error exceeds 15% under sudden operating conditions; Insufficient signal interference suppression: Oil flow turbulence and power grid harmonics cause time-delay fluctuations in temperature signals, and traditional filtering algorithms have difficulty distinguishing between effective signals and high-frequency noise; Poor adaptability to operating conditions: The static thermal parameter model does not take into account dynamic factors such as oil quality deterioration and harmonic heating, resulting in a mismatch between the compensation model and the actual heat transfer process, and a significant decrease in the monitoring accuracy of aging equipment.

[0004] To address the aforementioned problems, this invention proposes a novel method integrating dynamic sensing, intelligent denoising, and physical-driven neural networks. This method optimizes spatial coverage through a reconfigurable sensor array, combines time-frequency domain joint denoising to suppress complex interference, and introduces time-varying thermal parameters and harmonic loss models to construct an adaptive compensation framework. This approach overcomes the monitoring inaccuracies of traditional methods under dynamic operating conditions, providing high-precision data support for intelligent transformer operation and maintenance. Summary of the Invention

[0005] The purpose of this invention is to solve existing problems and to propose a method and system for real-time monitoring of hot spot temperature in oil-immersed transformers.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for real-time monitoring of hot spot temperature in an oil-immersed transformer includes the following steps: T1, Multimodal sensing fusion and oil flow field reconstruction: Through dynamic sensor array deployment and three-dimensional oil flow field modeling, the winding temperature field distribution is captured in real time, and the hot spot tracking capability is optimized; T2, adaptive Kalman filter-wavelet packet joint denoising, combined with time-frequency domain joint denoising algorithm, effectively eliminates temperature signal distortion caused by oil flow disturbance and electromagnetic interference; T3, physical information neural network dynamic compensation, embedding time-varying thermal parameters and harmonic loss models, achieves high-precision temperature compensation through physical constraint neural network, adapting to complex working conditions.

[0007] Furthermore, the specific process of T1 is as follows: Deploy a multi-parameter embedded sensor array and install composite sensor modules in key parts of the transformer winding. Each module integrates: a miniature fiber Bragg grating temperature sensor, a MEMS miniature eddy current sensor, and a piezoelectric vibration sensor. The sensor array is arranged in a honeycomb topology, with spacing that satisfies the Reynolds number similarity criterion. , Let Reynolds number be 1. These are oil density, flow rate, and characteristic length, respectively. Dynamic viscosity; A three-dimensional reconstruction model of the oil flow field was constructed, and the oil flow velocity vector was obtained through an eddy current sensor. Discrete orthogonal decomposition was used to extract the principal modes of oil flow, and a reduced-order model was established: ,in Here, k represents the time-dependent modal coefficients, and k is the number of retained modes. Characteristic basis functions are used; vibration sensor data are fused to calculate turbulence intensity. Used to quantify the level of disturbance. For the speed standard deviation, The average flow velocity; Dynamically correlate oil flow and temperature field, and establish oil flow-temperature coupling equation: ,in Where is the thermal diffusivity, As the heat source of the winding, Let T be the oil flow velocity vector, T be the temperature field, and t be time. The temperature field correction coefficient matrix is ​​output in real time using the finite volume method (FVM).

[0008] Furthermore, the positioning process for key components of the transformer winding in T1 is as follows: Define a key index, which is calculated by weighting three types of parameters: Current correlation : The correlation coefficient between the temperature rise at a certain point in the winding and the change in load current; gradient strength The percentage of the system's temperature rise range relative to the maximum temperature difference between this point and the six adjacent measuring points in the same direction. Heat dissipation inertia The ratio of the time required for the temperature to drop to 63.2% of its initial value after a power outage to the oil circulation cycle; Key Judgment Logic: , For key scores; when At that time, it was determined to be a level one critical point. It was determined to be a secondary critical point; The winding is divided into cubic units with a side length of 5cm, and a three-dimensional coordinate system is established: Axial dimension Z-axis: divided into 10 equal parts according to the winding height; Radial dimension R-axis: divided into 4 concentric rings from the inside to the outside; Circumferential dimension θ-axis: divided into a sector every 45°; Establish a deployment priority scoring system; deployment rules are as follows: Units containing primary key points must be deployed; each sector must have at least two monitoring points deployed; and the spacing between adjacent sensors must meet the following requirements. ,in These are oil flow rate and system response time, respectively. Reconstruct a four-dimensional spatiotemporal feature matrix; a dynamic weight adjustment algorithm performs the following operations every 15 minutes: Calculate the rate of change of thermal state for each unit. ,in The rate of change of thermal state, The current temperature. The temperature is the value at the previous moment, and t is the time; update deployment priority: ,in The initial deployment priority is set; when the priority difference exceeds 20%, sensor position adjustment is triggered.

