Electromagnetic transient simulation system for power systems containing multiple power electronic switches

Abstract

The present application relates to the technical field of electromagnetic transient simulation, in particular to an electromagnetic transient simulation system of a power system containing multiple power electronic switches, in which the parameters of multiple switch nodes, resistors, inductors and capacitors are numbered and topological serial numbers are analyzed, an impedance conversion and admittance construction method is used to form a connection mapping matrix, a genetic algorithm is applied to form a sequence structure by coding control variables, a variable sequence is constructed by switching time points and topological connections, and a combination is selected by performing a combined evaluation and screening of errors and time length, after collecting energy storage voltage and filter current, a time matrix is constructed, an element vector and a covariance projection coefficient are extracted, an input matrix is constructed, harmonic information is fused in a unified time step, a residual network structure is introduced to perform step-by-step iterative updating of initial weights, dynamic adjustment is completed by combining residual offset calculation and local error compression at each step, and a continuous and complete reconstructed waveform is output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetic transient simulation, in particular to an electromagnetic transient simulation system of a power system containing multiple power electronic switches. BACKGROUND

[0002] The technical field of electromagnetic transient simulation aims to model and simulate the fast electromagnetic processes in a power system caused by power electronic switches, faults, control behaviors, etc. through numerical calculation means, to accurately depict the dynamic change behavior of voltage, current, electromagnetic energy on the time scale of microseconds to milliseconds, to evaluate the stability and reliability of the system under transient disturbance, to analyze the electromagnetic interference propagation path, to verify the effectiveness of the protection strategy, and to assist the design optimization of power electronic devices and control systems.

[0003] The electromagnetic transient simulation system of a power system containing multiple power electronic switches aims to model and simulate the action process of these switching devices in the system, capture the transient changes of current, voltage, electromagnetic energy caused by changes in device driving signals, device conduction and turn-off, network topology transformation, etc. behavior, evaluate the adaptability of system voltage sag, current impact, electromagnetic interference and system protection strategy, and provide reliable support for system design, operation evaluation and fault analysis.

[0004] The prior art mainly simulates the response of power electronic switches through fixed modeling and static parameter setting in the construction of electromagnetic transient simulation process, and the processing logic does not include effective evolution path and feedback iteration mechanism, resulting in a serious dependence on manual setting in the control parameter matching process and a lack of optimization process, making it difficult to complete multi-index trade-off. The current simulation mostly uses high sampling frequency and long period calculation mode, which greatly increases the calculation load but still cannot guarantee the effectiveness of data extraction, and the dimension reduction stage only performs simplification processing based on the number of truncated principal components, ignoring the contribution proportion of each vector to the total amount of information, resulting in insufficient information retention. In the reconstruction process, a simple interpolation splicing method is often used to merge each feature segment, forming waveform fracture and boundary discontinuity problems, which systematically restricts the simulation accuracy, scheduling efficiency and parameter traceability. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and to provide an electromagnetic transient simulation system of a power system containing multiple power electronic switches.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: the electromagnetic transient simulation system of a power system containing multiple power electronic switches comprises:

[0007] The topology analysis module: based on the switching node, inductance, capacitance, resistance parameters in the multi-power electronic switching device, impedance calculation and admittance conversion are carried out by extracting the switching number and topology serial number, the matrix elements are normalized and the node connection structure is mapped, and the network topology matrix is generated;

[0008] The parameter evolution module: based on the network topology matrix, genetic algorithm is adopted, combined with the switching device time sequence, variable sequence is established by extracting the switching time point and connection information, error and time length are recorded in simulation operation, sorting is judged to retain and mutate, and the control parameter optimization combination is generated;

[0009] The order reduction mapping module: based on the control parameter optimization combination, the energy storage unit voltage and filter current sequence are collected, the time matrix is constructed and the singular value calculation and characteristic vector selection are carried out, the projection coefficient is calculated and the residual adjustment dimension is judged, and the order reduction low-dimensional expression is obtained;

[0010] The state reconstruction module: based on the order reduction low-dimensional expression, the residual network is adopted, the input matrix is constructed combined with the switching state and harmonic, the combination sequence is constructed matching the time step and coefficient, the offset difference value is calculated to adjust the weight value iteration update, and the reconstructed waveform sequence list is obtained;

[0011] The flow scheduling module: based on the reconstructed waveform sequence list, combined with the definition of task unit of capacitor charging and discharging and resonance stage, the task time length and waiting proportion are counted, the parallel condition and trigger sequence are judged, and the scheduling trigger structure is established.

[0012] As a further scheme of the application, the topology analysis module comprises:

[0013] The node parameter mapping submodule: based on the switching node, inductance, capacitance, resistance parameters in the multi-power electronic switching device, the switching number is extracted and the topology serial number is calibrated by matching the node information, the complex impedance value array is constructed by converting the electrical parameters, the admittance real part and imaginary part are output by completing the numerical conversion, the admittance elements are formed into a vector set by proportional scaling, and the normalized admittance dataset is generated;

[0014] The connection structure construction submodule: based on the normalized admittance dataset, the node number and vector correspondence are processed by index mapping, the node pair connection matrix is constructed by matrix splicing, the admittance and structure two-dimensional mapping diagram is presented by element fusion, the multi-node link information set is output, and the network topology matrix is obtained.

[0015] As a further scheme of the application, the parameter evolution module comprises:

[0016] Switching sequence generation submodule: based on the network topology matrix, the switching time points of the switching device and the corresponding connection node number are extracted, a node index mapping table is constructed by combining the topology column label and genetic algorithm, and time sequence axis alignment operation is performed to correct the time points of each channel, connection information screening operation is performed to remove invalid node state, and full time domain effective switching section is arranged, variable state list is spliced by time sequence, complete sequence structure is constructed and missing field is filled, and variable sequence set is obtained;

[0017] Error screening submodule: based on the variable sequence set, execute circuit model simulation calculation output waveform group, calculate the error curve between the target waveform and the actual response through residual extraction operation, generate sequence merit and demerit grade according to error integral area sorting, execute running time statistics and establish secondary index matrix, combine error grade and time grade by weight, extract the top-ranked sequence and add a retention marker, generate a mutation marker matrix synchronously, and obtain the preferred parameter index;

[0018] Preferred combination generation submodule: based on the preferred parameter index, extract the retention state and the corresponding variable number index, execute mapping matching operation to associate the original control parameter data, form a complete control vector set through combination reconstruction operation, superimpose the mutation marker matrix to generate multiple groups of structure disturbance candidates, aggregate different control vector sets and corresponding mutation flag labels, construct a two-dimensional control parameter set array, and obtain the control parameter preferred combination.

