Converter transformer saturation protection method and system based on fusion network

Abstract

The invention discloses a converter transformer saturation protection method and system based on a fusion network. The method comprises the following steps: acquiring three-phase current and neutral point zero-sequence current of a converter transformer network side; inputting the three-phase current and the neutral point zero-sequence current of the converter transformer network side into the trained LSTM-Transform model of the fusion network to obtain a DC magnetic bias prediction value; according to the direct current magnetic bias prediction value, when a single working condition criterion or a composite working condition criterion is met, saturation protection is locked immediately; wherein the single working condition criterion comprises that the direct current magnetic bias prediction values are all smaller than a set upper limit within a set time, and the discrete slope of the direct current magnetic bias prediction values is smaller than a set slope threshold value; the composite working condition criterion includes that the proportion of the number of the direct current magnetic bias prediction values smaller than a set threshold value in the total number of all the direct current magnetic bias prediction values is larger than a set proportion threshold value. According to the method, when the converter transformer only has the magnetizing inrush current or the magnetizing inrush current accounts for the main component, the locking criterion is effectively utilized, and saturation protection maloperation is prevented.

Description

Technical Field

[0001] The present invention belongs to the technical field of direct current transmission converter transformer protection, and more specifically, relates to a converter transformer saturation protection method and system based on a fusion network. Background Art

[0002] UHV DC transmission technology has the characteristics of long distance, large capacity, low loss and good controllability. It is an effective way to realize the intensive development and transmission of clean energy and the cross-regional optimization of energy allocation. The UHV DC transmission system uses a three-phase split converter transformer. During the operation of the converter transformer, DC bias may occur, causing the converter transformer core to be continuously saturated and causing local overheating. Therefore, the converter transformer is equipped with saturation protection to prevent the DC bias from affecting the converter transformer core.

[0003] In actual operation, since the saturation protection calculates the peak value of the neutral point magnetizing current on the grid side to indirectly calculate the magnitude of the neutral point DC current, the excitation surge current during the empty charging and switching of the converter transformer is easily affected by the closing angle, residual magnetism, etc., which can easily lead to the saturation protection malfunction, seriously affecting the reliability of the saturation protection action. In the past three years, many converter stations have experienced excessive excitation surge currents, which have caused the converter transformer saturation protection to operate, and then caused the equipment to trip.

[0004] Therefore, a method for identifying the excitation inrush current and DC bias of a converter transformer is needed, and on this basis, an improvement on the saturation protection in the prior art is implemented. Summary of the invention

[0005] In order to solve the deficiencies in the prior art, the present invention provides a converter transformer saturation protection method and system based on a fusion network, which effectively locks the saturation protection when only the excitation inrush current exists in the converter transformer or the excitation inrush current accounts for the main component, so as to prevent the saturation protection from malfunctioning due to excessive excitation inrush current.

[0006] The present invention adopts the following technical solution.

[0007] A first aspect of the present invention provides a commutation variable saturation protection method based on a fusion network, comprising:

[0008] Collect three-phase current and neutral point zero-sequence current on the commutation transformer grid side;

[0009] The three-phase current and neutral point zero-sequence current on the converter transformer side are input into the trained fusion network LSTM-Transformer model to obtain the DC bias magnetic prediction value;

[0010] According to the DC bias magnetization prediction value, when a single operating condition criterion or a composite operating condition criterion is met, the saturation protection is immediately locked; wherein, the single operating condition criterion includes that the DC bias magnetization prediction values ​​are all less than the set upper limit within the set time, and the discrete slope of the DC bias magnetization prediction value is less than the set slope threshold; the composite operating condition criterion includes that the number of DC bias magnetization prediction values ​​less than the set threshold accounts for a proportion of the total number of all DC bias magnetization prediction values ​​greater than the set proportion threshold.