[0009] Furthermore, the specific operation steps of T2 are as follows: Hybrid denoising algorithm, primary filtering through the original temperature signal Perform adaptive Kalman filtering: Equations of state: ; Observation equation: ;in Let the temperature be the state temperature at time k. Let A be the observation value at time k, B be the state transition matrix, B be the control input matrix, and H be the observation matrix. These are process noise and observation noise, respectively. Process noise covariance and observation noise covariance Based on oil flow turbulence intensity Dynamic adjustment: ; Secondary denoising: Denoising the Kalman filter residuals Processing is carried out, among which For residual signals, These are the original temperature signal and the temperature after Kalman filtering, respectively. Online performance evaluation, defining the degree of signal-to-noise ratio improvement ,in These are noise energy and denoised signal energy, respectively. When necessary, it automatically switches to the backup empirical mode decomposition algorithm.

[0010] Furthermore, the specific steps for processing the Kalman filter residuals in T2 are as follows: The residual signal after Kalman filtering is processed by frame segmentation: frame length L = 256 sampling points; frame shift S = 64 sampling points; Hanning window function is applied: , ; Short-time Fourier transform analysis, spectral feature extraction, and calculation of the STFT for each frame of the signal: , ,in The result of STFT transformation. The Kalman filter residual signal sequence, Here, L is the Hanning window function, L is the frame length (256 sampling points), m is the frame number, and c is the frequency index. Construct a noise feature map and perform periodic detection: Calculate the autocorrelation function for frequencies above 500Hz. When the autocorrelation function value exist When a peak value appears, it is determined that there is power grid harmonic interference; among which For time delay; Frequency-time joint threshold design: dynamic threshold adjustment strategy, frequency domain weight calculation: for frequency bands where strong noise is detected. Calculate the noise energy percentage: ; Threshold reinforcement rule: when When the wavelet packet threshold multiplication factor corresponding to this frequency band is set to 1.5; when When the multiplicative factor is set to 1.2, the standard threshold is maintained in other cases. ; For the threshold, Where N is the standard deviation of the signal, and N is the number of samples. Wavelet packet decomposition and joint denoising: adaptive wavelet packet processing, frequency band mapping: establishing the correspondence between wavelet packet tree nodes and STFT frequency bands; Frequency division threshold processing: An improved threshold is applied to node 15. , The adjusted threshold For frequency band weighting coefficients, The number of samples for wavelet packet node 15 is given, while the standard threshold is maintained for node 14. Signal reconstruction: The denoised signal is reconstructed using inverse wavelet packet transform and then superimposed with the Kalman filter result. ; Online performance evaluation, defining the degree of signal-to-noise ratio improvement ,in These are noise energy and denoised signal energy, respectively. When necessary, it automatically switches to the backup empirical mode decomposition algorithm.

