An ai-driven transformer load prediction and fault early warning monitoring system and method

Abstract

The application discloses an AI-driven transformer load prediction and fault early warning monitoring system and method, relates to the technical field of monitoring and analysis, and comprises the following steps: in a direct current bias magnetic coupling diagnosis process, based on neutral point direct current data, magnetic flux offset data of a core working point is generated by mapping, the magnetic flux offset data and current load current amplitude data are input into a preset reference response neural network model, and then theoretical reference vibration data under a current working condition are output; a difference between body surface measured vibration data and the theoretical reference vibration data is calculated, and vibration residual factor data are generated; the vibration residual factor data are compared with a preset structure sensitivity threshold value, if the vibration residual factor data are greater than the preset structure sensitivity threshold value, a mechanical looseness early warning signal is generated and output. The application has the effect of improving fault early warning monitoring efficiency.

Description

Technical Field

[0001] This application relates to the field of monitoring and analysis technology, and in particular to an AI-driven transformer load prediction and fault early warning monitoring system and method. Background Technology

[0002] As power systems evolve towards a source-grid-load-storage interactive model, distribution transformers face increasingly complex operating environments. In particular, the bidirectional power flow and nonlinear load impacts brought about by distributed energy access and electric vehicle-grid interaction technologies cause transformers to operate under conditions of dynamic electromagnetic stress changes for extended periods. While existing transformer condition monitoring technologies can monitor conventional electrical and thermal indicators such as oil temperature and load factor, significant technical blind spots remain in the field of mechanical condition monitoring, especially in identifying latent mechanical faults such as loose core clamps and winding deformation. Traditional vibration monitoring methods often struggle to adapt to complex grid background noise, leading to a significant decrease in the accuracy and robustness of monitoring systems under non-standard operating conditions, failing to meet the requirements of refined equipment operation and maintenance in new power systems.

[0003] In actual operation, the operation of geomagnetically induced currents or the monopolar loop operation of DC transmission systems often causes DC components to intrude into the transformer neutral point, inducing DC bias. DC bias causes the transformer core operating point to shift and enter a half-wave saturation state, resulting in a significant enhancement of the core's magnetostrictive effect, which in turn leads to a significant increase in the transformer's vibration amplitude and noise level. Existing monitoring technologies typically rely on a single vibration threshold criterion for alarms, lacking the ability to decouple and analyze the vibration generation mechanism. This means the system cannot distinguish whether the detected high vibration is a normal physical response caused by DC bias or a pathological response caused by loose internal mechanical structures. This confusion leads to false alarms from the maintenance system when DC bias occurs, or to overlook the actual potential for loose mechanical fasteners, increasing the risk of the transformer operating with defects.

[0004] Furthermore, when a transformer operates in reverse power flow mode, its internal leakage magnetic field distribution and stress conditions differ from those in the traditional forward power supply mode. Most existing fault diagnosis models are based on the assumption of unidirectional power flow, lacking consideration of the combined field effect of bidirectional power flow and DC bias magnetization. Under these combined operating conditions, relying solely on electrical quantity monitoring is insufficient to detect minute changes in the mechanical structure, while simple acoustic and vibration monitoring lacks an electrical reference, making it impossible to establish a dynamic health baseline. This separation between electrical and mechanical characteristic analysis makes it difficult for existing technologies to quantitatively assess the theoretically expected state under specific operating conditions, and to accurately pinpoint early mechanical faults through residual analysis, resulting in significant lag in equipment monitoring.

[0005] To address the aforementioned issues, there is an urgent need in this field for an intelligent monitoring method that can combine real-time power flow direction and DC bias degree to construct a dynamic physical response benchmark, thereby effectively eliminating operating condition interference and accurately identifying abnormal states of the internal mechanical structure of transformers. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides an AI-driven transformer load prediction and fault early warning monitoring system and method.

