A method and system for early warning of subsequent commutation failure in converter stations, taking into account the action behavior of relay protection.

Abstract

This invention discloses a method and system for early warning of subsequent commutation failure in converter stations, considering the action behavior of relay protection devices. The method involves obtaining electrical and control quantities before and after a fault through simulation; training a machine learning early warning model and an autoregressive prediction model using different parameters before and after the fault; predicting the action time of the AC system relay protection device using the autoregressive prediction model after the fault; determining in real time whether the relay protection device has acted and correcting the predicted value; obtaining a risk factor based on the corrected value; and multiplying the predicted value of the machine learning early warning model by the risk factor to obtain the prediction result. This method uses AC and DC electrical and control quantities and their structural characteristics within a certain period after the fault as criteria to analyze the action type and action time of the AC system relay protection device, calculate the risk factor, and consider its impact on the risk of subsequent commutation failure. Through the machine learning model, a high-accuracy early warning of subsequent commutation failure is provided within tens of milliseconds.

Description

Technical Field

[0001] This invention relates to the intersection of power system protection and control and artificial intelligence, and in particular to a method and system for early warning of subsequent commutation failure in converter stations that takes into account the action behavior of relay protection. Background Technology

[0002] The blocking of LCC-HVDC (Line Commutated Converter-High Voltage Direct Current) converter stations will momentarily interrupt the transmission of large-capacity power, severely impacting the safety of the AC / DC hybrid power grid at both the sending and receiving ends. Subsequent commutation failures (according to actual power system operating practices, multiple commutation failures occurring within 200ms after the initial failure are counted as one failure, and subsequent failures occurring after 200ms are called subsequent commutation failures, again counted as one failure within a 200ms timeframe) will lead to the blocking of the inverter-side converter station after a certain number of failures, which is a significant contributing factor to the blocking of currently operational high-voltage direct current transmission converter stations.

[0003] Since high-voltage direct current (HVDC) transmission relies on the three-phase symmetrical voltage of the receiving-end AC grid for normal commutation, three-phase voltage asymmetry caused by AC grid faults at the receiving end is a significant cause of commutation failures at HVDC converter stations. In particular, prolonged AC grid faults can directly lead to subsequent commutation failures, potentially triggering single-pole or double-pole blocking at the converter station. Predicting subsequent commutation failures within tens of milliseconds after the initial failure would provide a basis for the DC controller response and auxiliary equipment protection actions at the converter station, significantly reducing the risk of converter station blocking.

[0004] However, due to the highly time-varying and nonlinear nature of high-voltage direct current (HVDC) transmission control systems, it is difficult to obtain accurate analytical solutions through time-series analysis. Most traditional studies have opted to reproduce numerical results through simulation, without further exploring the physical mechanisms of the time series, let alone providing early warning of subsequent commutation failures. Specifically, current research falls into two categories:

[0005] The first category summarizes the influencing factors of subsequent commutation failure through theoretical derivation or simulation analysis. YUH, WU Z, LI G, et al. analyzed the changes in commutation voltage and turn-off angle under common fault conditions such as single-phase grounding fault and three-phase short-circuit fault, and combined with simulation to give the factors that have a greater impact on subsequent commutation failure (Simulation of the commutation failure when faults occur in AC side of the HVDC system[C] / / 2017IEEE 2nd information technology, networking, electronic and automationcontrol conference (ITNEC). 2017: 1576–1580). Ou Kaijian, Ren Zhen, et al. analyzed the commutation voltage, transformer ratio, DC current, etc., and explored the system's ability to resist commutation failure when these factors change (Research on commutation failure of DC transmission system (I) - Analysis of influencing factors of commutation failure[J]. Electric Power Automation Equipment, 2003, 23(5): 5–8). Rahimi E, Filizadeh S, et al. proposed a multiple-run simulation and genetic computation method for simulating subsequent commutation failure (Commutation failure analysis in HVDC systems using advanced multiple-run methods[C] / / International conference on power systems transients. 2005: 19–23). This type of study qualitatively analyzes the impact of different electrical quantities and control parameters on subsequent commutation failure, but does not provide criteria for determining whether subsequent commutation failure will occur.

