Improved iTransform-based voltage sag collaborative traceability method under multi-source information fusion

Abstract

The invention discloses a voltage sag collaborative traceability method under multi-source information fusion based on improved iTransform. The method specifically comprises the following steps: acquiring voltage amplitude data under a system operation condition; designing an influence factor of the voltage amplitude and the strong positive correlation; performing normalization preprocessing on the input data set; marking the data set and dividing the data set into a training set and a test set; an improved iTransform model is built, and an improved iTransform model is built; and training the improved iTransform model by using the training set, and testing the model by using the test set. According to the method, on one hand, the influence of the multi-source information influence factors on the voltage sag traceability result is considered, and voltage sag collaborative traceability under multi-source information fusion is achieved; and on the other hand, the traditional iTransform is fused with the Bi-IndRNN, so that the local feature extraction capability is increased while the global view is reserved, and the accuracy of the voltage sag traceability result is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system analysis, and in particular relates to a voltage sag collaborative tracing method under multi-source information fusion based on an improved iTransformer. Background Art

[0002] Accurate voltage sag tracing is the main problem that needs to be solved after a voltage sag event occurs. This is of great significance for clarifying the responsibility for the event and improving the power quality of the power grid. With the increasing trend of power electronics in power systems and the growing development of AC / DC hybrid power grids, the scale of power systems has increased and the complexity has increased. The factors affecting voltage sags have become more complicated. For example, multi-source information such as meteorology and operating conditions will affect the voltage sag process. Therefore, the voltage sag process is diverse, complex, uncertain, and deeply coupled. Voltage sag tracing needs to consider the impact of multi-source information.

[0003] Most of the existing research on voltage sag tracing methods is based on physical model-driven methods, that is, by constructing a mathematical mechanism model for analysis, the physical characteristic quantities of the system voltage or current signal when a voltage sag occurs are established, and the voltage sag is traced after the threshold is manually set. However, the method based on physical characteristics is overly dependent on expert experience, and it is necessary to manually extract features and set thresholds in advance. As more and more new equipment is gradually connected to the power grid, the voltage waveform during the voltage sag process has a strong correlation with the equipment operating characteristics, tolerance characteristics, etc. Therefore, voltage sags have taken on a complex and changeable form. It is difficult to establish accurate and universal features based on physical characteristics, and the threshold setting is difficult, resulting in poor universality of such methods and prominent defects.

[0004] Different from the above-mentioned physical feature methods, the data feature-based method does not rely on physical feature quantities, providing a new research perspective for voltage sag tracing. With the construction and development of smart grids, power grids at all levels have built a certain scale of power information platforms and power quality monitoring systems. For each voltage sag, monitoring stations can collect a large amount of on-site monitoring data. In particular, these monitoring data themselves carry a lot of valuable information, and behind them are hidden certain operating states or laws of voltage sags. On the surface, independent voltage sags in regional power grids occur randomly, isolated from each other and unrelated. However, after analyzing a large amount of monitoring data, it is found that there are some meaningful laws and mutual connections with multi-source information. Using these laws, voltage sag tracing with multi-source information coordination can be realized. Summary of the invention

[0005] The purpose of the present invention is to provide a voltage sag collaborative tracing method under multi-source information fusion based on an improved iTransformer, to achieve voltage sag collaborative tracing under multi-source information fusion, and to improve the accuracy of voltage sag tracing results.

[0006] The technical solution adopted by the present invention is a voltage sag collaborative tracing method based on multi-source information fusion of an improved iTransformer, which is specifically implemented in the following steps:

[0007] Step 1: Collect voltage amplitude signals under 6 system operating conditions;

[0008] Step 2: Collect multi-source information influencing factors corresponding to each operating condition;

[0009] Step 3: Use the Pearson correlation coefficient to compare and analyze the correlation between the voltage sag amplitude and the multi-source influencing factors in step 2;

[0010] Step 4: Compare the Pearson correlation coefficient calculated in step 3 and select the influencing factors with strong positive correlation;

[0011] Step 5: Establishing a model input data set: For each of the six system operating conditions in step 1, the model input data set includes the voltage amplitude data in step 1 and the strong positive correlation influencing factors in step 4 corresponding to each of the voltage amplitude data in step 1, thereby obtaining the model input data set;

[0012] Step 6: Normalize and preprocess the input data set;

[0013] Step 7: Label the model input data set and divide it into training set and test set;

[0014] Step 8: Build and improve the iTransformer model;

[0015] Step 9: Train the improved iTransformer model using the training set to obtain the improved iTransformer model after training;

[0016] Step 10: After training, use the test set to test the improved iTransformer model.