[0007] In a first aspect, this application provides an AI-driven method for transformer load prediction and fault early warning monitoring, comprising the following steps: Acquire electrical data of the high-voltage side, electrical data of the low-voltage side, DC current data of the neutral point, and measured vibration data of the transformer body surface; Calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data; When the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold, the DC bias coupling diagnostic process is triggered. In the DC bias coupling diagnostic process, based on the neutral point DC current data, the magnetic flux offset data of the core operating point is mapped and generated. The magnetic flux offset data and the current load current amplitude data are input into a preset reference response neural network model, and then the theoretical reference vibration data under the current operating condition is output. The preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. Calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data to generate vibration residual factor data; The vibration residual factor data is compared with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

[0008] Preferably, the step of calculating the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data to generate power flow direction identification data specifically includes: Extract the corresponding high-voltage side voltage phase data and high-voltage side current phase data from the high-voltage side electrical data, and the low-voltage side voltage phase data and low-voltage side current phase data from the low-voltage side electrical data; Calculate the power factor angle data for the high-voltage side and the low-voltage side; Compare the polarity relationship between the high-voltage side power factor angle data and the low-voltage side power factor angle data; If the polarity is opposite and the active power component in the low-voltage side electrical data flows to the node corresponding to the high-voltage side electrical data, the power flow direction identification data is marked as reversed.

[0009] Preferably, the step of mapping and generating magnetic flux offset data of the core operating point based on the neutral point DC current data specifically includes: Retrieve pre-stored transformer core excitation characteristic curve data; The neutral point DC current data is converted into equivalent DC magnetic field strength data; The equivalent DC magnetic field strength data is superimposed on the real-time AC magnetic field data and projected onto the excitation characteristic curve data to calculate the maximum offset peak value of the core magnetic flux density. The maximum offset peak value is used as the magnetic flux offset data.

[0010] Preferably, the construction and training process of the pre-set baseline response neural network model specifically includes: Acquire historical DC bias magnetic test data and historical vibration response data of the transformer during its historical factory testing phase; Construct a deep neural network that includes an input layer, hidden layers, and an output layer; The historical DC bias test data is used as input features, and the historical vibration response data is used as target labels to supervise the training of the deep neural network. The network parameters are optimized using a loss function minimization algorithm until the error between the predicted vibration data output by the model and the target label converges to a preset range, thus obtaining a benchmark response neural network model.

[0011] Preferably, the step of calculating the difference between the measured vibration data of the body surface and the theoretical reference vibration data to generate vibration residual factor data specifically includes: The measured vibration data of the body surface are converted into the time-frequency domain to extract the measured energy spectral density data; Obtain the theoretical energy spectral density data corresponding to the theoretical reference vibration data; Calculate the Euclidean distance between the measured energy spectral density data and the theoretical energy spectral density data; The Euclidean distance is used as vibration residual factor data, which characterizes the abnormal vibration energy components caused by non-magnetostrictive effects.

[0012] Preferably, it also includes processing steps for DC bias conditions: If the vibration residual factor data is less than or equal to the preset structural sensitivity threshold, but the amplitude of the measured vibration data on the body surface is greater than the preset normal operation vibration threshold; A DC bias alarm signal is generated, which includes the current neutral point DC current data, indicating that the transformer's mechanical structure is intact but it is in an abnormal magnetic saturation operating state.

[0013] Preferably, the method also includes a step of introducing no-load loss data for auxiliary verification: When the mechanical loosening warning signal is generated, the real-time input power data and real-time output power data of the transformer are acquired simultaneously. Calculate the difference between the real-time input power data and the real-time output power data to obtain real-time excitation loss data; Determine whether the real-time excitation loss data shows a step-up trend; If so, the mechanical loosening warning signal will be upgraded to an emergency shutdown signal for core insulation fault and output.

[0014] Preferred options also include: When the power flow direction identification data is marked as reversed, the current harmonic distortion rate data of each branch feeder connected to the low-voltage side is obtained; The feeder with the highest current harmonic distortion rate was selected as the dominant injection source for the reverse power flow. The identification information of the dominant injection source is attached to the mechanical loosening warning signal and output together.

[0015] Secondly, this application provides an AI-driven transformer load prediction and fault early warning monitoring system, comprising: The data acquisition module is used to acquire electrical data of the high-voltage side, electrical data of the low-voltage side, DC current data of the neutral point, and measured vibration data of the transformer body surface. The identification generation module is used to calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data. The process triggering module is used to trigger the DC bias coupling diagnostic process when the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold. The mapping module is used to generate magnetic flux offset data of the core operating point based on the neutral point DC current data in the DC bias coupling diagnostic process. The output module is used to input the magnetic flux offset data and the current load current amplitude data into a preset reference response neural network model, and then output the theoretical reference vibration data under the current working condition; the preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. The calculation module is used to calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data, and generate vibration residual factor data; The signal warning module is used to compare the vibration residual factor data with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

[0016] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform any of the above-described AI-driven transformer load prediction and fault early warning monitoring methods.