[0011] Furthermore, the specific operation steps of T3 are as follows: Construct a PINN compensation model. Network structure: adopts a residual network framework. The input layer includes: denoised temperature... Oil flow velocity and turbulence intensity Load current and ambient temperature It also includes monitoring data; The output layer shows the compensated hotspot temperature. ; Physical constraint embedding: Adding partial differential equation residual terms to the loss function: Gradient terms are calculated through automatic differentiation. ;in These are the predicted temperature and the actual temperature, respectively. The physical residual weighting coefficient. The thermal diffusivity is a time-varying coefficient. It is a time-varying heat source; Online incremental learning: When oil quality deterioration is detected, i.e., the dielectric loss factor tanδ > 0.03, the model update mechanism is triggered; The Fisher information matrix F is retained by using an elastic weight solidification algorithm to preserve the important parameter θ: ; Distributed updates of model parameters are achieved using edge computing nodes, where, These are the updated loss function and the original loss function, respectively. These are the weighting coefficients. The value of the Fisher information matrix corresponding to the i-th parameter is... These are the current updated value and the old value of the i-th parameter, respectively.

[0012] Furthermore, the monitoring data in T3 includes: Oil dielectric loss value tanδ: obtained through an online oil chromatography monitoring device, reflecting the aging state of the insulating oil; Harmonic current components: extracted from the power quality analysis module to quantify the nonlinear distortion of the load current; Top oil temperature T oil The oil circulation heat dissipation efficiency is characterized by real-time data acquisition using distributed fiber optic sensors. Dynamic parameter modeling: The time-varying thermal diffusivity is dynamically adjusted based on oil dielectric loss and oil temperature. ,in These are the initial calibration values; the correction terms reflect the effects of oil aging and temperature on heat transfer. This is the oil dielectric loss value; Harmonic heating source correction: Harmonic component contributions are added to the Joule heating model. ,in As the total heat source, These are the fundamental current and the winding resistance, respectively. For the nth harmonic current, This is the harmonic loss coefficient.

[0013] A real-time monitoring system for hot spot temperature of an oil-immersed transformer includes: The multimodal sensing and oil flow field reconstruction module is used to deploy a composite sensor array, dynamically optimize the sensor layout based on key indices, construct a three-dimensional oil flow field model in real time, and correlate oil flow with temperature field distribution. An adaptive signal denoising module is used to fuse Kalman filtering with time-frequency domain joint denoising, dynamically adjust the threshold strategy, and suppress temperature signal noise caused by oil flow disturbances and power grid harmonics. The dynamic parameter modeling module is used to calculate the time-varying heat diffusion coefficient and harmonic heat source in real time, and embeds the oil flow-temperature coupling equation. The physical information neural network compensation module is used to construct the PINN model with a residual network architecture. It takes multi-source data as input and outputs the compensated hotspot temperature through physical constraint residual terms. The online learning and calibration module is used to dynamically update model parameters and sensor layout based on the elastic weight solidification and federated learning framework, supporting real-time adaptation to oil quality deterioration or harmonic abrupt changes.

[0014] Compared with the prior art, the beneficial effects of the present invention are: (1) Dynamic layout and precise positioning of sensors: By defining key indices and weighting the calculation of three types of parameters, namely current correlation, gradient intensity and heat dissipation inertia, the key parts of the transformer winding are precisely located, and a deployment priority scoring system is established accordingly, realizing the dynamic optimization layout of the sensor array. This layout method solves the problems of rigid sensor layout and insufficient spatial resolution in traditional monitoring methods. It can track the dynamic migration of hotspots in real time, greatly improving the accuracy of hotspot location identification and the real-time performance of monitoring, and significantly reducing the positioning error under sudden working conditions. (2) Time-frequency domain joint denoising and signal purification: An adaptive Kalman filter-wavelet packet joint denoising algorithm was used. First, the original temperature signal was filtered by Kalman filter. Then, the residual signal was analyzed by short-time Fourier transform to construct a noise feature map. Wavelet packet decomposition and denoising were performed in combination with a dynamic threshold adjustment strategy. This effectively distinguished the effective signal from the noise and suppressed the temperature signal noise caused by oil flow disturbance and power grid harmonics. This overcame the shortcomings of traditional filtering algorithms in effectively separating the effective signal from high-frequency noise, significantly improved the accuracy and reliability of the signal, and provided a purer and more realistic signal basis for subsequent temperature monitoring and analysis, thereby improving the overall performance and accuracy of the monitoring system. (3) Adaptive dynamic compensation and intelligent adaptation to operating conditions: By using the Physical Information Neural Network (PINN) compensation model, time-varying thermal parameter modeling and harmonic loss quantification are integrated into it. By embedding physical constraints, high-precision dynamic compensation of monitoring data is achieved. At the same time, the online incremental learning mechanism can update the model parameters in real time according to oil quality deterioration and operating condition changes, giving the monitoring system a strong adaptive capability. It can effectively cope with complex operating condition changes such as oil quality deterioration and harmonic heating, and solve the problems of compensation model mismatch and decreased monitoring accuracy that occur in traditional monitoring methods when equipment ages and operating conditions change. This extends the service life of the equipment, enhances the stability and safety of power system operation, and provides more accurate and reliable data support for the intelligent operation and maintenance of transformers. Attached Figure Description

