SDAE-BP voltage sag type identification method, device and system based on improved PSO optimization

Abstract

The application discloses an SDAE-BP voltage sag type identification method, device and system based on improved PSO optimization, which comprises the following steps: original sampling processing, calculating three-phase voltage effective value and phase; extracting data of a short-circuit fault time period, unifying time length, resampling amplitude and phase data and performing normalization processing; randomly distributing preprocessed voltage sag data into a training set and a test set; initializing improved particle swarm parameters, and using the improved PSO algorithm and the training set data to complete optimization training of SDAE-BP parameters; when the voltage sag type identification error rate reaches an expected target or the training reaches a maximum iteration number, terminating the training and determining SDAE-BP parameters; and inputting the test set into SDAE-BP for voltage sag type identification classification. The improved particle swarm algorithm is used to find the optimal parameters of the SDAE-BP network, so that the problems of low efficiency and low accuracy caused by improper selection of parameters by people can be effectively avoided, and the accuracy of voltage sag type identification is improved.

Description

Technical Field

[0001] This invention relates to the field of power quality analysis, and specifically to a method, apparatus, and system for identifying SDAE-BP sag types based on improved PSO optimization. Background Technology

[0002] Different short-circuit faults cause different types of voltage sags, which have different impacts on users. Accurately identifying the type of voltage sag allows for analysis, compensation, and suppression of the actual voltage sag situation, which is of great significance for the management of voltage sags. It can also serve as a basis for resolving disputes between power supply departments and users.

[0003] Currently, voltage sag source identification mainly relies on a "feature extraction + type recognition" model. However, with more and more power electronic devices connected to the power grid, the actual waveforms are not ideal, affecting the extracted features and consequently impacting sag source identification. Deep learning can extract effective features from complex data, achieving higher sag classification accuracy. However, the number of hidden layers and the settings of network parameters significantly influence the classification results. Currently, most deep networks rely on prior experience or multiple experiments to determine relevant parameters, which to some extent reduces the accuracy and efficiency of sag classification. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a method, apparatus, and system for identifying voltage sag types based on an improved PSO-optimized SDAE-BP method. This scheme utilizes a stacked denoising autoencoder (SDAE) in deep learning to extract voltage sag signal features, and classifies and identifies sag data through a BP neural network, thus solving the problem of being affected by unknown features and noise when manually extracting features. The improved PSO algorithm is used to optimize the SDAE-BP network parameters, adaptively selecting optimization parameters to improve the network's generalization ability and the accuracy of sag type identification.

[0005] Technical solution: This invention includes an SDAE-BP descent type identification method based on improved PSO optimization, comprising the following steps:

[0006] (1) Preprocess the raw sampled data to calculate the effective value and phase of the three-phase voltage;

[0007] (2) Extract data during the period when the short circuit fault occurred, unify the data duration, and resample the amplitude and phase data to reduce the data dimensionality;

[0008] (3) Form a union sequence according to the order of phase voltage and phase of phases A, B, C, and C, and perform data normalization processing;

[0009] (4) The preprocessed voltage sag data are randomly assigned to the training set and the test set as input data for the SDAE-BP network;

[0010] (5) Initialize the particle swarm, set the target threshold and iteration number parameters, and use the improved PSO algorithm to complete the adaptive selection of SDAE-BP network parameters for temporary feature extraction and classification training.

[0011] (6) When the classification error rate reaches the expected target or the maximum number of iterations is reached, the training is terminated and the optimized network parameters are output to determine the SDAE-BP network model parameters.

[0012] (7) Use the test set as input to the SDAE-BP network to obtain the classification results of the temporary descent type and test and verify the classification accuracy of the model.

[0013] In step (4), the training set and test set include data samples of sag types obtained when a three-phase short circuit, single-phase ground fault, two-phase short circuit, or two-phase ground fault occurs in the corresponding neutral point effectively grounded system, and the corresponding sag type labels.