[0006] The second category focuses on the changes of certain physical quantities during subsequent commutation failures through theoretical and simulation analysis, thereby proposing optimization schemes for the control strategy of DC transmission systems to avoid subsequent commutation failures after the first commutation failure. Wang Haijun, Huang Yilong, and others focused on the arc extinction angle control in DC transmission system control, analyzed the system response characteristics under different control methods, and proposed arc extinction angle predictive control technology to avoid subsequent commutation failures (Analysis of Response Strategy and Predictive Control Technology Route for Commutation Failure in High Voltage DC Transmission [J]. Power System Protection and Control, 2014(21): 124–131). Zheng Chao, Zhou Jingmin, and others focused on the reactive power consumption characteristics during the commutation failure process and proposed predictive control technology to significantly improve the reactive power demand on the inverter side under large disturbances (Impact of Commutation Failure Predictive Control on Voltage Stability and Optimization Measures [J]. Automation of Electric Power Systems, 2016, 40(12):5). Zhang Guohui, Jing Liuming, and others analyzed the harmonic characteristics of the system during subsequent commutation failures and proposed a control strategy to improve the system's harmonic characteristics and suppress subsequent commutation failures (Research on the Suppression Method of Continuous Commutation Failures in HVDC Transmission Caused by Harmonics [J]. Power System Protection and Control, 2020, 048(005): 33–41). Li Chunhua, Huang Ying, and others focused on the current change characteristics during subsequent commutation failures and proposed limiting the maximum short-circuit current to suppress continuous commutation failures (A Method for Suppressing Continuous Commutation Failures in HVDC Transmission Based on Maximum Short-Circuit Current Limitation [J]. Southern Power Grid Technology, 2018, 012(005): 10–16). These studies reduce the probability of subsequent commutation failures as much as possible through DC system control methods, but do not provide an estimate of the risk of subsequent commutation failures.

[0007] To address the obstacles encountered by traditional analysis methods in subsequent commutation failure analysis, Zhang Guohui, Li Zhizhong, Wang Bin, and others proposed an Adaboost-based early warning method for subsequent commutation failures (An Adaboost-based Early Warning Method for Continuous Commutation Failures in High-Voltage DC Lines [J]. Power System Protection and Control, 2019, 47(19):9). This method uses a data-driven approach to extract key features and patterns from massive amounts of fault data, without relying on the analysis of internal mechanisms or requiring analytical solutions. However, this research only uses electrical quantities such as the firing angle, commutation voltage, and DC current at the time of the first commutation failure as features. It does not analyze the correlation between the first and subsequent commutation failures, particularly neglecting the action behavior of the relay protection of the AC fault line at the receiving end after the first commutation failure, and its impact on subsequent commutation failures. The later the relay protection device operates, the slower the DC control system recovers, making subsequent commutation failures more likely—a significant and non-negligible factor.

[0008] While traditional analysis methods also use system electrical quantities as features for prediction, limitations in model complexity and training principles prevent them from uncovering relay protection behavior patterns and from considering their impact on subsequent commutation failures through prior knowledge. Furthermore, simply using machine learning fails to reflect the physical essence; the patterns extracted by machine learning do not accurately reflect reality, are prone to overfitting, and exhibit poor generalization ability and low accuracy in practical applications. Summary of the Invention

[0009] The technical problem to be solved by this invention is how to improve the prediction accuracy while fully considering the action behavior of the AC fault line relay protection at the receiving end after the first commutation failure.

[0010] The present invention solves the above-mentioned technical problems through the following technical means:

[0011] 1. A method for early warning of subsequent commutation failure in converter stations considering the action behavior of relay protection, characterized by comprising the following steps:

[0012] Step 1: Set up multiple fault types and use simulation software to perform transient simulation of the LCC-HVDC high-voltage direct current transmission system to obtain the original data of AC and DC electrical quantities and control quantities before and after the fault occurs; use the electrical state quantities of the AC and DC system before and after the fault and the control parameters of the DC control system to train the machine learning early warning model, and use the voltage and current electrical state quantities of the AC system to train the autoregressive prediction model.