[0019] As a further scheme of the application, the genetic algorithm first takes the variable sequence set obtained after time sequence axis alignment and invalid node screening as the initial population, with each sequence corresponding to a chromosome, then takes the weighted sum of simulation error integral area and running time as the fitness function, and evaluates the fitness of each chromosome in the population, then selects the parent from the high fitness chromosome through the selection operator, and recombines the time point and node number mapping table of the parent sequence through single-point or multi-point crossover operator to obtain the offspring chromosome, then uses the mutation operator to make a slight disturbance to the switching time or connection information of part of the offspring chromosomes, and sorts the parent and offspring according to the fitness, and retains the top N optimal chromosomes as the new generation population, and repeats the iteration until the convergence condition is met and the preset upper limit of the number of generations is reached, and outputs the optimal control parameter combination scheme.

[0020] As a further scheme of the application, the genetic algorithm is according to the formula:

[0021] ;

[0022] Wherein: represents the updated arrangement result of the candidate node index sequence in the th genetic iteration, represents the updated arrangement result of the candidate node index sequence in the The current permutation state in the next iteration. Represents the global adjustment coefficient. This indicates the weight of the fitness difference component in the index update. This indicates the weight of the channel switching time consistency component. The weights representing the node topological connectivity scores The weight representing the frequency of choices made by historical elites. Indicates the first The fitness value of the individual with the minimum number of switching state conflicts and optimal connectivity performance in the next iteration. Indicates the first Individual in the first The current fitness value in the next iteration Indicates the first The mean squared error of the channel switching time sequence for an individual. Indicates the first The score representing the percentage of connected edges in the node connection matrix constructed by an individual. Indicates the first The frequency ratio of individuals selected as elite individuals in each genetic iteration.

[0023] As a further aspect of the present invention, the order reduction mapping module includes:

[0024] Feature matrix generation submodule: Based on the preferred combination of the control parameters, the voltage and current values ​​of the energy storage unit and the filter are read by voltage and current acquisition, a two-dimensional sequence matrix is ​​constructed by time indexing, the energy distribution is divided by column vector energy decomposition, the principal vector set is screened and identified by peak contribution, and the principal component feature matrix is ​​generated.

[0025] The low-dimensional mapping derivation submodule: Based on the principal component feature matrix, matrix mapping is used to calculate the output coefficient list, residual variance signal accumulation is used to determine the projection dimension, combination screening is used to extract the minimum dimension vector array, and time rearrangement is used to reconstruct the low-dimensional feature sequence to obtain the reduced-order low-dimensional expression.

[0026] As a further aspect of the present invention, the state reconstruction module includes:

[0027] Input matrix construction submodule: Based on the reduced-order low-dimensional expression, the switching device status data and harmonic interference channel signals corresponding to each stage are collected. The data of different frequency bands are uniformly resampled by the equal time step amplitude interpolation method. An index table is constructed to synchronously label the time label and variable name. A three-dimensional matrix structure is formed by mapping the node number and splitting the harmonic frequency band. Tensor expansion operation is performed to compress it into a two-dimensional input structure to obtain the input matrix.

[0028] Weight Iteration Submodule: Based on the input matrix, a residual network is used, an initial weight vector is introduced as a reference value for the superposition coefficient, residual calculation operation is performed to evaluate the waveform offset under the current combination, the alignment index is updated using the maximum offset position, the weight difference between adjacent time steps is corrected by the local error compression method, iterative update is performed and the result of each step is retained in the record sequence, the iteration is repeated until the offset change is lower than the threshold, and the weight iteration sequence is obtained.

[0029] Waveform output submodule: Based on the weighted iterative sequence, feature segments are extracted from the input matrix according to the time dimension and weighted superimposed to generate initial waveform segments. Continuous frame splicing operation is performed to restore the complete waveform sequence. The consistency of overlapping segments in the critical region is verified through the sliding window mode. Boundary transition correction and length normalization processing are completed to generate a stable and traceable waveform output table and obtain the reconstructed waveform sequence table.

[0030] As a further aspect of the present invention, the residual network first inputs a sample set and introduces an initialized weight vector as the initial parameters of each connection layer in the residual network. Then, it is expanded according to a time series structure. Each neural unit performs a combination of linear transformation and activation function operation to generate a predicted output sequence and calculates the difference with the target waveform to obtain the residual sequence. Then, the mean squared error is used as the loss function to quantify the residual size, backpropagates the error information, and updates the network weights. In each iteration, the offset between the current output waveform and the target reference waveform is compared, the offset point is extracted and the time alignment index is adjusted. At the same time, the local weight difference between the previous and next time steps is compressed and adjusted to reduce oscillation. The iteration process continues until the residual change is lower than a set threshold, and the converged weight iteration sequence is output as a stable parameter set.

[0031] As a further aspect of the present invention, the residual network is configured according to the formula:

[0032] ;

[0033] in: Indicates the first The weight vector generated in the next iteration Indicates the first The current weight vector in the next iteration This represents the learning rate of the residual network. This represents the power system input matrix obtained from sampling. Represents the input matrix The transpose operation, Represents the desired target waveform vector. The weighting coefficients of the residual smoothing term are represented. Indicates the first The residual vector of the next iteration This represents the weight coefficient of the regularization term. represents a regularization gradient imposed on the current weight vector, represents a penalty coefficient of a residual variance term, represents a variance value of a residual vector in the i-th iteration, represents a variance value of a residual vector in the i-th iteration, represents a weight coefficient of a history momentum term, represents a momentum term accumulated in the i-th iteration, represents a momentum term accumulated in the i-th iteration.