[0017] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides an AI-driven method for transformer load prediction and fault early warning monitoring. By introducing a DC bias magnetic coupling diagnostic process and constructing a benchmark response neural network model, it effectively separates normal magnetostrictive vibration from pathological vibration caused by abnormal internal mechanical structure. By calculating the vibration residual factor and comparing it with a threshold, it can directly point to non-magnetostrictive mechanical faults. This fundamentally solves the core pain points of the traditional single vibration threshold method under DC bias magnetic conditions, which has a high false alarm rate and cannot identify real mechanical hidden dangers. It greatly improves the pertinence and accuracy of fault early warning. 2. By utilizing a neural network model, it is possible to dynamically generate theoretical benchmark vibration data for the transformer's health status under the current operating conditions based on real-time load current and accurately calculated magnetic flux offset. This enables the monitoring system to detect subtle deviations in the mechanical state under the same electrical conditions. Through the keen capture of residual signals, it achieves early and advanced warnings for latent mechanical faults such as loosening of iron core clamps, changing the situation of lagging fault identification in traditional methods. 3. By linking power flow direction identification with DC bias judgment, the accuracy and timing of the activation of the diagnostic benchmark model are ensured, enabling the monitoring system to maintain high robustness and diagnostic reliability in the complex operating environment of the new power system with source-grid-load-storage interaction. 4. Based on vibration monitoring, this invention further introduces real-time excitation loss data for auxiliary verification. When abnormal vibration residual is detected and accompanied by a step increase in excitation loss, it can accurately determine that the fault has developed from simple mechanical loosening to core insulation failure, thereby triggering an emergency shutdown signal. This vibration-loss dual criterion mechanism effectively prevents the omission of serious hidden dangers. At the same time, for operating conditions with only large vibration but normal loss, only a bias alarm is reported, realizing graded control of fault severity and ensuring the safe operation of the equipment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the AI-driven transformer load prediction and fault early warning monitoring method according to an embodiment of this application.

[0020] Figure 2 This is a system schematic diagram of the AI-driven transformer load prediction and fault early warning monitoring method according to an embodiment of this application. Detailed Implementation

[0021] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0022] Application Overview: In existing technologies, transformer fault monitoring largely relies on single electrical quantity threshold alarms or offline oil chromatography analysis, making it difficult to balance real-time performance with operational condition decoupling capabilities. Traditional methods, when transformers are subjected to DC bias interference, significantly amplify vibration and noise signals, masking subtle characteristics caused by early mechanical faults, leading to high false alarm rates or a significant risk of missed detections. Existing monitoring systems cannot distinguish between normal magnetostriction caused by DC bias and abnormal vibrations caused by mechanical loosening. Especially in complex scenarios with bidirectional power flow and superimposed DC components, a single threshold criterion model will suffer severe logical failures, failing to meet the demands of modern power systems for refined equipment operation and maintenance.

[0023] To address the aforementioned issues, the inventors discovered a nonlinear mapping between the degree of DC bias and the theoretical magnetostrictive vibration energy of a transformer. They then established a magneto-mechanical coupling benchmark model to isolate fault characteristics. During the research, it was found that the neutral point DC current causes a shift in the core magnetic flux density. This shift can be quantified and used to deduce the theoretical vibration benchmark, leading to the proposal of using residual analysis to identify mechanical anomalies. Further data verification, using the vibration residual factor as the core criterion and combining it with power flow information, resulted in a dynamic early warning mechanism for specific operating conditions.

[0024] Specifically, the monitoring system first synchronously collects electrical data from the high and low voltage sides of the transformer, neutral point DC current, and body vibration data. It calculates the power factor angle to determine if reverse power flow exists and triggers a specific diagnostic process based on the DC current amplitude. The system uses an AI model to map the DC component to the core flux offset and inputs it into a neural network to predict the theoretically expected baseline vibration energy under the current operating conditions. When the residual between the measured vibration and the theoretical baseline exceeds the structural sensitivity threshold, the system determines it to be a loose mechanical structure, rather than simply a result of bias magnetization. During the diagnostic process, the system also introduces the no-load loss increment as an auxiliary verification indicator; if it is accompanied by an abnormal increase in loss, the alarm level is upgraded. For cases where the residual does not exceed the standard but the vibration amplitude is large, only a bias magnetization operation prompt is output.