[0017] The present invention is also characterized in that:

[0018] In step 1, specifically: collect voltage amplitude data in the power system to be analyzed, where each line needs to include 6 system operating conditions including normal operation, single-phase grounding, two-phase grounding, three-phase grounding, high-power induction motor starting, and transformer input, and collect voltage amplitude data under these 6 system operating conditions.

[0019] In step 2, the multi-source information influencing factors include meteorological data, operating conditions and line length. The meteorological data includes temperature and humidity, and the operating conditions include voltage level and user category.

[0020] In step 3, the Pearson correlation coefficient ρ p,q The calculation formula is shown in formula (1):

[0021]

[0022] Where n is the number of voltage amplitude sampling points collected in one power frequency voltage cycle; p i and q i is the i-th observation value of two variables p and q; For all p i The average value of For all q i The average value of p,q The value range is [-1, 1].

[0023] In step 4, compare the Pearson correlation coefficient calculated in step 3 and select ρ in step 3 p,q The impact factors >0.5 together constitute the multi-source information impact factors with strong positive correlation that need to be considered in the subsequent model calculations.

[0024] In step 7, specifically:

[0025] All the lines in the power system to be analyzed that may experience voltage sag are numbered 1 to m, and the model input data set is labeled, where each sample data in the data set contains an input item and an output item, the input item is the voltage data preprocessed in step 6 and the multi-source information influencing factor with strong positive correlation; the output item is the line number i that experiences voltage sag in the working condition corresponding to the input item, and then 80% of the total sample is randomly selected as the training set, and the remaining 20% ​​is used as the test set.

[0026] In step 8, based on improving the iTransformer model envelope input part, improving the iTransformer structure layer and the output part;

[0027] The input part contains only one input layer, which converts the input data obtained in step 7 into fixed-dimensional matrix data. Its data format is [Batchsize, N, T]; where Batchsize is the number of samples input to the model each time, N is the dimension of the input data, and T is the sequence length;

[0028] The improved iTransformer structure layer includes multiple iTransformer structures, each of which includes two modules: a multi-head attention module and a Bi-IndRNN (BidirectionalIndependently Recurrent Neural Network) module; the multi-head attention module includes multiple single-head attention units, each of which performs a scaled dot product attention operation, and then concatenates the outputs of all single-head attention units to form a long vector; then, the concatenated vector is linearly transformed by a fully connected layer to integrate information from different single-head attention units to obtain the final output of the multi-head attention module; each Bi-IndRNN module includes a forward IndRNN and a backward IndRNN; both the multi-head attention module and the Bi-IndRNN module use residual connections, and the outputs are connected to the normalization layer;

[0029] The output part includes a fully connected layer and Softmax to obtain the line number i where the fault is located.

[0030] In step 9, specifically:

[0031] The improved iTransformer model is trained using the sample training set in step 7, the back propagation algorithm is used to update the parameters, the Adam optimizer is used for training, the loss function is the cross-entropy loss function, the input data is the input item of each sample data in the training set in step 7, and the output data is the output item of each sample data in the training set in step 7.

[0032] The beneficial effects of the present invention are as follows: on the one hand, the method of the present invention takes into account the influence of multi-source information influencing factors on the voltage sag tracing results, realizes the collaborative tracing of voltage sags under multi-source information fusion, and avoids the problem of low accuracy of voltage sag tracing results when only a single influencing factor is considered; on the other hand, the traditional iTransformer is fused with Bi-IndRNN to increase the local feature extraction capability while retaining the global perspective, and can accurately extract the global features and local features of the input signal for time-varying scenarios such as voltage sag tracing, thereby improving the accuracy of the voltage sag tracing results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a structural diagram of a power system used in an embodiment of the present invention;

