Flexible direct current transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU
By combining traveling wave characteristics and multi-strategy improvement optimization algorithm, the BiGRU model is optimized for parameters, and the fault feature quantity is extracted by empirical modal decomposition and the Teager energy operator, the complexity of manual tuning thresholds in the fault diagnosis of flexible DC transmission lines is solved. and the problem of difficulty in detecting high resistance faults, achieving high accuracy and reliability fault diagnosis.
Patent Information
- Application Number
- CN202411137902.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-06
AI Technical Summary
In the fault diagnosis of flexible DC transmission lines, the problem of manual threshold setting process is complicated and high-impedance faults are difficult to detect, and the dependence on preset expressions or thresholds, resulting in a decrease in diagnostic accuracy.
The BiGRU model is optimized by using the Duckling Optimization Algorithm (MIDBO) based on traveling wave features and multi-strategy improvement. Combining empirical modal decomposition (EMD) and the Teager energy operator, the fault feature quantity is adaptively extracted and the model is trained. For fault diagnosis of flexible DC transmission systems.
It improves the accuracy and reliability of fault diagnosis, especially in high-impedance fault conditions, which can effectively detect faults, ensuring the normal operation and protection reliability of the system.
Smart Images

Figure CN119087122A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flexible direct current transmission fault diagnosis, and in particular to a flexible direct current transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU. Background Art
[0002] my country's renewable energy distribution and electricity demand are inversely proportional. DC grid technology based on modular multilevel converters offers advantages such as flexible control and no commutation failures, making it suitable for long-distance, high-capacity power transmission. However, the development of multi-terminal flexible DC transmission faces several key technical challenges, one of which is how to quickly and reliably detect faults in DC lines.
[0003] After searching existing technical literature, we found that the document "Fault-type discrimination in HVDC transmission lines using rate of change of local currents [J]. IEEE Transactions on Power Delivery, 2019, 35(1):117-129" identifies the fault pole of flexible DC transmission lines based on the characteristics of the current change rate. Although this method has high sensitivity and reliability, it is difficult to detect and isolate faults in a timely manner under high resistance conditions. At the same time, in order to improve the sensitivity of protection, such methods usually require a higher sampling frequency.
[0004] The protection scheme proposed in the document "Longitudinal protection of high-voltage DC lines based on transient energy ratio of DC filter link [J]. Transactions of China Electrotechnical Society, 2020, 35(09):1985-1998" is based on the energy ratio of specific frequency bands at the line boundary, which can quickly identify faults inside and outside the area. However, this method requires two-terminal electrical quantities and manual calculation of the setting values.
[0005] The paper "Multi extreme learning machine approach for fault location in multi-terminal high-voltage direct current systems[J]. Computers & Electrical Engineering, 2019, 78: 313-327" uses S transform to extract high-frequency features and uses extreme learning machine for fault diagnosis. Although it avoids manual setting of the threshold, it only uses current data as input, and its performance is poor under high-resistance faults.
[0006] Prior art fault diagnosis for flexible DC transmission lines suffers from the complex process of manually setting thresholds, making high-resistance faults difficult to detect. Existing methods may simplify system parameters or discard key parameters when calculating set values to simplify the model or facilitate calculations, which can reduce the accuracy of diagnostic results. Furthermore, existing methods rely on pre-set expressions or thresholds for fault diagnosis. However, changes in system parameters can render these expressions or thresholds inapplicable, or interference factors such as transition resistance can make it difficult to accurately calculate thresholds, thus affecting the accuracy of the final diagnosis. Summary of the Invention
[0007] To solve the above technical problems, the present invention provides a flexible DC transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU. Due to its accurate and reliable fault diagnosis results, this method can improve the reliability of the flexible DC transmission system and ensure the normal operation of the system.
[0008] The technical solution adopted by the present invention is:
[0009] The fault diagnosis method for flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU includes the following steps:
[0010] Step 1: First, use empirical mode decomposition (EMD) to decompose the power at the busbar and DC line to obtain the intrinsic mode function (IMF) representing the high-frequency component, and calculate its energy as the fault characteristic;
[0011] Step 2: Secondly, the multi-strategy improved dungbeetle optimizer (MIDBO) algorithm is used to optimize the parameters of the BiGRU model;
[0012] Step 3: Then, obtain a dataset; use the training set to train the MIDBO-BiGRU model and use the validation set to test it until the MIDBO-BiGRU model reaches the expected accuracy. The trained MIDBO-BiGRU model is then used for fault diagnosis of the flexible DC transmission system.
[0013] Step 4: Finally, a simulation model was built in PSCAD / EMTDC. The results showed that this method can not only detect busbar faults and line faults, but also solve the problem of high-resistance fault protection failure while ensuring protection reliability and speed.