[0013] Step 2: After a fault occurs, the autoregressive prediction model is used to predict the operating time of the AC system relay protection device;

[0014] Step 3: Analyze the electrical quantities after the fault occurs in real time, determine whether the relay protection device has been activated, and correct the predicted values ​​in Step 2.

[0015] Step 4: Obtain the risk factor based on the corrected relay protection device operating time from Step 3;

[0016] Step 5: Multiply the predicted value of the machine learning early warning model by the risk factor in step 4 to obtain the prediction result.

[0017] This invention designs two models, trained with different data. An autoregressive prediction model is used to predict the operating time of relay protection devices after a fault occurs, and this prediction is used to correct the prediction results of the machine learning early warning model, fully demonstrating the impact of relay protection operating time on the outcome. This invention designs a separate relay protection operating time prediction model, avoiding the limitation of putting all electrical quantities into a single model for machine learning, which is constrained by the complexity of the model itself and cannot extract the pattern from the data, resulting in low accuracy. Experiments show that this invention can provide a highly accurate early warning of subsequent commutation failure within tens of milliseconds. Furthermore, by collecting data in real time and continuously updating the prediction results, the method can correct the results in real time based on the actual relay operation, continuously improving accuracy over time.

[0018] Furthermore, the method for constructing the input features of the early warning model in step 1 is as follows:

[0019] Input features are denoted as This represents a column vector where each feature is concatenated end-to-end, where each column vector... , N It is the length of the time series;

[0020] Construct the following function to transform the original data and form the training set of the early warning model;

[0021]

[0022]

[0023]

[0024]

[0025] .

[0026] Furthermore, in step 2, the operating behavior of the AC system relay protection device considers current differential protection, and the prediction method selected is autoregressive prediction, as follows:

[0027] (21) Based on the current sequence at both ends of the AC system after the fault occurs, an autoregressive prediction model is established and trained. For a p-order autoregressive prediction, its general term is: ,in This is the offset. As weight, This represents regression bias;

[0028] (22) Extrapolate the autoregressive prediction model in step (21) to obtain the result at any time. Two-terminal current ;

[0029] (23) Predicting action time ,in To meet the operating conditions of current differential protection At that moment, This is the set value for the current differential protection. This refers to the operating time of the circuit breaker.

[0030] Furthermore, the method for correcting the predicted value in step 3 is as follows:

[0031] (31) Assumption If the deviation of the three-phase AC voltage from the steady-state value before the fault occurs is less than the set threshold, the relay protection device is considered to have successfully operated and the fault has been cleared.

[0032] (32) If Then the correction value This indicates that the relay protection has acted ahead of schedule;

[0033] (33) If the predicted time is reached If the relay protection still fails to operate, it indicates that the relay protection may have failed or is delayed; when When, the correction value is ; To correct the time;

[0034] (34) If If the protection device is detected to activate, the correction value will be adjusted. .

[0035] Furthermore, the method for constructing the risk factor in step 4 can be... Where K is a proportionality coefficient greater than zero. This serves as a reference time for subsequent commutation failures. This indicates the minimum permissible time difference.

[0036] Furthermore, the training data for the machine learning early warning model in step 1 includes simulated AC bus voltage, commutation voltage, DC voltage and current, and advanced trigger angle electrical quantities; the training data for the autoregressive prediction model reports the power supply and current of the AC line.

[0037] Corresponding to the above method, the present invention also provides a converter station subsequent commutation failure early warning system that considers the relay protection action behavior, comprising:

[0038] The model training module is used to set various fault types and use simulation software to perform transient simulations on the LCC-HVDC high-voltage direct current transmission system to obtain the original data of AC and DC electrical quantities and control quantities before and after the fault occurs. The electrical state quantities of the AC and DC system before and after the fault occur and the control parameters of the DC control system are used to train the machine learning early warning model, and the voltage and current electrical state quantities of the AC system are used to train the autoregressive prediction model.