[0014] In step (5), the SDAE-BP network parameters include the number of hidden layers in the first layer of SDAE, the number of hidden layers in the second layer of SDAE, the number of iterations of the SDAE training model, the number of hidden layers in the BP neural network, the number of iterations of the BP neural network, the proportion of SDAE input data set to 0, the learning rate of the BP neural network, and the momentum parameter of the BP neural network.

[0015] In step (5), the improved PSO algorithm is used to complete the temporary descent feature extraction and adaptive selection of SDAE-BP network parameters for classification training. Specifically, the following steps are included:

[0016] (5.1) Based on the dimension of the union of the temporary data and the range of parameter values, set the range of values ​​of the parameters that need to be optimized in the SDAE-BP network, and use it as the range interval for updating particle velocity and position. When the value exceeds the range, take the maximum or minimum value of the interval.

[0017] (5.2) Initialize the particle swarm parameters, including the initial position and velocity of the particles, the population size, the inertia weight, the acceleration factor, the number of iterations of the particles, and the target threshold;

[0018] (5.3) The temporary reduction in classification error rate obtained from training SDAE-BP is used as the fitness value of the improved PSO algorithm. If the fitness value is better than the individual optimal value and the global optimal value, then update the values ​​of these two values. At the same time, in order to increase the diversity of the population, replace the 10% of particles with the worst fitness with randomly generated new particles. If the training reaches the accuracy requirement or the number of iterations, the obtained parameters are used as the optimal structure parameters of SDAE-BP. Otherwise, update the particle velocity and particle position, and repeat the step.

[0019] To prevent the particle swarm optimization algorithm from getting trapped in local optima, this invention fully utilizes the historical extreme value information of the same particle and enhances information exchange between particles. An improved particle swarm optimization velocity and position update formula is adopted, which are expressed as follows:

[0020]

[0021] x[t+1] = x[t] + v[t+1]

[0022] In the formula, v represents the particle's velocity; ω represents the inertial weight; x represents the particle's position; and pbest represents the individual extreme value. This represents the weighted average of the individual extreme values ​​in the particle's history. represents the average of all individual particle extreme values; gbest represents the global extreme value; r1, r2, r3, r4 represent random numbers between (0, 1); c1, c2 represent learning factors.

[0023] The weighted average of historical individual extreme values ​​is expressed as follows:

[0024] In the formula, ω i ω represents the weight value of the individual's extreme value during the i-th training iteration. i The value of ω i = i / (t-1).

[0025] The average of the extreme values ​​of all individual particles in this training is expressed as follows:

[0026] In the formula, pbest i [t] represents the individual extreme value of the i-th particle in this training, and N represents the population size, which ranges from 20 to 40.

[0027] The training of SDAE-BP includes training of a stacked noise-reducing autoencoder SDAE and training of a BP neural network model.

[0028] The training of the stacked noise reduction autoencoder (SDAE) includes: inputting voltage sag test set samples into the pre-built SDAE network, training each layer of the SDAE network in sequence, determining the optimal weights and optimal biases that minimize the reconstruction error of each layer of the DAE, and obtaining the feature parameters of the output of the last layer of the DAE.

[0029] The training of the BP neural network model includes: configuring corresponding temporary descent type labels for the final features extracted by the SDAE network according to the training samples corresponding to the features; inputting the features and corresponding temporary descent type labels into the pre-constructed BP neural network to complete the training of the BP neural network.

[0030] The present invention also includes an SDAE-BP sag type identification device based on improved PSO optimization, employing an SDAE-BP sag type identification method based on improved PSO optimization, comprising: a sampling data acquisition module configured to acquire voltage sag sampling data of the sag type to be identified; a preprocessing module for preprocessing the voltage sag sampling data obtained by the sampling data acquisition module; a parameter optimization module for obtaining the optimal parameters of the SDAE-BP network; a feature extraction module for inputting the voltage sag data obtained by the preprocessing module into a pre-trained stacked noise-reducing autoencoder SDAE of the parameter optimization module to obtain voltage sag feature data extracted by the SDAE; and a sag type identification module for inputting the sag feature data obtained by the voltage sag feature extraction module and its corresponding type label data into a pre-trained BP neural network model of the parameter optimization module, and determining the voltage sag category based on its output.