[0034] Figure 2 The overall framework diagram of the improved iTransformer model proposed in the present invention;

[0035] Figure 3This is a structural diagram of the improved iTransformer model proposed in the present invention;

[0036] Figure 4 A comparison chart of test results between the improved iTransformer model of the present invention and the traditional iTransformer model;

[0037] Figure 5 The accuracy test results of improving the iTransformer model at different numbers of iterations. DETAILED DESCRIPTION

[0038] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1

[0040] The present invention is based on the voltage sag collaborative tracing method under multi-source information fusion of improved iTransformer, such as Figure 1 As shown, first, the PSCAD simulation software is used to set 6 working conditions for each line in the IEEE-30 system, and the voltage amplitudes at nodes 2, 15, 21, and 25 in the IEEE-30 system are collected; then, the system voltage level, user category, and line length data are collected, and the Pearson correlation coefficient is calculated, and the quantity with a Pearson correlation coefficient greater than 0.5 is selected as the strong positive correlation multi-source information influencing factor determined in this embodiment. Secondly, the sample data set is manually annotated and divided into a training sample set and a test sample set; then, the built improved iTransformer model is trained; finally, the trained improved iTransformer model is tested.

[0041] Example 2

[0042] The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer is implemented in the following steps:

[0043] Step 1: Collect voltage amplitude signals under 6 system operating conditions;

[0044] Step 2: Collect multi-source information influencing factors corresponding to each operating condition;

[0045] Step 3: Use the Pearson correlation coefficient to compare and analyze the correlation between the voltage sag amplitude and the multi-source influencing factors in step 2;

[0046] Step 4: Compare the Pearson correlation coefficient calculated in step 3 and select the influencing factors with strong positive correlation;

[0047] Step 5: Establishing a model input data set: For each of the six system operating conditions in step 1, the model input data set includes the voltage amplitude data in step 1 and the strong positive correlation influencing factors in step 4 corresponding to each of the voltage amplitude data in step 1, thereby obtaining the model input data set;

[0048] Step 6: Normalize and preprocess the input data set;

[0049] Step 7: Label the model input data set and divide it into training set and test set;

[0050] Step 8: Build and improve the iTransformer model;

[0051] Step 9: Train the improved iTransformer model using the training set to obtain the improved iTransformer model after training;

[0052] Step 10: After training, use the test set to test the improved iTransformer model.

[0053] Example 3

[0054] The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer is specifically implemented according to the following steps:

[0055] Step 1: Collect the voltage amplitude signal in the IEEE-30 system, where each line needs to include six system operating conditions: normal operation, single-phase grounding, two-phase grounding, three-phase grounding, high-power induction motor starting, and transformer input, and collect the voltage amplitude data under these six system operating conditions;

[0056] The voltage amplitude data collection is specifically: combining the voltage transformer, voltage sensor or power quality measuring instrument installed in the power system to realize the voltage amplitude data collection.

[0057] Step 2: For the six operating conditions in step 1, collect multi-source information influencing factors corresponding to each operating condition, including meteorological data (including temperature and humidity), operating conditions (including voltage level and user category), and line length;

[0058] Meteorological data (including temperature and humidity) are collected specifically as follows: temperature and humidity data are collected through web crawlers or by installing meteorological data collection equipment; voltage levels, user categories, and line lengths are collected by querying the power company or after setting simulation conditions;

[0059] Step 3: Use the Pearson correlation coefficient to compare and analyze the correlation between the voltage sag amplitude and the multi-source influencing factors in step 2;

[0060] Pearson correlation coefficient ρ p,q The calculation formula is shown in formula (1):

[0061]

[0062] Where n is the number of voltage amplitude sampling points collected in one power frequency voltage cycle; p i and q i is the i-th observation value of two variables p and q; For all p i The average value of For all q i The average value of ρ p,q The value range is [-1, 1]. The closer it is to 0, the more it is considered that there is no correlation between the two influencing factors. The closer it is to 1, the more it is considered that there is a strong positive correlation between the two factors. The closer it is to -1, the stronger the negative correlation between the two factors.

