A method for obtaining an ELM network model for commutation identification and prediction

Abstract

A method for obtaining an ELM network model for commutation identification and prediction includes the following steps: Step 1. Training a model for identifying the first commutation failure; Step 2. Training a model for predicting subsequent commutation failures; The obtained trained model is the ELM network model for commutation identification and prediction. The purpose of this invention is to identify the first commutation failure and predict subsequent commutation failures, so as to solve the technical problems of insufficient applicability, poor accuracy and slow prediction speed of existing commutation failure identification and prediction methods.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, and in particular to a data-driven model and method for identifying and predicting commutation failures in high-voltage direct current transmission. This invention is a divisional application of the invention patent with application number 202210224288.5. Background Technology

[0002] In high-voltage direct current (HVDC) transmission systems, the converter valves of converters are highly susceptible to commutation failures due to AC system malfunctions. If subsequent commutation failures occur, they can lead to single-pole or double-pole blocking of the converter station, resulting in power surplus in the sending-end grid and power deficit in the receiving-end grid, posing a significant threat to the safety and stability of the power grid.

[0003] Initial commutation failure in high-voltage direct current (HVDC) transmission systems is generally unavoidable, but its impact on the power grid is limited and it can recover on its own. Subsequent commutation failures following the initial failure can cause multiple shocks to the power grid, threatening its safety. In the prior art, patent document CN110441658A discloses a method for determining HVDC commutation failure considering changes in DC current. It calculates the turn-off angle expressions for single-phase ground faults and three-phase short-circuit faults by establishing turn-off angle models for three-phase ground faults and single-phase ground faults. When the turn-off angle under a three-phase short-circuit fault is less than or equal to the minimum turn-off angle for inverter commutation failure, it is judged as a commutation failure; otherwise, it is judged as a successful commutation. When a single-phase ground fault occurs, the turn-off angle under the single-phase ground fault is compared with the minimum turn-off angle of the inverter when commutation fails under the DC standard test system. If the turn-off angle under the single-phase ground fault is less than or equal to the minimum turn-off angle of the inverter when commutation fails, it is judged as commutation failure; otherwise, it is judged as commutation success.

[0004] The aforementioned existing technologies can effectively identify commutation failures by comparing the shut-off angle under fault conditions with the minimum shut-off angle. However, these existing technologies cannot predict subsequent commutation failures, which are more dangerous. Summary of the Invention

[0005] The purpose of this invention is to identify the first commutation failure and predict subsequent commutation failures, so as to solve the technical problems of insufficient applicability, poor accuracy and slow prediction speed of existing commutation failure identification and prediction methods.

[0006] A model for identifying commutation failures in high-voltage direct current transmission includes a data acquisition module, a feature extraction module, a decision module, and an output module.

[0007] The data acquisition module is used to collect raw data and perform data preprocessing;

[0008] The feature extraction module is used to extract and classify features from the raw data, and to perform calculations on the extracted feature data to transform it into a sample set that can be used by the classifier;

[0009] The decision module is used to identify the first commutation failure and predict subsequent commutation failures;

[0010] The output module is used to output the prediction results.

[0011] The output of the data acquisition module is connected to the input of the feature extraction module, the output of the feature extraction module is connected to the input of the decision module, and the output of the decision module is connected to the input of the output module.

[0012] The feature extraction module includes a first feature calculation module and a second feature calculation module, and the decision module includes a first commutation failure identification module and a subsequent commutation failure prediction module.

[0013] The output of the data acquisition module is connected to the input of the first feature calculation module and the second feature calculation module. The output of the first feature calculation module is connected to the input of the first commutation failure identification module. The output of the first commutation failure identification module is connected to the output module and the input of the second feature calculation module. The output of the second feature calculation module is connected to the input of the subsequent commutation failure prediction module. The output of the subsequent commutation failure prediction module is connected to the input of the output module.

[0014] The first feature calculation module is used to convert the first segment of original features into features usable by the classifier. The second feature calculation module is used to convert the second segment of original features into features usable by the classifier. The first commutation failure identification module is used to identify the first commutation failure. The subsequent commutation failure prediction module is used to predict subsequent commutation failures.

[0015] The data acquisition module includes a data acquisition module and a preliminary calculation module. The output of the data acquisition module is connected to the input of the preliminary calculation module, and the output of the preliminary calculation module is connected to the input of the first feature calculation module and the second feature calculation module.

[0016] Regarding the data acquisition module, it collects real-time operating data, including: inverter-side converter bus voltage, DC current, early firing angle, and turn-off angle parameters. The collected data undergoes preliminary processing to obtain the effective value of the converter bus voltage and its instantaneous values ​​for each phase, the effective value of the DC current, the firing angle command value, and the actual turn-off angle value.