[0031] The present invention also includes an SDAE-BP slump type identification system based on improved PSO optimization, comprising a processor and a memory, wherein the memory stores computer instructions, characterized in that: the processor operates according to the computer instructions to execute the SDAE-BP slump type identification method based on improved PSO optimization.

[0032] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) The present invention realizes the identification of voltage sag types based on the improved PSO optimized SDAE-BP model. In the model construction and training stage, the SDAE-BP network is jointly trained by the improved PSO algorithm to avoid the problem of low efficiency and low accuracy of sag type identification caused by inappropriate parameter selection. The parameters of the deep neural network are adaptively selected to improve the network generalization ability; (2) The data used to input the SDAE-BP network is the data after calculating the effective value and phase of the original sampled waveform data and then resampling it. This can greatly reduce the sample length, reduce the complexity of the network model, shorten the training time and testing time, improve the training efficiency, and reduce the identification difficulty; (3) The present invention makes full use of the feature extraction capability of SDAE, transforms the manually set feature extraction into automatic feature extraction, and uses the BP neural network for classification and identification, which can realize the effective identification of voltage sag types and can be effectively applied to the actual power system. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention;

[0034] Figure 2 This is a schematic diagram of the SDAE-BP model training process in this invention;

[0035] Figure 3 This is a schematic diagram of the feature vectors for the seven types of voltage sags defined in an embodiment of the present invention;

[0036] Figure 4 The present invention uses improved PSO and standard PSO to optimize the SDAE-BP network parameters, wherein, Figure 4 (a) Training curves for optimizing SDAE-BP network parameters using improved PSO. Figure 4 (b) Training curves for optimizing SDAE-BP network parameters using standard PSO. Detailed Implementation

[0037] The technical solution of the present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0038] like Figure 1 As shown, the present invention includes an SDAE-BP descent type identification method based on improved PSO optimization, comprising the following steps:

[0039] (1) Based on multiple preset sag types, acquire multiple sets of historical voltage sag signals, preprocess the original sampled data, and calculate the effective value and phase of the three-phase voltage;

[0040] (2) Extract data during the period when the short circuit fault occurred, unify the data duration, and resample the amplitude and phase data to reduce the data dimensionality;

[0041] (3) Form a union sequence according to the order of phase voltage and phase of phases A, B, C, and C, and perform data normalization processing;

[0042] (4) The preprocessed voltage sag data are randomly assigned to the training set and the test set as input data for the SDAE-BP (a fusion algorithm of stacked denoising auto encoder (SDAE) and back propagation (BP) neural network) network; wherein, the training set and the test set include data samples of sag types and corresponding sag type labels obtained when three-phase short circuit, single-phase ground fault, two-phase short circuit, and two-phase ground fault occur in the neutral point effectively grounded system and after propagation through the transformer;

[0043] The voltage sag types corresponding to effectively grounded neutral systems are derived and summarized in this invention, taking into account the influence of phase jumps, based on the voltage sag classifications described in existing literature. The expressions, characteristics, and sources of the seven types of voltage sags caused by short-circuit faults in effectively grounded neutral systems are detailed in the table below. In the table, These are the voltages of phase A, phase B, and phase C, respectively. Z S0 Z S1 Z S2 and Z F0 Z F1 Z F2 These represent the equivalent system zero-sequence, positive-sequence, and negative-sequence impedances, and the equivalent line zero-sequence, positive-sequence, and negative-sequence impedances, respectively. α is determined by the system impedance angle and the fault impedance angle. Considering a neutral-point effectively grounded system, this paper assumes Z... S0 =Z S1 =Z S2 Z F0 =Z F1 =Z F2 Z S1 =R S1 +jX S1 Z F1 =R F1 +jX F1 R S1 X S1 R F1 and X F1 Let X be the equivalent system positive-sequence resistance, equivalent system positive-sequence reactance, equivalent line positive-sequence resistance, and equivalent line positive-sequence reactance, respectively. S1 / R S1 ≠X F1 / R F1 The voltage sag caused by a short-circuit fault will then undergo a phase jump. The calculation process of the voltage sag expression obtained through transformer propagation is consistent with existing literature, and this invention directly provides the calculation results.