[0014] In step 1, EMD has a good effect in decomposing multi-frequency superposition signals and can adaptively stabilize non-stationary signals without selecting basis functions. Therefore, the present invention selects EMD to process the signal. The process of EMD decomposition includes the following steps:
[0015] S1.1: Find the extreme point of the original signal s(t), t = 1, 2, ..., T, where T is the length of the original data s(t). If it does not exist, the decomposition ends.
[0016] The maximum value formula is as follows:
[0017] s max (t)>s(t-1)&s max (t)>s(t+1) (1);
[0018] In formula (1), s max (t) is the maximum value in the original signal s(t); s(t-1) is the t-1th data in the original signal s(t); s(t+1) is the t+1th data in the original signal s(t).
[0019] The minimum formula is as follows:
[0020] s min (t) <s(t-1)&s min (t) <s(t+1)(2);
[0021] In formula (2), s min (t) is the minimum value in the original signal s(t).
[0022] S1.2: Use the cubic spline function to fit the extreme points, and use equations (3) and (4) to fit the extreme points (t, s max (t)), and the minimum point (t,s min (t)) perform upper envelope fitting and lower envelope fitting respectively to obtain the upper envelope u(t) and lower envelope l(t);
[0023]
[0024] In formula (3), T max is the number of maximum values; and are the second-order derivatives of the i-th and i+1-th maximum points respectively; and are the serial numbers of the i-th and i+1-th maximum values in the original signal s(t); and are the i-th and i+1-th maxima respectively.
[0025]
[0026] In formula (4), T min is the number of minimum values; and are the second-order derivatives of the i-th and i+1-th minimum points respectively; and are the serial numbers of the i-th and i+1-th minimum values in the original signal s(t); and are the i-th and i+1-th minima respectively.
[0027] S1.3: Calculate the envelope mean;
[0028]
[0029] In formula (5), m(t) is the mean of the envelope, which represents the local trend of the signal.
[0030] S1.4: Calculate the difference between the original signal and the mean, i.e. the intermediate signal;
[0031] h(t)=s(t)-m(t) (6)
[0032] In formula (6), h(t) is the intermediate signal, which is a more stable signal.
[0033] S1.5: Determine whether the intermediate signal meets the IMF component condition. If so, it is used as the IMF component. If not, return to S1.1. The judgment conditions are as follows:
[0034]
[0035] In formula (7), n zero 、n max and n min are the number of zero-crossing points, the number of maximum points, and the number of minimum points of the intermediate signal respectively; s h (t) and l h (t) represent the upper envelope and lower envelope of the intermediate signal h(t), respectively.
[0036] S1.6: Remove the intermediate signal that meets the IMF component conditions from the original signal and use it as the original signal. Repeat S1.1 to S1.5 to continue decomposition.
[0037] s new (t) = s(t) - c j (t) (8)
[0038] In formula (8), s new (t) is the new original signal; c j(t) is the IMF component that satisfies the IMF component conditions obtained through the above steps.
[0039] In step 1, the original signal s(t) is the IMF component c of each order j (t) is the superposition of the residual function r(t), that is,
[0040]
[0041] In formula (9), j is the order of IMF; c j (t) characterizes the signal from high frequency to low frequency, and each component corresponds to a local oscillation mode in the signal; r(t) represents the average trend of the signal, that is, the direction of change of the fault signal; n is the maximum number of IMF components that can be decomposed from the original signal s(t).
[0042] When a fault occurs, the high-frequency components of the positive and negative lines in different intervals are quite different. The present invention uses the empirical mode decomposition EMD shown in formulas (1) to (8) to respectively calculate the positive power P of the busbar measurement point. P-bus , negative power P N-bus And the positive power P at the DC line measurement point P-line , negative power P N-line Decompose;
[0043] The highest frequency IMF1 component is used to calculate the high frequency energy E of the positive electrode of the bus P-bus , busbar negative pole high frequency energy E N-bus , DC line positive high frequency energy E P-line , DC line negative pole high frequency energy E N-line , the calculation formula is shown in formula (10);
[0044]
[0045] In formula (10), and are the elements in IMF1 of the bus positive pole, bus negative pole, line positive pole and line negative pole power respectively, and N is the number of elements.
[0046] The energy value obtained by formula (10) is input into the MIDBO-BiGRU model neural network as the fault feature.
[0047] In order to avoid the frequent activation of the protection system caused by normal fluctuations in the line. Considering that the voltage change rate as a fault trigger element cannot be properly activated under high resistance, the present invention uses the Teager Energy Operator (TEO) as the trigger criterion. By applying the Teager Energy Operator (TEO) to the DC voltage signal, the energy changes in the signal are highlighted.
[0048] The Teager energy operator of a discrete signal x(n) is defined as:
[0049]
[0050] In formula (11): is the instantaneous energy value; x(n), x(n+1) and x(n-1) are the signals at sampling points n, n+1 and n-1 respectively; when Exceeding the threshold After that, protection starts.