[0063] Step 4: Compare the Pearson correlation coefficient calculated in step 3. The closer the value is to 0, the more it is considered that there is no correlation between the voltage sag tracing result and the influencing factor. The closer the value is to 1, the more it is considered that there is a strong positive correlation between the voltage sag tracing result and the influencing factor. The closer the value is to -1, the more it is considered that there is a strong negative correlation between the voltage sag tracing result and the influencing factor. Select the influencing factors with strong positive correlation.

[0064] Select ρ in step 3 p,q The impact factors >0.5 together constitute the multi-source information impact factors with strong positive correlation that need to be considered in the subsequent model calculation;

[0065] Step 5: Establish a model input data set. For each of the six system operating conditions in step 1, the model input data set includes the voltage amplitude data in step 1 and the multi-source information influencing factors with strong positive correlation in step 4 corresponding to each of the voltage amplitude data in step 1, and obtain an initial model input data set.

[0066] Step 6: Perform normalization preprocessing on the input data set formed in step 5 to obtain a normalized preprocessed model input data set;

[0067] Normalization preprocessing, as shown in formula (2):

[0068]

[0069] Among them, y * is the preprocessed data output, y is the original collected data, and y max is the maximum value of the input sample data, y min is the minimum value in the input sample data.

[0070] Step 7: Label the model input data set obtained in step 6 and randomly divide it into a training set and a test set in a ratio of 80% and 20%;

[0071] All the lines in the power system to be analyzed that may experience voltage sag are numbered 1 to m, and the model input data set is labeled, where each sample data in the data set contains an input item and an output item, the input item is the voltage data preprocessed in step 6 and the strong positive correlation multi-source information influencing factor; the output item is the line number i (i∈(1,2,······,m)) where the voltage sag occurs in the working condition corresponding to the input item one by one to form the total sample; then, 80% of the total sample is randomly selected as the training set, and the remaining 20% ​​is used as the test set;

[0072] Step 8: Build and improve the iTransformer model;

[0073] Based on the improved iTransformer model envelope input part, the improved iTransformer structure layer, and the output part, wherein;

[0074] The input part contains only one input layer, which converts the input data obtained in step 7 into matrix data of fixed dimension. Its data format is [Batchsize, N, T]; Batchsize is the number of samples input to the model each time, N is the dimension of the input data, and T is the sequence length.

[0075] Improve the iTransformer structure layer, such as Figure 2 and Figure 3 As shown, it contains multiple iTransformer structures, and each iTransformer structure part includes two modules: a multi-head attention module and a bidirectional independent recurrent neural network Bi-IndRNN module. The multi-head attention module contains multiple single-head attention units, each of which performs a scaled dot product attention operation, and then concatenates the outputs of all single-head attention units to form a long vector. Then, the concatenated vector is linearly transformed with a fully connected layer to integrate information from different single-head attention units to obtain the final output of the multi-head attention module. Each Bi-IndRNN module includes a forward IndRNN and a backward IndRNN.

[0076] Specifically, the data x passed from the input layer to the improved iTransformer structure layer are respectively combined with the matrix W q , matrix W k and the matrix W v Multiply them together to get the query vector q, key vector k and value vector v, as shown in formula (3):

[0077]

[0078] For each single-head attention unit in the multi-head attention module, a scaled dot product attention operation is performed, as shown in formula (4):

[0079]

[0080] Where, d k is the dimension of the input layer.

[0081] In order to improve the generalization ability of the Transformer network, multiple single-head attention units are stacked side by side and then projected through a fully connected layer to complete the mapping of global features. This mechanism enables the network to focus on different features in the subspace, thereby improving its generalization ability.

[0082] Each Bi-IndRNN module includes a forward IndRNN and a backward IndRNN. The forward IndRNN calculation process is shown in formula (5), the backward IndRNN calculation process is shown in formula (6), and the bidirectional Bi-IndRNN calculation process is shown in formula (7).

[0083]

[0084] Among them, x t , is the input at time t, the forward IndRNN output state, and the backward IndRNN output state; The input weight, recurrent input weight, and bias matrix of the forward IndRNN; H is the input weight, recurrent input weight, and bias matrix of the backward IndRNN; t is the output of the Bi-IndRNN module.