[0015] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0018] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0019] like Figure 1 As shown, a method for real-time monitoring of hot spot temperature in an oil-immersed transformer includes the following steps: Step 1: Multimodal sensor fusion and oil flow field reconstruction. Through dynamic sensor array deployment and three-dimensional oil flow field modeling, the winding temperature field distribution is captured in real time, and the hot spot tracking capability is optimized. A multi-parameter embedded sensor array is deployed, and composite sensor modules are installed at key locations in the transformer windings. Each module integrates: a miniature fiber optic Bragg grating (FBG) temperature sensor (accuracy ±0.1℃), a MEMS miniature eddy current sensor (measuring oil flow velocity, range 0~2m / s), and a piezoelectric vibration sensor (monitoring oil flow turbulence intensity, frequency response range 10Hz~1kHz). The positioning process for the key locations in the transformer windings is as follows: Define a key index, which is calculated by weighting three types of parameters: Current correlation : Correlation coefficient between temperature rise at a certain point in the winding and change in load current (calculation window ≥ 30 minutes); Gradient strength The percentage of the system's temperature rise range relative to the maximum temperature difference between this point and the six adjacent measuring points; heat dissipation inertia. The ratio of the time required for the temperature to drop to 63.2% of its initial value after a power outage to the oil circulation cycle; Key Judgment Logic: , For critical scoring, when At that time, it was determined to be a level one critical point. It was determined to be a secondary critical point; The winding is divided into cubic units with a side length of 5cm, and a three-dimensional coordinate system is established: Axial dimension (Z-axis): divided into 10 equal parts according to the winding height; Radial dimension (R-axis): divided into 4 concentric rings from the inside to the outside; Circumferential dimension (θ-axis): divided into a sector every 45°. Establish a deployment priority scoring system, refer to Table 1:

[0020] The deployment rules are as follows: Units containing primary key points must be deployed; each sector must have at least two monitoring points deployed; and the spacing between adjacent sensors must meet the following requirements. ,in These are oil flow rate and system response time, respectively. Reconstruct the four-dimensional spatiotemporal feature matrix, as shown in Table 2:

[0021] The dynamic weight adjustment algorithm performs the following calculation every 15 minutes: calculate the rate of change of thermal state for each unit. ,in The rate of change of thermal state, The current temperature. The temperature is the value at the previous moment, and t is the time; update deployment priority: ,in The initial deployment priority is set; when the priority difference exceeds 20%, sensor position adjustment is triggered.

[0022] The sensor array is arranged in a honeycomb topology, with spacing that satisfies the Reynolds number similarity criterion. , Let Reynolds number be 1. These are oil density, flow rate, and characteristic length, respectively. Dynamic viscosity; A three-dimensional reconstruction model of the oil flow field was constructed, and the oil flow velocity vector was obtained through an eddy current sensor. Discrete orthogonal decomposition (POD) was used to extract the principal modes of oil flow, and a reduced-order model was established. ,in Here, k represents the time-dependent modal coefficients, and k is the number of retained modes. Characteristic basis functions are used; vibration sensor data are fused to calculate turbulence intensity. Used to quantify the level of disturbance. For the speed standard deviation, The average flow velocity; Dynamically correlate oil flow and temperature field, and establish oil flow-temperature coupling equation: ,in Where is the thermal diffusivity, As the heat source of the winding, Let T be the oil flow velocity vector, t be the temperature field, and t be time. The temperature field correction coefficient matrix is ​​output in real-time using the finite volume method (FVM). .