[0051] In step 2, the learning rate and neuron parameters of the BiGRU model are optimized using the multi-strategy improved dung beetle optimization algorithm;
[0052] The Dung Beetle Optimization Algorithm (DBO) is a metaheuristic algorithm based on the social behavior of dung beetle populations. It determines the optimal location of dung beetles by simulating their social behaviors, such as ball rolling, dancing, foraging, stealing, and reproduction. However, the Dung Beetle Optimization Algorithm (DBO) suffers from weak global exploration capabilities and is prone to falling into local optima. Therefore, this paper improves the Dung Beetle Optimization Algorithm by introducing an iterative chaotic map with infinite collapses (ICMIC), a spiral search strategy, and a randomized differential mutation strategy.
[0053] The initial population of the Dung Beetle Optimization Algorithm (DBO) is randomly generated, which affects the performance of the algorithm to a certain extent. This paper introduces ICMIC to generate the initial population, ensuring that the initial population is evenly distributed in the search space and improving the convergence speed. The ICMIC expression is as follows:
[0054]
[0055] In formula (12), n is the individual of the population; x n+1 is the nth mapping individual; a is the control parameter, which ranges from 0 to positive infinity.
[0056] The dung beetle optimization algorithm DBO only considers the worst case scenario when simulating the rolling ball behavior and lacks the ability to communicate with other dung beetles in a timely manner, which may lead to poor algorithm performance.
[0057] Therefore, the present invention uses a spiral search strategy to replace the position update method of the rolling ball behavior. The spiral search strategy formula is shown in formula (13).
[0058] The spiral search strategy can more effectively conduct global searches, avoid falling into local optima, and enhance the implicit information exchange between individuals, thereby improving the overall optimization effect. This improvement makes the DBO algorithm more stable and reliable in complex problems.
[0059] x(t+1)=e bk cos(2πk)*|x * (t)-x(t)|+x * (t) (13);
[0060] In formula (13), x(t+1) is the new position; b is a constant representing the spiral shape; k is an arbitrary number from 0 to 1; x * (t) is the current best position; x(t) is the current position; e bk It represents the expansion amplitude of the spiral search. By introducing randomness, the diversity of the search and the global search ability of the algorithm can be enhanced.
[0061] During the stealing phase, the dung beetle's position update is mainly based on the individual's current optimal value, which may cause the algorithm to enter a local optimal solution. Therefore, the present invention introduces a random difference mutation strategy to enhance population diversity. The formula is as follows:
[0062] P(t+1)=r1·(P * -P(t))+r2·(P′(t)-P(t)) (14);
[0063] In formula (14), P(t+1) is the position of the new individual obtained by random mutation; r1 and r2 are random values between 0 and 1; P * is the current best individual position; P′(t) is the position of an individual randomly selected from the population; P(t) is the current individual position.
[0064] In step 2, the improved Dung Beetle Optimization (DBO) algorithm, using equations (12) to (14), theoretically achieves a balance between global exploration and local search capabilities, ensuring that the algorithm is more stable and reliable when dealing with complex problems. These improvements give the algorithm stronger global exploration capabilities, faster convergence speed, and higher solution quality.
[0065] The initial population of the dung beetle optimization algorithm (DBO) is randomly generated, which may cause the initial population to be unevenly distributed within the search space, thereby affecting the algorithm's convergence speed and global exploration ability. ICMIC has a widely distributed chaotic characteristic, which enables the generated initial population to evenly cover the search space, contributing to a more comprehensive global search. At the same time, because the initial population covers a wider space, the algorithm can find high-quality solutions more quickly in the early stages, thereby improving the convergence speed. The present invention introduces the infinite folding iterative chaotic map ICMIC through formula (12), which enables the initial population to be more evenly distributed within the search space.
[0066] The original DBO algorithm only considers the worst-case scenario when simulating the ball-rolling behavior of dung beetles, lacks information exchange with other individuals, and is prone to falling into local optimal solutions. The spiral search strategy enables a more diverse and extensive search within the search space, reducing the risk of falling into local optimal solutions. At the same time, the spiral search strategy adjusts the search path, allowing each individual to implicitly exchange information with each other when updating their position, thereby improving the optimization effect of the overall algorithm. The present invention introduces the spiral search strategy through equation (13) to better perform global search.
[0067] During the stealing phase of the dung beetle optimization algorithm, individual positions are updated primarily based on the current optimal value. This approach can lead to a decrease in population diversity, making it prone to falling into a local optimum. The random differential mutation strategy ensures population diversity by introducing random mutations, allowing the algorithm to maintain good exploration capabilities in the later search phase. At the same time, by continuously introducing new mutant individuals, the algorithm can escape the local optimum, thereby improving global search capabilities and optimization effects. The present invention introduces a random differential mutation strategy through formula (14), which can introduce randomness at each update and enhance population diversity.