[0017] Regarding the feature extraction module, this invention implements a two-stage binary classification based on ELM for commutation failure identification and prediction. Two segments of electrical quantity data are extracted from the data output by the data acquisition module. The first segment extracts the effective values ​​of the inverter-side converter bus voltage and DC current after the time fault, serving as the original sample set for the first ELM classifier, which can be expressed as:

[0018]

[0019] In the formula, X 1-original This is the original sample set of the first segment collected. Let m be the b-th original electrical quantity of the a-th original sample, m be the number of samples in the original sample set, and n be the number of original electrical quantities contained in each sample.

[0020] The second extraction time is the time from the fault onwards until the shutdown angle recovers to the reference value. The extraction time for the instantaneous phase voltage of the inverter-side converter bus is the moment the fault begins. The collected data includes: RMS DC current, RMS converter bus voltage and instantaneous values ​​for each phase, and trigger angle command value. This serves as the original sample set for the second ELM classifier and can be represented as follows:

[0021]

[0022] In the formula, X 2-original This is the original sample set collected for the second segment. Let m be the b-th original electrical quantity of the a-th original sample, m be the number of samples in the original sample set, and n be the number of original electrical quantities contained in each sample.

[0023] The original feature values ​​are transformed into feature values, and the feature values ​​used in the first segment are:

[0024] I 1-max : Maximum current, U 1-min Minimum voltage;

[0025]

[0026] In the formula, X 1-data For the sample set of the first segment obtained by calculation, Let m be the b-th original electrical quantity of the a-th sample, m be the number of samples in the sample set, and n be the number of features contained in each sample.

[0027] The feature quantity used for the second segment is:

[0028] I 2-max : Maximum current, U 2-max : Maximum voltage, α 2-max : Maximum trigger angle, I 2-min Minimum current, U2-min : Minimum voltage, α 2-min Minimum trigger angle, I 2-average : Average current, U 2-averag : Average voltage, α 2-average : Average firing angle, THD u Total voltage harmonic distortion (THD);

[0029] △T: Sampling time interval

[0030]

[0031] In the formula, X 2-data To calculate the sample set for the second segment, Let m be the b-th original electrical quantity of the a-th sample, m be the number of samples in the sample set, and n be the number of features contained in each sample.

[0032] The first sample set is used to identify the first commutation failure, and the second sample set is used to predict subsequent commutation failures. Furthermore, the processing of the second sample set data is initiated based on the decision module's startup. If the decision module determines that the first commutation failure has occurred, the processing of the second sample set data is initiated, and the subsequent process continues; otherwise, it is not initiated.

[0033] The decision-making module consists of two parts: initial commutation failure identification and subsequent commutation failure prediction. Subsequent commutation failure prediction is implemented based on the initial commutation failure identification. First, the first sample set from the feature extraction module is input into a pre-trained initial commutation failure model to identify the initial commutation failure. If the model output indicates no commutation failure occurred, the identification result is output to the output module; if the result indicates an initial commutation failure, a signal is sent to the feature extraction module to initiate the processing of the second data set. Then, the transformed sample set is input into the pre-trained subsequent commutation failure prediction model in the decision-making module to predict subsequent commutation failures. Finally, the model prediction result is output to the output module.

[0034] When the feature extraction module and the decision-making module work together, the following steps are included:

[0035] Step 1: The feature extraction module extracts two segments of electrical feature data from the original electrical quantity data, which are needed for commutation failure identification and prediction;

[0036] Step 2: The first segment of data is processed by the first feature calculation module and transformed into a sample usable by the classifier. The sample is then input into the first commutation failure identification module in the decision module to perform the first commutation failure identification.

[0037] Step 3: Determine whether commutation has failed based on the output of the first commutation failure identification module. If the output indicates that no commutation failure has occurred, output the identification result to the output module.

[0038] Step 4: If the output result is that the first commutation failure has occurred, a signal is sent to the second feature calculation module. The second feature calculation module then calculates the second segment of data and converts it into a sample that can be used by the classifier. This sample is then input into the subsequent commutation failure prediction module in the decision module to perform subsequent commutation failure prediction.

[0039] Step 5: Output the model prediction results from the output module.

[0040] An ELM network model for commutation identification and prediction is established using the following steps:

[0041] Step 1. Train the first commutation failure identification model;

[0042] Step 1 includes the following steps:

[0043] 1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer nodes, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer nodes.