[0023] Step 2: Adaptive Kalman filtering-wavelet packet joint denoising, combined with time-frequency domain joint denoising algorithm, effectively eliminates temperature signal distortion caused by oil flow disturbance and electromagnetic interference; Hybrid denoising algorithm, primary filtering through the original temperature signal Perform adaptive Kalman filtering: Equations of state: ; Observation equation: ;in Let the temperature be the state temperature at time k. Let A be the observation value at time k, B be the state transition matrix, B be the control input matrix, and H be the observation matrix. These are process noise and observation noise, respectively; process noise covariance. and observation noise covariance Based on oil flow turbulence intensity Dynamic adjustment: ; Secondary denoising: Denoising the Kalman filter residuals Processing is carried out, among which For residual signals, The original temperature signal and the temperature after Kalman filtering are shown below: The residual signal after Kalman filtering is processed into frames: frame length L = 256 sampling points (corresponding to 5.12ms@50kHz sampling rate); frame shift S = 64 sampling points (overlap rate 75%); Hanning window function is applied. , ; Short-Time Fourier Transform (STFT) Analysis: Spectral Feature Extraction, Calculation of STFT for Each Frame of Signal: , ,in The result of STFT transformation. The Kalman filter residual signal sequence, The Hanning window function is used, where L is the frame length (256 sampling points), m is the frame number, and c is the frequency index. A noise feature map is constructed, and periodic detection is performed: the autocorrelation function is calculated for frequency bands above 500Hz (k≥26). When the autocorrelation function value exist When a peak value appears at (corresponding to a 100Hz interval), it is determined that there is power grid harmonic interference; among which... For time delay; Frequency-time joint threshold design: dynamic threshold adjustment strategy, frequency domain weight calculation: for frequency bands where strong noise is detected. Calculate the noise energy percentage: Threshold reinforcement rule: when When the wavelet packet threshold multiplication factor corresponding to this frequency band is set to 1.5; when When the multiplicative factor is set to 1.2, the standard threshold is maintained in other cases. ; For the threshold, Where N is the standard deviation of the signal, and N is the number of samples. Wavelet packet decomposition and joint denoising: Adaptive wavelet packet processing, frequency band mapping: Establishing the correspondence between wavelet packet tree nodes and STFT frequency bands, refer to Table 3:

[0024] Frequency band threshold processing: An improved threshold is applied to node 15 (high-frequency band). , The adjusted threshold For frequency band weighting coefficients, The number of samples for wavelet packet node 15 is given, while the standard threshold is maintained for node 14 (low-frequency band); Signal reconstruction: The denoised signal is reconstructed using inverse wavelet packet transform and then superimposed with the Kalman filter result. .

[0025] Online performance evaluation, defining the degree of signal-to-noise ratio improvement ,in These are noise energy and denoised signal energy, respectively. When necessary, it automatically switches to the backup Empirical Mode Decomposition (EMD) algorithm.

[0026] Step 3: Dynamic compensation using physical information neural network. By embedding time-varying thermal parameters and harmonic loss models, high-precision temperature compensation is achieved through physical constraint neural network to adapt to complex working conditions. Construct a PINN compensation model. Network structure: Employ the ResNet framework. Input layer includes: denoised temperature... Oil flow velocity and turbulence intensity Load current and ambient temperature It also includes monitoring data: Oil dielectric loss value (tanδ): obtained through an online oil chromatography monitoring device, reflecting the aging state of the insulating oil; Harmonic current components (2nd to 5th order): extracted from the power quality analysis module, quantifying the nonlinear distortion of the load current; Top oil temperature (T oil The oil circulation heat dissipation efficiency is characterized by real-time data acquisition using distributed fiber optic sensors; the output layer shows the compensated hotspot temperature. ; Dynamic parameter modeling: The time-varying thermal diffusivity is dynamically adjusted based on oil dielectric loss and oil temperature. ,in These are the initial calibration values; the correction terms reflect the effects of oil aging and temperature on heat transfer. For oil dielectric loss; harmonic heating source correction, harmonic component contributions are added to the Joule heating model: ,in As the total heat source, These are the fundamental current and the winding resistance, respectively. For the nth harmonic current, The harmonic loss coefficient is determined experimentally (typical value 0.1~0.2).