[0025] Compared to existing technologies, traditional methods only monitor the absolute value of vibration and lack a decoupling mechanism for operating conditions, making it impossible to identify the vibration source under DC bias conditions. This solution innovatively integrates multi-dimensional data on electrical flow direction, DC component, and mechanical vibration, achieving precise identification of abnormal features by establishing a magneto-mechanical reference neural network. Unlike existing static threshold alarm models, this solution dynamically generates judgment criteria based on real-time magnetic saturation levels and effectively shields against operating condition interference through residual analysis, significantly improving fault identification accuracy in complex scenarios.

[0026] Through the above technical solutions, this application effectively overcomes the problems of misjudgment and missed judgment of transformer mechanical faults caused by DC bias interference, and improves fault location accuracy while ensuring real-time monitoring. The dynamic benchmark generation mechanism takes into account both the tolerance to normal bias response and the sensitivity to mechanical loosening, and the multi-dimensional data fusion verification function ensures the reliability of the diagnostic results.

[0027] Example 1: This application discloses an AI-driven method for predicting transformer load and monitoring fault early warning.

[0028] Reference Figure 1 An AI-driven method for transformer load prediction and fault early warning monitoring includes the following steps: Acquire electrical data from the high-voltage side, low-voltage side, neutral point DC current, and measured vibration data from the transformer's surface. Specifically, a high-frequency data acquisition card (sampling frequency no less than 10kHz) synchronously connects the voltage transformers and current transformers on both the high-voltage and low-voltage sides to acquire instantaneous waveform data containing high-order harmonic information, i.e., high-voltage and low-voltage side electrical data. Simultaneously, a Hall sensor installed on the transformer's neutral point grounding lead collects neutral point DC current data, which serves as a direct quantitative indicator for subsequent DC bias assessment. Furthermore, a magnetically adsorbed piezoelectric accelerometer is attached to the transformer tank wall to collect measured vibration data of the transformer's surface. This step provides a comprehensive data foundation for subsequent physical field decoupling. Alternatively, vibration data can be acquired using a non-contact laser Doppler vibration meter. The technical advantage of this step is that it achieves synchronous alignment of electrical and mechanical quantities, avoiding the failure of multi-physics correlation analysis due to time asynchrony. For example, in a 500kVA distribution transformer monitoring scenario, the system synchronously acquires 10 sets of full-dimensional data packets per second, ensuring the capture of millisecond-level transient changes.

[0029] Calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data; When the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold, the DC bias coupling diagnostic process is triggered. Specifically, the system monitors the power flow identification data and neutral point DC current data generated in the preceding steps in real time. A high-performance diagnostic process is only initiated when both conditions are met simultaneously—that is, when the transformer is in reverse transmission mode and experiencing DC intrusion. The preset zero-point threshold is typically set to 0.5A or 1A to filter out sensor zero drift and environmental noise. For example, if reverse power flow is detected, but the neutral point DC current is only 0.1A (less than the threshold of 1A), the system determines it as normal reverse power feeding and does not trigger subsequent complex magneto-mechanical decoupling calculations; if the DC current suddenly increases to 3A, the diagnostic process is immediately triggered.

[0030] In the DC bias coupling diagnostic process, based on the neutral point DC current data, the magnetic flux offset data of the core operating point is mapped and generated. The magnetic flux offset data and the current load current amplitude data are input into a preset reference response neural network model, and then the theoretical reference vibration data under the current operating condition is output. The preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. Calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data to generate vibration residual factor data; The vibration residual factor data is compared with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

[0031] For example, the step of calculating the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data to generate power flow direction identification data specifically includes: Extract the corresponding high-voltage side voltage phase data and high-voltage side current phase data from the high-voltage side electrical data, and the low-voltage side voltage phase data and low-voltage side current phase data from the low-voltage side electrical data; Calculate the power factor angle data for the high-voltage side and the low-voltage side; Compare the polarity relationship between the high-voltage side power factor angle data and the low-voltage side power factor angle data; If the polarity is opposite and the active power component in the low-voltage side electrical data flows to the node corresponding to the high-voltage side electrical data, the power flow direction identification data is marked as reversed.