[0044] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0045] 3) Calculate the hidden layer output matrix H1;

[0046] 4) Calculate the Moor-Penrose generalized inverse H1 of the output matrix H1. + ;

[0047] 5) Calculate the output weight β1 to obtain the model output T1 = H1β1;

[0048] The above steps complete the training of the first commutation failure identification model.

[0049] Step 2. Train the subsequent commutation failure prediction model;

[0050] Step 2 includes the following steps:

[0051] 1) Select an infinitely differentiable function g2(x) as the activation function of the hidden layer nodes, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer nodes;

[0052] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0053] 3) Calculate the hidden layer output matrix H2;

[0054] 4) Calculate the Moor-Penrose generalized inverse H2 of the output matrix H2. + ;

[0055] 5) Calculate the output weight β2 to obtain the model output T2 = H2β2;

[0056] The above steps complete the training of the subsequent commutation failure prediction model.

[0057] In step 1.3), matrix H1 is:

[0058]

[0059] Where r is the number of samples, l is the number of hidden layer nodes, and F 1j ω 1i b 1i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., r; i = 1, 2, ..., l);

[0060] In step 1.4), the Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using the singular value decomposition method. + ;

[0061] In step 1.5), the formula for calculating β1 is as follows:

[0062] β1=H1 + T1.

[0063] In step 2.3), matrix H2 is:

[0064]

[0065] Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i b 2i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., R; i = 1, 2, ..., L);

[0066] In step 2.4), the Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using the singular value decomposition method. + ;

[0067] In step 2.5), the formula for calculating β2 is as follows:

[0068] β2=H2 + T2.

[0069] In the first commutation failure identification model, after obtaining the classification result T1, the electrical characteristic quantities under the fault are used to identify whether commutation failure has already occurred under each fault.

[0070] In the subsequent commutation failure prediction model, after obtaining the classification result T2, the electrical characteristics under each fault are used to predict whether subsequent commutation failure will occur.

[0071] A method for identifying and predicting commutation failures in high-voltage direct current transmission includes the following steps:

[0072] Step 1: Continuously monitor the voltage of the inverter-side converter bus. When a sudden voltage change is detected, perform data acquisition and data preprocessing to obtain the raw electrical quantity data.

[0073] Step 2: Extract two electrical feature quantities from the original electrical quantity data for commutation failure identification and prediction. Calculate the first segment of data and convert it into samples usable by the ELM classifier, then input it into the decision module.

[0074] Step 3: Use the first sample from the feature extraction module as input to the first commutation failure identification model to determine whether a first commutation failure has occurred. If not, send the result to the output module.

[0075] Step 4: If the first commutation failure occurs, a signal is sent to the feature extraction module to perform the second stage of data calculation and processing, and input into the subsequent commutation failure prediction model of the decision module to predict subsequent commutation failures, and the prediction result is input into the output module.

[0076] Step 5: The output module outputs the final commutation failure information.

[0077] The ELM network model is used as the first commutation failure detection model. The following steps are taken when training the first commutation failure detection model:

[0078] 1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer nodes, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer nodes.

[0079] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0080] 3) Calculate the hidden layer output matrix H1;

[0081] 4) Calculate the Moor-Penrose generalized inverse H1 of the output matrix H1. + ;

[0082] 5) Calculate the output weight β1 to obtain the model output T1 = H1β1;

[0083] The above steps complete the training of the first commutation failure recognition model.

[0084] In step 3), matrix H1 is:

[0085]

[0086] Where r is the number of samples, l is the number of hidden layer nodes, and F 1j ω 1i b 1i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., r; i = 1, 2, ..., l);

[0087] In step 4), the Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using the singular value decomposition method. + ;

[0088] In step 5), the formula for calculating β1 is as follows:

[0089] β1=H1 + T1.

[0090] The ELM network model is used as the subsequent commutation failure prediction model. The following steps are taken when training the subsequent commutation failure prediction model:

[0091] 1) Select an infinitely differentiable function g2(x) as the activation function of the hidden layer nodes, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer nodes;

[0092] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0093] 3) Calculate the hidden layer output matrix H2;

[0094] 4) Calculate the Moor-Penrose generalized inverse H2 of the output matrix H2. + ;

[0095] 5) Calculate the output weight β2 to obtain the model output T2 = H2β2; complete the subsequent training of the commutation failure recognition model through the above steps.

[0096] The above steps complete the training of the subsequent commutation failure prediction model.

[0097] In step 3), matrix H2 is:

[0098]

[0099] Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i b 2i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., R; i = 1, 2, ..., L);

[0100] In step 4), the Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using the singular value decomposition method. + ;

[0101] In step 5), the formula for calculating β2 is as follows:

[0102] β2=H2 + T2.