[0011] Preferably, the training of the fusion network LSTM-Transformer model includes:

[0012] Step A.1, determining multiple excitation inrush current simulation scenarios and parameters with the set closing angle as the step length;

[0013] Step A.2, determining a plurality of DC bias simulation scenarios and parameters within a set neutral point current range on the commutation transformer side and with a set neutral point current step; forming a training set with a plurality of excitation inrush current simulation scenarios and parameters and a plurality of DC bias simulation scenarios and parameters;

[0014] Step A.3: Construct the feature vector x using the state vectors in the training set before time t t , expressed as the following formula,

[0015] x t ={f t ,v t ,k t ,s t ,p t} T

[0016] Where:

[0017] f t , v t , k t ,s t , p t All represent the state vectors before time t;

[0018] Step A.4, normalize and segment the feature vector, and convert the segmented time series x n = {X1, X2, ..., X n} is input into the fusion network LSTM-Transformer model for feature learning, where normalization and segmentation processing are expressed as follows:

[0019]

[0020] Where:

[0021] x t ' represents the normalized feature vector;

[0022] min(x t )、max(x t ) are respectively a vector consisting of the minimum values ​​of each state vector in the eigenvector and a vector consisting of the maximum values ​​of each state vector;

[0023] x1′, x′2, …, x′ n+k-1 are the normalized eigenvectors at the first moment, the second moment, ... the n+k-1th moment respectively;

[0024] n is the number of time periods;

[0025] k is the number of moments contained in each time period;

[0026] X1…X n represents the normalized characteristic value sequence of the neutral point current on the grid side of the commutation transformer in n time periods;

[0027] Step A.5, output the DC bias magnetic prediction value based on the feature learning result of the previous step and the feature vector of the current step.

[0028] Preferably, the fusion network LSTM-Transformer model includes an LSTM module, a fusion layer and a multi-head self-attention mechanism, which are respectively expressed by the following formulas:

[0029]

[0030] Where:

[0031] Represents the output vector of the LSTM network;

[0032] Represents the current hidden layer vector;

[0033] G t Represents the output vector of the multi-head attention operation;

[0034] g(h t-1 ) represents the hidden layer vector after the sigmoid function operation;

[0035] L represents recursive operation based on long short-term memory, F represents feature fusion operation, and T represents multi-head attention operation.

[0036] Preferably, the single operating condition criterion is used to lock the saturation protection when it is determined that the neutral point zero-sequence current is an excitation surge current, and is expressed by the following formula:

[0037]

[0038] Where:

[0039] F(t) is the DC bias magnetic prediction value;

[0040] a is the first threshold;

[0041] k1 represents the discrete slope;

[0042] N is the total number of DC bias magnetic prediction values;

[0043] F(t=t i ) indicates t i The DC bias magnetic prediction value at the moment;

[0044] b is the second threshold;

[0045] t s Indicates the set time.

[0046] Preferably, the setting time is 6s, and the three-phase current of the bushing on the converter transformer grid side and the neutral point zero-sequence current are collected every 0.02s and input into the fusion network for prediction; the first threshold is set to 0.1, and the second threshold is set to 0.05.

[0047] Preferably, when the composite working condition criterion is applied, it also includes: the DC bias magnetic content is lower than the maximum DC current value allowed to flow through the neutral point.

[0048] Preferably, the composite operating condition criterion is used to lock the saturation protection when it is determined that both the excitation inrush current and the DC bias magnetism exist in the neutral point zero-sequence current, but the proportion of the excitation inrush current is greater than the critical value, and is expressed by the following formula:

[0049]

[0050] Where:

[0051] F(t=t i ) indicates t i The DC bias magnetic prediction value at the moment;

[0052] c is the third threshold;

[0053] IF[*] is a conditional function, the value is 1 when the condition* is met, and 0 when the condition* is not met;

[0054] N is the total number of DC bias magnetic prediction values;

[0055] n* is the number of DC bias magnetic prediction values ​​less than the third threshold;

[0056] m is the fourth threshold;

[0057] k2 represents the proportion of predicted value points whose DC bias magnetic prediction values ​​are less than the third threshold value to the total predicted value points.

[0058] Preferably, the third threshold is set to 0.05, and the fourth threshold is set to 0.93.

[0059] A second aspect of the present invention provides a commutation transformer saturation protection system based on a fusion network, and the commutation transformer saturation protection method based on a fusion network is executed, comprising:

[0060] Signal acquisition module, used to collect three-phase current of bushing on the grid side of converter transformer and zero-sequence current of neutral point;

[0061] The model training module is used to train the LSTM-Transformer model of the fusion network and calculate the DC bias magnetic prediction value based on the three-phase current of the bushing on the grid side and the neutral point zero-sequence current actually collected;

[0062] The logic judgment module is used to compare the DC bias magnetic prediction value according to the set judgment conditions. When any condition is met, the saturation protection is immediately locked.

