A control method of a high-power isolated power supply

Abstract

The application relates to the technical field of power supply control, and particularly discloses a control method of a high-power isolation power supply, which realizes intelligent management of the whole life cycle of an oil-paper insulation system of the isolation power supply by constructing a complete technical chain from medium loss factor monitoring to PDIV drift inversion, to insulation life prediction and operation control, and by collecting a key insulation performance index of the medium loss factor in real time through an insulation state online monitoring unit, the aging degree and performance change of the oil-paper insulation medium in the operation process can be accurately reflected, compared with a traditional offline detection method, the real-time performance and the accuracy of monitoring are greatly improved. According to the real-time evaluation result of the insulation state, the equipment operation parameters can be dynamically adjusted, sudden failures caused by insulation failure can be effectively avoided, and the safety and reliability of equipment operation are significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power supply control and relates to a control method of a high-power isolated power supply. BACKGROUND

[0002] In plateau regions above 3,000 meters, due to the significant reduction of atmospheric pressure, the partial discharge inception voltage (PDIV) of the oil-paper insulation system decreases by 8%-15% compared with that in plain regions. This phenomenon cannot be detected under the standard atmospheric pressure test conditions of the equipment at the factory, so that there is a potential insulation safety hazard when the equipment is operated in the plateau field. Due to the harsh environmental conditions in the plateau, it is difficult to obtain data from field monitoring, and the data is sparse and has a low signal-to-noise ratio. The operation and maintenance personnel can only rely on the voltage, current, temperature and a small amount of partial discharge pulse data on the secondary side to evaluate the state. When the PDIV gradually approaches or is lower than the system operating voltage, the insulation system will enter a critical discharge state, which is prone to cause short-circuit faults, resulting in equipment damage and power supply interruption. However, the traditional monitoring method based on fixed threshold cannot effectively distinguish between weak partial discharge signals during normal operation of the equipment and abnormal discharge signals in the critical discharge state, resulting in high false positive and false negative rates. At the same time, due to the lack of deep analysis capability for long-term operation data, the existing technology is difficult to establish an effective prediction model to identify the drift trend of PDIV and the insulation aging process in advance. In addition, the incompleteness of field data and the complexity of environmental interference factors make it impossible for the monitoring method based on a single electrical parameter to accurately reflect the actual state of the insulation system and to ensure the safe and stable operation of the high-power isolated power supply in the plateau region. SUMMARY

[0003] In view of the problems existing in the prior art, the application provides a control method of a high-power isolated power supply, which is used to solve the above technical problems.

[0004] In order to achieve the above purpose and other purposes, the technical scheme adopted by the application is as follows:

[0005] The application provides a control method of a high-power isolated power supply, which comprises the following steps:

[0006] Step S1: collecting the dielectric loss factor of the oil-paper insulation medium inside the isolated power supply in real time through an insulation state online monitoring unit to obtain dielectric loss online monitoring data; based on the dielectric loss online monitoring data, combining a dielectric loss factor and partial discharge inception voltage drift correlation curve calibrated in advance through an accelerated aging experiment, performing PDIV dynamic inversion calculation to obtain current insulation PDIV drift data;

[0007] Step S2: based on the PDIV drift data, time sequence coupling processing is performed with historical operating environment parameters and power load curve to obtain insulation performance time sequence degradation data; an insulation life remaining prediction model is established according to the insulation performance time sequence degradation data to obtain an insulation remaining life prediction result;

[0008] Step S3: based on the insulation remaining life prediction result, operating control is performed on the isolation power supply.

[0009] As described above, the control method of the high-power isolation power supply provided by the application has at least the following beneficial effects:

[0010] The application realizes intelligent management of the insulation system of the isolation power supply throughout the life cycle by constructing a complete technical chain from medium loss factor monitoring to PDIV drift inversion, and then to insulation life prediction and operating control. The key insulation performance index of the medium loss factor can be collected in real time by the insulation state online monitoring unit, which can accurately reflect the aging degree and performance change of the oil-paper insulation medium during the operation process. Compared with the traditional offline detection method, the real-time performance and accuracy of the monitoring are greatly improved. Secondly, based on the medium loss factor and PDIV drift correlation curve calibrated in advance through the accelerated aging experiment, PDIV dynamic inversion calculation is performed, which effectively solves the technical problem that the PDIV drop is difficult to be found in the factory test under the plateau environment, so that the insulation performance of the equipment can be accurately predicted before the equipment is operated in the field. Thirdly, by coupling the PDIV drift data with the historical operating environment parameters and the power load curve in time sequence, a multi-dimensional analysis model of insulation performance degradation is established, which fully considers the influence of multiple environmental factors such as temperature, humidity and load fluctuation on insulation aging, so that the insulation life prediction result is more close to the actual operating condition. Finally, based on the insulation remaining life prediction result, operating control is performed on the isolation power supply, which realizes the transformation from passive maintenance to predictive maintenance, can dynamically adjust the equipment operating parameters according to the real-time evaluation result of the insulation state, effectively avoids the sudden failure caused by insulation failure, and significantly improves the safety and reliability of the equipment operation. This closed-loop control strategy based on online monitoring data not only prolongs the service life of the equipment and reduces the maintenance cost, but also provides important technical support for the safe and stable operation of high-power electrical equipment in plateau areas. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed for the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0012] Figure 1A schematic diagram for connecting each step of the method of the present application. DETAILED DESCRIPTION