[0034] As a further scheme of the present application, the flow scheduling module comprises:

[0035] The task unit division sub-module: based on the reconstructed waveform sequence table, a continuous amplitude rising section detection method is adopted, a continuous amplitude falling section identification method is adopted for superimposed discharge section extraction operation, a zero-crossing interval extraction method is adopted for combined resonance section extraction operation, peak and valley point columns are extracted by superimposed equal-interval sampling operation, a time length statistical table is formed by matching the start and end times of each section, and a task timing distribution is obtained.

[0036] The trigger structure construction sub-module: based on the task timing distribution, parallel section numbers are extracted, time sequence number groups are constructed by superimposed trigger sequence generation operation, a sequence chain is established by adjacent task numbers by executing a dependency path arrangement operation, a trigger priority grouping table is constructed, a priority index is generated according to the waiting proportion, and a scheduling trigger structure is obtained.

[0037] Compared with the prior art, the present application has the following advantages and positive effects:

[0038] In the present application, the multi-switch node, resistance, inductance and capacitance parameters are numbered and topological sequence number is analyzed, an impedance conversion and admittance construction method is adopted to form a connection mapping matrix, a genetic algorithm is applied to form a sequence structure by encoding control variables, a variable sequence is constructed by switching time points and topological connections, and a combination is selected by executing error and time length joint evaluation, so that the control parameter evolution process has an automatic screening ability.

[0039] In the present application, after collecting the energy storage voltage and the filter current, a time matrix is constructed, a principal element vector and a covariance projection coefficient are extracted, a projection error residual is dynamically dimensioned to obtain a compressed effective low-dimensional sequence, an input matrix is constructed and harmonic information is fused in a unified time step, a residual network structure is introduced to perform iterative updating processing of the initial weight value at each time step, dynamic adjustment is completed by combining residual offset calculation and local error compression at each step, a weight sequence more consistent with the real response characteristics is obtained, and a continuous and complete reconstructed waveform is output. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The system flowchart of the present application. DETAILED DESCRIPTION

[0041] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0042] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0043] Embodiment: Please refer to Figure 1 The present application provides a technical solution: an electromagnetic transient simulation system of a power system containing multiple power electronic switches comprises:

[0044] Topology analysis module: based on the parameters of each switching node, inductance, capacitance and resistance in the multiple power electronic switching devices, impedance calculation and admittance conversion are performed by extracting switching number and topology number, matrix elements are normalized and node connection structure is mapped, and a network topology matrix is generated;

[0045] Parameter evolution module: based on the network topology matrix, genetic algorithm is adopted, combined with switching device timing, variable sequence is established by extracting switching time point and connection information, error and time length are recorded during simulation operation, sorting is used to determine retention and variation, and control parameter optimization combination is generated;

[0046] Order reduction mapping module: based on the control parameter optimization combination, energy storage unit voltage and filter current sequence are collected, time matrix is constructed and singular value calculation and characteristic vector selection are performed, projection coefficient is calculated and residual adjustment dimension is judged, and order reduction low-dimensional expression is obtained;

[0047] State reconstruction module: based on the order reduction low-dimensional expression, residual network is adopted, input matrix is constructed combined with switching state and harmonics, combination sequence is constructed matching time step and coefficient, offset difference value is calculated to adjust weight value iteration update, and reconstructed waveform sequence table is obtained;

[0048] Process scheduling module: based on the reconstructed waveform sequence table, task units are defined combined with capacitor charging and discharging and resonance stage, task time length and waiting proportion are counted, parallel condition and trigger sequence are judged, and scheduling trigger structure is established.

[0049] The topology analysis module comprises:

[0050] The node parameter mapping submodule: based on the switching node, inductance, capacitance, and resistance parameters in the multi-power electronic switching device, the switching number is extracted by matching the node information, and the topology serial number is calibrated, the complex impedance value array is constructed by converting the electrical parameters, the admittance real part and imaginary part are output by numerical conversion, the admittance elements are processed by proportional scaling to form a vector set, and the normalized admittance dataset is generated;

[0051] The connection structure construction submodule: based on the normalized admittance dataset, the node number and vector correspondence are processed by index mapping, the node pair connection matrix is constructed by matrix assembly, the admittance and structure two-dimensional mapping diagram are presented by element fusion, the multi-node link information set is output, and the network topology matrix is obtained;

[0052] The node parameter mapping submodule: based on the switching node, inductance, capacitance, and resistance parameters in the multi-power electronic switching device, the switching number is extracted by matching the node information, and the topology serial number is calibrated, the complex impedance value array is constructed by converting the electrical parameters, the admittance real part and imaginary part are output by numerical conversion, the admittance elements are processed by proportional scaling to form a vector set, and the normalized admittance dataset is generated;

[0053] The connection structure construction submodule, based on the normalized admittance dataset, uses an index mapping method to process the correspondence between node numbers and vectors. During the binding operation, the node numbers are mapped one-to-one with the node number list in vector order. Each node number is used as the key, and the corresponding admittance vector is written as the value into the mapping dictionary. A matrix assembly method is used to construct the node pair connection matrix. All node pair combinations are traversed, and for each pair, the number information is extracted, and the corresponding admittance vector value is obtained from the mapping dictionary. An assignment operation is performed, and the value is written to the corresponding position in the two-dimensional matrix. After the matrix is ​​constructed, an element fusion method is used to present the two-dimensional mapping diagram of admittance and structure. The real and imaginary parts of the admittance vector are used as input values ​​for the two-dimensional image coordinates, and a corresponding coordinate matrix is ​​constructed in conjunction with the node pair numbers. The image rendering window size and resolution are set, and all valid admittance points are drawn graphically to generate a two-dimensional distribution map. A multi-node link information set is summarized and output. All node pair information involved in the construction is numbered and encapsulated, and written into a triplet structure set in numerical order to obtain the network topology matrix.