[0032] In one embodiment of the application, during the specific implementation of generating power flow direction identification data, the phase angle characteristics are first analyzed and the power factor angle is calculated based on the synchronously acquired instantaneous waveform data of the high-voltage and low-voltage sides. This step aims to quantify the phase shift characteristics of electrical quantities in the time domain. First, the phase angle of the fundamental component is extracted using a fast Fourier transform or Hilbert transform, and then the power factor angle is calculated using the following formula: ,in, The power factor angle, Reactive power The active power is obtained by integrating instantaneous voltage and current data. The subscript "side" represents the high-voltage side (HV) or low-voltage side (LV), respectively. This calculation generates independent power factor angle data for the high-voltage and low-voltage sides. Then, based on the calculated power factor angle data, polarity correlation analysis and active power flow direction determination are performed to establish a reverse power flow discrimination model. The core discrimination formula is as follows: ,in, This is data indicating the direction of power flow (logical value, 1 represents reverse, 0 represents forward). This is a sign function used to extract the polarity characteristics of the power factor. and These are the power factor angles (in radians) for the high-voltage side and the low-voltage side, respectively. This characterizes the phase coupling relationship between voltage and current. When the product of their polarities is -1, it indicates that one side exhibits power supply characteristics (generating power) and the other side exhibits load characteristics (absorbing power). This is the topological premise for identifying power flow reversal. This represents the measured active power amplitude on the low-voltage side. The transformer transmission efficiency coefficient is taken from the equipment's nameplate data. The total power loss of the transformer under current operating conditions is obtained through real-time estimation of the load factor. The safety margin coefficient for determining the flow direction is set to 1.05 to filter out measurement noise interference. This discrimination logic not only compares the phase polarity, but also physically verifies whether the energy injected on the low-voltage side still has residual energy to be sent back to the high-voltage side after deducting the transformer's own losses. This rigorously marks the power flow direction identification data as reversed, ensuring the physical authenticity of the triggering conditions for the subsequent DC bias coupling diagnostic process.

[0033] For example, the step of mapping and generating magnetic flux offset data of the core operating point based on the neutral point DC current data specifically includes: Retrieve pre-stored transformer core excitation characteristic curve data; The neutral point DC current data is converted into equivalent DC magnetic field strength data; The equivalent DC magnetic field strength data is superimposed on the real-time AC magnetic field data and projected onto the excitation characteristic curve data to calculate the maximum offset peak value of the core magnetic flux density. The maximum offset peak value is used as the magnetic flux offset data.

[0034] In one embodiment of the application, in the specific implementation of generating magnetic flux offset data of the core operating point based on the neutral point DC current data mapping, firstly, a multi-physics correlation foundation of magneto-electric-mechanical is established based on the transformer physical design parameters and material properties. Then, pre-stored transformer core excitation characteristic curve data is called. This curve data is constructed by fitting the measured BH scatter points using a piecewise cubic Hermitian interpolation algorithm, and the function expression is denoted as... ,in, The nonlinear magnetization mapping function accurately characterizes the nonlinear decay characteristics of the permeability of the core material in the saturation and deep saturation regions. Next, a DC bias magnetic field strength conversion calculation is performed, converting the monitored neutral point DC current data into equivalent DC magnetic field strength data acting on the core magnetic circuit. The calculation formula is as follows: ,in, The equivalent DC magnetic field strength, The number of turns in the high-voltage side winding is taken from the transformer design file. This is the neutral point DC current data. The average magnetic circuit length of the iron core. The magnetic circuit topology coefficient, taken as 0.85 for a three-phase three-limb transformer, is used to correct the uneven distribution of magnetic flux between different limbs. This step, based on Ampere's circuital law, transforms the physical quantity of current dimension into an excitation source of magnetic field dimension. Then, the equivalent DC magnetic field strength data is superimposed on the real-time AC magnetic field data for dynamic operating point projection analysis. Nonlinear mapping is performed on the excitation characteristic curve data to calculate the maximum offset peak value of the core magnetic flux density. This process essentially simulates the translation behavior of the core operating point on the BH curve, and the calculation formula is as follows: ,in, This represents the peak magnetic flux density under biased magnetization. The peak AC magnetic field strength during normal rated operation is obtained by inversely calculating from the rated voltage. Based on this, the magnetic flux offset data is further calculated using the following formula: ,in, This is magnetic flux offset data. The rated operating magnetic flux density of the transformer is taken from the nameplate parameter (usually 1.7T). This calculation process quantifies the depth at which the core enters the half-wave saturation region from the linear region due to DC bias. Finally, the maximum offset peak value and the calculated offset are output as the magnetic flux offset data, providing key physical boundary conditions for the subsequent benchmark response neural network model to characterize the degree of magnetic saturation nonlinearity.