[0044] Table 1

[0045]

[0046]

[0047]

[0048] In the table, D = (Z S1 +Z F1 (Z) S0 +Z F0 +Z S2 +Z F2 )+(Z S0 +Z F0 (Z) S2 +Z F2 ).

[0049] This invention considers the voltage and phase angle characteristics of the seven types of sags to be distinguished. In order to shorten the model training time, improve the training efficiency, and reduce the recognition difficulty, the preprocessing method of the original sampled data is as follows: The data used for inputting the SDAE-BP network in this invention is the data after calculating the effective value and phase of the original sampled waveform data and then resampling it.

[0050] Specifically, power quality sampling data typically consists of 512 points per cycle, resulting in a massive amount of data. Considering the data processing volume and the relative stability of RMS and phase data within a single cycle, this invention first calculates the RMS and phase of the fault waveform data. This calculation only needs to be performed once per cycle, significantly reducing the sample length, network model complexity, training time, and testing time. Then, data from the time period of the short-circuit fault is extracted and resampled. Even with different fault types and durations, the same number of points can be sampled. This ensures that the length and format of the pre-processed data input to the model are consistent across various fault types and durations, further reducing the difficulty of model training and recognition, and improving efficiency.

[0051] (5) Initialize the particle swarm and set parameters such as the target threshold and number of iterations. Use the improved PSO (Particle Swarm Optimization) algorithm to perform adaptive selection of parameters for the transient feature extraction and classification training of the SDAE-BP network. The SDAE-BP network parameters include the number of hidden layers in the first and second layers of the SDAE network, the number of iterations for training the SDAE model, the number of hidden layers in the BP neural network, the number of iterations in the BP neural network, the proportion of SDAE input data set to zero, the learning rate of the BP neural network, and the momentum parameter of the BP neural network. This step specifically includes the following:

[0052] (5.1) Based on the dimension of the union of the transient data and the range of parameter values, set the range of values ​​for the parameters that need to be optimized in the SDAE-BP network, and use this range as the interval for updating particle velocity and position. When the value exceeds the interval, take the maximum or minimum value of the interval accordingly; set the particle position range interval as follows:

[0053] plimit = [150,384; 50,150; 50,150; 20,80; 50,300; 0,0.5; 0,1; 0,1]; the particle velocity range is set to vlimit = [-10,10; -10,10; -10,10; -5,5; -10,10; -0.1,0.1; -0.1,0.1; -0.1,0.1].

[0054] (5.2) Initialize the particle swarm parameters, including the initial position and velocity of the particles, the number of iterations, the population size, the inertia weight, and the acceleration factor.

[0055] PSO is initialized at position x[1] = P min +rand()*(P max -P min ), initialize the particle velocity as v[1]=V max -rand()*(V max -V min The population size is 20, the number of iterations is 20, the initial inertia weight is 0.9, the termination weight is 0.4, the learning factor c1 = c2 = 2, and the expected target is set to 1e. -4 .

[0056] Where v[1] is the initial velocity of the particle, x[1] is the initial position of the particle, and P max and P min These are the upper and lower limits of the particle's position, V. max and V min These are the upper and lower limits of the particle velocity, respectively, and rand() is a random number between (0, 1).