[0063] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0064] A fourth aspect of the present invention provides a computer device / equipment / system, comprising:

[0065] Memory, for storing computer programs / instructions;

[0066] A processor is used to execute the computer program / instructions to implement the steps of the method.

[0067] A fifth aspect of the present invention provides a computer program product, comprising a computer program / instruction, characterized in that the computer program / instruction implements the steps of the method when executed by a processor.

[0068] Compared with the prior art, the beneficial effects of the present invention include at least: the present invention provides a commutation transformer saturation protection method and system based on the fusion network LSTM-Transformer, which uses a neural network algorithm to effectively identify the scenarios of pure excitation inrush current, pure DC bias, and excitation inrush current combined with DC bias, and can effectively avoid the saturation protection action caused by pure excitation inrush current or excitation inrush current as the main component, thereby improving the reliability of the saturation protection action. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of a commutation variable saturation protection method based on a fusion network LSTM-Transformer provided by an embodiment of the present invention;

[0070] Figure 2It is a schematic diagram of the locking criterion of the LSTM-Transformer fusion network provided by an embodiment of the present invention;

[0071] Figure 3 This is a schematic diagram of LSTM-Transformer data training of the fusion network provided by an embodiment of the present invention;

[0072] Figure 4 It is a schematic diagram of a simple excitation inrush current prediction provided by an embodiment of the present invention;

[0073] Figure 5 This is a schematic diagram of a simple DC bias magnetic prediction provided by an embodiment of the present invention;

[0074] Figure 6 It is a schematic diagram of composite prediction of DC bias magnetization and excitation inrush current provided by an embodiment of the present invention;

[0075] Figure 7 It is a schematic diagram of waveform prediction of a field example provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0077] The present invention provides a method for preventing the saturation protection of a converter transformer from misoperating during an excitation inrush current. The method adds a blocking criterion that the excitation inrush current accounts for a necessary component on the basis of the original saturation protection criterion of the converter transformer. When the zero-sequence current at the grid side of the converter transformer satisfies the prediction relationship of the fusion data network, the saturation protection is blocked. The innovation of the present invention lies in: 1. Identifying the excitation inrush current and the distorted current caused by DC bias magnetization through the change of the predicted value of the neutral point zero-sequence current in the fusion network LSTM-Transformer model; 2. Calculating whether the magnitude and slope of the predicted value of the neutral point zero-sequence current meet the criterion of the simple excitation inrush current; 3. Calculating whether the proportion of the predicted value of the neutral point zero-sequence current meets the criterion of the DC bias composite excitation inrush current and the excitation inrush current accounts for the main component.

[0078] like Figure 1 As shown, embodiment 1 of the present invention provides a commutation variable saturation protection method based on a fusion network LSTM-Transformer, comprising the following steps:

[0079] Step 1: collect the three-phase current and neutral point zero-sequence current on the grid side of the converter transformer, and start timing when the start condition of the converter transformer saturation protection is met.

[0080] Preferably, the peak current of each unit power frequency cycle is obtained based on the zero-sequence current of the neutral point of the converter transformer, and when the average value of the peak current of 6 cycles is greater than the set value of the inverse time curve, the saturation protection starts timing.

[0081] Step 2: The actual collected three-phase current I of the bushing on the grid side of the converter transformer acA ,I acB ,I acC and the grid-side neutral point zero-sequence current I fz0 , and input it into the trained fusion network LSTM-Transformer model to obtain the DC bias magnetic prediction value F(t), whose size represents the probability that the neutral point zero-sequence current predicted by the model is DC bias magnetic or the DC bias magnetic content in the neutral point zero-sequence current.

[0082] There is a difference between the distorted current when the converter transformer has an excitation surge current and a DC bias. When there is no DC bias, the excitation current of the transformer is symmetrical on the positive and negative half axes and contains only odd harmonics. When DC flows through the transformer winding, the DC and AC magnetic fluxes are superimposed, and the magnetic flux of the half cycle that is consistent with the DC bias direction is greatly increased, while the magnetic flux of the other half cycle is reduced, causing the excitation current to work at the saturation of the core magnetization curve, resulting in the excitation current waveform showing an extremely asymmetrical shape of the positive and negative half waves.