[0013] The above content will be combined with the embodiments of the present application, which are only examples and illustrations of the concept of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace, as long as they do not deviate from the concept of the present application or exceed the scope defined by the present claims, which shall belong to the protection scope of the present application.

[0014] Embodiment 1, please refer to Figure 1 As shown in the figure, a control method of a high-power isolation power supply, the method comprises:

[0015] Step S1: collecting the dielectric loss factor of the oil-paper insulation medium inside the isolation power supply in real time through the insulation state online monitoring unit to obtain dielectric loss online monitoring data; based on the dielectric loss online monitoring data, combining the dielectric loss factor and the partial discharge inception voltage drift correlation curve calibrated through the accelerated aging experiment in advance, performing PDIV dynamic inversion calculation to obtain the current insulation PDIV drift data;

[0016] Step S1 comprises:

[0017] Step S11: performing millisecond-level current phase angle tracking on the oil-paper insulation medium through the insulation state online monitoring unit to obtain the medium polarization current amplitude data in the power frequency period, and generating the dielectric loss online monitoring data after Fourier fundamental frequency filtering processing;

[0018] Step S12: compensating the dielectric loss online monitoring data for environmental error, using the preset temperature loss gradient matrix and humidity compensation factor to eliminate the dielectric loss measurement deviation, and outputting the corrected dielectric loss characteristic data;

[0019] Step S13: calling the dielectric loss factor and PDIV drift correlation curve library calibrated through the accelerated aging experiment, performing loss offset mapping matching on the corrected dielectric loss characteristic data, and establishing a dynamic quantization matrix of loss factor and voltage drift through the bilinear interpolation operation of adjacent aging nodes;

[0020] Step S14: combining the aging cumulative rate history data of the oil-paper insulation, inputting the dynamic quantization matrix into the preset aging effect time integral model, performing time deconvolution operation on the partial discharge inception voltage based on the dielectric loss change gradient, and outputting the current insulation PDIV drift data.

[0021] In the embodiment of the present application, first, the high-precision current phase angle sensor carried by the insulation state online monitoring unit is used to capture the polarization current waveform of the oil-paper insulation medium under 50Hz power frequency at a fixed sampling frequency, the synchronous phase-locked amplification technology is used to separate the medium polarization current and the capacitive current component, the original current signal is subjected to Hanning window function truncation processing to eliminate frequency spectrum leakage, the base frequency current amplitude phase difference is extracted through fast Fourier transform, and the online monitoring data of the medium loss factor is calculated in combination with the reactive power reference value of the standard capacitor. Subsequently, the monitoring data is corrected based on the temperature-dielectric loss gradient matrix in the environmental compensation database: the oil temperature data of the top layer of the transformer oil tank is collected in real time through the embedded temperature sensor, the temperature compensation coefficient is obtained by indexing the gradient matrix at an interval of 0.1℃; the relative humidity value measured by the dew point sensor is used to call the humidity compensation factor lookup table for secondary compensation; the original dielectric loss value is subtracted by the temperature compensation offset and then multiplied by the humidity correction factor to generate the medium loss characteristic data eliminating environmental interference. Then, in the accelerated aging correlation curve matching stage, the corrected dielectric loss characteristic value is imported into the correlation database containing 200 aging samples, the three-dimensional coordinate projection algorithm is used to determine the time and space position of the dielectric loss value in the aging curve cluster, the four adjacent aging nodes are selected to form a bilinear interpolation unit, the dielectric loss change gradient slope is calculated in the time domain aging axis direction, and the voltage attenuation amount corresponding to each 10% dielectric loss increment is calculated in the voltage drift axis direction, and the dielectric loss-voltage drift quantitative matrix is generated through double-axis weighted average. Finally, when performing time domain deconvolution calculation, the historical aging cumulative rate spectrum in the device operation log is loaded, 6 minutes is set as the basic integration unit in the time integral model, the dielectric loss-voltage relationship parameters in the quantitative matrix are sequentially convolved with the aging rate, the contribution weight of the aging effect at different historical times to the current voltage drift is decomposed by using the reverse iteration method, the early aging influence is exponentially weighted and attenuated by establishing the attenuation memory function, and finally the PDIV drift data at the current time is output.