[0103] Compared with the prior art, the present invention has the following technical effects:

[0104] This invention is based on a data-driven ELM classification and prediction model, and considers multiple factors affecting commutation failure as features, thus overcoming the shortcomings of previous methods that mainly relied on current and voltage factors for quantification and calculation. This model can identify the first commutation failure and predict subsequent commutation failures, solving the problems of insufficient applicability, poor accuracy, and slow prediction speed of existing commutation failure identification and prediction methods. Attached Figure Description

[0105] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0106] Figure 1 This is a structural diagram of the commutation failure identification and prediction model of the present invention;

[0107] Figure 2 This is a schematic diagram of the commutation failure identification and prediction model of the present invention;

[0108] Figure 3 This invention relates to a modified IEEE 39-node system incorporating DC power transmission. Detailed Implementation

[0109] like Figure 1 As shown, a model for identifying commutation failure in high-voltage direct current transmission includes a data acquisition module 1, a feature extraction module 2, a decision module 3, and an output module 4.

[0110] Data acquisition module 1 is used to acquire raw data and perform data preprocessing;

[0111] Feature extraction module 2 is used to extract and classify features from the raw data, and to calculate and transform the extracted feature data into a sample set that can be used by the classifier.

[0112] Decision module 3 is used to identify the first commutation failure and predict subsequent commutation failures;

[0113] Output module 4 is used to output the prediction results.

[0114] The output of data acquisition module 1 is connected to the input of feature extraction module 2, the output of feature extraction module 2 is connected to the input of decision module 3, and the output of decision module 3 is connected to the input of output module 4.

[0115] Feature extraction module 2 includes a first feature calculation module 2-1 and a second feature calculation module 2-2; decision module 3 includes a first commutation failure identification module 3-1 and a subsequent commutation failure prediction module 3-2.

[0116] The output of data acquisition module 1 is connected to the input of the first feature calculation module 2-1 and the second feature calculation module 2-2. The output of the first feature calculation module 2-1 is connected to the input of the first commutation failure identification module 3-1. The output of the first commutation failure identification module 3-1 is connected to the input of the output module 4 and the second feature calculation module 2-2. The output of the second feature calculation module 2-2 is connected to the input of the subsequent commutation failure prediction module 3-2. The output of the subsequent commutation failure prediction module 3-2 is connected to the input of the output module 4.

[0117] The first feature calculation module 2-1 is used to convert the first segment of original features into features usable by the classifier. The second feature calculation module 2-2 is used to convert the second segment of original features into features usable by the classifier. The first commutation failure identification module 3-1 is used to identify the first commutation failure. The subsequent commutation failure prediction module 3-2 is used to predict subsequent commutation failures.

[0118] The data acquisition module 1 includes a data acquisition module 1-1 and a preliminary calculation module 1-2. The output end of the data acquisition module 1-1 is connected to the input end of the preliminary calculation module 1-2, and the output end of the preliminary calculation module 1-2 is connected to the input ends of the first feature calculation module 2-1 and the second feature calculation module 2-2.

[0119] When feature extraction module 2 and decision module 3 work together, the following steps are included:

[0120] Step 1: Feature extraction module 2 extracts two segments of electrical feature data from the original electrical quantity data, which are needed for commutation failure identification and prediction;

[0121] Step 2: The first segment of data is converted into a sample usable by the classifier by the first feature calculation module 2-1 and input into the first commutation failure identification module 3-1 in the decision module 3 for first commutation failure identification;

[0122] Step 3: Determine whether commutation has failed based on the output of the first commutation failure identification module 3-1. If the output result is that no commutation failure has occurred, then output the identification result to the output module 4.

[0123] Step 4: If the output result is that the first commutation failure has occurred, a signal is sent to the second feature calculation module 2-2. The second feature calculation module 2-2 then calculates the second segment of data and converts it into a sample that can be used by the classifier. The sample is then input into the subsequent commutation failure prediction module 3-2 in the decision module 3 to perform subsequent commutation failure prediction.

[0124] Step 5: Output the model prediction results from output module 4.

[0125] This invention also includes an ELM network model for commutation recognition and prediction.

[0126] Its purpose is to address the technical problems of insufficient applicability, poor accuracy, and slow prediction speed of existing network models for identifying and predicting commutation failures in high-voltage direct current transmission.

[0127] It employs the following steps during its creation:

[0128] Step 1. Train the first commutation failure identification model;

[0129] Step 1 includes the following steps:

[0130] 1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer nodes, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer nodes.