[0039] The relay protection device operation prediction module is used to predict the operation time of the AC system relay protection device after a fault occurs, using the autoregressive prediction model.

[0040] The correction module is used to analyze the electrical quantities after a fault occurs in real time, determine whether the relay protection device has been activated, and correct the predicted value in step 2.

[0041] The risk factor calculation module is used to obtain the risk factor based on the corrected relay protection device operating time in the correction module.

[0042] The prediction module is used to multiply the predicted value of the machine learning early warning model with the risk factor to obtain the prediction result.

[0043] Furthermore, the method for constructing the input features of the early warning model in the model training module is as follows:

[0044] Input features are denoted as This represents a column vector where each feature is concatenated end-to-end, where each column vector... , N It is the length of the time series;

[0045] Construct the following function to transform the original data and form the training set of the early warning model;

[0046]

[0047]

[0048]

[0049]

[0050] .

[0051] Furthermore, in the relay protection device operation prediction module, the operation behavior of the AC system relay protection device is considered in light of current differential protection. The prediction method selected is autoregressive prediction, as follows:

[0052] (21) Based on the current sequence at both ends of the AC system after the fault occurs, an autoregressive prediction model is established and trained. For a p-order autoregressive prediction, its general term is: ,in This is the offset. As weight, This represents regression bias;

[0053] (22) Extrapolate the autoregressive prediction model in step (21) to obtain the result at any time. Two-terminal current ;

[0054] (23) Predicting action time ,in To meet the operating conditions of current differential protection At that moment, This is the set value for the current differential protection. This refers to the operating time of the circuit breaker.

[0055] Furthermore, the method for correcting the predicted value in the correction module is as follows:

[0056] (31) Assumption If the deviation of the three-phase AC voltage from the steady-state value before the fault occurs is less than the set threshold, the relay protection device is considered to have successfully operated and the fault has been cleared.

[0057] (32) If Then the correction value This indicates that the relay protection has acted ahead of schedule;

[0058] (33) If the predicted time is reached If the relay protection still fails to operate, it indicates that the relay protection may have failed or is delayed; when When, the correction value is ; To correct the time;

[0059] (34) If If the protection device is detected to activate, the correction value will be adjusted. .

[0060] Furthermore, the method for constructing risk factors in the risk factor calculation module can be... Where K is a proportionality coefficient greater than zero. This serves as a reference time for subsequent commutation failures. This indicates the minimum permissible time difference.

[0061] Furthermore, in the model training module, the training data for the machine learning early warning model includes simulated AC bus voltage, commutation voltage, DC voltage and current, and advanced trigger angle electrical quantities; the training data for the autoregressive prediction model reports the power supply and current of the AC line.

[0062] The advantages of this invention are:

[0063] This invention designs two models, trained with different data. An autoregressive prediction model is used to predict the operating time of relay protection devices after a fault occurs, and this prediction is used to correct the prediction results of the machine learning early warning model, fully demonstrating the impact of relay protection operating time on the outcome. This invention designs a separate relay protection operating time prediction model, avoiding the limitation of putting all electrical quantities into a single model for machine learning, which is constrained by the complexity of the model itself and cannot extract the pattern from the data, resulting in low accuracy. Experiments show that this invention can provide a highly accurate early warning of subsequent commutation failure within tens of milliseconds. Furthermore, by collecting data in real time and continuously updating the prediction results, the method can correct the results in real time based on the actual relay operation, continuously improving accuracy over time.

[0064] The method proposed in this invention can provide a basis for the response of DC controllers and the protection actions of auxiliary equipment, thereby greatly reducing the risk of subsequent commutation failure, improving the safety of power grid operation, and having broad engineering prospects. Attached Figure Description

[0065] Figure 1 This is a flowchart of the early warning process of the method in Embodiment 1 of the present invention.