[0057] (5.3) The temporary reduction in classification error rate obtained from training SDAE-BP is used as the fitness value of the improved PSO. If the fitness value is better than the individual optimal value and the global optimal value, then these two values ​​are updated. At the same time, in order to increase the diversity of the population, the 10% of particles with the worst fitness are replaced with randomly generated new particles. In this invention, the population size is selected as 20, and the two particles with the worst fitness in this training are replaced. The update formula is the particle initial position formula. If the training reaches the accuracy requirement or the number of iterations, the obtained parameters are used as the optimal structure parameters of SDAE-BP. Otherwise, the particle velocity and particle position are updated, and this step is repeated.

[0058] To prevent the particle swarm optimization algorithm from getting trapped in local optima, this invention fully utilizes the historical extreme value information of the same particle and enhances information exchange between particles. An improved particle swarm optimization velocity and position update formula is adopted, expressed as follows:

[0059]

[0060] x[t+1] = x[t] + v[t+1]

[0061] In the formula, v represents the particle's velocity; ω represents the inertial weight, the value of which decreases linearly from 0.9 to 0.4 as the number of iterations increases; x represents the particle's position (considering practical situations, 1-5 dimension data need to be rounded); pbest represents the individual extreme value; The weighted average of the historical individual extreme values ​​of a particle; represents the average of all individual particle extreme values; gbest represents the global extreme value; r1, r2, r3, r4 represent random numbers between (0, 1); c1, c2 represent learning factors.

[0062] The expression is

[0063]

[0064] In the formula, ω i ω represents the weight value of the individual's extreme value during the i-th training iteration. i The value of ω i = i / (t-1).

[0065] The expression is

[0066] In the above formula, pbest i [t] represents the individual extreme value of the i-th particle in this training, and N is the population size, ranging from 20 to 40. This scheme uses 20.

[0067] (6) When the classification error rate reaches the expected target or the maximum number of iterations is reached, the training is terminated and the optimized network parameters are output to determine the SDAE-BP network model parameters.

[0068] (7) Use the test set as input to the SDAE-BP network to obtain the classification results of the temporary descent type and test and verify the classification accuracy of the model.

[0069] In application, the SDAE-BP network is optimized and trained based on the improved PSO for sag type identification. First, the network parameters of SDAE-BP are optimized using the improved PSO. Then, the SDAE is used to extract voltage sag signal features. Finally, the BP neural network is used to identify the sag type. This can avoid the problems of low efficiency and low accuracy caused by inappropriate parameter selection. It can also solve the problem of being affected by unknown features and noise when manually extracting features, thus improving the accuracy of sag type identification.

[0070] This embodiment has seven preset descent types, see reference. Figure 3 As shown, where, Figure 3 (a) Figure 3 (b) Figure 3 (c) Figure 3 (d) are vector diagrams showing three-phase short circuits, single-phase ground faults, two-phase short circuits, and two-phase ground faults in a neutral-point effectively grounded system, respectively. Figure 3 (e) Figure 3 (f) Figure 3 (g) is the vector diagram of voltage sag types obtained after propagation through the transformer. The corresponding historical voltage sag data to be obtained are the voltage sag signals caused by different types of short circuit faults such as three-phase short circuit, single-phase ground fault, two-phase short circuit, and two-phase ground fault, as well as the voltage sag signal after propagation through the transformer.

[0071] For each type of temporary descent, multiple sets of temporary descent signals need to be obtained to form sufficient samples to achieve the purpose of effectively training the network model.

[0072] The preprocessing of historical voltage sag signals in this embodiment includes:

[0073] For each voltage sag signal, calculate the effective value and phase of its three-phase voltage;

[0074] Extract voltage sag data during the fault period and downsample the effective value and phase of the three-phase voltage;

[0075] The effective values ​​and phase data of the three-phase voltage obtained by downsampling are normalized.

[0076] The normalized data are combined into a union sequence, which is then used as the input data for the network.

[0077] The above-mentioned normalized data can be used to form a union sequence by arranging the voltage and phase data of phases A, B, and C into a sequence. For example, each sag is sampled 64 times on average according to its duration, and then a 1×384 sequence is formed according to the order of voltage and phase of phases A, B, and C as the input of SDAE.