[0027] Physical constraint embedding: Adding partial differential equation residual terms to the loss function: Gradient terms are calculated through automatic differentiation. ;in These are the predicted temperature and the actual temperature, respectively. The physical residual weighting coefficient. The thermal diffusivity is a time-varying coefficient. It is a time-varying heat source; Online incremental learning: When oil quality deterioration is detected (substance loss factor tanδ > 0.03), a model update mechanism is triggered; the Elastic Weight Consolidation (EWC) algorithm is used to retain the Fisher information matrix F of the important parameter θ. ; Distributed updates of model parameters are achieved using edge computing nodes, where, These are the updated loss function and the original loss function, respectively. These are the weighting coefficients. The value of the Fisher information matrix corresponding to the i-th parameter is... These are the current updated value and the old value of the i-th parameter, respectively.

[0028] A real-time monitoring system for hot spot temperature of an oil-immersed transformer includes a multimodal sensing and oil flow field reconstruction module, an adaptive signal denoising module, a dynamic parameter modeling module, a physical information neural network compensation module, and an online learning and calibration module. The multimodal sensing and oil flow field reconstruction module is used to deploy a composite sensor array (temperature, flow velocity, vibration), dynamically optimize the sensor layout based on key indices, construct a three-dimensional oil flow field model in real time, and correlate the oil flow and temperature field distribution. The adaptive signal denoising module is used to fuse Kalman filtering and time-frequency domain joint denoising (STFT + wavelet packet), dynamically adjust the threshold strategy, suppress temperature signal noise caused by oil flow disturbance and power grid harmonics, and improve the signal-to-noise ratio by ≥15dB. The dynamic parameter modeling module is used to calculate the time-varying heat diffusion coefficient (oil dielectric loss / oil temperature correction) and harmonic heat source (fundamental wave + harmonic loss) in real time, and embeds the oil flow-temperature coupling equation to solve the problems of aging and operating condition mismatch. The physical information neural network compensation module is used to construct the PINN model with ResNet architecture. It takes multi-source data (temperature, oil flow, harmonics) as input and outputs the compensated hot spot temperature through physical constraint residual terms. The online learning and calibration module is used to dynamically update model parameters and sensor layout based on the Elastic Weights Consolidation (EWC) and federated learning framework, supporting real-time adaptation to oil quality deterioration / harmonic abrupt changes, with a response latency of <50ms.

[0029] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for real-time monitoring of hot-spot temperature of oil-immersed transformer, characterized in that, The method comprises the following steps: T1, multi-modal sensor fusion and oil flow field reconstruction, through dynamic sensor array deployment and three-dimensional oil flow field modeling, real-time capture of winding temperature field distribution, optimization of hotspot tracking capability; T2, adaptive Kalman filter-wavelet packet joint denoising, combined with time-frequency domain joint denoising algorithm, effectively eliminating temperature signal distortion caused by oil flow disturbance and electromagnetic interference; T3, physical information neural network dynamic compensation, embedding time-varying thermal parameters and harmonic loss model, realizing high-precision temperature compensation through physical constraint neural network, adapting to complex working conditions.