[0066] Figure 2 This is a waveform diagram of the three-phase AC voltage before and after a fault in one embodiment of the method of Embodiment 1 of the present invention.

[0067] Figure 3 This is a prediction curve for subsequent commutation failure in one embodiment of the method of Embodiment 1 of the present invention.

[0068] Figure 4 This is a waveform diagram of the arc extinction angle before and after the fault in one embodiment of the method of Embodiment 1 of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1

[0071] A method for early warning of subsequent commutation failure in converter stations that considers the action behavior of relay protection includes the following steps:

[0072] Step 1: Set up multiple fault types and use simulation software to perform transient simulation on the LCC-HVDC system to obtain the original data of AC and DC electrical quantities and control quantities before and after the fault occurs. The fault types set include different transition resistances, fault initiation angles, short circuits, and grounding types, as shown in the table below.

[0073] Table 1. Fault Parameter Setting Table

[0074]

[0075] A total of 320 sets of data were obtained from the simulation. The raw data for AC and DC electrical and control quantities were selected from the three-phase voltage of the AC side, the DC current firing angle, and the arc extinction angle. In this embodiment, the electrical state quantities of the AC and DC systems before and after the fault occurred, and the control parameters of the DC control system were used to train the machine learning early warning model, and the voltage and current electrical state quantities of the AC system were used to train the autoregressive prediction model.

[0076] The input features of the early warning model are constructed as follows:

[0077] The input features can be denoted as This represents a column vector where each feature is concatenated end-to-end. Each column vector... , N It is the length of the time series.

[0078] Construct the following function to transform the original data and form the training set for the subsequent phase-change failure random forest early warning model.

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] In this embodiment, a random forest model is selected as the machine learning model. The dataset described above is used for training.

[0085] The following uses a specific fault scenario as an example to illustrate the intermediate calculation results of the method. The fault parameters for the simulation are set as follows: a three-phase short-circuit fault occurs on the inverter side at 25ms of the simulation, the transition resistance is 30 ohms, and the fault is cleared after 100ms.

[0086] The prediction method chosen is autoregressive prediction, and the method is as follows:

[0087] (21) Based on the current sequence at both ends of the AC system after the fault occurs, an autoregressive prediction model is established and trained. For a p-order autoregressive prediction, its general term is: ,in This is the offset. As weight, This represents regression bias;

[0088] (22) Extrapolate the autoregressive prediction model in step (21) to obtain the result at any time. Two-terminal current ;

[0089] (23) Predicting action time ,in To meet the operating conditions of current differential protection At that moment, This is the set value for the current differential protection. This refers to the operating time of the circuit breaker.

[0090] Step 2: Based on the electrical quantities after the fault occurs, predict the operating time of the AC system relay protection device. The steps are as follows: Autoregressive prediction model. Extrapolation yields the following results: At that time, under the per-unit value ,satisfy ,in This is the setting value for the current differential protection.

[0091] Therefore, predicting action time .in For the operating time of the circuit breaker, take .

[0092] Step 3: Analyze the electrical quantities after the fault occurs in real time, determine whether the relay protection device has operated, and correct the predicted value in step (2). The method is as follows:

[0093] From the appendix Figure 2 The three-phase AC voltage waveform can be analyzed in real time. The time-relay protection device successfully activated, and the fault was cleared. This means the predicted time was reached. If the relay protection still fails to operate, it indicates that the relay protection is delayed. Time correction value .

[0094] If the protection device is detected to activate, the correction value will be adjusted. .

[0095] In summary, the correction value (Unit: ms)

[0096] Step 4: Obtain the risk factor based on the corrected relay protection device operating time in Step 3.

[0097] The construction method of risk factors can be Where K is a proportionality coefficient greater than zero, taken as... , As a reference time for subsequent commutation failures, take , This represents the minimum permissible time difference. .

[0098] Substituting the corrected relay protection device operating time from step 3, we obtain the risk factor.

