Capacitor optimization switching method and system based on timing control

Abstract

The present invention provides a capacitor optimization switching method and system based on timing control, which relates to the field of circuit protection technology, including establishing a health status database of a capacitor group, using a deep learning algorithm to construct a remaining life prediction model of the capacitor group, and establishing a backup protection strategy for a standby capacitor group; monitoring the voltage fluctuation rate and capacitance loss rate of the main working capacitor group to generate a health status assessment report of the main working capacitor group, and establishing a load prediction model in combination with the real-time working parameters of the duplex capacitor group; when the state of the main working capacitor group is abnormal, after the standby capacitor group meets the switching conditions, determining the best switching time and using soft switching technology to achieve disturbance-free switching; optimizing the switching time based on the historical data of the updated health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of any capacitor group is lower than a preset threshold, triggering a warning signal and prompting replacement.

Description

Technical Field

[0001] The present invention relates to circuit protection technology, and in particular to a capacitor optimization switching method and system based on timing control. Background Art

[0002] As a key energy storage element, capacitor banks play an important role in traction power supply systems. To ensure the reliable operation of the system, a duplex capacitor bank configuration is usually adopted. However, the existing technology has the following problems: the health status monitoring method of the capacitor bank is relatively simple, mainly relying on the over-limit alarm of voltage and current, and lacks the ability to systematically evaluate and predict the performance degradation of the capacitor bank.

[0003] The capacitor bank switching strategy lacks intelligence and is often switched passively after a fault occurs. It is unable to select the optimal switching time according to the load characteristics and system operating status, which can easily cause system disturbances. The soft switch control effect during the switching process is not ideal, and there are problems such as large voltage fluctuations and high switching impact current, which affect the stable operation of the system.

[0004] The management strategy of the backup capacitor bank is relatively extensive, lacking effective status detection and self-recovery mechanisms, making it difficult to ensure reliability after switching. Existing methods generally lack comprehensive data collection and analysis of the switching process, and cannot provide effective support for system optimization and maintenance decisions. Summary of the invention

[0005] The embodiments of the present invention provide a capacitor optimization switching method and system based on timing control, which can solve the problems in the prior art.

[0006] In a first aspect of an embodiment of the present invention, a capacitor optimization switching method based on timing control is provided, including: collecting voltage signals, current signals and temperature signals of a duplex capacitor group in a railway system, and establishing a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a standby capacitor group; based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model of the duplex capacitor group, and a backup protection strategy for the standby capacitor group is established; real-time monitoring of the voltage fluctuation rate and capacitance loss rate of the main working capacitor group, generating a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establishing a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report When the state of the main working capacitor group is displayed to be abnormal, a pre-switching state check is performed on the standby capacitor group according to the backup protection strategy. After the standby capacitor group meets the switching conditions, a switching strategy is determined according to the load prediction model. The switching strategy includes the optimal switching timing, and soft switching technology is used to achieve disturbance-free switching of the duplex capacitor group. The switched standby capacitor group is set to the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, the state data of the switching process is updated to the health status database, and the switching strategy is optimized based on the historical data of the health status database and the calculation results of the remaining life prediction model. When the remaining life prediction model predicts that the remaining life of the main working capacitor group or the standby capacitor group is lower than the preset threshold, an early warning signal is automatically triggered and a replacement suggestion is prompted.

[0007] In an optional embodiment, based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model for a duplex capacitor group, and the steps of establishing a backup protection strategy for a standby capacitor group include: extracting degradation characteristics of the duplex capacitor group from the health status database, the degradation characteristics including the equivalent series resistance change rate calculated by the voltage and current phase difference, the capacitance loss rate based on the charge and discharge curve slope analysis, the leakage current density using the ratio of steady-state leakage current to rated voltage, and the temperature stress index based on the cumulative effect of thermal cycles; performing accelerated aging on the duplex capacitor group using temperature stress cycle tests, voltage stress tests, and ripple current superposition tests to obtain actual remaining life data under the end-of-life criterion, establishing a mapping relationship between the degradation characteristics and the actual remaining life, and forming a remaining life prediction. A training data set for the model; constructing a remaining life prediction model based on deep learning, the remaining life prediction model includes a multi-layer stacked bidirectional long short-term memory network layer for capturing temporal feature dependencies, and a multi-head attention mechanism layer for adaptive adjustment of feature weights, the multi-head attention mechanism layer includes a time attention unit and a feature attention unit; using a weighted mean square error loss function to train the remaining life prediction model, the loss function includes a remaining life prediction error term and a capacitor group degradation rate error term, and determining a dynamic adaptive weight coefficient of the loss function through iterative optimization; based on the output result of the trained remaining life prediction model, establishing a backup protection strategy for the standby capacitor group, the backup protection strategy includes a hardware overcurrent protection and overvoltage protection circuit, and a thermistor latching protection circuit.

[0008] In an optional implementation, the remaining life prediction model is trained using a weighted mean square error loss function, the loss function includes a remaining life prediction error term and a capacitor group degradation rate error term, and the step of determining a dynamic adaptive weight coefficient of the loss function by iterative optimization includes: the remaining life prediction error term is the square difference between the predicted remaining life value and the actual remaining life value, the capacitor group degradation rate error term is the square difference between the predicted degradation rate and the actual degradation rate, the predicted degradation rate is obtained by dividing the difference between the predicted remaining life values ​​at adjacent moments by the time interval, and the actual degradation rate is obtained by dividing the difference between the actual remaining life values ​​at adjacent moments by the time interval; iteratively optimizing the dynamic adaptive weight coefficient The method comprises calculating the partial derivative of the loss function with respect to the dynamic adaptive weight coefficient, updating the dynamic adaptive weight coefficient according to the direction of the partial derivative and the learning rate, and determining the optimal dynamic adaptive weight coefficient when the error change rate of the validation set in the training data set of multiple consecutive iterations is less than a preset threshold; training the remaining life prediction model based on the loss function with the optimal dynamic adaptive weight coefficient, selecting training samples in batch iteration, and adaptively adjusting the learning rate of the network parameters of each layer of the remaining life prediction model according to the gradient information of each batch, and optimizing and updating the network parameters of the remaining life prediction model; introducing an early stopping mechanism in the training process, and stopping the training process when the loss value on the validation set has no convergence trend after multiple consecutive rounds of training.

[0009] In an optional implementation, the voltage fluctuation rate and capacitance loss rate of the main working capacitor group are monitored in real time, and a health status assessment report of the main working capacitor group is generated according to the voltage fluctuation rate and capacitance loss rate. Based on the health status assessment report and the real-time working parameters of the duplex capacitor group, the steps of establishing a load prediction model include: collecting voltage data of the main working capacitor group, performing noise reduction processing on the voltage data through a sliding window, and calculating the ratio of the voltage difference between adjacent sampling points and the rated voltage to obtain the voltage fluctuation rate; at the end of each working cycle, the actual capacitance is calculated based on the slope of the voltage-time curve during the charging process of the main working capacitor group, and the capacitance loss rate is obtained by taking the difference ratio between the actual capacitance and the rated capacitance; the voltage fluctuation rate and the capacitance loss rate are multiplied by preset weight coefficients respectively and then summed to obtain a health index, and a health status assessment report including the current health index value, change trend and alarm level is generated according to the health index; the real-time working parameters of the duplex capacitor group are obtained. The invention relates to a load prediction model, wherein the load prediction model comprises a plurality of load prediction elements, a plurality of load prediction elements, and a plurality of load prediction elements. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result.