[0131] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0132] 3) Calculate the hidden layer output matrix H1;

[0133] 4) Calculate the Moor-Penrose generalized inverse H1 of the output matrix H1. + ;

[0134] 5) Calculate the output weight β1 to obtain the model output T1 = H1β1;

[0135] The above steps complete the training of the first commutation failure identification model.

[0136] Step 2. Train the subsequent commutation failure prediction model;

[0137] Step 2 includes the following steps:

[0138] 1) Select an infinitely differentiable function g2(x) as the activation function of the hidden layer nodes, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer nodes;

[0139] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0140] 3) Calculate the hidden layer output matrix H2;

[0141] 4) Calculate the Moor-Penrose generalized inverse H2 of the output matrix H2. + ;

[0142] 5) Calculate the output weight β2 to obtain the model output T2 = H2β2;

[0143] The above steps complete the training of the subsequent commutation failure prediction model.

[0144] In step 1.3), matrix H1 is:

[0145]

[0146] Where r is the number of samples, l is the number of hidden layer nodes, and F 1j ω 1i b 1i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., r; i = 1, 2, ..., l);

[0147] In step 1.4), the Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using the singular value decomposition method. + ;

[0148] In step 1.5), the formula for calculating β1 is as follows:

[0149] β1=H1 + T1.

[0150] In step 2.3), matrix H2 is:

[0151]

[0152] Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i b 2i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., R; i = 1, 2, ..., L);

[0153] In step 2.4), the Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using the singular value decomposition method. + ;

[0154] In step 2.5), the formula for calculating β2 is as follows:

[0155] β2=H2 + T2.

[0156] In the first commutation failure identification model, after obtaining the classification result T1, the electrical characteristic quantities under the fault are used to identify whether commutation failure has already occurred under each fault.

[0157] In the subsequent commutation failure prediction model, after obtaining the classification result T2, the electrical characteristics under each fault are used to predict whether subsequent commutation failure will occur.

[0158] This invention also includes a method for identifying and predicting commutation failures in high-voltage direct current transmission.

[0159] Its purpose is to solve the technical problems of insufficient applicability, poor accuracy and slow prediction speed of existing high voltage direct current transmission commutation failure identification and prediction methods;

[0160] Includes the following steps:

[0161] Step 1: Continuously monitor the voltage of the inverter-side converter bus. When a sudden voltage change is detected, perform data acquisition and data preprocessing to obtain the raw electrical quantity data.

[0162] Step 2: Extract two electrical feature quantities from the original electrical quantity data for commutation failure identification and prediction. Calculate the first segment of data and convert it into samples usable by the ELM classifier, then input it into the decision module.

[0163] Step 3: Use the first sample from the feature extraction module as input to the first commutation failure identification model to determine whether a first commutation failure has occurred. If not, send the result to the output module.

[0164] Step 4: If the first commutation failure occurs, a signal is sent to the feature extraction module to perform the second stage of data calculation and processing, and input into the subsequent commutation failure prediction model of the decision module to predict subsequent commutation failures, and the prediction result is input into the output module.

[0165] Step 5: The output module outputs the final commutation failure information.

[0166] The ELM network model is used as the first commutation failure detection model. The following steps are taken when training the first commutation failure detection model:

[0167] 1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer nodes, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer nodes.

[0168] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0169] 3) Calculate the hidden layer output matrix H1;

[0170] 4) Calculate the Moor-Penrose generalized inverse H1 of the output matrix H1. + ;

[0171] 5) Calculate the output weight β1 to obtain the model output T1 = H1β1;

[0172] The above steps complete the training of the first commutation failure recognition model.

[0173] In step 3), matrix H1 is:

[0174]

[0175] Where r is the number of samples, l is the number of hidden layer nodes, and F 1j ω 1i b 1i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., r; i = 1, 2, ..., l);

[0176] In step 4), the Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using the singular value decomposition method. + ;

[0177] In step 5), the formula for calculating β1 is as follows:

[0178] β1=H1 + T1.

[0179] The ELM network model is used as the subsequent commutation failure prediction model. The following steps are taken when training the subsequent commutation failure prediction model:

[0180] 1) Select an infinitely differentiable function g2(x) as the activation function of the hidden layer nodes, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer nodes;

[0181] 2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes.

[0182] 3) Calculate the hidden layer output matrix H2;

[0183] 4) Calculate the Moor-Penrose generalized inverse H2 of the output matrix H2. + ;

[0184] 5) Calculate the output weight β2 to obtain the model output T2 = H2β2;

[0185] The above steps complete the training of the subsequent commutation failure prediction model.