[0099]

[0100] Step 5: Build a machine learning early warning model and multiply it with the risk factor in Step 4 to obtain the prediction result.

[0101] The model predicts that the probability of subsequent commutation failure is: Multiply by the risk factor in step (4):

[0102] when , The prediction result is that subsequent commutation failure will occur.

[0103] when , The prediction result is that subsequent commutation failure will occur.

[0104] when , The prediction result is that subsequent commutation failure will occur.

[0105] As time progresses after the failure occurs, the prediction results are updated in real time, as shown in the attached figure. Figure 3 As shown.

[0106] The prediction results for the entire process are the same as the simulation results, as shown in the attached figure. Figure 4 As shown.

[0107] Example 2

[0108] Corresponding to Example 1, the present invention also provides a method for early warning of subsequent commutation failure in converter stations that considers the action behavior of relay protection, comprising the following steps:

[0109] The model training module is used to set various fault types and perform transient simulations of the LCC-HVDC system using simulation software to obtain raw data of AC and DC electrical quantities and control quantities before and after a fault occurs. The fault types set include different transition resistances, fault initiation angles, short circuits, and grounding types, as shown in the table below.

[0110] Table 1. Fault Parameter Setting Table

[0111]

[0112] A total of 320 sets of data were obtained from the simulation. The raw data for AC and DC electrical and control quantities were selected from the three-phase voltage of the AC side, the DC current firing angle, and the arc extinction angle. In this embodiment, the electrical state quantities of the AC and DC systems before and after the fault occurred, and the control parameters of the DC control system were used to train the machine learning early warning model, and the voltage and current electrical state quantities of the AC system were used to train the autoregressive prediction model.

[0113] The input features of the early warning model are constructed as follows:

[0114] The input features can be denoted as This represents a column vector where each feature is concatenated end-to-end. Each column vector... , N It is the length of the time series.

[0115] Construct the following function to transform the original data and form the training set for the subsequent phase-change failure random forest early warning model.

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] In this embodiment, a random forest model is selected as the machine learning model. The dataset described above is used for training.

[0122] The following uses a specific fault scenario as an example to illustrate the intermediate calculation results of the method. The fault parameters for the simulation are set as follows: a three-phase short-circuit fault occurs on the inverter side at 25ms of the simulation, the transition resistance is 30 ohms, and the fault is cleared after 100ms.

[0123] The prediction method chosen is autoregressive prediction, and the method is as follows:

[0124] (21) Based on the current sequence at both ends of the AC system after the fault occurs, an autoregressive prediction model is established and trained. For a p-order autoregressive prediction, its general term is: ,in This is the offset. As weight, This represents regression bias;

[0125] (22) Extrapolate the autoregressive prediction model in step (21) to obtain the result at any time. Two-terminal current ;

[0126] (23) Predicting action time ,in To meet the operating conditions of current differential protection At that moment, This is the set value for the current differential protection. This refers to the operating time of the circuit breaker.

[0127] The relay protection device operation prediction module uses an autoregressive prediction model to predict the operation time of AC system relay protection devices. The steps are as follows: Autoregressive prediction model... Extrapolation yields the following results: At that time, under the per-unit value ,satisfy ,in This is the setting value for the current differential protection.

[0128] Therefore, predicting action time .in For the operating time of the circuit breaker, take .

[0129] The correction module is used to analyze electrical quantities in real time after a fault occurs, determine whether the relay protection device has operated, and correct the predicted values ​​in the relay protection device operation prediction module. The method is as follows:

[0130] Appendix Figure 2 In the image, a, b, and c represent the three-phase AC voltage waveforms. From these waveforms, real-time analysis can be performed... When the three-phase AC voltage recovers, it indicates that the relay protection device is functioning correctly. Timely action, fault clearing. That is, reaching the predicted time. If the relay protection still fails to operate, it indicates that the relay protection is delayed. Time correction value .

[0131] If the protection device is detected to activate, the correction value will be adjusted. .