[0054] The parameter evolution module includes:

[0055] The switching sequence generation submodule extracts the switching time points of switching devices and their corresponding connection node numbers based on the network topology matrix. It uses a genetic algorithm to construct a node index mapping table by combining topology row and column labels, performs time axis alignment to correct the time points of each channel, performs connection information filtering to remove invalid node states, organizes the effective switching segments in the entire time domain, and constructs a complete sequence structure by splicing the variable state list in time order and filling in missing fields to obtain the variable sequence set.

[0056] Error Filtering Submodule: Based on the variable sequence set, it performs circuit model simulation calculation to output waveform groups, calculates the error curve between the target waveform and the actual response through residual extraction operation, generates sequence quality levels by sorting by error integral area, performs runtime statistics and establishes a secondary index matrix, merges error level and duration level through weight allocation, extracts the top-ranked sequence and adds retention marks, synchronously generates a mutation mark matrix, and obtains the optimal parameter index.

[0057] The optimal combination generation submodule: Based on the optimal parameter index, extract the retained state and the corresponding variable number index, perform mapping matching operation to associate the original control parameter data, form a structurally complete control vector set through combination reconstruction operation, generate multiple sets of structural disturbance candidates by superimposing the mutation label matrix, summarize different control vector sets and corresponding mutation labels, construct a two-dimensional control parameter set array, and obtain the optimal combination of control parameters;

[0058] Switching sequence generation submodule: based on the network topology matrix, the switching time points of the switching device and the corresponding connection node number are extracted, the genetic algorithm is adopted, the population size is set to 100 chromosomes, the length of each chromosome is set to 20 bits, the encoding mode is binary encoding, each chromosome is randomly generated as 0 or 1 when initializing the population, the error sum of squares is used as the basis for calculating the fitness function, each chromosome is traversed item by item to control the variable and the error is calculated in the simulation platform to generate the fitness value, the roulette selection method is used for selection operation, the fitness array is normalized and weighted sampling is performed to generate a new generation of chromosome index, the single-point crossover operation is performed with a crossover probability of 0.8, a crossover point index is randomly generated in each pair of parents and the gene segments after the crossover point are exchanged, the mutation rate is set to 0.02, all bit positions in each chromosome are traversed bit by bit and the bit is flipped with a probability of 0.02 to form a new set of mutated chromosomes, a node index mapping table is constructed combining the topology row label, a mapping dictionary is established by one-to-one correspondence between the node number and the row and column position number in the topology matrix, and the timing axis alignment operation is performed to correct the time points of each channel, all time channels are aligned with the minimum unit and the missing time point state is filled to form a consistent time step list, the connection information screening operation is performed to remove invalid node states, an invalid node list is set and the corresponding records are removed in the state sequence according to the number, the valid switching section in the whole time domain is arranged, the corresponding state set is spliced to generate a unified state sequence after sorting each time point, the missing fields are filled by the field definition template to generate a complete state sequence structure, and a variable sequence set is generated;

[0059] Error screening submodule: based on the variable sequence set, the circuit model simulation calculation output waveform group is executed, the simulation model input parameters include control vector, power condition, sampling time step 1e-6 seconds, simulation total time 0.01 seconds, each variable sequence is executed independently and the waveform output array under the corresponding time axis is collected, the difference between the target waveform and the actual simulation response is calculated by residual extraction operation, the error signal list is obtained by subtracting the simulation value from the current target value at each time point, the error integral area array is formed by numerical integration of all time periods according to the error absolute value, the error level list is generated by sorting the error integral value from small to large, the running time length statistical operation is executed, the total time length of each sequence is calculated and the running time length list is generated, the two-level index matrix is established, the error level weight is set to 0.7 and the running time length weight is set to 0.3, the weighted sum of error level value and time length level value is executed to generate the total score value, the top-ranked sequence index is extracted according to the score and the reserved state is marked, the reserved mark array is constructed and the mutation mark matrix is generated synchronously according to the index, and the optimal parameter index is obtained.

[0060] The optimal combination generation submodule extracts the retained state and corresponding variable number index based on the optimal parameter index. It extracts the control vector corresponding to the variable number in the original control parameter set, performs a mapping matching operation to perform a one-to-one lookup between the variable index and the record in the control parameter table and extracts the corresponding control value, performs a combination reconstruction operation to reassemble the extracted control vector list in a unified order to generate a control vector set, superimposes the mutation marker matrix to generate multiple sets of structural perturbation candidates, performs parameter perturbation operation on the positions with a value of 1 in the marker matrix, adds a perturbation step size of 0.01 in the positive and negative directions on the basis of the corresponding parameters to construct a new parameter set, and forms a candidate vector set together with the original parameters. It summarizes all control vectors and mutation marker labels, organizes them in order into a two-dimensional array structure, and constructs the optimal combination of control parameters.

[0061] The genetic algorithm first uses the set of variable sequences obtained after time axis alignment and invalid node removal as the initial population, with each sequence corresponding to a chromosome. Then, the fitness function is used as the weighted sum of the simulation error integral area and the runtime to evaluate the fitness of each chromosome in the population. Next, the selection operator selects the parent from the high-fitness chromosomes, and the mapping table of time points and node numbers of the parent sequence is recombined by the single-point or multi-point crossover operator to obtain the offspring chromosomes. Subsequently, the mutation operator is used to slightly perturb the switching time or connection information of some offspring chromosomes. At the same time, the parent and offspring are sorted by fitness, and the top N best chromosomes are retained as the new generation population. The iteration is repeated until the convergence condition is met and the preset generation limit is reached, and the optimal combination of control parameters is output.