[0010] In an optional implementation, when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, a pre-switching state check is performed on the backup capacitor group. After the backup capacitor group meets the switching conditions, the switching strategy is determined according to the load prediction model, and the switching strategy includes the optimal switching timing. The steps of using soft switching technology to achieve disturbance-free switching of the duplex capacitor group include: when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, the insulation resistance, equivalent series resistance and leakage current of the backup capacitor group are measured through a self-test circuit; the backup capacitor group is pre-charged with a constant current source, the charging curve is recorded, and the first-order derivative change rate of the charging curve is analyzed to obtain a curve smoothness index; the backup capacitor group is determined according to the insulation resistance, equivalent series resistance, leakage current and curve smoothness index. The invention discloses a method for determining the switching state of a capacitor group; obtaining a load power prediction sequence within a future time window output by a load prediction model; calculating an average load power and a power variance based on the load power prediction sequence, constructing a switching condition discrimination function including a load power deviation term, a power change rate term and a power fluctuation term, and determining an optimal switching time by minimizing the switching condition discrimination function; based on the optimal switching time, connecting a resonant inductor in series in the main circuit, wherein the value of the resonant inductor is related to the resonant angular frequency and the equivalent capacitance; controlling the turn-on time of the auxiliary switch to lag behind the turn-on time of the main switch; configuring an absorption circuit in the cut-off branch of the main working capacitor group, wherein the parameters of the absorption circuit are related to the voltage rise time during the shutdown process; realizing a zero current switch through the energy oscillation between the resonant inductor and the standby capacitor group, realizing a zero voltage switch through the absorption circuit, and completing the switching operation of the duplex capacitor group.

[0011] In an optional embodiment, the average load power and power variance are calculated based on the load power prediction sequence, and a switching condition judgment function including a load power deviation term, a power change rate term, and a power fluctuation term is constructed. The step of determining the optimal switching timing by minimizing the switching condition judgment function includes: using an exponentially weighted moving average method to smooth the load power prediction sequence to obtain a smoothed average load power; based on the smoothed average load power, a power variance is calculated recursively; a switching condition judgment function including a power deviation term, a power change rate term, and a power fluctuation term is constructed, wherein the power deviation term is calculated based on the deviation between the current load power and a preset optimal switching power threshold, the power change rate term is calculated based on the time derivative of the load power, and the power fluctuation term is calculated based on the power variance; by setting an objective function for maximizing the switching success rate, a switching condition judgment function is used. The particle swarm optimization algorithm iteratively optimizes the non-negative normalized weight coefficients of the power deviation term, the power change rate term and the power fluctuation term; the product of the optimized non-negative normalized weight coefficient and the corresponding term is summed to obtain the final switching condition discriminant function; with minimization of the switching condition discriminant function as the optimization goal, an improved gradient descent method is used to solve the optimal switching timing, and the improved gradient descent method includes calculating the adaptive step size by dividing the initial step size by the increasing term of the number of iterations; in each iteration, the gradient of the switching condition discriminant function is calculated based on the current moment, and the switching moment is updated according to the calculated gradient direction and the adaptive step size; the switching condition discriminant function value corresponding to the updated switching moment is calculated, and when the difference between the switching condition discriminant function values ​​of two adjacent iterations is less than the function value convergence threshold and the difference between the switching moments of two adjacent iterations is less than the time convergence threshold, the switching moment obtained in the current iteration is determined as the optimal switching moment.

[0012] In an optional implementation, the switched standby capacitor group is set to the main working mode, the original main working capacitor group enters the standby state and executes a self-recovery program, the state data of the switching process is updated to the health state database, and based on the historical data of the health state database and the calculation results of the remaining life prediction model, the step of optimizing the switching strategy includes: after completing the switching operation, the switched standby capacitor group is set to the main working mode; the original main working capacitor group is switched to the standby state and a self-recovery program is executed, and the self-recovery program includes executing a charge and discharge balance test to evaluate the consistency of the single capacitor capacitance, eliminating residual charge through a constant current charge and discharge cycle, and executing an internal resistance test to evaluate the change of the equivalent series resistance. status, and perform capacitor balancing compensation; collect status data during the switching process and update it to a health status database, wherein the status data during the switching process includes voltage and current waveform data, transient response characteristics, switching operation timing information, temperature change data, and test results of a self-recovery program; based on the status data during the switching process, update the health index of the original main working capacitor group, and based on the health index, optimize the execution parameters of the self-recovery program; combine the historical data in the health status database and the calculation results of the remaining life prediction model to optimize the timing selection standard of the switching strategy and the non-negative normalized weight coefficient of the switching condition discrimination function; and use the optimized switching strategy to guide the next capacitor group switching operation.

[0013] According to a second aspect of an embodiment of the present invention, a capacitor optimization switching system based on timing control is provided, including: a first unit, used to collect voltage signals, current signals and temperature signals of a duplex capacitor group in a railway system, and establish a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a standby capacitor group; based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model of the duplex capacitor group, and a backup protection strategy for the standby capacitor group is established; a second unit is used to monitor the voltage fluctuation rate and capacitance loss rate of the main working capacitor group in real time, generate a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establish a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report is used, a load prediction model is established; and When the estimation report shows that the state of the main working capacitor group is abnormal, the pre-switching state check is performed on the standby capacitor group according to the backup protection strategy. After the standby capacitor group meets the switching conditions, the switching strategy is determined according to the load prediction model. The switching strategy includes the optimal switching time, and the soft switching technology is used to achieve disturbance-free switching of the duplex capacitor group; the third unit is used to set the switched standby capacitor group as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, and the state data of the switching process is updated to the health status database. The switching strategy is optimized based on the historical data of the health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of the main working capacitor group or the standby capacitor group is lower than the preset threshold, an early warning signal is automatically triggered and a replacement suggestion is prompted.

[0014] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] The present invention realizes high-precision prediction of the life of the capacitor group through a deep learning algorithm, and effectively controls the prediction error; the soft switching technology is used to significantly suppress the voltage fluctuation and current impact during the switching process, and greatly improves the dynamic characteristics of the switching process; the self-recovery program greatly improves the availability of the standby capacitor group and effectively enhances the system reliability. In addition, the intelligent switching strategy reduces human intervention and maintenance costs, provides a strong guarantee for the safe and stable operation of the power supply system, and has important engineering application value. The present invention significantly improves the prediction accuracy of the model and significantly reduces the average prediction error on the test set. The dynamic adaptive weight mechanism enables the model to better balance the needs of life prediction and degradation trend fitting, and effectively improves the temporal consistency of the prediction results. The adaptive learning rate strategy significantly accelerates the convergence speed of the model and shortens the training time. The early stopping mechanism effectively prevents overfitting. The model shows excellent generalization performance in practical applications, and shows a high accuracy rate of capacitor life prediction under different working conditions, providing a reliable basis for the preventive maintenance of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a flow chart of a capacitor optimization switching method based on timing control according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a capacitor optimization switching system based on timing control according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 FIG. 1 is a flow chart of a capacitor optimization switching method based on timing control according to an embodiment of the present invention. Figure 1As shown, the method includes: S1. collecting voltage signals, current signals and temperature signals of a duplex capacitor group in a railway system, and establishing a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a standby capacitor group; based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model of the duplex capacitor group, and a backup protection strategy for the standby capacitor group is established; S2. real-time monitoring of the voltage fluctuation rate and capacitance loss rate of the main working capacitor group, generating a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establishing a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report shows the status of the main working capacitor group When the state is abnormal, the standby capacitor group is pre-switched according to the backup protection strategy. After the standby capacitor group meets the switching conditions, the switching strategy is determined according to the load prediction model. The switching strategy includes the optimal switching time, and the soft switching technology is used to achieve disturbance-free switching of the duplex capacitor group; S3. The switched standby capacitor group is set as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, the state data of the switching process is updated to the health status database, and the switching strategy is optimized based on the historical data of the health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of the main working capacitor group or the standby capacitor group is lower than the preset threshold, the early warning signal is automatically triggered and a replacement suggestion is prompted.