[0035] For example, the construction and training process of a pre-defined baseline response neural network model specifically includes: Acquire historical DC bias magnetic test data and historical vibration response data of the transformer during its historical factory testing phase; Construct a deep neural network that includes an input layer, hidden layers, and an output layer; The historical DC bias test data is used as input features, and the historical vibration response data is used as target labels to supervise the training of the deep neural network. The network parameters are optimized using a loss function minimization algorithm until the error between the predicted vibration data output by the model and the target label converges to a preset range, thus obtaining a benchmark response neural network model.

[0036] Specifically, in this embodiment, to ensure that the model predicts the theoretical value of a healthy transformer, the training data must be strictly limited to specific sources: namely, DC bias magnetic withstand test data from when the transformer was first manufactured, or operational data after a major overhaul confirming the mechanical structure is secure. Historical DC bias magnetic test data includes DC current values ​​and load current values ​​at different levels, while historical vibration response data consists of the corresponding measured vibration amplitudes. The constructed deep neural network (such as a BP network with 3 hidden layers) continuously adjusts the weights through a backpropagation algorithm, aiming to minimize the mean square error between the predicted value and the actual healthy value. For example, training stops when the training error is less than 0.01. The purpose of this step is to allow the AI ​​to remember the vibration behavior that a transformer should exhibit when facing DC bias magnetic shock in a healthy state, thus serving as a standard for measuring whether the current vibration is abnormal in subsequent online monitoring.

[0037] For example, the step of calculating the difference between the measured vibration data of the body surface and the theoretical reference vibration data to generate vibration residual factor data specifically includes: The measured vibration data of the body surface are converted into the time-frequency domain to extract the measured energy spectral density data; Obtain the theoretical energy spectral density data corresponding to the theoretical reference vibration data; Calculate the Euclidean distance between the measured energy spectral density data and the theoretical energy spectral density data; The Euclidean distance is used as vibration residual factor data, which characterizes the abnormal vibration energy components caused by non-magnetostrictive effects.

[0038] Specifically, the vibration signal is first converted into energy spectral density data using short-time Fourier transform. The Euclidean distance between the measured energy spectral density data and the theoretical energy spectral density data output by the model is then calculated using the following formula: ,in, and These represent the measured energy spectral density data and the theoretical energy spectral density data, respectively. This distance value is directly used as the vibration residual factor data. The advantage of this technique is that mechanical loosening often generates additional harmonic energy in specific high-frequency bands (such as at 2 or 3 times the fundamental frequency). By calculating the spectral distance, these minute energy distortions caused by non-magnetostrictive effects (i.e., mechanical collisions and friction) can be captured more sensitively, significantly improving the detection rate of early minor loosening faults.

[0039] For example, it also includes processing steps for DC bias conditions: If the vibration residual factor data is less than or equal to the preset structural sensitivity threshold, but the amplitude of the measured vibration data on the body surface is greater than the preset normal operation vibration threshold; A DC bias alarm signal is generated, which includes the current neutral point DC current data, indicating that the transformer's mechanical structure is intact but it is in an abnormal magnetic saturation operating state.

[0040] Specifically, when the transformer vibrates violently, but the aforementioned residual analysis shows that the deviation between the measured and theoretical values ​​is within the allowable range, this indicates that the current severe vibration is entirely in accordance with physical laws and is caused by increased magnetostriction due to DC bias, rather than loose screws. In this case, the system does not report a mechanical fault but generates a DC bias alarm signal. This step effectively avoids the drawback of traditional systems that falsely report transformer faults upon encountering bias.

[0041] For example, it also includes the step of introducing no-load loss data for auxiliary verification: When the mechanical loosening warning signal is generated, the real-time input power data and real-time output power data of the transformer are acquired simultaneously. Calculate the difference between the real-time input power data and the real-time output power data to obtain real-time excitation loss data; Determine whether the real-time excitation loss data shows a step-up trend; If so, the mechanical loosening warning signal will be upgraded to an emergency shutdown signal for core insulation fault and output.