[0062] Genetic algorithm, according to the formula:

[0063] ;

[0064] in: Indicates the candidate node index sequence at the th The updated permutation results in each genetic iteration. Indicates the candidate node index sequence at the th The current permutation state in the next iteration. Represents the global adjustment coefficient. This indicates the weight of the fitness difference component in the index update. This indicates the weight of the channel switching time consistency component. The weights representing the node topological connectivity scores The weight representing the frequency of choices made by historical elites. Indicates the first The fitness value of the individual with the minimum number of switching state conflicts and optimal connectivity performance in the next iteration. Indicates the first Individual in the first The current fitness value in the secondary iteration, The mean square error of the channel switching time point sequence of the individual, The proportion of connected edges in the node connection matrix constructed by the individual, The proportion of connected edges in the node connection matrix constructed by the individual, The proportion of connected edges in the node connection matrix constructed by the individual, The proportion of connected edges in the node connection matrix constructed by the individual, The frequency ratio of the individual being selected as an elite individual in the previous genetic iteration;

[0065] Execution process: first, extract the key event timestamp recorded by each sample in multiple processing channels from multiple monitoring nodes, and construct the initial node index sequence , then calculate the fitness value of the current sequence under the indexes of task flow integrity, sampling path rationality, and record time consistency , and select the individual with the best fitness value from all candidates to obtain , then evaluate the time balance of each node on the event flow axis, generate the time consistency parameter through the standard deviation of each channel key time point , calculate the actual connectivity composition index of the node in the sample processing path , that is, whether the node pair in the task has established a physical or logical transmission path, and at the same time, count the frequency ratio of the node sequence being selected as a high-quality sample in the previous evolution process , form the historical selection weight, and multiply it by the normalized weight 、 、 、 , comprehensively express the performance ability of a candidate sample processing sequence in traceability, stability, structure rationality and time continuity, multiply the weighted result by , and add it to the current node index sequence , and finally generate the updated sequence , realize the evolution optimization of the sample processing path and the high-precision construction of the traceable sequence.

[0066] The order reduction mapping module includes:

[0067] Feature matrix generation submodule: based on the optimal combination of control parameters, the voltage and current acquisition method is used to read the energy storage unit voltage and filter current values, a two-dimensional sequence matrix is constructed using time indexing, energy distribution is divided using column vector energy decomposition, a principal component feature matrix is generated by identifying the principal vector set using peak value contribution screening.

[0068] Low-dimensional mapping derivation submodule: based on the eigenmatrix, the output coefficient list is calculated by matrix mapping, the projection dimension is judged by residual variance signal accumulation, the minimum dimension vector array is refined by combination screening, the low-dimensional feature sequence is reconstructed by time rearrangement, and the reduced low-dimensional expression is obtained;

[0069] Eigenmatrix generation submodule: based on the control parameter optimization combination, the fast Fourier transform algorithm is adopted, the sampling rate is set to 1000 Hz, the sampling window length is 256 samples, the overlap is 128 samples, the energy storage unit voltage value and filter current value are read and stored in the time sequence buffer, a two-dimensional sequence matrix is constructed using the sampling timestamp, the fast Fourier transform operation is performed on the extracted vectors by column, the amplitude and phase of each frequency component are calculated, the peak value detection algorithm is used to set the threshold value to 0.8 times the maximum amplitude to screen the main frequency index, the corresponding frequency spectrum vector is extracted according to the screening index to form a main frequency vector set, and the eigenmatrix is generated;

[0070] Low-dimensional mapping derivation submodule: based on the eigenmatrix, the principal component analysis algorithm is adopted, the maximum principal component number is set to 5, the characteristic matrix is first processed by row centering, the covariance matrix of the centering matrix is calculated, the eigenvalue and eigenvector solving operation is performed, the eigenvalue list and eigenvector matrix are obtained, the eigenvalues are arranged in descending order to calculate the cumulative variance contribution rate, the cumulative variance threshold is set to 0.90 to screen the minimum principal component number and extract the corresponding eigenvector to generate the projection matrix, the projection matrix is used to perform matrix multiplication operation on the centering matrix, and the low-dimensional feature sequence is generated by rearranging in the original time sequence order, and the reduced low-dimensional expression is obtained.

[0071] The state reconstruction module includes:

[0072] Input matrix construction submodule: based on the reduced low-dimensional expression, the corresponding switching device state data and harmonic interference channel signals of each stage are collected, the different frequency band data is uniformly resampled by using the equal time step value interpolation method, the index table is constructed to synchronize the time label and variable name, the three-dimensional matrix structure is formed by node number mapping and harmonic frequency band splitting, the tensor expansion operation is performed to compress into a two-dimensional input structure, and the input matrix is obtained;

[0073] Weight iteration submodule: based on the input matrix, the residual network is adopted, the initial weight vector is introduced as the reference value of the superposition coefficient, the residual calculation operation is performed to evaluate the waveform offset degree under the current combination, the alignment index is updated using the maximum offset position, the local error compression method is used to correct the weight difference of adjacent time steps, the iteration is updated and the result of each step is retained in the record sequence, and the iteration is repeated until the offset change is lower than the threshold value, and the weight iteration sequence is obtained;

[0074] Waveform output submodule: based on the weight iteration sequence, the feature paragraphs are extracted from the input matrix in the time dimension and weighted and superimposed to generate the initial waveform segment, the continuous frame splicing operation is performed to restore the complete waveform sequence, the sliding window mode is used to check the consistency of the overlapping segments in the critical area, the boundary transition correction and length normalization processing are completed, the stable and traceable waveform output table is generated, and the reconstructed waveform sequence table is obtained;

[0075] Input matrix construction submodule: based on the reduced low-dimensional expression, the state data of the switching devices and the harmonic interference channel signals at each stage are collected, the linear amplitude interpolation method is used to uniformly resample the data of different frequency bands, the uniform time step is set to 1 millisecond, the original sampling time sequence is converted to a standard time axis, the time steps are aligned in the channel order in the channel data and the corresponding amplitude points are inserted, the interpolation method uses equidistant interpolation, the amplitude at the target time point is obtained by performing linear piecewise interpolation on the amplitudes of adjacent time points, an index table is constructed to synchronize the time labels and variable names, the time labels are recorded in millisecond units from the starting time, the variable names are uniformly paired with the corresponding channel data using the number sequence, the channel number is mapped to the corresponding topology node index using the node number mapping method, the channel number is used to distinguish the frequency band channels, each signal data is divided into frequency bands, the signal values of the same frequency band are classified into a unified frequency channel set, and a three-dimensional matrix structure is formed using the frequency band grouping structure, the three-dimensional matrix dimension is set to the product of the number of frequency bands, the time step and the number of channels, the tensor expansion operation is performed to compress the two-dimensional input structure in the channel priority mode, the dimension is set to the product of the total number of channels and the time step, and the input matrix is obtained.