[0186] In step 3), matrix H2 is:

[0187]

[0188] Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i b 2i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., R; i = 1, 2, ..., L);

[0189] In step 4), the Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using the singular value decomposition method. + ;

[0190] In step 5), the formula for calculating β2 is as follows:

[0191] β2=H2 + T2.

[0192] To facilitate a better understanding of the present invention by those skilled in the art, further explanations and descriptions are provided below:

[0193] A model for identifying commutation failures in high-voltage direct current (HVDC) transmission includes a data acquisition module, a feature extraction module, a decision module, and an output module. The data acquisition module collects and preprocesses raw data; the feature extraction module extracts and classifies features from the raw data, and simultaneously calculates and transforms the extracted features into a sample set usable by the classifier; the decision module uses an ELM network trained on simulation data to identify the first commutation failure and predict subsequent commutation failures; the output module outputs the prediction results. The output of the data acquisition module is connected to the input of the feature extraction module, the output of the feature extraction module is connected to the input of the decision module, and the output of the decision module is connected to the input of the output module. Additionally, the feature extraction module has another input from the decision module.

[0194] Specifically, the functions of the above four modules are as follows:

[0195] 1. Data Acquisition Module

[0196] Real-time operating data is collected at a sampling interval of 0.0002s. The data includes inverter-side converter bus voltage, DC current, early firing angle, and turn-off angle parameters. The collected data is preliminarily processed to obtain the effective value of the converter bus voltage and the instantaneous value of each phase, the effective value of the DC current, the firing angle command value, and the actual value of the turn-off angle.

[0197] 2. Feature Extraction Module

[0198] This invention implements a two-stage binary classification based on ELM (Elastic Computation Model) for commutation failure identification and prediction. Two segments of electrical quantity data are extracted from the data output by the data acquisition module. The first segment extracts the effective values ​​of the inverter-side converter bus voltage and DC current after the fault (which can be 0.03 s), serving as the original sample set for the first segment's ELM classifier, which can be expressed as:

[0199]

[0200] In the formula, X 1-original This is the original sample set of the first segment collected. Let m be the b-th original electrical quantity of the a-th original sample, m be the number of samples in the original sample set, and n be the number of original electrical quantities contained in each sample.

[0201] The second extraction time is the time from the fault (which can be 0.03s) to the recovery of the turn-off angle to the reference value. The extraction time for the instantaneous phase voltage of the inverter-side converter bus is the moment the fault begins. The collected data includes: RMS DC current, RMS converter bus voltage and instantaneous values ​​for each phase, and trigger angle command value. This serves as the original sample set for the second ELM classifier and can be represented as follows:

[0202]

[0203] In the formula, X 2-original This is the original sample set collected for the second segment. Let m be the b-th original electrical quantity of the a-th original sample, m be the number of samples in the original sample set, and n be the number of original electrical quantities contained in each sample.

[0204] The original feature values ​​are transformed into feature values, and the feature values ​​used in the first segment are:

[0205] I 1-max : Maximum current; U 1-min Minimum voltage;

[0206]

[0207] In the formula, X 1-data For the sample set of the first segment obtained by calculation, Let m be the b-th original electrical quantity of the a-th sample, m be the number of samples in the sample set, and n be the number of features contained in each sample.

[0208] The feature quantity used for the second segment is:

[0209] I 2-max : Maximum current; U 2-max : Maximum voltage; α 2-max : Maximum trigger angle; I 2-min Minimum current; U 2-min : Minimum voltage; α 2-min Minimum trigger angle; I 2-average : Average current; U 2-averag : Average voltage; α 2-average : Average firing angle; THD u Total voltage harmonic distortion (THD); ΔT: Sampling time interval;

[0210]

[0211] In the formula, X 2-data To calculate the sample set for the second segment, Let m be the b-th original electrical quantity of the a-th sample, m be the number of samples in the sample set, and n be the number of features contained in each sample.

[0212] The first sample set is used to identify the first commutation failure, and the second sample set is used to predict subsequent commutation failures. Furthermore, the processing of the second sample set data is initiated based on the decision module's startup. If the decision module determines that the first commutation failure has occurred, the processing of the second sample set data is initiated, and the subsequent process continues; otherwise, it is not initiated.

[0213] 3. Decision Module

[0214] The decision-making module consists of two parts: initial commutation failure identification and subsequent commutation failure prediction. Subsequent commutation failure prediction is implemented based on the initial commutation failure identification. First, the first sample set from the feature extraction module is input into a pre-trained initial commutation failure model to identify the initial commutation failure. If the model output indicates no commutation failure occurred, the identification result is output to the output module; if the result indicates an initial commutation failure, a signal is sent to the feature extraction module to initiate the processing of the second data set. Then, the transformed sample set is input into a pre-trained subsequent commutation failure prediction model in the decision-making module to predict subsequent commutation failures. Finally, the model prediction result is output to the output module.