[0078] The aforementioned SDAE-BP network can construct a two-layer SDAE network. It requires training and optimizing eight parameters: the number of hidden layers in the first and second layers of the SDAE, the proportion of input data set to zero, the number of training model iterations, the number of hidden layers in the BP neural network, the number of iterations, the learning rate, and the momentum parameter. Therefore, the PSO space has an 8-dimensionality. The parameters from the first to the eighth dimension are, respectively, the number of hidden layers in the first and second layers of the SDAE, the number of training model iterations for the SDAE, the number of hidden layers in the BP neural network, the number of iterations in the BP neural network, the proportion of input data set to zero for the SDAE, the learning rate of the BP neural network, and the momentum parameter of the BP neural network.

[0079] like Figure 2 As shown, in this scheme, training the SDAE-BP includes training the stacked denoising autoencoder (SDAE) and the BP neural network model. Training the stacked denoising autoencoder (SDAE) involves inputting voltage sag test set samples into the pre-built SDAE network, sequentially training each layer of the SDAE network's DAE, determining the optimal weights and biases that minimize the reconstruction error of each DAE layer, and obtaining the feature parameters output by the last DAE layer. Training the BP neural network model involves configuring corresponding sag type labels for the final features extracted by the SDAE network based on the corresponding training samples; inputting the features and corresponding sag type labels into the pre-built BP neural network to complete the training of the BP neural network.

[0080] The following example illustrates this embodiment:

[0081] The training set of 7000 voltage sag data sets (1000 sets per class) was preprocessed and then input into the model for parameter optimization, feature extraction, and classification training. The improved PSO population size was 20, the number of iterations was 20, the inertia decreased linearly from 0.9 to 0.4 with increasing iterations, and the learning factors c1 = c2 = 2. The trained SDAE-BP model parameters are as follows: the number of units in the input layer is the same as the number of voltage sag data inputs, 384; the number of hidden layers in the first SDAE layer is 203, the number of hidden layers in the second SDAE layer is 69, the proportion of input data set to 0 is 0.231, and the number of training iterations is 122; the BP neural network has 69 input layers, 48 ​​hidden layers, and 7 output layers corresponding to the number of voltage sag types, with 225 iterations, a learning rate of 0.779, and a momentum parameter of 0.793. The optimization curves of the SDAE-BP network parameters optimized using the improved PSO and the standard PSO are shown below. Figure 4 (a) and Figure 4 (b) The optimization results show that the improved PSO algorithm for optimizing the SDAE-BP network parameters results in a lower sag classification error rate, which is superior to the standard PSO algorithm. Considering the different probabilities of various faults in actual power systems, 3000 test sets (100 sets of Class A, 1100 sets of Class B, 600 sets of Class C, 400 sets of Class D, 400 sets of Class E, 200 sets of Class F, and 200 sets of Class G) were preprocessed and input into the trained SDAE-BP model for feature extraction and classification. The accuracy of voltage sag type identification on the test set is shown in Table 2. The experimental results show that the average identification rate of the sag type identification method using the SDAE-BP network after parameter optimization using the improved PSO algorithm is 99.4%. The identification accuracy of voltage sag signals with different signal-to-noise ratios was also tested. Simulation results show that the model can maintain good accuracy in different levels of noise.

[0082] Table 2

[0083]

[0084] The present invention also includes an SDAE-BP slump type identification device based on improved PSO optimization, which employs an SDAE-BP slump type identification method based on improved PSO optimization, including:

[0085] The sampling data acquisition module is configured to acquire voltage sag sampling data of the type of sag to be identified.

[0086] The preprocessing module is used to preprocess the voltage sag sampling data obtained by the sampling data acquisition module;

[0087] The parameter optimization module is used to obtain the optimal parameters of the SDAE-BP network.

[0088] The feature extraction module is used to input the voltage sag data obtained by the preprocessing module into the stacked noise reduction autoencoder SDAE pre-trained by the parameter optimization module to obtain the voltage sag feature data extracted by the SDAE.