[0083] According to the distinguishing characteristics of the distorted current, a prediction model based on the LSTM-Transformer fusion network is obtained by inputting typical DC bias magnetization and excitation inrush current waveforms for training. Figure 3 As shown, the training of the fusion network LSTM-Transformer model includes the following steps:

[0084] Step A.1, set a typical excitation inrush current scenario, considering the comprehensiveness of the closing angle, with a step size of 1 degree, and design a total of 360 simulation scenarios and parameters as the network training set;

[0085] Step A.2, set a typical DC bias magnetic scenario, set the neutral point current injected into the commutation transformer grid side to be between 0 and 300 A, with 1 A as a step size, and obtain a total of 300 simulation scenarios and parameters as a training set for the network;

[0086] It is worth noting that other step sizes, such as 2 degrees, 3 degrees, 2A, etc., can be selected according to actual needs to obtain different numbers of excitation inrush currents or DC bias simulation scenarios, which are all within the protection scope of the present invention based on the spirit of the present invention.

[0087] Step A.3: Construct the feature vector x using the state vectors in the training set before time t t , expressed as the following formula,

[0088] x t ={f t ,v t ,k t ,s t ,p t} T

[0089] Where:

[0090] f t , v t , k t ,s t , p t They all represent the state vectors before time t.

[0091] Step A.4, normalize and segment the feature vector, and convert the segmented time series x n = {X1, X2, ..., X n} is input into the fusion network LSTM-Transformer model for feature learning, where normalization and segmentation processing are expressed as follows:

[0092]

[0093] Where:

[0094] x t ' represents the normalized eigenvalue;

[0095] min(x t )、max(x t ) are respectively a vector consisting of the minimum values ​​of each state vector in the eigenvector and a vector consisting of the maximum values ​​of each state vector;

[0096] x1′, x′2, …, x′ n+k-1 are the normalized eigenvectors at the first moment, the second moment, ... the n+k-1th moment respectively;

[0097] n is the number of time periods;

[0098] k is the number of moments contained in each time period;

[0099] X1…X n represents the normalized characteristic value sequence of the neutral point current on the grid side of the commutation transformer in n time periods;

[0100] Step A.5, in the feature layer, based on the feature learning result of the previous step and combined with the feature vector of the current step, output the DC bias magnetic prediction value.

[0101] Preferably, in the feature layer, the LSTM module, fusion layer and multi-head self-attention mechanism are expressed as:

[0102]

[0103] Where:

[0104] Represents the output vector of the LSTM network;

[0105] Represents the current hidden layer vector;

[0106] G t Represents the output vector of the multi-head attention operation;

[0107] g(h t-1 ) represents the hidden layer vector after the sigmoid function operation;

[0108] L represents recursive operation based on long short-term memory, F represents feature fusion operation, and T represents multi-head attention operation.

[0109] Step 3: according to the DC bias magnetic prediction value, when the criterion for locking saturation protection is met, the saturation protection is immediately locked.

[0110] In a preferred but non-limiting embodiment of the present invention, the criterion for blocking saturation protection is added on the basis of the original criterion for converter transformer saturation protection as follows: Figure 2 As shown, it includes a single operating condition criterion (locking criterion 1) and a compound operating condition criterion (locking criterion 2).

[0111] The single working condition criterion is for the working condition under the simple excitation inrush current, and is used to lock the saturation protection when it is determined that the neutral point zero-sequence current component is a simple excitation inrush current, specifically including: when the DC bias magnetic prediction value F(t) is continuously less than the first threshold a within the set time, and the discrete slope k1 is less than the second threshold b, the saturation protection is locked. Under the fusion network LSTM-Transformer, the DC bias magnetic prediction value of the simple excitation inrush current working condition is equal to 0 or close to 0, and the slope hardly changes.