[0021] For example, first, high-precision data acquisition equipment is deployed to conduct comprehensive monitoring of the duplex capacitor group. The sampling frequency is set to 10kHz to achieve high-speed sampling of voltage and current signals; the temperature signal sampling frequency is 1Hz. The collected data is filtered and denoised and stored in the health status database to construct a multidimensional data matrix containing electrical parameters and environmental parameters.

[0022] Based on the health status database, a deep learning network structure is designed for remaining life prediction. The network adopts a 4-layer bidirectional long short-term memory network, each layer contains 128 neurons. The input layer receives standardized time series features, including equivalent series resistance change rate, capacitance loss rate, leakage current density and temperature stress index. The attention mechanism layer consists of 8 time attention units and 8 feature attention units to achieve adaptive weight allocation of key time series features.

[0023] The real-time monitoring system uses a sliding window method to process voltage data, with the window length set to 100ms and an overlap rate of 50%. The voltage fluctuation rate is obtained by calculating the ratio of the voltage change at adjacent sampling points to the rated voltage. At the end of each working cycle, the actual capacitance is calculated based on the slope of the charging curve to evaluate the capacitance loss. When the voltage fluctuation rate exceeds 3% or the capacitance loss rate exceeds 15%, the system determines that the capacitor bank is in an abnormal state.

[0024] The load prediction model adopts a dual-channel architecture. The time series channel uses a one-dimensional convolution layer with a convolution kernel size of 3 to extract historical load features, and the state channel uses a three-layer multi-layer perceptron to process health status data. The feature fusion layer realizes dynamic feature fusion by calculating the correlation weights. The fused features are passed through a feature extraction network with residual connections to obtain the final prediction results.

[0025] The core of the switching strategy is the selection of the best switching time. The system first performs a pre-switching check on the backup capacitor bank, measuring its insulation resistance, equivalent series resistance, and leakage current. A constant current source is used for pre-charging test, and the smoothness of the charging curve is analyzed. When the indicators of the backup capacitor bank meet the switching requirements, the switching time is determined based on the load prediction results.

[0026] In the specific switching operation, soft switching technology is used to achieve disturbance-free switching. A resonant inductor is connected in series in the main circuit, and its value is determined according to the system resonant frequency. Zero current switching is achieved by controlling the conduction timing of the auxiliary switch. At the same time, an absorption circuit is configured in the cut-off branch to ensure zero voltage shutdown and effectively suppress voltage and current shocks during the switching process.

[0027] After the switch is completed, the system automatically executes the self-recovery procedure. The original main working capacitor group is charged and discharged to balance the test to evaluate the consistency of the capacitance value; the residual charge is eliminated through constant current charge and discharge cycles; and the internal resistance test is performed to evaluate the performance change. The test data is recorded and updated to the health status database for optimizing the subsequent switching strategy.

[0028] In this embodiment, a high-precision prediction of the life of the capacitor bank is achieved through a deep learning algorithm, effectively controlling the prediction error; the soft switching technology is used to significantly suppress the voltage fluctuation and current impact during the switching process, greatly improving the dynamic characteristics of the switching process; the self-recovery program greatly improves the availability of the backup capacitor bank and effectively enhances the system reliability. In addition, the intelligent switching strategy reduces human intervention and maintenance costs, provides a strong guarantee for the safe and stable operation of the power supply system, and has important engineering application value.

[0029] In an optional embodiment, based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model for a duplex capacitor group, and the steps of establishing a backup protection strategy for a standby capacitor group include: extracting degradation characteristics of the duplex capacitor group from the health status database, the degradation characteristics including the equivalent series resistance change rate calculated by the voltage and current phase difference, the capacitance loss rate based on the charge and discharge curve slope analysis, the leakage current density using the ratio of steady-state leakage current to rated voltage, and the temperature stress index based on the cumulative effect of thermal cycles; performing accelerated aging on the duplex capacitor group using temperature stress cycle tests, voltage stress tests, and ripple current superposition tests to obtain actual remaining life data under the end-of-life criterion, establishing a mapping relationship between the degradation characteristics and the actual remaining life, and forming a remaining life prediction. A training data set for the model; constructing a remaining life prediction model based on deep learning, the remaining life prediction model includes a multi-layer stacked bidirectional long short-term memory network layer for capturing temporal feature dependencies, and a multi-head attention mechanism layer for adaptive adjustment of feature weights, the multi-head attention mechanism layer includes a time attention unit and a feature attention unit; using a weighted mean square error loss function to train the remaining life prediction model, the loss function includes a remaining life prediction error term and a capacitor group degradation rate error term, and determining a dynamic adaptive weight coefficient of the loss function through iterative optimization; based on the output result of the trained remaining life prediction model, establishing a backup protection strategy for the standby capacitor group, the backup protection strategy includes a hardware overcurrent protection and overvoltage protection circuit, and a thermistor latching protection circuit.

[0030] Exemplarily, the degradation characteristics of the duplex capacitor group are first extracted from the health status database. The equivalent series resistance change rate is obtained by collecting the voltage and current waveforms in real time, calculating the phase difference angle, and comparing it with the initial value. Taking a certain capacitor group as an example, the initial phase difference is 2.1 degrees, which increases to 2.8 degrees after 6 months of operation, and the corresponding equivalent series resistance change rate is 33.3%. The capacitance loss rate is based on the analysis of the voltage change rate during the charging and discharging process, with a sampling interval of 0.1ms, and the actual capacitance is calculated by the slope of the curve. The leakage current density uses a high-precision current sensor to measure the steady-state leakage current at the rated voltage. The leakage current density of a test sample increases from the initial 0.02mA / V to 0.035mA / V. The temperature stress index takes into account the cumulative effect of thermal cycles and records the temperature change amplitude and the number of cycles.

[0031] When conducting accelerated aging tests, the temperature stress cycle test cycles in the range of -40°C to 85°C, with each cycle lasting 4 hours, including 30 minutes of temperature holding. The voltage stress test applies 1.2 times the rated voltage for 168 hours. The ripple current superposition test superimposes a high-frequency ripple component of 0.5 times the rated value on the rated current, with a frequency of 10kHz. Through these tests, life termination data is obtained, and the end-of-life criteria are defined as a capacitance loss of more than 20% or an increase in equivalent series resistance of more than 200%.

[0032] The remaining life prediction model adopts a 4-layer stacked bidirectional long short-term memory network structure, each layer contains 128 neurons. The input layer receives the standardized degradation features, and the middle layer captures the bidirectional temporal dependencies through the gating mechanism. The multi-head attention mechanism layer contains 8 time attention units and 8 feature attention units. The time attention unit calculates the correlation weights of different time steps, and the feature attention unit evaluates the importance of different degradation features.

[0033] The model training uses weighted mean square error as the loss function. The remaining life prediction error term reflects the deviation between the predicted value and the actual value, and the capacitor group degradation rate error term characterizes the degree of fit of the degradation trend. The dynamic adaptive weight coefficient is initially set to 0.6 and 0.4, and is dynamically adjusted according to the performance of the validation set during training. The training data contains 50 groups of complete life cycle data of capacitor samples under different working conditions, and 5-fold cross validation is used to evaluate the model performance.

[0034] The hardware implementation of the backup protection strategy includes a triple protection mechanism. The overcurrent protection uses a Hall current sensor for real-time sampling, and the protection is triggered when the current exceeds the rated value by 120% for 100ms. The overvoltage protection has a double threshold and is activated when the instantaneous overvoltage exceeds 130% or the continuous overvoltage exceeds 115%. The thermistor lockout protection cuts off the main circuit when the temperature exceeds 75°C to ensure system safety.