[0132] In summary, the correction value (Unit: ms)

[0133] The risk factor calculation module is used to obtain the risk factor based on the operation time of the relay protection device after correction in the correction module.

[0134] The construction method of risk factors can be Where K is a proportionality coefficient greater than zero, taken as... , As a reference time for subsequent commutation failures, take , This represents the minimum permissible time difference. .

[0135] Substituting the corrected relay protection device operating time into the risk factor yields the risk factor.

[0136]

[0137] The prediction module is used to multiply the predicted value of the machine learning early warning model with the risk factor to obtain the prediction result.

[0138] The model predicts that the probability of subsequent commutation failure is: Multiply by the risk factor in step (4):

[0139] when , The prediction result is that subsequent commutation failure will occur.

[0140] when , The prediction result is that subsequent commutation failure will occur.

[0141] when , The prediction result is that subsequent commutation failure will occur.

[0142] The prediction results are updated in real time as time passes after the failure occurs. (Appendix) Figure 3 The example demonstrates how the prediction results change over time. It can be seen that as the prediction time... If the relay protection fails to activate, the delay in activation significantly increases the risk of commutation failure. Once the fault is cleared, the risk of commutation failure stops increasing. This demonstrates that the invention can correct the results in real time based on the actual relay operation, continuously improving accuracy over time.

[0143] The predicted results for the entire process are consistent with the simulation results. (Attached) Figure 4 The waveforms of the arc extinction angle before and after the fault in the embodiment are shown. As can be seen from the figure, the arc extinction angle drops back to zero after 200ms, indicating that a subsequent commutation failure occurred, which is consistent with the prediction result. The model proposed in the invention achieves accurate prediction.

[0144] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting commutation failure of a converter station after a relay protection action, characterized in that, The method comprises the following steps: Step 1, setting multiple fault types, using simulation software to perform transient simulation on the LCC-HVDC high-voltage direct-current power transmission system, and obtaining original data of electrical quantities and control quantities of the AC / DC system before and after the fault occurs; The electrical state quantities of the AC / DC system before and after the fault occurs and the control parameters of the DC control system are used to train a machine learning early warning model, and the voltage and current electrical state quantities of the AC system are used to train an autoregressive prediction model; Step 2, after the fault occurs, the autoregressive prediction model is used to predict the action time of the AC system relay protection device; In step 2, the action behavior of the AC system relay protection device considers current differential protection, and the prediction method selects autoregressive prediction, and the method is as follows: (21) According to the current sequence of the double terminal of the alternating current system after the fault occurs, an autoregressive prediction model is established and trained, and for a p-order autoregressive prediction, the general term is: wherein is an offset, is a weight, is a regression bias; (22) extrapolating the bi-directional current at any time instant from the auto-regressive prediction model in step (21) ; (23) Predicted action time wherein to meet the current differential protection action condition at the time, is a set value of the current differential protection, is the operating time of the circuit breaker; Step 3, real-time analysis of electrical quantities after the fault occurs is performed to determine whether the relay protection device acts, and the predicted value in step 2 is corrected; Step 4, a risk factor is obtained according to the corrected action time of the relay protection device in step 3; Step 5, the machine learning early warning model predicts the probability of commutation failure, and the probability of commutation failure is multiplied by the risk factor in step 4 to obtain a prediction result; The input feature construction method of the early warning model in step 1 is as follows: The input features are denoted as represent a column vector with each column vector , N is the length of the time series; The following functions are constructed to transform the original data to form a training set of the early warning model; 。 2. The method for HVDC converter station subsequent commutation failure early warning considering the action behavior of relay protection according to claim 1, characterized in that, The method for correcting the predicted value in the correction module is as follows: (31) Assume If the deviation of the three-phase alternating voltage at the moment from the steady-state value before the fault is less than a set threshold value, it is considered that the relay protection device successfully acts, and the fault is removed. (32) If , the correction value , indicates that the relay protection action is advanced; (33) If the relay protection still does not act at the predicted time , it is indicated that the relay protection may be failed or delayed; when , the correction value is ; the correction time is (34) If the protection device is detected to act, the correction value .