[0076] Weight iteration submodule: based on the input matrix, a multi-layer step-by-step correction operation is performed using a residual network structure, an initial weight vector is introduced as a superposition coefficient reference value, the length of the initial weight vector is consistent with the time step of the input matrix, all values are set to 1, a residual calculation operation is performed to evaluate the waveform offset degree under the current combination, the difference between the fitted waveform and the original waveform at each time step is calculated, the result is arranged in the time dimension as an offset sequence, the time alignment index is updated using the position of the maximum offset in the offset sequence, a sliding window with a window length of 5 time steps is used to traverse the offset sequence, the error mean value in each window is calculated, the weight between adjacent windows with large error difference is compressed and adjusted, the adjustment method is to compare the average offset difference between the two time periods, introduce an adjustment amount in the opposite direction in the weight vector for fine tuning, the adjustment step is set to 0.01, the current weight vector is recorded after each iteration and retained in the record sequence, and the iteration is repeated until the change amplitude of the offset sequence in the last two iterations is less than the set threshold value 0.001, and the weight iteration sequence is obtained.

[0077] Waveform Output Submodule: Based on the weighted iterative sequence, feature segments are extracted from the input matrix along the time dimension and weighted superimposed to generate initial waveform segments. Each input signal segment is mapped one-to-one with the corresponding time point weight at a unified time step, and the product of the corresponding values ​​is used as the weighted amplitude. All time segments are arranged in the original time order. A sliding splicing method is used to perform continuous frame splicing operation. The sliding window length is set to 50 time steps, and the step size is 25. A linear interpolation fusion operation is performed on the overlapping segments between two consecutive frames. The fusion method is to linearly superimpose the tail of the front segment and the head of the back segment with equal weight ratio to generate a smooth transition segment of the splicing area. All spliced ​​waveform segments are sequentially merged into a complete waveform sequence. The length of the waveform sequence is normalized. By setting the target length to 1024 time steps and redistributing the original waveform sequence data points at equal time intervals, a unified duration reconstruction is completed, generating a reconstructed waveform sequence table.

[0078] The residual network first takes an input sample set and introduces an initial weight vector as the initial parameters of each connection layer in the residual network. Then, it is expanded according to the time series structure. Each neural unit performs a combination of linear transformation and activation function operation to generate a predicted output sequence and calculates the difference with the target waveform to obtain the residual sequence. Then, the mean squared error is used as the loss function to quantify the residual size. Error information is backpropagated to update the network weights. In each iteration, the offset between the current output waveform and the target reference waveform is compared, the offset point is extracted and the time alignment index is adjusted. At the same time, the local weight difference between the previous and next time steps is compressed and adjusted to reduce oscillation. The iteration process continues until the residual change is lower than the set threshold. The converged weight iteration sequence is output as a stable parameter set.

[0079] Residual networks, according to the formula:

[0080] ;

[0081] in: Indicates the first The weight vector generated in the next iteration Indicates the first The current weight vector in the next iteration This represents the learning rate of the residual network. This represents the power system input matrix obtained from sampling. Represents the input matrix The transpose operation, Represents the desired target waveform vector. The weighting coefficients of the residual smoothing term are represented. Indicates the first The residual vector of the next iteration This represents the weight coefficient of the regularization term. This represents the regularized gradient applied to the current weight vector. This represents the penalty coefficient for the residual variance term. Indicates the first The variance of the residual vector in each iteration This represents the weighting coefficient of the historical momentum term. Indicates the first The momentum term accumulated in each iteration;

[0082] Execution process: First, the input matrix is ​​obtained from multiple detection channels. Each column represents the response waveform of the radiation sample at a specific process node or measurement stage, and the desired output waveform is extracted. As a reference standard, reflecting the ideal response trajectory of the target sample throughout the entire process, the weight vector is initialized. And calculate the current residual. Substituting the residuals into the formula yields the gradient update direction. Combined with the weighted adjustment of the residual term itself To mitigate the impact of local fluctuations, a regularized gradient is then introduced. This is reflected in the constraints on the controllability and physical rationality of the weights at each stage in the traceability system, and multiplied by a coefficient. To suppress overfitting, then introduce As the first The residual variance reflects the degree of error dispersion of the system in the time dimension. The adjustments improve the overall stability of sequence fitting, and finally, the momentum term is incorporated. Reflecting the cumulative trend of the weight evolution direction in previous rounds, through Weighted balancing convergence speed and correction step size are calculated by factoring the above factors with the learning rate. Multiply and then add to the current weight. To obtain the updated weights in the next round The process is repeated and the results of each round are recorded until the residual fluctuations converge to the system-set threshold, generating a complete weight iteration sequence for traceable reconstruction of each signal node during the radiation experiment.

[0083] The process scheduling module includes:

[0084] Task unit division into sub-modules: Based on the reconstructed waveform sequence list, the continuous amplitude rising segment detection method is adopted, the discharge segment extraction operation adopts the continuous amplitude falling segment identification method, and the resonant segment extraction operation adopts the zero-crossing interval extraction method to complete the division. The peak and valley point series are extracted by superimposing equal interval sampling operation, and the start and end times of each segment are matched to form a duration statistics table to obtain the task time sequence distribution.