[0035] In this embodiment, a high-precision prediction of the remaining life of the capacitor bank is achieved, and the reliability of the life prediction is improved through the comprehensive analysis of multi-dimensional degradation characteristics. The accelerated aging test method shortens the data acquisition cycle and provides sufficient training samples. The multi-layer deep learning structure has excellent feature extraction capabilities, and the attention mechanism improves the model's perception of key features. Triple hardware protection ensures reliable switching of the backup capacitor bank.

[0036] In an optional implementation, the remaining life prediction model is trained using a weighted mean square error loss function, the loss function includes a remaining life prediction error term and a capacitor group degradation rate error term, and the step of determining a dynamic adaptive weight coefficient of the loss function by iterative optimization includes: the remaining life prediction error term is the square difference between the predicted remaining life value and the actual remaining life value, the capacitor group degradation rate error term is the square difference between the predicted degradation rate and the actual degradation rate, the predicted degradation rate is obtained by dividing the difference between the predicted remaining life values ​​at adjacent moments by the time interval, and the actual degradation rate is obtained by dividing the difference between the actual remaining life values ​​at adjacent moments by the time interval; iteratively optimizing the dynamic adaptive weight coefficient The method comprises calculating the partial derivative of the loss function with respect to the dynamic adaptive weight coefficient, updating the dynamic adaptive weight coefficient according to the direction of the partial derivative and the learning rate, and determining the optimal dynamic adaptive weight coefficient when the error change rate of the validation set in the training data set of multiple consecutive iterations is less than a preset threshold; training the remaining life prediction model based on the loss function with the optimal dynamic adaptive weight coefficient, selecting training samples in batch iteration, and adaptively adjusting the learning rate of the network parameters of each layer of the remaining life prediction model according to the gradient information of each batch, and optimizing and updating the network parameters of the remaining life prediction model; introducing an early stopping mechanism in the training process, and stopping the training process when the loss value on the validation set has no convergence trend after multiple consecutive rounds of training.

[0037] Exemplarily, in the process of constructing the loss function, the remaining life prediction error term is first calculated. For example, when the predicted remaining life is 1200 hours and the actual remaining life is 1150 hours, the prediction error term is the square of the difference, that is, 2500 square hours. For the degradation rate error term, take the data of two adjacent sampling points. If the predicted life of the first sampling point is 1200 hours, and the predicted life of the second sampling point 10 hours later is 1180 hours, then the predicted degradation rate is 2 hours of life per hour; similarly, the actual degradation rate is calculated, and the square of the difference between the two is used as the degradation rate error term.

[0038] The iterative optimization of the dynamic adaptive weight coefficient adopts the gradient descent method. The initial setting of the remaining life prediction error term weight is 0.6, and the degradation rate error term weight is 0.4. The rate of change of the loss function to the weight coefficient is calculated for each iteration, and the learning rate is set to 0.01. During the iteration process, if it is found that increasing the weight of the remaining life prediction error term can reduce the overall loss, the weight distribution is adjusted accordingly. When the error change rate of the validation set for 5 consecutive iterations is less than 0.1%, it is considered that the optimal weight coefficient has been found.

[0039] The training process uses a small batch method with a batch size of 32, and each training cycle contains 200 batches. An adaptive learning rate strategy is used for network parameters of different layers, and the initial learning rate is set to 0.001. For example, in the early stage of training, when the parameter gradient of a convolutional layer is large, its learning rate may be reduced to 0.0005, while the fully connected layer with a smaller gradient may maintain a learning rate of 0.001. After each batch is processed, the gradient is calculated according to the loss value of the batch and the network parameters are updated.

[0040] An adaptive moment estimation optimizer was introduced for parameter update, where the momentum parameter was set to 0.9 and the second-order momentum parameter was set to 0.999. During the training process, the validation set performance was monitored in real time. When the validation set loss value did not decrease significantly (the rate of change was less than 0.05%) for 10 consecutive training cycles, the early stopping mechanism was triggered. In one training, the model achieved the best performance of the validation set in the 87th training cycle, with a prediction error of 5.2%. After that, the validation set loss value fluctuated no more than 0.04% for 10 consecutive training cycles, at which time the training was stopped.

[0041] In order to enhance the generalization ability of the model, the cross-validation method was used to evaluate the model performance. The data set was divided into training set, validation set and test set in a ratio of 8:1:1. During the cross-validation process, the model performed stably under different data partitions, and the standard deviation of the prediction error was controlled within 1.5%. At the same time, a random dropout mechanism was introduced to prevent overfitting, and the random dropout rate was set to 0.2. In the parameter optimization process, the grid search method was used to determine the key hyperparameters of the network structure.

[0042] In this embodiment, the prediction accuracy of the model is significantly improved, and the average prediction error on the test set is greatly reduced. The dynamic adaptive weight mechanism enables the model to better balance the needs of life prediction and degradation trend fitting, and effectively improves the temporal consistency of the prediction results. The adaptive learning rate strategy significantly accelerates the convergence of the model and shortens the training time. The early stopping mechanism effectively prevents overfitting. The model shows excellent generalization performance in practical applications, and shows a high accuracy rate of capacitor life prediction under different working conditions, providing a reliable basis for preventive maintenance of power systems.

[0043] In an optional implementation, the voltage fluctuation rate and capacitance loss rate of the main working capacitor group are monitored in real time, and a health status assessment report of the main working capacitor group is generated according to the voltage fluctuation rate and capacitance loss rate. Based on the health status assessment report and the real-time working parameters of the duplex capacitor group, the steps of establishing a load prediction model include: collecting voltage data of the main working capacitor group, performing noise reduction processing on the voltage data through a sliding window, and calculating the ratio of the voltage difference between adjacent sampling points and the rated voltage to obtain the voltage fluctuation rate; at the end of each working cycle, the actual capacitance is calculated based on the slope of the voltage-time curve during the charging process of the main working capacitor group, and the capacitance loss rate is obtained by taking the difference ratio between the actual capacitance and the rated capacitance; the voltage fluctuation rate and the capacitance loss rate are multiplied by preset weight coefficients respectively and then summed to obtain a health index, and a health status assessment report including the current health index value, change trend and alarm level is generated according to the health index; the real-time working parameters of the duplex capacitor group are obtained. The invention relates to a load prediction model, wherein the load prediction model comprises a plurality of load prediction elements, a plurality of load prediction elements, and a plurality of load prediction elements. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result. The load prediction model comprises a plurality of load prediction elements, and a plurality of load prediction elements are combined with each other to form a load prediction result.

[0044] For example, a high-precision sampling system is used for voltage data acquisition and processing, and the sampling frequency is set to 10kHz to ensure signal integrity. The original voltage data is processed by a 20-point sliding window for noise reduction, which effectively suppresses high-frequency interference. For example, when the voltage at adjacent sampling points drops from 380V to 375V and the rated voltage is 400V, the voltage fluctuation rate is calculated to be 1.25%.

[0045] Capacitance loss detection is performed at the end of the working cycle, and the voltage-time data pairs during the charging process are recorded. For example, the charging process lasts for 2ms, 100 voltage values ​​are collected, and the voltage rising slope is obtained through data fitting. The actual capacitance value is calculated to be 95μF, which is 5% compared to the rated value of 100μF.