3. The method of claim 1, wherein the method further comprises: determining a time period for the first and second time periods; and determining a time period for the third time period. The method for constructing the risk factor in step 4 can be wherein K is a proportional coefficient greater than zero, is a reference time for a subsequent commutation failure, denotes a minimum allowed difference time.

4. The method for HVDC converter station subsequent commutation failure early warning considering the action behavior of relay protection according to any one of claims 1 to 3, characterized in that, The training data of the machine learning early warning model in step 1 includes simulated AC bus voltage, commutation voltage, DC voltage and current, and electrical quantities of the lead trigger angle; and the training data of the autoregressive prediction model reports the power and current of the AC line.

5. A converter station subsequent commutation failure early warning system considering the behavior of the relay protection action, characterized in that, It comprises: A model training module is configured to set multiple fault types, use simulation software to perform transient simulation on the LCC-HVDC high-voltage direct-current power transmission system, and obtain original data of electrical quantities and control quantities of the AC / DC system before and after the fault occurs; The electrical state quantities of the AC / DC system before and after the fault occurs and the control parameters of the DC control system are used to train a machine learning early warning model, and the voltage and current electrical state quantities of the AC system are used to train an autoregressive prediction model; A relay protection device action prediction module is configured to, after the fault occurs, use the autoregressive prediction model to predict the action time of the AC system relay protection device; in the relay protection device action prediction module, the action behavior of the AC system relay protection device considers current differential protection, and the prediction method selects autoregressive prediction, and the method is as follows: (21) According to the current sequence of the double terminal of the alternating current system after the fault occurs, an autoregressive prediction model is established and trained, and for a p-order autoregressive prediction, the general term is: wherein is an offset, is a weight, is a regression deviation; (22) extrapolating the two-terminal current at any time instant from the autoregressive prediction model in step (21) ; (23) Predicted action time wherein is the set value of the current differential protection, is the time at which the current differential protection action condition is met, is the operating time of the circuit breaker, is the time at which the current differential protection action condition is met, A correction module is configured to real-time analyze electrical quantities after the fault occurs, determine whether the relay protection device acts, and correct the predicted value in step 2; A risk factor calculation module is configured to obtain a risk factor according to the corrected action time of the relay protection device in the correction module; A prediction module is configured to multiply the probability of commutation failure predicted by the machine learning early warning model by the risk factor to obtain a prediction result; The input feature construction method of the machine learning early warning model in the model training module is as follows: The input features are denoted by is a column vector of length T, where each column vector , N is the length of the time series; The following functions are constructed to transform the original data to form a training set of the early warning model; 。 6. The HVDC converter station subsequent commutation failure early warning system considering the behavior of the protection action of claim 5, wherein, The method for correcting the predicted value in the correction module is as follows: (31) Assume If the deviation of the three-phase AC voltage at the moment from the steady-state value before the fault is less than a set threshold value, it is considered that the relay protection device successfully acts, and the fault is cleared. (32) If , the correction value , indicates that the relay protection action is advanced; (33) If the relay protection still does not act at the predicted time , it is indicated that the relay protection may be disabled or delayed; when , the correction value is ; the correction time is (34) If the protection device is detected to act, the correction value .

7. The HVDC converter station subsequent commutation failure early warning system considering the action behavior of the relay protection according to claim 6, characterized in that, The construction method of the risk factor in the risk factor calculation module can be wherein K is a proportional coefficient greater than zero, is a reference time of a subsequent commutation failure, represents a minimum allowable difference time.

8. The HVDC converter station post-commutation failure early warning system considering the action behavior of the relay protection according to any one of claims 5 to 7, characterized in that, The training data of the machine learning early warning model in the model training module include simulated alternating current bus voltage, commutation voltage, direct current voltage and current, and lead trigger angle electrical quantity; and the training data of the autoregressive prediction model reports power and current of an alternating current line.

Images