[0089] And a voltage sag type identification module, which is used to input the feature data obtained by the voltage sag feature extraction module and its corresponding type label data into the BP neural network model pre-trained by the parameter optimization module, and determine the voltage sag category based on its output.

[0090] The present invention also includes an SDAE-BP slump type identification system based on improved PSO optimization, comprising a processor and a memory, the memory storing computer instructions, and the processor operating according to the computer instructions to execute the SDAE-BP slump type identification method based on improved PSO optimization.

[0091] In summary, this invention can adaptively select network parameters, avoiding inefficiency and low accuracy caused by inappropriate parameter selection, and improving the network's generalization ability; it can automatically extract the feature parameters of voltage sag signals during application, thereby realizing the identification of sag types and avoiding information loss during manual feature extraction; the model generated by its training has better noise resistance and can be effectively applied to actual power systems.

Claims

1. A method for identifying SDAE-BP descent type based on improved PSO optimization, characterized in that, Includes the following steps: (1) Preprocess the raw sampled data to calculate the effective value and phase of the three-phase voltage; (2) Extract data during the period when the short circuit fault occurred, unify the data duration, and resample the amplitude and phase data to reduce the data dimensionality; (3) Form a union sequence according to the order of phase voltage and phase of phases A, B, C, and C, and perform data normalization; (4) The preprocessed voltage sag data are randomly assigned to the training set and the test set as input data for the SDAE-BP network; (5) Initialize the particle swarm optimization and set the target threshold and iteration number parameters. Use the improved PSO algorithm to complete the adaptive selection of parameters for transient feature extraction and classification training of the SDAE-BP network. Specifically, the adaptive selection of parameters for transient feature extraction and classification training of the SDAE-BP network using the improved PSO algorithm includes the following steps: (5.1) Based on the dimension of the union of the temporary data and the range of parameter values, set the range of values ​​of the parameters that need to be optimized in the SDAE-BP network, and use it as the range interval for updating particle velocity and position. When the value exceeds the range, take the maximum or minimum value of the interval. (5.2) Initialize the particle swarm parameters, including the initial position and velocity of the particles, the population size, inertia weight, acceleration factor, number of iterations of the particles, and target threshold; (5.3) The temporary reduction in classification error rate obtained from training SDAE-BP is used as the fitness value of the improved PSO algorithm. If the fitness value is better than the individual optimal value and the global optimal value, then update the values ​​of these two values. At the same time, in order to increase the diversity of the population, replace the 10% of particles with the worst fitness with randomly generated new particles. If the training reaches the accuracy requirement or the number of iterations, the obtained parameters are used as the optimal structure parameters of SDAE-BP. Otherwise, update the particle velocity and particle position, and repeat the step. (6) When the classification error rate reaches the expected target or the maximum number of iterations is reached, the training is terminated and the optimized network parameters are output to determine the SDAE-BP network model parameters. (7) Use the test set as input to the SDAE-BP network to obtain the classification results of the temporary descent type and test and verify the classification accuracy of the model.

2. The SDAE-BP slump type identification method based on improved PSO optimization according to claim 1, characterized in that: In step (4), the training set and test set include data samples of sag types obtained when a three-phase short circuit, single-phase ground fault, two-phase short circuit, or two-phase ground fault occurs in the corresponding neutral point effectively grounded system, and the corresponding sag type labels.

3. The SDAE-BP descent type identification method based on improved PSO optimization according to claim 1, characterized in that: In step (5), the SDAE-BP network parameters include the number of hidden layers in the first layer of SDAE, the number of hidden layers in the second layer of SDAE, the number of iterations of the SDAE training model, the number of hidden layers in the BP neural network, the number of iterations of the BP neural network, the proportion of SDAE input data set to 0, the learning rate of the BP neural network, and the momentum parameter of the BP neural network.