[0042] Specifically, this embodiment introduces thermal parameters as a cross-verification method for mechanical faults. Severe mechanical loosening may lead to wear of the insulating varnish film between the iron core laminations, resulting in multi-point grounding or eddy current circulation, manifested as a surge in iron loss (excitation loss). Once the system has been determined to have mechanical loosening based on vibration residuals, real-time excitation loss data is further calculated. ,in, This is represented as real-time input power data. This is represented as real-time output power data. Expressed as real-time load current, This loss is expressed as the equivalent resistance of the winding. If this loss value experiences a sudden jump (e.g., from 1kW to 3kW) within a short period, it indicates that the loosening has damaged the insulation structure and developed into an electrical fault. At this point, the system no longer merely issues a warning but directly outputs an emergency shutdown signal for a core insulation fault. This step, using a dual criterion of vibration and loss, greatly improves the response speed to severe destructive faults and prevents the transformer from burning out.

[0043] For example, it also includes: When the power flow direction identification data is marked as reversed, the current harmonic distortion rate data of each branch feeder connected to the low-voltage side is obtained; The feeder with the highest current harmonic distortion rate was selected as the dominant injection source for the reverse power flow. The identification information of the dominant injection source is attached to the mechanical loosening warning signal and output together.

[0044] Specifically, when the system detects reverse power flow causing abnormal transformer operation, it analyzes the total harmonic distortion (THD) data of each feeder branch on the low-voltage side. Since power electronic converters often emit high harmonics during inversion, the branch with the highest THD is typically the primary power injection source. The system designates this feeder (e.g., "Charging Pile Feeder No. 3") as the dominant injection source and includes it in the warning signal. The technical advantage of this step is that it not only informs maintenance personnel that the transformer is loose but also identifies the cause of the deterioration, facilitating targeted load management or harmonic mitigation on the grid side.

[0045] Example 2: This application also discloses an AI-driven transformer load prediction and fault early warning monitoring system.

[0046] Reference Figure 2 An AI-driven transformer load prediction and fault early warning monitoring system includes: The data acquisition module is used to acquire electrical data of the high-voltage side, electrical data of the low-voltage side, DC current data of the neutral point, and measured vibration data of the transformer body surface. The identification generation module is used to calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data. The process triggering module is used to trigger the DC bias coupling diagnostic process when the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold. The mapping module is used to generate magnetic flux offset data of the core operating point based on the neutral point DC current data in the DC bias coupling diagnostic process. The output module is used to input the magnetic flux offset data and the current load current amplitude data into a preset reference response neural network model, and then output the theoretical reference vibration data under the current working condition; the preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. The calculation module is used to calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data, and generate vibration residual factor data; The signal warning module is used to compare the vibration residual factor data with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

[0047] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0048] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0049] 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 any specific implementation. 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.

Claims

1. An AI-driven method for transformer load prediction and fault early warning monitoring, characterized in that, Includes the following steps: Acquire electrical data of the high-voltage side, electrical data of the low-voltage side, DC current data of the neutral point, and measured vibration data of the transformer body surface; Calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data; When the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold, the DC bias coupling diagnostic process is triggered. In the DC bias coupling diagnostic process, based on the neutral point DC current data, the magnetic flux offset data of the core operating point is mapped and generated. The magnetic flux offset data and the current load current amplitude data are input into a preset reference response neural network model, and then the theoretical reference vibration data under the current operating condition is output. The preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. Calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data to generate vibration residual factor data; The vibration residual factor data is compared with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

2. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, The step of calculating the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data to generate power flow direction identification data specifically includes: Extract the corresponding high-voltage side voltage phase data and high-voltage side current phase data from the high-voltage side electrical data, and the low-voltage side voltage phase data and low-voltage side current phase data from the low-voltage side electrical data; Calculate the power factor angle data for the high-voltage side and the low-voltage side; Compare the polarity relationship between the high-voltage side power factor angle data and the low-voltage side power factor angle data; If the polarity is opposite and the active power component in the low-voltage side electrical data flows to the node corresponding to the high-voltage side electrical data, the power flow direction identification data is marked as reversed.

3. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, The step of mapping and generating magnetic flux offset data of the core operating point based on the neutral point DC current data specifically includes: Retrieve pre-stored transformer core excitation characteristic curve data; The neutral point DC current data is converted into equivalent DC magnetic field strength data; The equivalent DC magnetic field strength data is superimposed on the real-time AC magnetic field data and projected onto the excitation characteristic curve data to calculate the maximum offset peak value of the core magnetic flux density. The maximum offset peak value is used as the magnetic flux offset data.

4. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, The construction and training process of the pre-defined baseline response neural network model specifically includes: Acquire historical DC bias magnetic test data and historical vibration response data of the transformer during its historical factory testing phase; Construct a deep neural network that includes an input layer, hidden layers, and an output layer; The historical DC bias test data is used as input features, and the historical vibration response data is used as target labels to supervise the training of the deep neural network. The network parameters are optimized using a loss function minimization algorithm until the error between the predicted vibration data output by the model and the target label converges to a preset range, thus obtaining a benchmark response neural network model.

5. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, The step of calculating the difference between the measured vibration data of the body surface and the theoretical reference vibration data, and generating vibration residual factor data, specifically includes: The measured vibration data of the body surface are converted into the time-frequency domain to extract the measured energy spectral density data; Obtain the theoretical energy spectral density data corresponding to the theoretical reference vibration data; Calculate the Euclidean distance between the measured energy spectral density data and the theoretical energy spectral density data; The Euclidean distance is used as vibration residual factor data, which characterizes the abnormal vibration energy components caused by non-magnetostrictive effects.

6. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, It also includes processing steps for DC bias conditions: If the vibration residual factor data is less than or equal to the preset structural sensitivity threshold, but the amplitude of the measured vibration data on the body surface is greater than the preset normal operation vibration threshold; A DC bias alarm signal is generated, which includes the current neutral point DC current data, indicating that the transformer's mechanical structure is intact but it is in an abnormal magnetic saturation operating state.

7. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 1, characterized in that, It also includes the step of introducing no-load loss data for auxiliary verification: When the mechanical loosening warning signal is generated, the real-time input power data and real-time output power data of the transformer are acquired simultaneously. Calculate the difference between the real-time input power data and the real-time output power data to obtain real-time excitation loss data; Determine whether the real-time excitation loss data shows a step-up trend; If so, the mechanical loosening warning signal will be upgraded to an emergency shutdown signal for core insulation fault and output.

8. The AI-driven transformer load prediction and fault early warning monitoring method according to claim 2, characterized in that, Also includes: When the power flow direction identification data is marked as reversed, the current harmonic distortion rate data of each branch feeder connected to the low-voltage side is obtained; The feeder with the highest current harmonic distortion rate was selected as the dominant injection source for the reverse power flow. The identification information of the dominant injection source is attached to the mechanical loosening warning signal and output together.

9. An AI-driven transformer load prediction and fault early warning monitoring system, applied to the AI-driven transformer load prediction and fault early warning monitoring method described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire electrical data of the high-voltage side, electrical data of the low-voltage side, DC current data of the neutral point, and measured vibration data of the transformer body surface. The identification generation module is used to calculate the phase difference based on the high-voltage side electrical data and the low-voltage side electrical data, and generate power flow direction identification data. The process triggering module is used to trigger the DC bias coupling diagnostic process when the power flow direction identification data indicates reverse power flow and the neutral point DC current data is greater than a preset zero-point threshold. The mapping module is used to generate magnetic flux offset data of the core operating point based on the neutral point DC current data in the DC bias coupling diagnostic process. The output module is used to input the magnetic flux offset data and the current load current amplitude data into a preset reference response neural network model, and then output the theoretical reference vibration data under the current working condition; the preset reference response neural network model is generated by training based on the nonlinear mapping relationship between DC bias and magnetostrictive vibration learned by the transformer under a preset healthy mechanical state. The calculation module is used to calculate the difference between the measured vibration data of the body surface and the theoretical reference vibration data, and generate vibration residual factor data; The signal warning module is used to compare the vibration residual factor data with a preset structural sensitivity threshold. If the vibration residual factor data is greater than the preset structural sensitivity threshold, a mechanical loosening warning signal is generated and output.

10. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform an AI-driven transformer load prediction and fault early warning monitoring method as described in any one of claims 1 to 8.

Images