[0046] The generation of a health status assessment report involves three key elements. The first is the presentation of the health index value, which directly displays the calculated index value. For example, the voltage fluctuation rate has a weight of 0.4, the capacitance loss rate has a weight of 0.6, and the voltage fluctuation rate of 1.25% and the capacitance loss rate of 5% are weighted to obtain a health index value of 3.5%. The second is the analysis of the change trend. By comparing the previous health index data, it is determined whether the equipment status is stable, slowly deteriorating, or rapidly deteriorating. Finally, the alarm level is determined, and the health index is divided into three intervals: 0-2% is normal, 2-5% is a warning, and above 5% is an alarm.

[0047] The working parameter collection of the duplex capacitor bank includes three types of real-time data: input current, switching frequency, and ambient temperature, with a sampling interval of 1s. At the same time, historical load data is collected at intervals of 5 minutes within 24 hours, with a total of 288 data points. For example, the input current is 15A, the switching frequency is 20kHz, the ambient temperature is 35℃, and the load rate fluctuation range is 45%-85%.

[0048] The feature processing layer normalizes all input data. The input current 15A is mapped to 0.75, the switching frequency 20kHz is mapped to 0.8, the ambient temperature 35℃ is mapped to 0.6, and the health index 3.5% is mapped to 0.35. The processed feature data are all in the range of 0-1, eliminating the dimension effect.

[0049] The dual-channel feature extraction layer processes the timing features and state features respectively. The timing channel uses a three-layer one-dimensional convolution structure with convolution kernel sizes of 5, 3, 3 and channels of 64, 128, and 256. Actual measurements show that this structure can extract periodic and trend features from 288 historical load data points. The state channel uses a three-layer fully connected network with 128, 256, and 256 nodes to process health indexes and real-time parameters.

[0050] The feature fusion layer processes the feature information of the two input channels and uses the attention mechanism to calculate the correlation weight matrix between the feature vectors of the two channels. The time series feature vector and the state feature vector are dot-producted and normalized to obtain the weight matrix. When performing feature fusion based on the correlation weight matrix, the weight matrix is ​​multiplied by the feature vectors of the two channels respectively, and then the weighted feature vectors are concatenated to obtain the fused feature vector. The fused features are input into a multi-layer cross feature extraction network with residual connections. The network contains multiple cross feature extraction layers, each of which is composed of a fully connected layer, and residual connections are used between layers to alleviate the gradient vanishing problem. Batch normalization is performed after each layer of feature extraction to standardize the feature distribution. The last layer of features is mapped to obtain the load forecast value for the future period.

[0051] The model training uses a batch size of 32 and an initial learning rate of 0.001. If the validation set loss does not decrease for 5 consecutive cycles, the learning rate is reduced to 0.1 times the original value. An early stopping mechanism is introduced, and training is stopped when the validation set loss improvement is less than 0.1% for 10 consecutive cycles. The training data set is divided into training set, validation set, and test set in a ratio of 8:1:1.

[0052] In this embodiment, the voltage fluctuation detection method based on the sliding window effectively reduces the influence of sampling noise and improves the monitoring accuracy. The capacitance loss calculation method based on the slope of the voltage-time curve realizes the accurate assessment of the health status of the capacitor. The automatic generation function of the health status assessment report provides an important basis for equipment maintenance. The dual-channel feature extraction structure innovatively combines timing information and status information, enabling the model to capture the load change law and equipment operation status at the same time. The dynamic weighted fusion mechanism enhances the adaptive ability of the model and significantly improves the load prediction accuracy. The residual connection structure and batch normalization processing effectively solve the problem of deep network training and ensure the stability of the prediction results.

[0053] In an optional implementation, when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, a pre-switching state check is performed on the backup capacitor group. After the backup capacitor group meets the switching conditions, the switching strategy is determined according to the load prediction model, and the switching strategy includes the optimal switching timing. The steps of using soft switching technology to achieve disturbance-free switching of the duplex capacitor group include: when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, the insulation resistance, equivalent series resistance and leakage current of the backup capacitor group are measured through a self-test circuit; the backup capacitor group is pre-charged with a constant current source, the charging curve is recorded, and the first-order derivative change rate of the charging curve is analyzed to obtain a curve smoothness index; the backup capacitor group is determined according to the insulation resistance, equivalent series resistance, leakage current and curve smoothness index. The invention discloses a method for determining the switching state of a capacitor group; obtaining a load power prediction sequence within a future time window output by a load prediction model; calculating an average load power and a power variance based on the load power prediction sequence, constructing a switching condition discrimination function including a load power deviation term, a power change rate term and a power fluctuation term, and determining an optimal switching time by minimizing the switching condition discrimination function; based on the optimal switching time, connecting a resonant inductor in series in the main circuit, wherein the value of the resonant inductor is related to the resonant angular frequency and the equivalent capacitance; controlling the turn-on time of the auxiliary switch to lag behind the turn-on time of the main switch; configuring an absorption circuit in the cut-off branch of the main working capacitor group, wherein the parameters of the absorption circuit are related to the voltage rise time during the shutdown process; realizing a zero current switch through the energy oscillation between the resonant inductor and the standby capacitor group, realizing a zero voltage switch through the absorption circuit, and completing the switching operation of the duplex capacitor group.

[0054] For example, when the health status assessment report shows that the health index of the main working capacitor group exceeds 5% and enters the alarm state, the system automatically starts the pre-switching state check process of the backup capacitor group. The key parameters of the backup capacitor group are measured by a high-precision self-test circuit, and the insulation resistance value is 500MΩ, the equivalent series resistance is 25mΩ, and the leakage current is 0.5mA. These parameter values ​​are all within the safe operating range and preliminarily meet the switching conditions.

[0055] The pre-charge test is performed using a constant current source, the charging current is set to 2A, and the sampling frequency is 10kHz to record the voltage data during the charging process. The first-order derivative change rate of the charging curve is analyzed, and the curve smoothness index of the change rate is calculated.

[0056] The load prediction model outputs a load power prediction sequence with a 1-minute interval within the next 2 hours. For example, the average load power is 50kW and the power variance is 2.5kW. The switching condition discriminant function comprehensively considers three aspects: the deviation between the predicted load and the rated load does not exceed 20%, the power change rate is less than 5% per minute, and the power fluctuation amplitude is within 10% of the rated value. Through optimization calculation, it is determined that the best switching time is in the period when the load is relatively stable.

[0057] The soft switching technology uses resonant inductor to assist switching. For the 500V / 100μF capacitor bank, a 200μH resonant inductor is selected with a resonant frequency of 5kHz. The main switch uses IGBT (insulated gate bipolar transistor) with a rated voltage of 1200V and a rated current of 400A. The auxiliary switch turn-on delay is set to 50μs to ensure that the main switch is fully turned on. The absorption circuit is composed of an RC parallel branch with a resistance value of 10Ω and a capacitance value of 1μF, which can control the voltage rise time within 2μs.

[0058] The switching process is divided into three stages: first, the main switch is turned on, and the resonant inductor and the backup capacitor group begin to oscillate energy; when the current passes through zero, the auxiliary switch is turned on, and the current of the main working capacitor group naturally decays; finally, the main working capacitor group branch is disconnected, and the absorption circuit absorbs the remaining energy.

[0059] The switching control is implemented by a digital signal processor, with a sampling frequency of 100kHz and a control period of 10μs. The dead time of the drive circuit is set to 1μs to prevent the bridge arm from being connected. The gate drive of the switching device is isolated by an optocoupler, with a drive voltage of +15V / -5V, and the rise time and fall time are both less than 100ns.

[0060] The system sets multiple protection mechanisms: the overcurrent protection threshold is 1.5 times the rated current, the overvoltage protection threshold is 1.2 times the rated voltage, and the temperature protection threshold is 85°C. Once the protection is triggered, the system immediately locks and cuts off the PWM signal, and records the fault information at the same time.