2. The method according to claim 1, wherein, The specific process of T1 is as follows: Deploy a multi-parameter embedded sensor array, install a composite sensor module at the key position of the transformer winding, and each module integrates: a miniature fiber Bragg grating temperature sensor, a MEMS miniature eddy current sensor, and a piezoelectric vibration sensor; The sensor array is arranged in a honeycomb topology with a spacing that satisfies a Reynolds number similarity criterion: , is the Reynolds number, are the oil density, flow velocity, and characteristic length, respectively, is the dynamic viscosity; A three-dimensional reconstruction model of oil flow field is constructed, and the oil flow velocity vector is obtained by eddy current sensor The main mode of oil flow is extracted by discrete orthogonal decomposition, and a reduced order model is established: Where is the time-dependent modal coefficient, k is the number of retained modes, is the characteristic basis function; Fusion of vibration sensor data, calculation of turbulence intensity , used to quantify the disturbance level, is the standard deviation of velocity, is the average flow velocity; Dynamic correlation of oil flow and temperature field, the oil flow-temperature coupling equation is established: Wherein is the thermal diffusion coefficient, is the winding heat source, is the oil flow velocity vector, T is the temperature field, and t is the time; Real-time solution by finite volume method (FVM), output temperature field correction coefficient matrix.

3. The method according to claim 2, wherein, The positioning process of the key position of the transformer winding in T1 is as follows: Define the criticality index, which is calculated by weighting three types of parameters: Current correlation degree : correlation coefficient of temperature rise of a point of winding and change of load current; Gradient strength : The percentage of the maximum difference between the temperature of the point and the temperature of the six adjacent points in the system temperature range. Heat dissipation inertia : ratio of time required for temperature to drop to 63.2% of initial value after power-off to oil circulation period; Criticality determination logic: , is the criticality score; When a primary key point, a secondary key point; Divide the winding into cubic units with a side length of 5 cm, and establish a three-dimensional coordinate system: Axial dimension Z axis: divide 10 equal parts according to the height of the winding; radial dimension R axis: divide 4 concentric rings from inside to outside; circumferential dimension θ axis: divide a sector every 45°; A deployment priority scoring system is established; the deployment rules are as follows: the unit where the first-class key point is located must be deployed, at least 2 monitoring points are deployed in each sector, and the spacing between adjacent sensors meets wherein are the oil flow rate and the system response time, respectively; Then construct a four-dimensional space-time feature matrix; dynamic weight adjustment algorithm, execute the following operations every 15 minutes: Computing the rate of change of thermal state of each cell wherein is the rate of change of thermal state, is the current temperature, is the temperature at the previous time instant, t is time; updating the deployment priority: wherein is the initial deployment priority; when the priority difference exceeds 20%, triggering sensor position adjustment.

4. The method of claim 1, wherein the method comprises: The specific operation steps of T2 are as follows: The mixed denoising algorithm, the primary filtering is performed by adaptive Kalman filtering on the original temperature signal ​ Equation of state: ; Observation equation: ; where is the state temperature at time k, is the observation at time k, A is the state transition matrix, B is the control input matrix, H is the observation matrix, are the process noise and observation noise, respectively; Process noise covariance and observation noise covariance According to the turbulence intensity of the oil flow Dynamic adjustment: ; Secondary denoising: on Kalman filter residuals is processed, wherein is the residual signal, are the original temperature signal and the Kalman filtered temperature, respectively; Online performance evaluation, defining signal-to-noise ratio improvement wherein Enoise and Edenoise are the noise energy and the denoised signal energy, respectively, when an automatic switch to a backup empirical mode decomposition algorithm is made.

5. The method of claim 4, wherein the method further comprises: The specific operation steps of the Kalman filter residual error processing in T2 are as follows: The residual signal after Kalman filtering is processed by frame: frame length L = 256 sampling points; frame shift S = 64 sampling points; Hanning window function is applied: , ; Short-time Fourier transform analysis, spectral feature extraction, STFT of each frame signal is calculated: , where is the STFT transform result, is the Kalman filter residual error signal sequence, is the Hanning window function, L is the frame length (256 sampling points), m is the frame number, and c is the frequency index. Constructing noise feature map, periodic detection: calculate autocorrelation function for frequency band above 500Hz: When autocorrelation function value Peaks appear at , it is determined that there is power grid harmonic interference; wherein is time delay; Frequency-time domain joint threshold design: dynamic threshold adjustment strategy, frequency domain weight calculation: for the frequency band with strong noise detected , calculate the proportion of noise energy: ; Threshold enhancement rule: when , the wavelet packet sub-band threshold multiplicative factor corresponding to the frequency band is set to 1.5; when , the multiplicative factor is set to 1.2; in other cases, the standard threshold is kept; , where is the signal standard deviation, and N is the sample number; Wavelet packet decomposition and joint denoising: adaptive wavelet packet processing, frequency band mapping: establish the correspondence between wavelet packet tree nodes and STFT frequency bands; Sub-band thresholding: improved thresholding is used for node 15: , is the adjusted threshold, is the sub-band weight coefficient, is the number of samples of the wavelet packet node 15, and the standard threshold is maintained for node 14; Signal reconstruction: the de-noised signal is reconstructed by inverse wavelet packet transform and superimposed with the Kalman filtering result: ; Online performance evaluation, defining signal-to-noise ratio improvement wherein Enoise and Edenoise are the noise energy and the denoised signal energy, respectively, when an automatic switch to a backup empirical mode decomposition algorithm is made.