[0022] Step S13 comprises:

[0023] The dielectric loss factor and PDIV drift correlation curve library established by the accelerated aging experiment is called to perform dynamic matching on the corrected dielectric loss characteristic data in the aging stage, and the coordinate projection interval of the current dielectric loss value on the correlation curve is determined;

[0024] Two adjacent aging stage nodes in the coordinate projection interval are selected, the dielectric loss factor threshold and the voltage drift reference value corresponding to the nodes are extracted, and the aging stage node matching data is generated; the bidirectional interpolation operation is performed based on the aging stage node matching data, the dielectric loss change gradient interpolation is performed along the time aging axis, the drift amount linear compensation is performed along the voltage drift axis, and the quantization relationship parameters after interpolation and compensation are output;

[0025] The quantization relationship parameters after the interpolation compensation are cross-dimensionally fused to generate a dynamic quantization matrix of the dielectric loss factor and the voltage drift.

[0026] In the embodiment of the present application, firstly, a three-dimensional correlation curve cluster containing the dielectric loss-PDIV drift corresponding relationship under different aging periods is called from the accelerated aging experiment database, the modified dielectric loss characteristic value is projected along the time aging axis, the dielectric loss change axis and the voltage drift axis in the three-dimensional space, the projection area of the dielectric loss value on the aging surface is determined by the nearest neighbor density estimation algorithm, and a cubic search space containing eight adjacent aging sample points is constructed with the projection point as the center. Subsequently, according to the principle of thin plate spline interpolation, the time attenuation gradient of the current dielectric loss value and the adjacent aging nodes is calculated along the time aging axis, and the correlation coefficient of the dielectric loss increment and the voltage drift is calculated along the dielectric loss change axis. Through the double constraint conditions, two most adjacent aging stage nodes are screened out, and the dielectric loss threshold boundary value and the corresponding voltage drift reference value are extracted respectively. When performing bidirectional interpolation operation, a nonlinear interpolation method is used for the time aging axis, and according to the second derivative change characteristics of the aging curve, a logarithmic decay function is used to calculate the dielectric loss gradient interpolation coefficient between two aging nodes. For the voltage drift axis, a piecewise linear interpolation strategy is used, and according to the linear regression slope of the dielectric loss increment and the voltage drift, the drift amount compensation calculation is performed according to the position ratio of the current dielectric loss value relative to the adjacent node. After completing the bidirectional interpolation, the interpolation results of the time aging axis are assigned with a weight coefficient of 60%, and the interpolation results of the voltage drift axis are assigned with a weight coefficient of 40% by establishing a weight distribution model of dielectric loss-voltage drift. The cross-dimensionally fused quantization relationship parameter matrix is generated by using the weighted average algorithm. In particular, an abnormal parameter checking mechanism is set, when the dispersion of the bidirectional interpolation results exceeds 15%, the cubic spline interpolation algorithm is used to recalculate. In the weight distribution stage, a dielectric loss change rate correction factor is introduced, and an additional 20% gain coefficient is given to the voltage drift amount in the fast degradation stage, so as to ensure the parameter accuracy in the aging acceleration period. The finally generated dynamic quantization matrix contains the dielectric loss change step, the corresponding voltage drift amount and the confidence interval data.

[0027] Step S14 includes:

[0028] The aging cumulative rate historical data set of the oil paper insulation material is extracted, the dynamic quantization relationship matrix generated in step S13 is loaded, and the input parameter group of the aging effect time integration is established;

[0029] The input parameter group is introduced into the preset aging effect time integration model, the dielectric loss gradient is iteratively calculated in the cumulative amount according to the hour-level time unit, and the cumulative dielectric loss drift sequence on the time axis is output;

[0030] The PDIV starting voltage is decoupled in reverse time based on the accumulated dielectric loss drift sequence, and an exponential decay weighting algorithm is used to separate the voltage drift contribution of different aging stages to generate periodized voltage drift components;

[0031] The time domain normalization processing is performed on the periodized voltage drift components, and the current insulation PDIV drift data is output by component amplitude superposition and phase compensation calculation.