[0061] In this embodiment, disturbance-free switching of the duplex capacitor bank is achieved, effectively solving the voltage spike and current impact problems in traditional hard switch switching. The pre-switching state check of the standby capacitor bank ensures the reliability of the switching operation. The optimal switching timing selection based on load prediction significantly reduces the impact of the switching process on the system. The application of soft switching technology reduces switching losses and improves system efficiency. Multiple protection mechanisms and real-time monitoring functions ensure the safety of the switching process.

[0062] In an optional embodiment, the average load power and power variance are calculated based on the load power prediction sequence, and a switching condition judgment function including a load power deviation term, a power change rate term, and a power fluctuation term is constructed. The step of determining the optimal switching timing by minimizing the switching condition judgment function includes: using an exponentially weighted moving average method to smooth the load power prediction sequence to obtain a smoothed average load power; based on the smoothed average load power, a power variance is calculated recursively; a switching condition judgment function including a power deviation term, a power change rate term, and a power fluctuation term is constructed, wherein the power deviation term is calculated based on the deviation between the current load power and a preset optimal switching power threshold, the power change rate term is calculated based on the time derivative of the load power, and the power fluctuation term is calculated based on the power variance; by setting an objective function for maximizing the switching success rate, a switching condition judgment function is used. The particle swarm optimization algorithm iteratively optimizes the non-negative normalized weight coefficients of the power deviation term, the power change rate term and the power fluctuation term; the product of the optimized non-negative normalized weight coefficient and the corresponding term is summed to obtain the final switching condition discriminant function; with minimization of the switching condition discriminant function as the optimization goal, an improved gradient descent method is used to solve the optimal switching timing, and the improved gradient descent method includes calculating the adaptive step size by dividing the initial step size by the increasing term of the number of iterations; in each iteration, the gradient of the switching condition discriminant function is calculated based on the current moment, and the switching moment is updated according to the calculated gradient direction and the adaptive step size; the switching condition discriminant function value corresponding to the updated switching moment is calculated, and when the difference between the switching condition discriminant function values ​​of two adjacent iterations is less than the function value convergence threshold and the difference between the switching moments of two adjacent iterations is less than the time convergence threshold, the switching moment obtained in the current iteration is determined as the optimal switching moment.

[0063] Exemplarily, the smoothing of the load power forecast sequence uses an exponentially weighted moving average method to obtain the average load power. For example, the forecast sequence contains 120 power values ​​at one-minute intervals, and the initial smoothing factor is set to 0.9. Each new forecast value is assigned a weight of 0.9, and the historical smoothed value is assigned a weight of 0.1. This method can effectively suppress short-term fluctuations and retain long-term trends.

[0064] The power variance calculation adopts a recursive method to avoid storing all historical data. Every time a new predicted power value is received, the mean is updated and the square of the deviation from the new mean is calculated, and a weight of 0.9 is assigned to the historical variance value. In actual measurements, the power variance is maintained at around 2.5kW under stable conditions, and can reach more than 10kW when the load changes suddenly.

[0065] The switching condition discrimination function integrates three key indicators. The power deviation term reflects the gap between the current load and the optimal switching power. The optimal switching power threshold is set to 70% of the rated power. The power change rate term represents the load change speed, which is obtained by dividing the power difference between adjacent moments by the time interval. The power fluctuation term is represented by the standard deviation, which is calculated by the square root of the variance.

[0066] The particle swarm algorithm is used for weight optimization, with the population size set to 50 and the maximum number of iterations set to 200. The particle position represents three weight coefficients, with a value range of 0-1 and a sum of 1. The fitness function is designed based on the historical switching success rate, and switching success is defined as a voltage fluctuation of less than 2% during the process. After optimization, the weight of the power deviation term is 0.5, the weight of the power change rate term is 0.3, and the weight of the power fluctuation term is 0.2.

[0067] The improved gradient descent method is used to solve the optimal switching time. The initial step size is set to 60 seconds, and each iteration is divided by the square root of the number of iterations for adaptive adjustment. The gradient calculation uses the central difference method to calculate the difference between the discriminant function values ​​1 minute before and after the current moment.

[0068] The iteration termination condition sets two thresholds: the discriminant function value convergence threshold is 0.001, and the time convergence threshold is 1 second. In a certain optimization, starting from the initial moment, after 15 iterations, the discriminant function value dropped from 0.25 to 0.008, and the difference between adjacent iteration times was 0.5 seconds, which met the termination condition. The optimal switching time finally determined is in the stable range of load power 50kW, power change rate less than 1kW / min, and power fluctuation less than 2kW.

[0069] In this embodiment, a complete mechanism for determining the optimal switching timing is established. The exponentially weighted moving average method effectively smooths the power prediction sequence and improves the reliability of decision-making. The comprehensive discriminant function of the three indicators comprehensively considers the switching conditions. The weight distribution optimized by the particle swarm algorithm improves the accuracy of the discrimination. The improved gradient descent method ensures the efficiency and convergence of the solution. This scheme provides an important decision-making basis for the safe and reliable switching of the duplex capacitor group and has significant engineering practical value.

[0070] In an optional implementation, the switched standby capacitor group is set to the main working mode, the original main working capacitor group enters the standby state and executes a self-recovery program, the state data of the switching process is updated to the health state database, and based on the historical data of the health state database and the calculation results of the remaining life prediction model, the step of optimizing the switching strategy includes: after completing the switching operation, the switched standby capacitor group is set to the main working mode; the original main working capacitor group is switched to the standby state and a self-recovery program is executed, and the self-recovery program includes executing a charge and discharge balance test to evaluate the consistency of the single capacitor capacitance, eliminating residual charge through a constant current charge and discharge cycle, and executing an internal resistance test to evaluate the change of the equivalent series resistance. status, and perform capacitor balancing compensation; collect status data during the switching process and update it to a health status database, wherein the status data during the switching process includes voltage and current waveform data, transient response characteristics, switching operation timing information, temperature change data, and test results of a self-recovery program; based on the status data during the switching process, update the health index of the original main working capacitor group, and based on the health index, optimize the execution parameters of the self-recovery program; combine the historical data in the health status database and the calculation results of the remaining life prediction model to optimize the timing selection standard of the switching strategy and the non-negative normalized weight coefficient of the switching condition discrimination function; and use the optimized switching strategy to guide the next capacitor group switching operation.

[0071] Exemplarily, after the switching operation is completed, the system sends a working mode switching instruction to the main controller through the communication bus, and configures the newly put into use backup capacitor bank as the main working mode. The controller updates the working status flag, adjusts the sampling channel configuration, and ensures the real-time collection and processing of the measurement data. During the working mode switching process, the system maintains a stable voltage of 500V and a ripple factor of less than 1%.

[0072] The original main working capacitor group enters the self-recovery program. First, a charge and discharge balance test is performed. A 2A constant current is used to charge each monomer capacitor to the rated voltage, and the voltage rise curve is recorded. The residual charge elimination adopts a cyclic charge and discharge method. The charging current is limited to 1A, and the discharge resistance is selected as 100Ω. Each cycle lasts for 10 minutes, and the residual voltage is monitored by a high-precision voltmeter. The internal resistance test adopts the AC impedance method with a test frequency of 1kHz. Record the equivalent series resistance value of each monomer capacitor. The equivalent series resistance value of a normal capacitor is about 25mΩ. Monomer capacitors exceeding 35mΩ need to be balanced and compensated, and the balanced compensation is achieved through a parallel balancing circuit.