6. The method of claim 1, wherein the method further comprises: The specific operation steps of T3 are as follows: Construct PINN compensation model, network structure: adopt residual network framework, input layer contains: denoising temperature , oil flow speed And turbulence intensity , load current And ambient temperature ; Also includes monitoring data; The output layer is the compensated hot spot temperature ; Physical constraint embedding: adding partial differential equation residual term in loss function , compute gradient term by automatic differentiation ; wherein are predicted temperature and real temperature, respectively, is a physical residual weight coefficient, is a time-varying thermal diffusion coefficient; is a time-varying heat source; Online incremental learning: when oil degradation is detected, i.e. dielectric loss factor tanδ>0.03, trigger the model update mechanism; An elastic weight consolidation algorithm is used to preserve the Fisher information matrix F of the important parameters θ: ; The distributed update of model parameters is realized by using an edge computing node, wherein, are respectively an updated loss function and an original loss function, is a weight coefficient, is a Fisher information matrix value corresponding to the i th parameter, are respectively a current update value of the i th parameter and an old value of the i th parameter.

7. The method and system for real-time monitoring of hot-spot temperature of oil-immersed transformer according to claim 6, characterized in that, The monitoring data in T3 includes: Oil dielectric loss value tanδ: obtained by an oil chromatography online monitoring device, reflecting the aging state of the insulating oil; Harmonic current component: extracted from the power quality analysis module, quantifying the nonlinear distortion of the load current; Top oil temperature T oil : Real-time acquisition by distributed optical fiber sensor, representing oil circulation heat dissipation efficiency; Dynamic parameter modeling, time-varying thermal diffusivity coefficient is dynamically adjusted according to the oil dielectric loss value and oil temperature, wherein is an initial calibration value, and the correction term reflects the influence of oil aging and temperature on heat transfer, is the oil dielectric loss value; Harmonic heat source correction, the harmonic component contribution is added in the joule heat model: where is the total heat source, is the fundamental current and winding resistance, respectively, is the n-th harmonic current, is the harmonic loss coefficient.

8. The system for real-time monitoring of hot-spot temperature of an oil-immersed transformer according to any one of claims 1 to 7, characterized in that, It includes: Multi-modal sensing and oil flow field reconstruction module, for deploying a composite sensor array, dynamically optimizing the sensor layout based on the criticality index, real-time constructing a three-dimensional oil flow field model, and correlating the oil flow and temperature field distribution; Adaptive signal denoising module, for fusing Kalman filter and time-frequency domain joint denoising, dynamically adjusting threshold strategy, and suppressing temperature signal noise caused by oil flow disturbance and grid harmonics; Dynamic parameter modeling module, for real-time calculation of time-varying thermal diffusion coefficient and harmonic heat source, embedding oil flow-temperature coupling equation; Physical information neural network compensation module, for constructing a PINN model with a residual network architecture, inputting multi-source data, and outputting compensated hotspot temperature through physical constraint residual term; Online learning and calibration module, for dynamically updating model parameters and sensor layout based on elastic weight solidification and federated learning framework, supporting real-time adaptation to oil degradation or harmonic mutation.

Images