[0112] Further preferably, the single operating condition criterion is expressed by the following formula:

[0113]

[0114] Where:

[0115] a is a first threshold, more preferably, set to 0.1;

[0116] b is a second threshold, more preferably, set to 0.05;

[0117] N is the total number of DC bias magnetic prediction values, that is, the total number of prediction value points;

[0118] F(t=t i ) indicates t i The DC bias magnetic prediction value at the moment;

[0119] k1 represents the discrete slope;

[0120] t s Indicates the set time.

[0121] In an exemplary but non-limiting implementation, the time is set to 6 seconds, and the three-phase current of the bushing on the grid side of the converter transformer and the zero-sequence current of the neutral point are collected every 0.02 seconds and input into the fusion network for prediction. A total of 300 prediction value points are obtained within 6 seconds. At this time, there are t s =6, N=300.

[0122] The composite working condition criterion is aimed at the working condition of excitation surge current and DC bias, and is used to lock the saturation protection when it is determined that both excitation surge current and DC bias exist in the neutral point zero-sequence current, but the excitation surge current accounts for the main component. Specifically, it includes: when the number of predicted value points where the DC bias predicted value F(t) is less than the third threshold c within the set time accounts for a ratio k2 of the total number of predicted value points greater than the fourth threshold m, the saturation protection is locked. Under the fusion network LSTM-Transformer, when the initial excitation surge current of the composite working condition is large, the DC bias predicted value is close to 0. As the excitation surge current decays, the DC bias predicted value gradually increases, and the higher the DC bias component, the faster the rate of change of the DC bias predicted value.

[0123] Further preferably, under the excitation inrush current combined with DC bias working condition, when applying the combined working condition criterion, the DC bias content should be lower than the maximum DC current value allowed to flow through the neutral point, for example, 30A.

[0124] Further preferably, the composite operating condition criterion is expressed by the following formula:

[0125]

[0126] Where:

[0127] k2 represents the proportion of predicted value points whose DC bias magnetic prediction value is less than the third threshold value to the total predicted value points;

[0128] c is a third threshold, more preferably, set to 0.05;

[0129] IF[*] is a conditional function, the value is 1 when the condition* is met, and 0 when the condition* is not met;

[0130] n* is the number of DC bias magnetic prediction values ​​less than the third threshold;

[0131] m is the fourth threshold.

[0132] More preferably, the value of m in the composite operating condition criterion is set to 0.93. According to simulation experimental data, m = 0.93 corresponds to the proportion of DC bias magnetization prediction values ​​less than 0.05 when the residual magnetism of a phase is 1.0, the closing angle is 0°, the residual magnetism of the other two phases is 0, the single-phase excitation inrush current is the largest, and the DC current injected into the grid-side neutral point is 30A.

[0133] Embodiment 2 of the present invention provides a commutation variable saturation protection system based on a fusion network LSTM-Transformer, characterized in that it includes:

[0134] Signal acquisition module, used to collect three-phase current of bushing on the grid side of converter transformer and zero-sequence current of neutral point;

[0135] The model training module is used to train the LSTM-Transformer model of the fusion network and calculate the DC bias magnetic prediction value based on the three-phase current of the bushing on the grid side and the neutral point zero-sequence current actually collected;

[0136] The logic judgment module is used to compare the DC bias magnetic prediction value according to the set judgment conditions. When any condition is met, the saturation protection is immediately locked.

[0137] Embodiment 3 of the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of any one of the methods described in Embodiment 1 are implemented.

[0138] Embodiment 4 of the present invention provides a computer device / equipment / system, including:

[0139] Memory, for storing computer programs / instructions;

[0140] A processor is used to execute the computer program / instructions to implement the steps of any method described in Example 1.

[0141] Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which implements the steps of any method described in Embodiment 1 when executed by a processor.

[0142] Application examples:

[0143] In order to further clearly describe the technical solution of the present invention and the beneficial technical effects brought about by it, an application example is introduced below.

[0144] like Figure 4As shown in the figure, the transformer is closed without load when the initial phase angle of phase A voltage is 0°, 90°, 180° and 270° respectively. The neutral point excitation inrush current waveform is shown in the figure, where the ordinate value represents the instantaneous value of the current, and the data generated by the network is identified and predicted through the fusion network LSTM-Transformer, and the corresponding results are marked in the form of scattered points. Figure 4 From the above simulation results, it can be seen that, considering the influence of the change in the closing phase angle, the predicted values ​​of the fusion network LSTM-Transformer model are all 0 or close to 0, which is consistent with the data label of the excitation inrush current sample set.