[0073] The acquisition of voltage and current waveform data is realized through a high-speed sampling device. The instantaneous voltage and current values ​​during the switching process are sampled and recorded to obtain a complete waveform curve. The transient response characteristics mainly include key indicators such as overvoltage amplitude, overcurrent amplitude, and oscillation decay time during the switching action. By analyzing the collected waveform data, these characteristic quantities are extracted and recorded in the database characteristic indicator table to evaluate the dynamic performance of the switching process. The switching operation timing information records the time nodes of each link in the entire process from the issuance of the control signal to the completion of the switching, including the control command issuance time, mechanism action time, main circuit conduction time, auxiliary contact action time, etc. The temperature change data is collected by an infrared temperature measuring device to collect the temperature rise of the contacts and other key parts during the switching process. The test results of the self-recovery program include the self-test status of each function after the switching. The voltage and current waveform data, transient response characteristics, switching operation timing information, temperature change data, and the test results of the self-recovery program are updated to the health status database.

[0074] The health status database is built using a relational database, which mainly includes equipment basic information table, operation status table, switching record table and alarm information table. The data generated by each switching operation is recorded, including waveform data, operation records and test results. The health index is calculated based on these data. The self-recovery program parameter optimization is implemented based on the health index, and the number of charge and discharge cycles and low balance compensation current are adjusted according to the health index.

[0075] The remaining life prediction model is trained based on historical data from the health status database. For example, the remaining life of a capacitor bank under normal use is about 20,000 hours, and the remaining life drops to 5,000 hours after reaching the warning threshold. When optimizing the switching strategy, for capacitor banks with a remaining life of less than 10,000 hours, the weight of the power fluctuation term in the switching condition discriminant function is increased to 0.3.

[0076] In this embodiment, a complete capacitor bank switching post-processing mechanism is established. The self-recovery program effectively improves the performance status of the standby capacitor bank and extends the service life of the equipment. The real-time collection and storage of status data provides a reliable basis for equipment diagnosis. Parameter optimization based on the health index improves the self-recovery effect. The combination of remaining life prediction and switching strategy optimization ensures the long-term stable operation of the system. This solution has important application value in the field of large-scale power electronic equipment maintenance and significantly improves equipment reliability.

[0077] Figure 2 FIG. 1 is a schematic diagram of a capacitor optimization switching system based on timing control according to an embodiment of the present invention. Figure 2As shown, the system includes: a first unit, which is used to collect voltage signals, current signals and temperature signals of a duplex capacitor group in a railway system, and establish a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a standby capacitor group; based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model of the duplex capacitor group, and a backup protection strategy for the standby capacitor group is established; a second unit is used to monitor the voltage fluctuation rate and capacitance loss rate of the main working capacitor group in real time, generate a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establish a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report shows that the main working capacitor group When the state of is abnormal, according to the backup protection strategy, the pre-switching state check is performed on the standby capacitor group. After the standby capacitor group meets the switching conditions, the switching strategy is determined according to the load prediction model. The switching strategy includes the optimal switching time, and the soft switching technology is used to realize the disturbance-free switching of the duplex capacitor group; the third unit is used to set the switched standby capacitor group as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, the state data of the switching process is updated to the health status database, and the switching strategy is optimized based on the historical data of the health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of the main working capacitor group or the standby capacitor group is lower than the preset threshold, the early warning signal is automatically triggered and the replacement suggestion is prompted.