[0032] In the embodiment of the present application, first, the oil-paper insulation material past five years of aging rate historical data set is called from the insulation life management system, which contains the dielectric loss change gradient, load rate and environmental temperature and humidity parameters recorded every hour. These historical data are matched with the dynamic quantization relationship matrix in space-time, and the data block is divided into 72 hours units by using the sliding time window technology. The input parameter group containing the initial value of dielectric loss, aging gradient and voltage drift rate is established in each data block, and the numerical jump caused by inconsistent data acquisition interval is eliminated by using cubic spline interpolation. When the parameter group is introduced into the aging effect time integral model, the adaptive step Runge-Kutta method is used for iterative calculation: in each hour time unit, the contribution value of the dielectric loss increment to the voltage drift in this period is calculated according to the corresponding dielectric loss-voltage drift coefficient in the dynamic quantization matrix, and the cumulative drift is updated in real time by the integral accumulator. The convergence condition is set as the relative error of the results of adjacent two iterations is less than 0.05%, and the cumulative dielectric loss drift sequence on the continuous time axis is output when the model is iterated to the current time node. When the reverse time decoupling is performed, the deconvolution model based on the exponential decay kernel function is constructed, and the influence weight of each historical period on the current voltage drift is traced back along the time axis: 90% weight coefficient is reserved for the aging effect of the last 30 days, and the weight distribution decreases according to the natural logarithm law with the increase of time span. The voltage recovery component caused by the material relaxation effect is considered synchronously in the decoupling calculation, and the improved Wiener filtering algorithm is used to separate the net voltage drift contribution value of each period. When the time domain normalization processing is completed, first, the voltage drift components of each period are standardized by the maximum amplitude, and the drift values of different orders are uniformly mapped to the interval of 0-1; then, based on the phase spectrum characteristics of the cumulative dielectric loss sequence, the Hilbert transform is used to extract the phase delay parameters of each component, and the time domain misplacement error is eliminated by constructing a phase compensation vector; finally, the standardized amplitude and phase correction value are vector superimposed, and the current insulation PDIV drift data is output.

[0033] Step S2: based on the PDIV drift data, time sequence coupling processing is performed with historical operating environment parameters and power load curve to obtain insulation performance time sequence degradation data; an insulation life remaining prediction model is established according to the insulation performance time sequence degradation data to obtain insulation remaining life prediction results;

[0034] Step S2 includes:

[0035] The PDIV drift data, historical environmental temperature data and power load curve are time-stamped aligned to establish a time-stamped unified dataset based on the operation cycle of the equipment;

[0036] Based on the time-stamped unified dataset, environmental and electrical stress coupling analysis is performed, and through temperature load factor weighting algorithm and power fluctuation gradient compensation calculation, a composite aging stress parameter sequence is generated; the composite aging stress parameter sequence is subjected to insulation medium time-varying response modeling, and a medium polarization reconstruction method based on current density distribution is adopted to output insulation performance time sequence degradation data;

[0037] The insulation material accelerated aging sample library is called, the time sequence degradation data is input into the LSTM neural network training model, and an insulation life remaining prediction model is established through a forgetting gate threshold dynamic adjustment mechanism; the dielectric response spectrum data of the current oil-paper medium is used as the model input parameter, the recursive forward calculation of the insulation life remaining prediction model is performed, and the insulation remaining life prediction result is output.

[0038] In the embodiment of the present application, first, the collected PDIV drift data, historical environmental temperature data and power load curve are aligned according to a unified timestamp, and through an interpolation algorithm and data resampling technology, a timestamp-unified data set based on the running cycle of the equipment is established, so as to ensure that the data of different sources have accurate corresponding relationship in the time dimension. On this basis, the environmental stress coupling analysis is performed, and the environmental temperature data are quantitatively processed through a temperature load factor weighting algorithm. According to the thermal aging characteristic curve of the oil-paper insulation material, the temperature load factor weighting algorithm converts different temperature values into corresponding aging weight coefficients, and combines the fluctuation gradient of the power load curve to perform compensation calculation. Specifically, the power change rate is processed through a sliding window average, the time period of sharp power change is identified, and the temperature load factor is dynamically adjusted in these time periods, so as to generate a composite aging stress parameter sequence which can accurately reflect the actual running condition of the equipment. Subsequently, the composite aging stress parameter sequence is subjected to insulation medium time-varying response modeling, and a medium polarization reconstruction method based on current density distribution is adopted. The method analyzes the current response characteristics of the oil-paper insulation system under different aging stresses, establishes a current density space-time distribution model, and combines the medium polarization theory to reconstruct the polarization state change process of the insulation material in the long-term running process, so as to output time sequence degradation data which can accurately describe the degradation law of insulation performance with time. When establishing the insulation life remaining prediction model, first, the pre-constructed insulation material accelerated aging sample library is called, which contains the long-term aging experimental data of the oil-paper insulation material under different temperature, humidity and electric field intensity conditions. The aforementioned obtained insulation performance time sequence degradation data are input into the pre-trained LSTM neural network training model as input parameters. The model adopts a forgetting gate threshold dynamic adjustment mechanism, dynamically adjusts the control threshold of the forgetting gate by monitoring the change of the internal state of the network in real time, so that the model can better adapt to the nonlinear characteristics and long-term dependence in the insulation aging process. In the model training process, the historical aging data are used to iteratively optimize the network parameters, so as to ensure that the model has good generalization ability. Finally, the dielectric response spectrum data of the current oil-paper medium are used as real-time input parameters of the model, which are obtained by real-time acquisition through an online monitoring system and contain the current frequency response characteristics of the insulation system. Through the recursive forward calculation process of the insulation life remaining prediction model, that is, the input data are processed step by step according to the time sequence processing mechanism of the LSTM network, the insulation remaining life prediction result is finally output.