[0145] like Figure 5 As shown, the DC currents are 0, 100, 200 and 300A respectively. The DC bias simulation model and the predicted value of neutral point current under these four different currents are shown in the figure, where the ordinate value represents the instantaneous value of the current. Due to the large difference between the excitation inrush current and the DC bias waveform, the predicted values ​​of the fusion network LSTM-Transformer model are all 1 or close to 1, which is consistent with the data label of the DC bias sample set. It is preliminarily proved that the fusion network LSTM-Transformer can correctly identify the DC bias.

[0146] like Figure 6 As shown in the figure, it is the working condition of DC bias magnetization superimposed on excitation inrush current. When the closing angle of phase A is set to 0°, and the DC current is 100 and 300A respectively. The composite simulation model of DC bias magnetization and excitation inrush current in these two different scenarios, as well as its neutral point current waveform and judgment results are shown in the figure, where the ordinate value represents the instantaneous value of the current. As can be seen from the figure, since DC bias magnetization almost does not exist in the first few seconds, the model judges it as 0 in the first part of the time, which is consistent with the data label of the excitation inrush current sample set. As the DC bias magnetization content increases and the excitation inrush current content decreases, the waveform gradually approaches the DC bias magnetization waveform, so the algorithm gradually approaches 1 in probability. As time goes on, the DC bias magnetization rises sharply, and the prediction value of the fusion network LSTM-Transformer model is gradually judged to be 1 or close to 1, which is consistent with the data label of the DC bias magnetization sample set.

[0147] like Figure 7 The figure shows the waveform prediction results of a DC bias magnetization and excitation inrush current at a certain station. Figure 7 It can be seen that the DC bias magnetic prediction value of a certain station is stable at 0.08, and the discrete point slope is 0.02, which meets the single operating condition criterion described in Example 1 of the present invention, and can normally lock the saturation protection to avoid protection tripping; the DC bias magnetic prediction value of a certain station is 0.92, at this time the saturation protection is normally opened, and the action is performed according to the time calculated according to the inverse time curve.

[0148] Compared with the prior art, the beneficial effects of the present invention include at least: the present invention provides a commutation transformer saturation protection method and system based on the fusion network LSTM-Transformer, which uses a neural network algorithm to effectively identify the scenarios of pure excitation inrush current, pure DC bias, and excitation inrush current combined with DC bias, and can effectively avoid the saturation protection action caused by pure excitation inrush current or excitation inrush current as the main component, thereby improving the reliability of the saturation protection action.

[0149] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A commutation variable saturation protection method based on a fusion network, characterized in that: include: Collect three-phase current and neutral point zero-sequence current on the commutation transformer grid side; The three-phase current and neutral point zero-sequence current on the converter transformer side are input into the trained fusion network LSTM-Transformer model to obtain the DC bias magnetic prediction value; According to the DC bias magnetization prediction value, when a single operating condition criterion or a composite operating condition criterion is met, the saturation protection is immediately locked; wherein, the single operating condition criterion includes that the DC bias magnetization prediction values ​​are all less than the set upper limit within the set time, and the discrete slope of the DC bias magnetization prediction value is less than the set slope threshold; the composite operating condition criterion includes that the number of DC bias magnetization prediction values ​​less than the set threshold accounts for a proportion of the total number of all DC bias magnetization prediction values ​​greater than the set proportion threshold.