[0085] Trigger structure construction submodule: based on the task timing distribution, extract the parallel segment number, superimpose the trigger sequence to generate operation construction time sequence number group, execute the dependency path arrangement operation to establish the sequence chain according to the adjacent task number, construct the trigger priority grouping table to generate the priority index according to the waiting ratio, and obtain the scheduling trigger structure;

[0086] Based on the reconstructed waveform sequence table, the amplitude continuous rising segment detection method is adopted, the detection window length is set to 5 time steps, and the traversal comparison operation is performed for each point. If the current amplitude is greater than the previous amplitude, the index is added to the rising sequence buffer. Otherwise, it is judged whether the buffer length is greater than or equal to 5. If the condition is met, the starting and ending indexes of the buffer are recorded as the rising segment, and the buffer is emptied. Continue to traverse, and the discharge segment extraction operation adopts the amplitude continuous descending segment identification method. The descending detection window length is set to 5 time steps. The amplitude comparison is performed for each point. If the current amplitude is less than the previous amplitude, the index is added to the descending sequence buffer. Otherwise, when the buffer length is greater than or equal to 5, the starting and ending indexes of the buffer are recorded as the descending segment, and the buffer is emptied. The resonance segment extraction operation adopts the zero-crossing interval extraction method. Traverse the complete waveform sequence table. If the signs of the adjacent samples are opposite, the phase crossover time point index is recorded, and the resonance segment is defined according to the interval of two crossover points. The peak and valley point list is extracted by the equal interval sampling operation. The sampling interval is set to 10 time steps. The amplitude maximum point and minimum point indexes are taken out from each rising segment and descending segment in a fixed step length to form a peak and valley index list. The rising segment, the descending segment, the resonance segment, the peak point index and the valley point index are matched to merge the starting index and the ending index. The time length is calculated and written into the time length statistical table to obtain the task timing distribution.

[0087] Based on the task timing distribution, the parallel segment number is extracted, the topological sorting algorithm is adopted, the in-degree mapping of all task segment nodes is initialized, and the adjacency list is constructed. For each task segment number, the dependency relationship is traversed. If the predecessor task segment end time is less than the successor task segment start time, a directed edge is added in the adjacency list, and the in-degree of the successor task segment is increased by one. The trigger sequence generation operation constructs a time sequence number group in the in-degree mapping. The in-degree of zero is filtered, and the task segment number is added to the queue and dequeued in turn. When writing, the in-degree of all successor nodes in the adjacency list is reduced by one. If the in-degree is zero after reduction, the node is enqueued for further processing. The dependency path arrangement operation establishes a sequence chain according to the queue dequeue order. The in-degree of the node is recorded as the dependency path table at the time of dequeuing. The trigger priority grouping table is constructed to generate the priority index according to the waiting ratio. The waiting ratio threshold is set to 0.5. The normalized value is calculated by dividing the waiting time by the longest waiting time. If the normalized value is greater than the threshold, it is marked as high priority. Otherwise, it is marked as low priority. The task segment number and the priority index are mapped and written into the priority grouping table to obtain the scheduling trigger structure.

[0088] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in other forms. Any skilled person in the art can modify or change the disclosed technical content into equivalent embodiments with equivalent changes, and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical solution content of the present application, still falls within the protection scope of the present application.

Claims

1. A power system electromagnetic transient simulation system containing multiple power electronic switches, characterized in that, The system includes: Topology analysis module: Based on the parameters of each switch node, inductor, capacitor and resistor in a multi-power electronic switching device, it performs impedance calculation and admittance conversion by extracting the switch number and topology number, normalizes matrix elements and maps node connection structure to generate network topology matrix; Parameter evolution module: Based on the network topology matrix, a genetic algorithm is used, combined with the timing of switching devices, to establish a variable sequence by extracting switching time points and connection information. The simulation runs and records the error and duration, sorts and judges the retention and mutation, and generates the optimal combination of control parameters. The order reduction mapping module: Based on the preferred combination of the control parameters, it collects the voltage and current sequences of the energy storage unit and the filter, constructs a time matrix and performs singular value calculation and feature vector selection, calculates the projection coefficients and determines the residual adjustment dimension, and obtains the order reduction low-dimensional expression; State reconstruction module: Based on the reduced-order low-dimensional expression, a residual network is used to construct an input matrix by combining switch states and harmonics, and a combination sequence is constructed by matching time steps and coefficients. The offset difference is calculated, the weight is adjusted and iteratively updated to obtain the reconstructed waveform sequence table. Process scheduling module: Based on the reconstructed waveform sequence list, combined with the definition of task units in the capacitor charging and discharging and resonance stages, the task duration and waiting ratio are statistically analyzed, the parallel conditions and triggering order are determined, and a scheduling triggering structure is established; The parameter evolution module includes: The switching sequence generation submodule extracts the switching time points of the switching devices and their corresponding connection node numbers based on the network topology matrix. It then uses a genetic algorithm to construct a node index mapping table by combining the topology row and column labels. The module performs a timing axis alignment operation to correct the time points of each channel, performs a connection information filtering operation to remove invalid node states, organizes the effective switching segments in the entire time domain, and constructs a complete sequence structure by splicing the variable state list in time order and filling in missing fields to obtain the variable sequence set. Error Filtering Submodule: Based on the variable sequence set, it performs circuit model simulation calculation to output waveform group, calculates the error curve between the target waveform and the actual response through residual extraction operation, generates sequence quality level by sorting by error integral area, performs runtime statistics and establishes a secondary index matrix, merges error level and duration level through weight allocation, extracts the top-ranked sequence and adds retention mark, synchronously generates mutation mark matrix, and obtains the optimal parameter index. The optimal combination generation submodule: Based on the preferred parameter index, extract the retained state and the corresponding variable number index, perform a mapping matching operation to associate the original control parameter data, form a structurally complete control vector set through a combination reconstruction operation, generate multiple sets of structural disturbance candidates by superimposing a mutation label matrix, summarize different control vector sets and corresponding mutation labels, construct a two-dimensional control parameter set array, and obtain the preferred combination of control parameters.

2. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 1, characterized in that, The topology resolution module includes: Node parameter mapping submodule: Based on the switching nodes, inductors, capacitors and resistors in the multi-power electronic switching device, the module extracts the switch number and marks the topology number by matching node information, constructs a complex impedance value array by converting electrical parameters, completes the output of the real and imaginary parts of the admittance by numerical conversion, and generates a normalized admittance dataset by scaling the admittance elements to form a vector set. The connection structure construction submodule: Based on the normalized admittance dataset, it uses index mapping to process the correspondence between node numbers and vectors, uses matrix assembly to construct the node pair connection matrix, uses element fusion to present the two-dimensional mapping diagram of admittance and structure, summarizes and outputs a set of multi-node link information, and obtains the network topology matrix.

3. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 2, characterized in that, The genetic algorithm first uses the set of variable sequences obtained after time-axis alignment and invalid node removal as the initial population, with each sequence corresponding to a chromosome. Then, the fitness function is used as the weighted sum of the simulation error integral area and the runtime to evaluate the fitness of each chromosome in the population. Next, a selection operator is used to select parents from high-fitness chromosomes, and the mapping table of time points and node numbers of the parent sequences is recombined using single-point or multi-point crossover operators to obtain offspring chromosomes. Subsequently, a mutation operator is used to slightly perturb the switching time or connection information of some offspring chromosomes. At the same time, the parents and offspring are sorted by fitness, and the top N optimal chromosomes are retained as the new generation population. This process is repeated until the convergence condition is met and the preset generation limit is reached, and the optimal combination of control parameters is output.

4. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 2, characterized in that, The genetic algorithm is based on the formula: ； in: Indicates the candidate node index sequence at the th The updated permutation results in each genetic iteration. Indicates the candidate node index sequence at the th The current permutation state in the next iteration. Represents the global adjustment coefficient. This indicates the weight of the fitness difference component in the index update. This indicates the weight of the channel switching time consistency component. The weights representing the node topological connectivity scores The weight representing the frequency of choices made by historical elites. Indicates the first The fitness value of the individual with the minimum number of switching state conflicts and optimal connectivity performance in the next iteration. Indicates the first Individual in the first The current fitness value in the next iteration Indicates the first The mean squared error of the channel switching time sequence for an individual. Indicates the first The score representing the percentage of connected edges in the node connection matrix constructed by an individual. Indicates the first The frequency ratio of individuals selected as elite individuals in each genetic iteration.

5. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 1, characterized in that, The order reduction mapping module includes: Feature matrix generation submodule: Based on the preferred combination of the control parameters, the voltage and current values ​​of the energy storage unit and the filter are read by voltage and current acquisition, a two-dimensional sequence matrix is ​​constructed by time indexing, the energy distribution is divided by column vector energy decomposition, the principal vector set is screened and identified by peak contribution, and the principal component feature matrix is ​​generated. The low-dimensional mapping derivation submodule: Based on the principal component feature matrix, it uses matrix mapping to calculate the output coefficient list, uses residual variance signal accumulation to determine the projection dimension, uses combination screening to extract the minimum dimension vector array, and uses time rearrangement to reconstruct the low-dimensional feature sequence to obtain the reduced-order low-dimensional expression.

6. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 1, characterized in that, The state reconstruction module includes: Input matrix construction submodule: Based on the reduced-order low-dimensional expression, the switching device status data and harmonic interference channel signals corresponding to each stage are collected. The data of different frequency bands are uniformly resampled by the equal time step amplitude interpolation method. An index table is constructed to synchronously label the time label and variable name. A three-dimensional matrix structure is formed by mapping the node number and splitting the harmonic frequency band. Tensor expansion operation is performed to compress it into a two-dimensional input structure to obtain the input matrix. Weight Iteration Submodule: Based on the input matrix, a residual network is used, an initial weight vector is introduced as a reference value for the superposition coefficient, residual calculation operation is performed to evaluate the waveform offset under the current combination, the alignment index is updated using the maximum offset position, the weight difference between adjacent time steps is corrected by the local error compression method, iterative update is performed and the result of each step is retained in the record sequence, the iteration is repeated until the offset change is lower than the threshold, and the weight iteration sequence is obtained. Waveform output submodule: Based on the weighted iterative sequence, feature segments are extracted from the input matrix according to the time dimension and weighted superimposed to generate initial waveform segments. Continuous frame splicing operation is performed to restore the complete waveform sequence. The consistency of overlapping segments in the critical region is verified through the sliding window mode. Boundary transition correction and length normalization processing are completed to generate a stable and traceable waveform output table and obtain the reconstructed waveform sequence table.

7. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 6, characterized in that, The residual network first takes an input sample set and introduces an initialized weight vector as the initial parameters for each connection layer in the residual network. Then, it is expanded according to a time series structure. Each neural unit performs a combination of linear transformation and activation function operation to generate a predicted output sequence and calculates the difference between the predicted output sequence and the target waveform to obtain the residual sequence. Then, the mean squared error is used as the loss function to quantify the residual size. Error information is backpropagated to update the network weights. In each iteration, the offset between the current output waveform and the target reference waveform is compared, the offset point is extracted and the time alignment index is adjusted. At the same time, the local weight difference between the previous and next time steps is compressed and adjusted to reduce oscillation. The iteration process continues until the residual change is lower than a set threshold. The converged weight iteration sequence is output as a stable parameter set.

8. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 6, characterized in that, The residual network is defined according to the formula: ； in: Indicates the first The weight vector generated in the next iteration Indicates the first The current weight vector in the next iteration This represents the learning rate of the residual network. This represents the power system input matrix obtained from sampling. Represents the input matrix The transpose operation, Represents the desired target waveform vector. The weighting coefficients of the residual smoothing term are represented. Indicates the first The residual vector of the next iteration This represents the weight coefficient of the regularization term. This represents the regularized gradient applied to the current weight vector. This represents the penalty coefficient for the residual variance term. Indicates the first The variance of the residual vector in each iteration This represents the weighting coefficient of the historical momentum term. Indicates the first The momentum term accumulated in each iteration.

9. The electromagnetic transient simulation system for a power system containing multiple power electronic switches according to claim 1, characterized in that, The process scheduling module includes: Task unit division sub-module: Based on the reconstructed waveform sequence list, the amplitude continuously rising segment detection method is adopted, the discharge segment extraction operation adopts the amplitude continuously falling segment identification method, the resonant segment extraction operation adopts the zero crossover interval extraction method to complete the division, the equal interval sampling operation is superimposed to extract the peak and valley point series, the start and end times of each segment are matched to form a duration statistics table, and the task time sequence distribution is obtained. Trigger structure construction submodule: Based on the task time sequence distribution, extract the parallelizable segment numbers, superimpose the trigger sequence generation operation to construct a time sequence number group, perform dependency path sorting operation to establish a sequence chain according to adjacent task numbers, construct a trigger priority grouping table to generate a priority index based on the waiting ratio, and obtain the scheduling trigger structure.

Images