[0215] This invention is based on the ELM two-stage binary classification algorithm to realize commutation failure identification and prediction. Therefore, two different ELM network models need to be trained in advance. The training process is as follows:

[0216] The training model is trained using simulation data, which is generated by setting faults on different bus nodes of the AC system on the inverter side. Relevant parameters are shown in Table 1.

[0217] Table 1 Fault Setting Related Parameters

[0218]

[0219] The two sample sets are labeled with corresponding commutation failure information for model training. In the first sample set, the sample without the first commutation failure is labeled as 0, and the sample with the first commutation failure is labeled as 1. In the second sample set, the sample without subsequent commutation failure is labeled as 2, and the sample with subsequent commutation failure is labeled as 3.

[0220] The training steps for the first commutation failure identification model are as follows:

[0221] (1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer node, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer node.

[0222] The activation functions used in the ELM algorithm are: sigmoid function, sinusoidal function, triangular basis function, hard threshold function, and radial basis function.

[0223] The sample set used for the first segment classification was used as input to the ELM algorithm. The number of hidden layer nodes was selected to be 10-60, and the accuracy of different activation functions on the training and test sets was compared. Different activation functions had different accuracies, and the activation function with the highest accuracy was selected as the activation function for the first commutation failure recognition model.

[0224] (2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm (the number of hidden layer nodes is generally less than the number of training set nodes), compare the accuracy with different numbers of hidden layer nodes, and thus select the optimal number of hidden layer nodes.

[0225] (3) Calculate the hidden layer output matrix H1:

[0226]

[0227] Where r is the number of samples, l is the number of hidden layer nodes, and F 1j ω 1i b 1i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., r; i = 1, 2, ..., l);

[0228] (4) The Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using the singular value decomposition method. + .

[0229] (5) Calculate the output weight β1 to obtain the model output T1 = H1β1. The formula for calculating β is as follows:

[0230] β1=H1 + T1

[0231] (6) Complete the training of the first phase commutation failure identification model.

[0232] The subsequent commutation failure prediction model is trained, and the specific training steps are as follows:

[0233] (1) Select g2(x) which is infinitely differentiable as the activation function of the hidden layer node, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer node.

[0234] The activation functions used in the ELM algorithm are: sigmoid function, sinusoidal function, triangular basis function, hard threshold function, and radial basis function.

[0235] The sample set used for the second segment classification was used as input to the ELM algorithm. The number of hidden layer nodes was selected to be 10-60, and the accuracy of different activation functions on the training and test sets was compared. Different activation functions had different accuracies, and the activation function with the highest accuracy was selected as the activation function for the first commutation failure recognition model.

[0236] (2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm (the number of hidden layer nodes is generally less than the number of training set nodes), compare the accuracy with different numbers of hidden layer nodes, and thus select the optimal number of hidden layer nodes.

[0237] (3) Calculate the hidden layer output matrix H2:

[0238]

[0239] Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i b 2i These are the input vector, input weight vector, and threshold vector of the hidden layer neurons, respectively (j = 1, 2, ..., R; i = 1, 2, ..., L);

[0240] (4) The Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using the singular value decomposition method. + .

[0241] (5) Calculate the output weight β2 to obtain the model output T2 = H2β2. The formula for calculating β is as follows:

[0242] β2=H2 + T2

[0243] (6) Complete the training of the subsequent commutation failure prediction model.

[0244] Given input data F r F R The classification results T1 and T2 can be represented as:

[0245] T1=H1(F 11 ,…F 1r ,ω 11 ,…ω 1l ,b 11 ,…b 1l )

[0246] T2=H2(F 21 ,…F 2R ,ω 21 ,…ω 2L ,b 21 ,…b 2L )

[0247] In the initial commutation failure identification model, after obtaining the classification result T1, the electrical characteristic quantities under each fault are used to identify whether a commutation failure has already occurred.

[0248] In the subsequent commutation failure prediction model, after obtaining the classification result T2, the electrical characteristics under each fault are used to predict whether subsequent commutation failure will occur.

[0249] Next, the effectiveness of the commutation failure identification and prediction model in this invention will be evaluated, using classification accuracy (CA) as the performance evaluation metric for the classifier. The higher the accuracy of the two classification results, the better the performance of the segmented classification model proposed in this invention.

[0250] Classification accuracy (CA) is a traditional and widely used evaluation metric. The calculation formula is as follows:

[0251]

[0252] For a given test set, TP and TN are the number of samples that are classified as positive and negative, respectively, and P and N are the total number of samples in the positive and negative classes, respectively.