[0039] The sub-steps in step S2 include:

[0040] The environmental temperature sequence and the power load gradient sequence are extracted from the timestamp-unified data set, the Arrhenius equation equivalent aging amount calculation is performed on the temperature sequence through the temperature-aging rate conversion coefficient, and the temperature load factor weighting data are obtained;

[0041] Perform peak holding rate analysis and valley recovery compensation on the power load gradient sequence, generate power gradient compensation data according to the preset load fluctuation sensitivity parameter;

[0042] Time domain convolution fusion of temperature load factor weighted data and power gradient compensation data, using sliding time window integral algorithm to generate composite aging stress parameter sequence;

[0043] Reconstruct the space charge distribution model of insulating medium based on the composite aging stress parameter sequence, and establish the medium polarization response characteristic matrix through the Lorentz distribution fitting of current density vector;

[0044] Perform time-varying degradation simulation on the medium polarization response characteristic matrix, use the preset medium relaxation time spectrum to perform depolarization current iterative inversion, and output the insulating performance time sequence degradation data.

[0045] In the embodiment of the present application, firstly, the ambient temperature sequence and the power load gradient sequence are extracted from the time scale unified data set, for the ambient temperature sequence, through the pre-calibrated temperature-aging rate conversion coefficient, the temperature value at each time is converted into the corresponding equivalent aging rate value, the conversion process is based on the physical principle of Arrhenius equation, that is, the chemical reaction rate increases exponentially with the increase of temperature by a certain amplitude, thereby calculating the equivalent aging amount cumulative value corresponding to the entire temperature sequence, forming the temperature load factor weighted data, which can accurately reflect the cumulative influence of temperature factor on the aging process of insulating materials. At the same time, the peak value retention rate of the power load gradient sequence is analyzed, through the statistical duration and amplitude characteristics of the power sharply rising period, the ability of the equipment to withstand the electric stress impact is evaluated, and the valley value of the power descending period is recovered and compensated, the compensation process considers the self-recovery characteristics of the insulating materials during low load period, combined with the pre-set load fluctuation sensitivity parameter, which is pre-determined according to the response characteristics of oil-paper insulating materials to electric stress changes, thereby generating power gradient compensation data which can accurately describe the influence of power fluctuation on aging. Subsequently, the temperature load factor weighted data and the power gradient compensation data are convolved and fused in the time domain, the fusion process adopts the sliding time window integral algorithm, through setting the sliding window with fixed time length, the weighted integral calculation of the two kinds of data is carried out in each time window, the weight coefficient is pre-determined according to the relative importance of temperature and electric stress on aging, thereby generating a composite aging stress parameter sequence which can comprehensively reflect the combined action of environment and electric stress. On this basis, the insulating medium space charge distribution model is reconstructed based on the composite aging stress parameter sequence, the reconstruction process establishes a three-dimensional distribution model of space charge density by analyzing the migration and accumulation rules of electric charges inside the insulating materials under the action of composite stress, and combines the Lorentz distribution fitting method of current density vector, that is, the distribution characteristics of current density in each direction in space are fitted into the form of Lorentz function through statistical method, thereby establishing a medium polarization response feature matrix which can describe the electrical response characteristics of insulating materials under the action of composite stress. Finally, the time-varying degradation simulation is performed on the medium polarization response feature matrix, the simulation process uses the pre-set medium relaxation time spectrum, which describes the time distribution characteristics required for the insulating materials to recover from the polarization state to the equilibrium state, through the depolarization current iterative inversion algorithm, that is, the corresponding relationship between the polarization state and the depolarization current is repeatedly calculated and adjusted, the real degradation process is gradually approached, thereby outputting the insulating performance time sequence degradation data.

[0046] The sub-steps in step S2 further include:

[0047] A plurality of dielectric loss-PDIV drift time sequence data sets in the insulating material accelerated aging sample library are called to perform sample equalization processing on the time sequence degradation data, and generate an LSTM neural network training data set;

[0048] An initial model of LSTM network with forget gate control is constructed, and the training data set is input into the model for forward propagation calculation, and the error of the output layer is back propagated to dynamically adjust the activation threshold of the forget gate;

[0049] According to the dielectric response characteristic deviation degree of the verification set, the forget gate threshold interval is optimized, the threshold dynamic adjustment interval is set as [0.2, 0.8], and an insulation life remaining prediction model is established;

[0050] The wideband dielectric response spectrum data of the current oil-paper medium is collected, the loss factor spectrum feature vector of the 10Hz-1MHz frequency band is extracted as the model input parameter; the dielectric spectrum feature vector is input into the trained prediction model, the forward propagation calculation based on time step recursion is executed, and the insulation remaining life prediction result in years is output.