[0085] Among them, both the multi-head attention module and the Bi-IndRNN module use residual connections, and the outputs are connected to the normalization layer to make all variables in a relatively uniform distribution, reduce the differences caused by different variable value ranges, and accelerate the convergence of the model. On the one hand, the multi-head attention module directly extracts the global features corresponding to the input after layer normalization and then passes through the fully connected layer; on the other hand, the multi-head attention model further extracts the local features corresponding to the input after layer normalization, Bi-IndRNN module, layer normalization and fully connected layer. After dimensional splicing of global features and local features, the complete features corresponding to the input can be extracted. The output features of the improved iTransformer structure layer are further sent to the output part of the next layer;

[0086] Although the traditional iTransformer structure can accurately extract global features, it lacks analysis of local features and is difficult to cope with time-varying voltage sag tracing scenarios. The improved iTransformer combines the traditional iTransformer with Bi-IndRNN, retaining the global view while increasing the local feature extraction capability, and can accurately extract the global and local features of the input signal for time-varying scenarios such as voltage sag tracing.

[0087] The output part, including the fully connected layer and Softmax, obtains the line number i (i∈(1,2,······,m)) where the fault is located.

[0088] Step 9: Use the sample training set in step 7 to train the improved iTransformer model to obtain the improved iTransformer model after training;

[0089] Specifically: the improved iTransformer model structure is trained using the sample training set in step 7, the back propagation algorithm is used to update the parameters, the Adam optimizer is used for training, the loss function is the cross-entropy loss function, the input data is the input item of each sample data in the training set in step 7, and the output data is the output item of each sample data in the training set in step 7.

[0090] Step 10: After training, use the test set in step 7 to test the improved iTransformer model.

[0091] Specifically, the improved iTransformer model structure obtained by training in step 9 is tested using the samples in step 7. The input data is the input item of each sample data in the test set in step 7, and the output data is the output item of each sample data in the test set in step 7.

[0092] Example 4

[0093] The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer is specifically implemented according to the following steps:

[0094] Step 1: Combine Figure 1 The IEEE-30 system topology shown is simulated using PSCAD simulation software when the system is running under 6 different working conditions, with line parameters, load capacity, and system impedance changed respectively. In this embodiment, the three-phase voltage amplitudes at nodes 2, 15, 21, and 25 are collected, for a total of 4 three-phase node voltage data.

[0095] Step 2: Collect the voltage level of each node and the line length of each line in the IEEE-30 system to form the multi-source information impact factor to be considered in this embodiment;

[0096] Step 3: Determine multi-source influencing factors;

[0097] Step 4: Count the results of step 3 and select ρ in step 3 p,q The voltage level and line length >0.5 together constitute the strong positive correlation multi-source information impact factor considered in this embodiment;

[0098] Step 5: For each of the six system operating conditions in step 1, collect the voltage amplitude data in step 1, the voltage level and line length data of the strong positive correlation multi-source information influencing factors determined in step 4, which correspond to each operating condition;

[0099] Step 6: pre-processing the voltage amplitude collected in step 1 to obtain pre-processed data;

[0100] Step 7: Each sample data established contains an input item and an output item. The input item is the voltage data and multi-source information influencing factors preprocessed in step 6; the output item is the line number i (i∈(1,2,······,m)) of the line where the voltage sag occurs in step 7, which corresponds to the input item one by one, to form the total sample; then, 80% of the total sample is randomly selected as the training set, and the remaining 20% ​​is used as the test set;

[0101] Step 8: Build the overall structure of the improved iTransformer model; the specific process is:

[0102] Based on improving the iTransformer model envelope input part, improving the iTransformer structure layer, and the output part;

[0103] The input part only includes one input layer, which converts the input data obtained in step 7 into matrix data of fixed dimension, and its data format is [Batchsize, N, T]; wherein Batchsize is the number of samples input to the model each time, N is the dimension of the input data, and T is the sequence length. The total amount of sampled data used in the present invention is 10240 sample data. Each sample data contains 256 time steps, T=256, and Batchsize is 512; each time step contains 4 (4 measurement nodes) × 3 (three-phase voltage) + 2 (multi-source information influencing factors composed of line length and voltage level), N=14.