2. The commutation variable saturation protection method based on the fusion network according to claim 1 is characterized in that: The training of the fused network LSTM-Transformer model includes: Step A.1, determining multiple excitation inrush current simulation scenarios and parameters with the set closing angle as the step length; Step A.2, determining a plurality of DC bias simulation scenarios and parameters within a set neutral point current range on the commutation transformer side and with a set neutral point current step; forming a training set with a plurality of excitation inrush current simulation scenarios and parameters and a plurality of DC bias simulation scenarios and parameters; Step A.3: Construct the feature vector x using the state vectors in the training set before time t t , expressed as the following formula, x t ={f t ,v t ,k t ,s t ,p t } T Where: f t , v t , k t ,s t , p t All represent the state vectors before time t; Step A.4, normalize and segment the feature vector, and convert the segmented time series x n = {X1, X2, ..., X n } is input into the fusion network LSTM-Transformer model for feature learning, where normalization and segmentation processing are expressed as follows: Where: x t ' represents the normalized feature vector; min(x t )、max(x t ) are respectively a vector consisting of the minimum values ​​of each state vector in the eigenvector and a vector consisting of the maximum values ​​of each state vector; x1′, x′2, …, x′ n+k-1 are the normalized eigenvectors at the first moment, the second moment, ... the n+k-1th moment respectively; n is the number of time periods; k is the number of moments contained in each time period; X1…X n represents the normalized characteristic value sequence of the neutral point current on the grid side of the commutation transformer in n time periods; Step A.5, output the DC bias magnetic prediction value based on the feature learning result of the previous step and the feature vector of the current step.

3. The commutation variable saturation protection method based on the fusion network according to claim 2 is characterized in that: The fusion network LSTM-Transformer model includes LSTM modules, fusion layers and multi-head self-attention mechanisms, which are respectively expressed by the following formulas: Where: Represents the output vector of the LSTM network; Represents the current hidden layer vector; G t Represents the output vector of the multi-head attention operation; g(h t-1 ) represents the hidden layer vector after the sigmoid function operation; L represents recursive operation based on long short-term memory, F represents feature fusion operation, and T represents multi-head attention operation.

4. The commutation variable saturation protection method based on the fusion network according to claim 1 is characterized in that: The single operating condition criterion is used to lock the saturation protection when it is determined that the neutral point zero-sequence current is an excitation surge current, and is expressed by the following formula: Where: F(t) is the DC bias magnetic prediction value; a is the first threshold; k1 represents the discrete slope; N is the total number of DC bias magnetic prediction values; F(t=t i ) indicates t i The DC bias magnetic prediction value at the moment; b is the second threshold; t s Indicates the set time.

5. The commutation variable saturation protection method based on the fusion network according to claim 4 is characterized in that: The time is set to 6s, and the three-phase current of the bushing on the converter transformer grid side and the neutral point zero-sequence current are collected every 0.02s and input into the fusion network for prediction; the first threshold is set to 0.1, and the second threshold is set to 0.

05.

6. The commutation variable saturation protection method based on the fusion network according to claim 1 is characterized in that: When applying the composite operating condition criterion, it also includes: the DC bias magnetic content is lower than the maximum DC current value allowed to flow through the neutral point.

7. The commutation variable saturation protection method based on the fusion network according to claim 1 is characterized in that: The composite working condition criterion is used to lock the saturation protection when it is determined that both the excitation inrush current and the DC bias magnetism exist in the neutral point zero-sequence current, but the proportion of the excitation inrush current is greater than the critical value, and is expressed by the following formula: Where: F(t=t i ) indicates t i The DC bias magnetic prediction value at the moment; c is the third threshold; IF[*] is a conditional function, the value is 1 when the condition* is met, and 0 when the condition* is not met; N is the total number of DC bias magnetic prediction values; n* is the number of DC bias magnetic prediction values ​​less than the third threshold; m is the fourth threshold; k2 represents the proportion of predicted value points whose DC bias magnetic prediction values ​​are less than the third threshold value to the total predicted value points.

8. The commutation variable saturation protection method based on the fusion network according to claim 7 is characterized in that: The third threshold is set to 0.05, and the fourth threshold is set to 0.

93.

9. A commutation variable saturation protection system based on a fusion network, running a commutation variable saturation protection method based on a fusion network as described in any one of claims 1 to 8, characterized in that: include: Signal acquisition module, used to collect three-phase current of bushing on the grid side of converter transformer and zero-sequence current of neutral point; The model training module is used to train the LSTM-Transformer model of the fusion network and calculate the DC bias magnetic prediction value based on the three-phase current of the bushing on the grid side and the neutral point zero-sequence current actually collected; The logic judgment module is used to compare the DC bias magnetic prediction value according to the set judgment conditions. When any condition is met, the saturation protection is immediately locked.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 8 are implemented.

11. A computer device / equipment / system, characterized in that: include: Memory, for storing computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images