[0051] In the embodiment of the present application, firstly, a pre-constructed insulation material accelerated aging sample library is called, which contains multiple sets of dielectric loss-PDIV drift time series data obtained under different temperature, humidity, electric field intensity and other multiple stress conditions, which covers the performance characteristics of oil-paper insulation materials at various aging stages, and the matching degree analysis is performed on the obtained time series degradation data and the data in the sample library, the sample data is balanced by calculating the similarity characteristics of data distribution, which includes interpolating and expanding the data of specific aging stages with insufficient samples, and downsampling the data with excessive samples, to ensure the uniformity of the distribution of training data at each aging stage, thereby generating a standardized data set suitable for LSTM neural network training. Subsequently, an initial model of LSTM network with forgetting gate control is constructed, which adopts a multi-layer recurrent neural network structure, and the forgetting gate control mechanism is used to determine the amount of historical information that needs to be retained or forgotten by the network at each time step. The balanced training data set is input into the initial model for forward propagation calculation, the output prediction result is calculated through the weight connection between network layers, and the output result is compared with the actual aging state label to calculate the prediction error. The error information is transmitted layer by layer in the network through the back propagation algorithm, and the internal parameters of the network are dynamically adjusted, especially the activation threshold of the forgetting gate, which controls the degree of forgetting of historical information by the network and is automatically adjusted according to the error change during training. During the model training process, the validation set data is introduced to monitor the model performance in real time, which contains dielectric response feature data independent of the training set. The deviation degree between the model prediction result and the actual label of the validation set is calculated to evaluate the generalization ability of the model. Based on the evaluation result of the deviation degree, the threshold interval of the forgetting gate is optimized and adjusted, and the dynamic adjustment interval of the threshold is set to the range of 0.2 to 0.8. This interval can balance the memory ability of the model to historical information and the adaptability to new information, avoid losing important historical information due to too large threshold, or the model cannot adapt to new aging characteristics due to too small threshold, thereby establishing an insulation life remaining prediction model with good prediction accuracy and stability. In the model application stage, the broadband dielectric response spectrum data of the current oil-paper medium are collected through the online monitoring system, which can fully reflect the dielectric property changes of the insulation material at different frequencies. In the data preprocessing process, the loss factor spectrum feature vector in this frequency band is extracted, which contains the key parameters that can represent the current insulation state. These feature vectors are used as real-time input parameters of the model and input into the trained prediction model to perform forward propagation calculation based on time step recursion, that is, the input features are processed step by step according to the time series processing mechanism of the LSTM network, making full use of the historical information and current input information of the network, and finally outputting the insulation remaining life prediction result in years.

[0052] Step S3: performing operation control on the isolated power supply based on the insulation residual life prediction result.

[0053] Step S3 comprises:

[0054] Obtaining the insulation residual life prediction result, extracting the residual life value and the reliability index corresponding to the current time, and obtaining the insulation state evaluation data;

[0055] Comparing the insulation state evaluation data with the preset life threshold interval to determine the operation level of the current insulation state, the operation level comprising normal operation, pre-warning operation and emergency de-rating operation, and obtaining the insulation operation level determination result;

[0056] According to the insulation operation level determination result, matching the corresponding power output control strategy from the pre-configured operation strategy library to generate the power adjustment instruction suitable for the current insulation state, and obtaining the power control strategy instruction set;

[0057] Downlinking the power control strategy instruction set to the power adjustment module of the isolated power supply to control the adjustment of the output power level and the operation timing, and collecting the adjusted power output parameters and the insulation state feedback data in real time to obtain the power adjustment execution data;

[0058] Based on the power adjustment execution data, the isolated power supply is controlled.