[0104] The improved iTransformer structure layer includes multiple iTransformer structures, each of which includes two modules: a multi-head attention module and a bidirectional independent recurrent neural network Bi-IndRNN (Bi-IndRNN, Bidirectional Independently Recurrent Neural Network) module. The multi-head attention module contains multiple single-head attention units, each of which performs a scaled dot product attention operation, and then concatenates the outputs of all single-head attention units to form a long vector. Then, the concatenated vector is linearly transformed with a fully connected layer to integrate information from different single-head attention units to obtain the final output of the multi-head attention module. Each Bi-IndRNN module includes a forward IndRNN and a backward IndRNN.

[0105] In this embodiment, six improved iTransformer structures are used to extract the features of input data, including the attention module and the Bi-IndRNN module. s} (s is the time step) is the matrix data obtained after the input part is processed, and its data format is [512,14,256]; then after the multi-head attention module, we get the same matrix as {v1,v2,…,v s}corresponding to the global features {M1,M2,…,M s};

[0106] In this embodiment, the multi-head attention module contains 6 single-head attentions. For each head, the scaled dot product attention operation shown in formula (4) is performed once, and the outputs of all single-head attention units are concatenated together to form a long vector. Then, the concatenated vector is linearly transformed by the fully connected layer to integrate the information from different heads, and the final multi-head attention output {G1, G2, ..., G s {G1,G2,……,G s} is further sent to the normalization layer and residual connection is performed, and the global features {M1,M2,…,M s}. At the same time, {M1,M2,……,M s} is further transmitted to the Bi-IndRNN module, and the output {H1, H2, ..., H s} is sent to the layer for normalization and residual connection, and then a fully connected layer is used to obtain the local features {L1, L2, ..., L s}.

[0107] Global features {M1,M2,...,Ms} and local features {L1,L2,...,L s} and concatenate them to get {v1,v2,…,v s The output complete features of the improved iTransformer structure layer are further sent to the output part of the next layer.

[0108] The output part includes a fully connected layer and a Softmax layer to obtain the line number i (i∈(1,2,······,m)) where the fault is located. In this embodiment, m is 41.

[0109] Step 9: Use the sample training set in step 7 to train the improved iTransformer overall model structure, use the back propagation algorithm to update the parameters, use the Adam optimizer for training, the loss function is the cross-entropy loss function, the input data is the input item of each sample data in the training set in step 7, and the output data is the output item of each sample data in the training set in step 7. Try to use different hyperparameters to adjust the model during training and obtain the model after training.

[0110] Step 10: Use the samples from step 7 to test the improved iTransformer overall model structure trained in step 9.

[0111] The input data is the input item of each sample data in the test set in step 7, and the output data is the output item of each sample data in the test set in step 7.

[0112] Example 5

[0113] further, Figure 4 The model test results of Example 4 are shown. When only single information is considered, the accuracy of the improved iTransformer model is increased from 91.27% of the traditional iTransformer model to 93.58%, an increase of 2.31%; when multi-source information is considered, the accuracy of the improved iTransformer model is increased from 92.71% of the traditional iTransformer model to 99.83%, an increase of 7.12%. Therefore, for voltage sag tracing, on the one hand, it is necessary to combine multi-source information to improve the accuracy of the tracing results; on the other hand, it is also necessary to adopt the improved iTransformer model proposed in the present invention to further improve the accuracy of the tracing results.

[0114] Figure 5 The accuracy test results of the model under different iteration times in this embodiment are as follows: Figure 5 It can be seen that the accuracy of the improved iTransformer model continues to increase, and there is no overfitting phenomenon.

[0115] Example 6

[0116] The present invention is based on the collaborative tracing method of voltage sag under multi-source information fusion of improved iTransformer. On the one hand, it considers the influence of multi-source information influencing factors on the voltage sag tracing results, realizes the collaborative tracing of voltage sag under multi-source information fusion, and avoids the problem of low accuracy of voltage sag tracing results when only considering a single influencing factor. On the other hand, the present invention fuses the traditional iTransformer with Bi-IndRNN, increases the local feature extraction capability while retaining the global field of view, and can accurately extract the global features and local features of the input signal for time-varying scenarios such as voltage sag tracing, thereby improving the accuracy of the voltage sag tracing results.