[0078] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0079] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0080] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A capacitor optimization switching method based on timing control, characterized in that: include: The voltage signal, current signal and temperature signal of the duplex capacitor group in the railway system are collected to establish a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a spare capacitor group; based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model of the duplex capacitor group, and a backup protection strategy for the spare capacitor group is established; Monitor the voltage fluctuation rate and capacitance loss rate of the main working capacitor group in real time, generate a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establish a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report shows that the state of the main working capacitor group is abnormal, perform a pre-switching state check on the standby capacitor group according to the backup protection strategy, and after the standby capacitor group meets the switching conditions, determine the switching strategy according to the load prediction model, the switching strategy includes the best switching time, and adopts soft switching technology to achieve disturbance-free switching of the duplex capacitor group; set the switched standby capacitor group as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, update the state data of the switching process to the health status database, and optimize the switching strategy based on the historical data of the health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of the main working capacitor group or the standby capacitor group is lower than the preset threshold, automatically trigger a warning signal and prompt a replacement suggestion; Based on the health status database, a deep learning algorithm is used to construct a remaining life prediction model for the duplex capacitor group, and the steps of establishing a backup protection strategy for the standby capacitor group include: extracting degradation characteristics of the duplex capacitor group from the health status database, the degradation characteristics including the equivalent series resistance change rate calculated by the voltage and current phase difference, the capacitance loss rate based on the charge and discharge curve slope analysis, the leakage current density using the ratio of steady-state leakage current to rated voltage, and the temperature stress index based on the cumulative effect of thermal cycles; using temperature stress cycle tests, voltage stress tests, and ripple current superposition tests to accelerate aging of the duplex capacitor group, obtaining actual remaining life data under the end-of-life criterion, and establishing a correlation between the degradation characteristics and the actual remaining life. The remaining life prediction model is constructed based on deep learning, wherein the remaining life prediction model includes a multi-layer stacked bidirectional long short-term memory network layer for capturing the temporal feature dependency relationship, and a multi-head attention mechanism layer for adaptively adjusting the feature weight, wherein the multi-head attention mechanism layer includes a time attention unit and a feature attention unit; the remaining life prediction model is trained using a weighted mean square error loss function, wherein the loss function includes a remaining life prediction error term and a capacitor group degradation rate error term, and a dynamic adaptive weight coefficient of the loss function is determined by iterative optimization; based on the output result of the trained remaining life prediction model, a backup A backup protection strategy for a capacitor group, the backup protection strategy including a hardware overcurrent protection and overvoltage protection circuit, and a thermistor lockout protection circuit; a weighted mean square error loss function is used to train the remaining life prediction model, the loss function including a remaining life prediction error term and a capacitor group degradation rate error term, and the step of determining a dynamic adaptive weight coefficient of the loss function through iterative optimization includes: the remaining life prediction error term is the square difference between the predicted remaining life value and the actual remaining life value, the capacitor group degradation rate error term is the square difference between the predicted degradation rate and the actual degradation rate, the predicted degradation rate is obtained by dividing the difference between the predicted remaining life values ​​at adjacent moments by the time interval, and the actual degradation rate is obtained by dividing the difference between the adjacent remaining life values ​​by the time interval. The difference between the actual remaining life values ​​at the moment and the time interval is obtained by dividing the difference by the time interval; iteratively optimizing the dynamic adaptive weight coefficient, including calculating the partial derivative of the loss function with respect to the dynamic adaptive weight coefficient, updating the dynamic adaptive weight coefficient according to the direction and learning rate of the partial derivative, and determining the optimal dynamic adaptive weight coefficient when the error change rate of the validation set in the training data set of multiple consecutive iterations is less than a preset threshold; training the remaining life prediction model based on the loss function with the optimal dynamic adaptive weight coefficient, selecting training samples in a batch iterative manner, and adaptively adjusting the learning rate of the network parameters of each layer of the remaining life prediction model according to the gradient information of each batch, and optimizing and updating the network parameters of the remaining life prediction model;An early stopping mechanism is introduced during the training process, and the training process is stopped when the loss value on the validation set has no convergence trend after multiple rounds of continuous training; the voltage fluctuation rate and capacitance loss rate of the main working capacitor group are monitored in real time, and a health status assessment report of the main working capacitor group is generated according to the voltage fluctuation rate and capacitance loss rate. Based on the health status assessment report and the real-time working parameters of the duplex capacitor group, the steps of establishing a load prediction model include: collecting voltage data of the main working capacitor group, performing noise reduction processing on the voltage data through a sliding window, and calculating the ratio of the voltage difference between adjacent sampling points to the rated voltage to obtain the voltage fluctuation rate; at the end of each working cycle, the actual capacitance is calculated based on the slope of the voltage-time curve during the charging process of the main working capacitor group, and the actual capacitance is converted into The difference ratio between the capacitance value and the rated capacitance value is used to obtain the capacitance loss rate; the voltage fluctuation rate and the capacitance loss rate are multiplied by the preset weight coefficient respectively and then summed to obtain the health index, and a health status assessment report containing the current health index value, change trend and alarm level is generated according to the health index; the real-time working parameters of the duplex capacitor group are obtained, including input current, switching frequency, ambient temperature and historical load data; the current health index value, change trend and alarm level in the health status assessment report are normalized with the real-time working parameters of the duplex capacitor group to obtain standardized feature data; a dual-channel feature extraction layer of the load prediction model is constructed, and the standardized feature data is input into the timing channel and the state channel respectively, wherein the timing channel A one-dimensional convolutional neural network is used to process historical load data, and the state channel uses a multi-layer perceptron to process the current health index value and real-time working parameter data; a feature fusion layer of the load prediction model is constructed, and the correlation weight matrix between the output features of the timing channel and the state channel is calculated, and the output features of the two channels are dynamically weighted and fused based on the correlation weight matrix; the weighted fused features are input into a multi-layer cross feature extraction network with a residual connection structure, and normalization is performed after each layer of feature extraction to output the load prediction result; when the health status assessment report shows that the state of the main working capacitor group is abnormal, the backup capacitor group is pre-switched according to the backup protection strategy, and after the backup capacitor group meets the switching conditions, the backup capacitor group is switched to the backup capacitor group. Determine the switching strategy according to the load prediction model, the switching strategy includes the best switching time, and the steps of using soft switching technology to achieve disturbance-free switching of the duplex capacitor group include: when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, measure the insulation resistance, equivalent series resistance and leakage current of the standby capacitor group through the self-test circuit; use a constant current source to pre-charge the standby capacitor group, record the charging curve, and analyze the first-order derivative change rate of the charging curve to obtain the curve smoothness index; determine the switching state of the standby capacitor group according to the insulation resistance, equivalent series resistance, leakage current and curve smoothness index; obtain the load power prediction sequence in the future time window output by the load prediction model;The average load power and power variance are calculated based on the load power prediction sequence, and a switching condition discrimination function including a load power deviation term, a power change rate term and a power fluctuation term is constructed, and the optimal switching timing is determined by minimizing the switching condition discrimination function; based on the optimal switching timing, a resonant inductor is connected in series in the main circuit, and the value of the resonant inductor is related to the resonant angular frequency and the equivalent capacitance; the turn-on time of the auxiliary switch is controlled to lag behind the turn-on time of the main switch; an absorption circuit is configured in the cut-off branch of the main working capacitor group, and the parameters of the absorption circuit are related to the voltage rise time during the shutdown process; zero current switching is achieved through the energy oscillation between the resonant inductor and the standby capacitor group, and zero voltage switching is achieved through the absorption circuit, so as to complete the duplex circuit. The switching operation of the capacity group; based on the load power prediction sequence, the average load power and the power variance are calculated, and a switching condition discrimination function including a load power deviation term, a power change rate term and a power fluctuation term is constructed, and the step of determining the best switching time by minimizing the switching condition discrimination function includes: using an exponentially weighted moving average method to smooth the load power prediction sequence to obtain a smoothed average load power; based on the smoothed average load power, the power variance is calculated in a recursive manner; constructing a switching condition discrimination function including a power deviation term, a power change rate term and a power fluctuation term, wherein the power deviation term is calculated based on the deviation between the current load power and a preset optimal switching power threshold, and the power change rate term is calculated based on the current load power and the preset optimal switching power threshold. The power fluctuation term is calculated based on the time derivative of the load power, and the power fluctuation term is calculated based on the power variance; by setting the objective function of maximizing the switching success rate, the particle swarm optimization algorithm is used to iteratively optimize the non-negative normalized weight coefficients of the power deviation term, the power change rate term and the power fluctuation term; the final switching condition discriminant function is obtained by summing the product of the optimized non-negative normalized weight coefficient and the corresponding term; with minimizing the switching condition discriminant function as the optimization goal, the improved gradient descent method is used to solve the optimal switching timing, and the improved gradient descent method includes calculating the adaptive step size by dividing the initial step size by the increasing term of the number of iterations; in each iteration, the gradient of the switching condition discriminant function is calculated based on the current moment, and the calculated The switching time is updated by the gradient direction and the adaptive step size; the switching condition discrimination function value corresponding to the updated switching time is calculated, and when the difference between the switching condition discrimination function values ​​of two adjacent iterations is less than the function value convergence threshold and the difference between the switching times of two adjacent iterations is less than the time convergence threshold, the switching time obtained by the current iteration is determined as the optimal switching time; the standby capacitor group after switching is set as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, and the state data of the switching process is updated to the health status database. Based on the historical data of the health status database and the calculation results of the remaining life prediction model, the steps of optimizing the switching strategy include: after completing the switching operation, the standby capacitor group after switching is set as the main working mode;Switch the original main working capacitor group to the standby state and execute the self-recovery program, which includes executing the charge and discharge balance test to evaluate the consistency of the single capacitor value, eliminating the residual charge through the constant current charge and discharge cycle, executing the internal resistance test to evaluate the change of the equivalent series resistance, and executing the capacitor balance compensation; collect the state data in the switching process and update it to the health state database, the state data in the switching process includes the voltage and current waveform data, transient response characteristics, switching operation timing information, temperature change data and the test results of the self-recovery program; update the health index of the original main working capacitor group based on the state data in the switching process, and optimize the execution parameters of the self-recovery program based on the health index; optimize the timing selection criteria of the switching strategy and the non-negative normalized weight coefficient of the switching condition discrimination function in combination with the historical data in the health state database and the calculation results of the remaining life prediction model; use the optimized switching strategy to guide the next capacitor group switching operation. ; 2. A capacitor optimization switching system based on timing control, used to implement the method described in claim 1, characterized in that: include: The first unit is used to collect voltage signals, current signals and temperature signals of a duplex capacitor group in a railway system, and establish a health status database of the duplex capacitor group, wherein the duplex capacitor group includes a main working capacitor group and a spare capacitor group; based on the health status database, a remaining life prediction model of the duplex capacitor group is constructed by using a deep learning algorithm, and a backup protection strategy for the spare capacitor group is established; The second unit is used to monitor the voltage fluctuation rate and capacitance loss rate of the main working capacitor group in real time, generate a health status assessment report of the main working capacitor group according to the voltage fluctuation rate and capacitance loss rate, and establish a load prediction model based on the health status assessment report and the real-time working parameters of the duplex capacitor group; when the health status assessment report shows that the state of the main working capacitor group is abnormal, according to the backup protection strategy, perform a pre-switching state check on the standby capacitor group, and after the standby capacitor group meets the switching conditions, determine the switching strategy according to the load prediction model, the switching strategy includes the best switching time, and adopts soft switching technology to achieve disturbance-free switching of the duplex capacitor group; The third unit is used to set the switched backup capacitor group as the main working mode, the original main working capacitor group enters the standby state and executes the self-recovery program, the state data of the switching process is updated to the health status database, and the switching strategy is optimized based on the historical data of the health status database and the calculation results of the remaining life prediction model; when the remaining life prediction model predicts that the remaining life of the main working capacitor group or the backup capacitor group is lower than the preset threshold, the warning signal is automatically triggered and a replacement suggestion is prompted.

3. An electronic device, characterized in that: include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the method described in claim 1.

4. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method described in claim 1 is implemented.

Images