[0059] In the embodiment of the present application, firstly, the prediction result output by the aforementioned insulation life remaining prediction model is acquired, which includes a remaining life value corresponding to the current time and a corresponding confidence index, wherein the remaining life value represents the expected time for which the device insulation system can still work normally in years, and the confidence index reflects the reliability of the prediction result, which is quantitatively obtained by analyzing the convergence of the internal state parameters of the prediction model and the historical prediction accuracy statistical data, thereby obtaining comprehensive insulation state evaluation data. Subsequently, the insulation state evaluation data is compared and analyzed with a pre-set life threshold interval, which is determined comprehensively according to the design life of the device, the safety margin requirement and the historical failure statistical data, and usually includes multiple grading standards, when the remaining life value is greater than the first threshold value, it is determined as a normal running state, when the remaining life value is between the first threshold value and the second threshold value, it is determined as a pre-warning running state, and when the remaining life value is less than the second threshold value, it is determined as an emergency derating running state, and the confidence index is combined to weight the determination result, ensuring the accuracy of the determination result, thereby determining the specific running grade of the current insulation state and obtaining the insulation running grade determination result. Based on the determination result, the corresponding power output control strategy is automatically matched from the pre-configured running strategy library, which includes detailed control schemes for different running grades, the standard power output mode is adopted in the normal running state, the restrictive power output mode is adopted in the pre-warning running state and the monitoring frequency is increased, the minimum safe power output mode is adopted in the emergency derating running state and the standby protection mechanism is started, the control strategy most suitable for the current state is retrieved from the strategy library through table lookup method or rule matching algorithm, the power adjustment instruction set suitable for the current insulation state is generated, and the complete power control strategy instruction set is formed. Then the power control strategy instruction set is issued to the power adjustment module of the isolation power supply through the communication interface, which includes core components such as power switching devices, voltage regulators and current controllers, and controls it to adjust the output power level and running time sequence according to the instruction requirements, for example, reducing the output power to 80% of the rated value and increasing the downtime interval in the pre-warning state, and reducing the output power to the safety limit value and starting the cycle start-stop mechanism in the emergency state, while real-time collection of the adjusted power supply output parameters including output voltage, current, power factor and other key indicators, and the insulation state feedback data obtained through the online monitoring system including real-time dielectric loss factor, partial discharge level and other parameters, thereby obtaining comprehensive power adjustment execution data. Finally, based on the power adjustment execution data, the isolation power supply is implemented closed-loop control, the deviation between the adjustment effect and the expected target is continuously monitored, the control parameters are dynamically adjusted, and it is ensured that the device can safely and stably run under the current insulation state.

[0060] It should be noted that the interval and threshold size is set for easy comparison, wherein the size of the threshold depends on how much sample data and the number of base set by those skilled in the art for each set of sample data, as long as it does not affect the proportion of the parameter and the quantized value. And the above formula is the calculation of the dimensionless value, the formula is obtained by collecting a large amount of data to simulate the formula of the nearest real situation, and the preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0061] It should be understood that the size of the sequence number of the above processes in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0062] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0063] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application, should be included in the protection scope of the present application.

Claims

1. A control method for a high-power isolated power supply, characterized in that, include: Step S1: The dielectric loss factor of the oil-paper insulation medium inside the isolation power supply is collected in real time through the insulation status online monitoring unit to obtain the dielectric loss online monitoring data; based on the dielectric loss online monitoring data, combined with the correlation curve between the dielectric loss factor and the partial discharge initiation voltage drift calibrated in advance through accelerated aging test, PDIV dynamic inversion calculation is performed to obtain the current insulation PDIV drift data. Step S2: Based on the PDIV drift data, perform time series coupling processing with historical operating environment parameters and power load curves to obtain insulation performance time-series degradation data; establish an insulation life remaining prediction model based on the insulation performance time-series degradation data to obtain the insulation life remaining prediction result; Step S3: Perform operation control of the isolation power supply based on the predicted remaining insulation life.

2. The control method for a high-power isolated power supply according to claim 1, characterized in that, Step S1 includes: Step S11: The online insulation status monitoring unit performs millisecond-level current phase angle tracking on the oil-paper insulation medium to obtain the dielectric polarization current amplitude data within the power frequency cycle. After Fourier fundamental frequency filtering, the data is used to generate online monitoring data of dielectric loss. Step S12: Perform environmental error compensation on the online monitoring data of dielectric loss. Use the preset temperature loss gradient matrix and humidity compensation factor to eliminate the dielectric loss measurement deviation and output the corrected dielectric loss characteristic data. Step S13: Call the library of correlation curves between dielectric loss factor and PDIV drift calibrated by accelerated aging test, perform loss offset mapping matching on the corrected dielectric loss characteristic data, and establish a dynamic quantization matrix between loss factor and voltage drift through bilinear interpolation operation of adjacent aging nodes. Step S14: Combining the historical data of the aging accumulation rate of the oil-paper insulation, the dynamic quantization matrix is ​​input into the preset aging effect time integral model, and the partial discharge initiation voltage is deconvolved with time based on the dielectric loss change gradient to output the current insulation PDIV drift data.

3. The control method for a high-power isolated power supply according to claim 2, characterized in that, Step S13 includes: The library of correlation curves between dielectric loss factor and PDIV drift established by accelerated aging experiments is called to dynamically match the corrected dielectric loss characteristic data for the aging stage, and the coordinate projection range of the current dielectric loss value on the correlation curve is determined. Within the coordinate projection range, select two adjacent aging stage nodes, extract the dielectric loss factor threshold and voltage drift reference value corresponding to the nodes, and generate aging stage node matching data; perform bidirectional interpolation operation based on the aging stage node matching data, perform dielectric loss change gradient interpolation along the time aging axis, perform drift amount linear compensation along the voltage drift axis, and output the interpolated and compensated quantized relationship parameters. Cross-dimensional weighted fusion is performed on the interpolated and compensated quantization parameters to generate a dynamic quantization matrix of dielectric loss factor and voltage drift.