Claims

1. A voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer, characterized in that: Follow the steps below to implement it: Step 1: Collect voltage amplitude data under 6 system operating conditions; Step 2: Collect multi-source information influencing factors corresponding to each operating condition; Step 3: Use the Pearson correlation coefficient to compare and analyze the correlation between the voltage sag amplitude and the multi-source influencing factors in step 2; Step 4: Compare the Pearson correlation coefficient calculated in step 3 and select the influencing factors with strong positive correlation; Step 5: Establishing a model input data set: For each of the six system operating conditions in step 1, the model input data set includes the voltage amplitude data in step 1 and the strong positive correlation influencing factors in step 4 corresponding to each of the voltage amplitude data in step 1, thereby obtaining the model input data set; Step 6: Normalize and preprocess the input data set; Step 7: Label the model input data set and divide it into training set and test set; Step 8: Build and improve the iTransformer model; Step 9: Train the improved iTransformer model using the training set to obtain the improved iTransformer model after training; Step 10: After training, use the test set to test the improved iTransformer model.

2. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 1, characterized in that: In the step 1, specifically: collecting voltage amplitude data in the power system to be analyzed, wherein each line needs to include six system operating conditions including normal operation, single-phase grounding, two-phase grounding, three-phase grounding, high-power induction motor starting, and transformer input, and collecting voltage amplitude data under these six system operating conditions.

3. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 2 is characterized in that: In step 2, the multi-source information influencing factors include meteorological data, operating conditions and line length, the meteorological data includes temperature and humidity, and the operating conditions include voltage level and user category.

4. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 3 is characterized in that: In step 3, the Pearson correlation coefficient ρ p,q The calculation formula is shown in formula (1): Where n is the number of voltage amplitude sampling points collected in one power frequency voltage cycle; p i and q i is the i-th observation value of two variables p and q; For all p i The average value of For all q i The average value of p,q The value range is [-1, 1].

5. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 4, characterized in that: In step 4, compare the Pearson correlation coefficient calculated in step 3 and select the ρ in step 3. p,q The impact factors >0.5 together constitute the strong positive correlation multi-source information impact factors that need to be considered in the subsequent model calculations.

6. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 5, characterized in that: In the step 7, specifically: All the lines with voltage sag in the power system to be analyzed are numbered 1 to m, and the model input data set is labeled, where each sample data in the data set contains an input item and an output item. The input item is the voltage data and multi-source information influencing factors preprocessed in step 6; the output item is the line number i of the voltage sag corresponding to the input item to form the total sample; then, 80% of the total sample is randomly selected as the training set, and the remaining 20% ​​is used as the test set.

7. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 6, characterized in that: In the step 8, based on improving the iTransformer model envelope input part, improving the iTransformer structure layer and the output part; The input part contains only one input layer, which converts the input data obtained in step 7 into fixed-dimensional matrix data. Its data format is [Batchsize, N, T]; where Batchsize is the number of samples input to the model each time, N is the dimension of the input data, and T is the sequence length; The improved iTransformer structure layer includes multiple iTransformer structures, each of which includes two modules: a multi-head attention module and a Bi-IndRNN module; the multi-head attention module includes multiple single-head attention units, each of which performs a scaled dot product attention operation, and then concatenates the outputs of all single-head attention units to form a long vector; then, the concatenated vector is linearly transformed by a fully connected layer to integrate information from different single-head attention units to obtain the final output of the multi-head attention module; each Bi-IndRNN module includes a forward IndRNN and a backward IndRNN; both the multi-head attention module and the Bi-IndRNN module use residual connections, and the outputs are connected to the normalization layer; The output part includes a fully connected layer and Softmax to obtain the line number i where the fault is located.

8. The voltage sag collaborative tracing method based on multi-source information fusion of improved iTransformer according to claim 7, characterized in that: In step 9, specifically: The improved iTransformer model is trained using the sample training set in step 7, the back propagation algorithm is used to update the parameters, the Adam optimizer is used for training, the loss function is the cross-entropy loss function, the input data is the input item of each sample data in the training set in step 7, and the output data is the output item of each sample data in the training set in step 7.

Images