4. The SDAE-BP descent type identification method based on improved PSO optimization according to claim 1, characterized in that, The update formulas for particle velocity and particle position are expressed as follows: ； ； In the formula, Indicates the velocity of the particle; Indicates inertia weight; Indicates the position of the particle; Indicates an individual extreme value; This represents the weighted average of the individual extreme values ​​in the particle's history. This represents the average of the individual extreme values ​​of all particles; Indicates the global extremum; This represents a random number between (0, 1); Indicates the learning factor; The weighted average of historical individual extreme values ​​is expressed as follows: ; In the formula, The weights represent the extreme values ​​of the i-th training instance. The value is ; The average of the extreme values ​​of all individual particles in this training is expressed as follows: ; In the formula, This represents the individual extreme value of the i-th particle in this training, and N represents the population size, which ranges from 20 to 40.

5. The SDAE-BP slump type identification method based on improved PSO optimization according to claim 1, characterized in that: The training of SDAE-BP includes training of a stacked noise-reducing autoencoder SDAE and training of a BP neural network model.

6. The SDAE-BP slump type identification method based on improved PSO optimization according to claim 5, characterized in that, The training of the stacked noise reduction autoencoder (SDAE) includes: inputting voltage sag test set samples into the pre-built SDAE network, training each layer of the SDAE network in sequence, determining the optimal weights and optimal biases that minimize the reconstruction error of each layer of the DAE, and obtaining the feature parameters of the output of the last layer of the DAE.

7. The SDAE-BP slump type identification method based on improved PSO optimization according to claim 5, characterized in that, The training of the BP neural network model includes: configuring corresponding temporary descent type labels for the final features extracted by the SDAE network according to the training samples corresponding to the features; inputting the features and corresponding temporary descent type labels into the pre-constructed BP neural network to complete the training of the BP neural network.

8. A device for identifying SDAE-BP descent type based on improved PSO optimization, characterized in that, The SDAE-BP descent type identification method based on improved PSO optimization as described in claim 1 includes: The sampling data acquisition module is configured to acquire voltage sag sampling data of the type of sag to be identified. The preprocessing module is used to preprocess the voltage sag sampling data obtained by the sampling data acquisition module; The parameter optimization module is used to obtain the optimal parameters of the SDAE-BP network. Specifically, it utilizes an improved PSO algorithm to perform transient feature extraction and adaptive selection of SDAE-BP network parameters for classification training. This includes the following steps: (5.1) Based on the dimension of the union of the temporary data and the range of parameter values, set the range of values ​​of the parameters that need to be optimized in the SDAE-BP network, and use it as the range interval for updating particle velocity and position. When the value exceeds the range, take the maximum or minimum value of the interval. (5.2) Initialize the particle swarm parameters, including the initial position and velocity of the particles, the population size, inertia weight, acceleration factor, number of iterations of the particles, and target threshold; (5.3) The voltage sag classification error rate obtained from training SDAE-BP is used as the fitness value of the improved PSO algorithm. If the fitness value is better than the individual optimal value and the global optimal value, then these two values ​​are updated. At the same time, in order to increase the diversity of the population, the 10% of particles with the worst fitness are replaced with randomly generated new particles. If the training reaches the accuracy requirement or the number of iterations, the obtained parameters are used as the optimal structure parameters of SDAE-BP. Otherwise, the particle velocity and particle position are updated, and this step is repeated. The feature extraction module is used to input the voltage sag data obtained from the preprocessing module into the stacked noise reduction autoencoder SDAE pre-trained by the parameter optimization module to obtain the voltage sag feature data extracted by SDAE. And a voltage sag type identification module, which is used to input the feature data obtained by the voltage sag feature extraction module and its corresponding type label data into the BP neural network model pre-trained by the parameter optimization module, and determine the voltage sag category based on its output.

9. A SDAE-BP descent type identification system based on improved PSO optimization, comprising a processor and a memory, wherein the memory stores computer instructions, characterized in that: The processor operates according to computer instructions to execute the SDAE-BP descent type identification method based on improved PSO optimization as described in any one of claims 1-7.

Images