4. The control method for a high-power isolated power supply according to claim 2, characterized in that, Step S14 includes: Extract the historical dataset of aging cumulative rate of oil paper insulation material, load the dynamic quantization relation matrix generated in step S13, and establish the input parameter set for the time integral of aging effect. The input parameter set is imported into the preset aging effect time integral model, and the cumulative dielectric loss gradient is iteratively calculated in hourly time units to output the cumulative dielectric loss drift sequence on the time axis. Based on the cumulative dielectric loss drift sequence, the partial discharge initiation voltage is decoupled in reverse time. An exponential decay weighted algorithm is used to separate the voltage drift contribution of different aging stages and generate time-phased voltage drift components. Time-domain normalization is performed on the time-varying voltage drift components, and the current insulation PDIV drift data is output through component amplitude superposition and phase compensation calculation.

5. The control method for a high-power isolated power supply according to claim 1, characterized in that, Step S2 includes: The PDIV drift data, historical ambient temperature data, and power load curve are timestamped to establish a unified time-stamped dataset based on the equipment operating cycle. Environmental and electrical stress coupling analysis is performed based on a time-scaled unified dataset. A composite aging stress parameter sequence is generated by using a temperature load factor weighting algorithm and power fluctuation gradient compensation calculation. The time-varying response model of the insulating medium is then performed on the composite aging stress parameter sequence. A dielectric polarization reconstruction method based on current density distribution is used to output the time-series degradation data of insulation performance. The model is trained by calling up the accelerated aging sample library of insulating materials and inputting the time-series degradation data into the LSTM neural network. An insulation lifetime prediction model is established through a forget gate threshold dynamic adjustment mechanism. The dielectric response spectrum data of the current oil-paper medium is used as the model input parameters to perform recursive forward calculation of the insulation lifetime prediction model and output the insulation lifetime prediction result.

6. The control method for a high-power isolated power supply according to claim 5, characterized in that, The sub-steps in step S2 include: The ambient temperature sequence and power load gradient sequence are extracted from the time-scaled unified dataset. The equivalent aging amount is calculated by performing the Arrhenius equation on the temperature sequence through the temperature-aging rate conversion coefficient to obtain the temperature load factor weighted data. Peak retention rate analysis and valley recovery compensation are performed on the power load gradient sequence, and power gradient compensation data is generated based on the preset load fluctuation sensitivity parameters. The temperature load factor weighted data and power gradient compensation data are fused by temporal convolution, and a sliding time window integration algorithm is used to generate a composite aging stress parameter sequence. The space charge distribution model of the insulating dielectric is reconstructed based on the composite aging stress parameter sequence, and the dielectric polarization response characteristic matrix is ​​established by fitting the Lorentz distribution of the current density vector. A time-varying degradation simulation is performed on the dielectric polarization response characteristic matrix. The depolarization current is iteratively inverted using a preset dielectric relaxation time spectrum, and the time-series degradation data of insulation performance is output.

7. The control method for a high-power isolated power supply according to claim 5, characterized in that, The sub-steps in step S2 also include: Multiple sets of dielectric loss-PDIV drift time series datasets from the accelerated aging sample library of insulating materials are called, and the time series degradation data are subjected to sample equalization processing to generate an LSTM neural network training dataset. Construct an initial LSTM network model with forget gate control, input the training dataset into the model for forward propagation computation, and dynamically adjust the forget gate activation threshold through backpropagation of the output layer error; The forget gate threshold range is optimized based on the dielectric response characteristic deviation of the validation set. The threshold dynamic adjustment range is set to [0.2, 0.8], and an insulation lifetime remaining prediction model is established. The broadband dielectric response spectrum data of the current oil-paper medium is collected, and the loss factor spectrum feature vector of the 10Hz-1MHz band is extracted as the model input parameter. The dielectric spectrum feature vector is input into the trained prediction model, and forward propagation calculation based on time step recursion is performed to output the insulation remaining lifetime prediction result in years.

8. The control method for a high-power isolated power supply according to claim 1, characterized in that, Step S3 includes: Obtain the insulation remaining lifetime prediction results, extract the remaining lifetime value and confidence index corresponding to the current moment, and obtain insulation condition assessment data; The insulation status assessment data is compared with a preset lifespan threshold range to determine the current operating level of the insulation status. The operating level includes normal operation, early warning operation, and emergency derating operation, and the insulation operating level determination result is obtained. Based on the insulation operation level determination result, the corresponding power output control strategy is matched from the pre-configured operation strategy library to generate a power adjustment command adapted to the current insulation state, thus obtaining a power control strategy instruction set; The power control strategy instruction set is sent to the power adjustment module of the isolated power supply to control it to adjust the output power level and operating sequence, and to collect the adjusted power output parameters and insulation status feedback data in real time to obtain power adjustment execution data. Control of the isolated power supply is implemented based on power regulation execution data.

Images