[0253] 4. Output module:

[0254] Output the decision results of the two ELM classifiers.

[0255] A data-driven method for identifying and predicting commutation failures in high-voltage direct current transmission is presented, with the following process:

[0256] First, the inverter-side converter bus voltage is continuously monitored. When a sudden voltage change is detected, the model is activated, and the data acquisition module is started to collect and preprocess the data to obtain raw electrical quantity data. Then, two electrical feature quantities required for commutation failure identification and prediction are extracted from the raw electrical quantity data. The first segment of data is calculated and converted into samples usable by the ELM classifier and input into the decision module. Next, the first segment of samples from the feature extraction module is used as input to the first commutation failure identification model to determine whether a first commutation failure has occurred. If not, the result is sent to the output module. If a failure has occurred, a signal is sent to the feature extraction module to perform the second segment of data calculation and processing, which is then input into the subsequent commutation failure prediction model of the decision module for subsequent commutation failure prediction. The prediction result is then input to the output module. Finally, the output module outputs the final commutation failure information.

[0257] Simulation model:

[0258] In the PSCAD / EMDTC electromagnetic transient simulation software, four DC circuits are added to the IEEE 39-bus standard system to form a configuration such as... Figure 3 The simulation model is shown. The LCC-HVDC system uses the CIGRE standard model, which is a unipolar HVDC system with a DC rated voltage of 500kV, a DC rated power of 1000MW, an AC rated voltage of 230kV on the inverter side, and a reference value of 18° for the inverter side turn-off angle.

Claims

1. A method for obtaining an ELM network model for commutation recognition and prediction, characterized in that, Includes the following steps: Step 1. Train the first commutation failure identification model; Step 1 includes the following steps: 1-1) Select an infinitely differentiable function g1(x) as the activation function of the hidden layer node, and randomly set the weight ω1 between the input layer and the hidden layer and the threshold b1 of the hidden layer node; 1-2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes. 1-3) Calculate the hidden layer output matrix H1; 1-4) Calculate the Moor-Penrose generalized inverse H1 of the output matrix H1. + ; 1-5) Calculate the output weight β1 to obtain the model output. ; The above steps complete the training of the first commutation failure identification model. Step 2. Train the subsequent commutation failure prediction model; Step 2 includes the following steps: 2-1) Select an infinitely differentiable function g2(x) as the activation function of the hidden layer nodes, and randomly set the weight ω2 between the input layer and the hidden layer and the threshold b2 of the hidden layer nodes. 2-2) Select the function with the highest accuracy as the activation function. By changing the number of hidden layer nodes in the ELM algorithm, compare the accuracy with different numbers of hidden layer nodes, and then select the optimal number of hidden layer nodes. 2-3) Calculate the hidden layer output matrix H2; 2-4) Calculate the Moor-Penrose generalized inverse H2 of the output matrix H2. + ; 2-5) Calculate the output weights β2 to obtain the model output. ; The above steps complete the training of the subsequent commutation failure identification model. The above steps are used to obtain the ELM network model for commutation recognition and prediction.

2. The method according to claim 1, characterized in that, In steps 1-3), matrix H1 is: ； Where r is the number of samples, F represents the number of hidden layer nodes. 1j ω 1i 、b 1i Let j be the input vector, input weight vector, and threshold vector of the hidden layer neuron, respectively. = 1, 2, ..., r; i= 1, 2, ..., l ; In steps 1-4), the Moor-Penrose generalized inverse H1 of the output matrix H1 is calculated using singular value decomposition. + ; In steps 1-5), the formula for calculating β1 is as follows: 。 3. The method according to claim 1, characterized in that, In steps 2-3), matrix H2 is: ； Where R is the number of samples, L is the number of hidden layer nodes, and F 2j ω 2i 、b 2i Let j be the input vector, input weight vector, and threshold vector of the hidden layer neuron, respectively. = 1, 2, ..., R; i= 1, 2, ..., L; In steps 2-4), the Moor-Penrose generalized inverse H2 of the output matrix H2 is calculated using singular value decomposition. + ; In steps 2-5), the formula for calculating β2 is as follows: 。 4. The model according to claim 2, characterized in that, In the first commutation failure identification model, after obtaining the classification result T1, the electrical characteristic quantities under each fault are used to identify whether a commutation failure has already occurred under each fault. In the subsequent commutation failure prediction model, after obtaining the classification result T2, the electrical characteristics under each fault are used to predict whether subsequent commutation failure will occur.

Images