[0068] In step 2, the long short-term memory (LSTM) is a recurrent neural network model, but its internal structure is complex and training time is long. The gated recurrent unit (GRU) has fewer parameters and faster convergence than the LSTM, but only considers unidirectional information.
[0069] The BiGRU model introduces a reverse cycle, which better captures the complex relationships in sequence data by considering both forward and backward information at each time step. The model is composed of two layers of gated recurrent units (GRUs). Figure 1 As shown. Figure 1As can be seen, the forward GRU unit starts at the beginning of the sequence and processes each time step of the input sequence in sequence. The forward hidden state at each time step depends on the state of the previous time step and the current input. In contrast to the forward GRU unit, the backward GRU unit starts at the end of the sequence and processes the input sequence in reverse order. The backward hidden state at each time step also depends on the state of the next time step and the current input.
[0070] The BiGRU model processes sequences from both the forward and reverse directions, fully capturing temporal information and concatenating the hidden vectors of hidden neurons as the final output. The BiGRU, with its gating mechanism, selectively transmits information during the learning process, including forget gates, update gates, and output gates. This effectively alleviates the vanishing gradient problem and improves the ability to understand and model complex relationships within the entire sequence, making it suitable for DC line fault diagnosis. The BiGRU model is expressed as:
[0071]
[0072] Where: and are the forward and backward outputs at time t respectively; and are the forward and backward outputs at time t-1 respectively; the GRU function represents the nonlinear transformation of the input; y t is the output at time t; x t is the input at time t; In the BiGRU model training process, the learning rate and the number of neurons are two key hyperparameters that have a direct impact on the training effect and final accuracy of the model.
[0073] The learning rate is a parameter that controls the step size with which the model adjusts weights in each iteration. It directly affects the model's convergence speed and final accuracy. If the learning rate is set too high, the model may skip the global optimal point during training, resulting in failure to converge or oscillation around the loss function. This often leads to unstable model training, low final accuracy, or even failure to converge to a suitable solution. If the learning rate is set too low, the model converges very slowly, requiring more iterations to reach a good solution. Furthermore, a too low learning rate may cause the model to become trapped in a local optimum, making it difficult to escape, thus affecting the model's final accuracy.
[0074] The number of neurons refers to the number of neurons in each layer of the network. The number of neurons affects the model's learning and expressive capabilities. If the number of neurons is too small, or if the number of neurons in each layer is too small, the model's expressiveness will be limited, and it may not be able to effectively capture the complex features of the input data. This often leads to underfitting of the model, meaning that the model cannot fit the training data well, resulting in low final accuracy. If the number of neurons is too large, although the model's expressiveness is enhanced, it may lead to overfitting, meaning that the model performs well on the training data but poorly on the validation set or test set. Furthermore, too many neurons increase computational complexity and memory consumption, slowing down the training process.
[0075] The learning rate and number of neurons are key hyperparameters in neural network training. A reasonable learning rate ensures the model converges stably and efficiently to a good solution, thereby improving model accuracy. An appropriate number of neurons ensures sufficient expressiveness without overfitting, resulting in good performance on both training and test data. Optimizing these two parameters using the Multi-Strategy Improved Dung Beetle Optimization (MIDBO) algorithm can effectively improve the final accuracy and generalization of the BiGRU model.
[0076] In step 3, a four-terminal flexible DC transmission system is built based on the PSCAD / EMTDC simulation platform, and a data set is obtained. A MIDBO-BiGRU model is trained using the training set, and a validation set is used to test the model until the model accuracy reaches the expected level. The trained MIDBO-BiGRU model is then used for fault diagnosis of the flexible DC transmission system.
[0077] When collecting the data set, the sampling frequency was 20kHz, and fault data under different conditions were collected. The fault locations included lines and buses inside and outside the area; the fault distances were 0%, 25%, 50%, 75% and 100% of each line; the transition resistance was 0.01Ω~1000Ω; the fault types included bipolar short circuit faults and positive and negative short circuit faults; the collected data set was divided into training set and validation set at a ratio of 4:1. The trained MIDBO-BiGRU model was used for online detection of faults in the flexible DC transmission system; Figure 3 As shown in the figure, online fault monitoring includes the following steps:
[0078] S3.1: The system continuously monitors DC voltage fluctuations. When the DC voltage fluctuation caused by a system fault reaches a threshold, the data acquisition device at the measurement point collects power data.
[0079] S3.2: Use EMD to decompose the power data, calculate the energy value of the decomposed IMF1, and input the energy value into the MIDBO-BiGRU model.
[0080] S3.3: Perform corresponding protection according to the diagnosis result of the MIDBO-BiGRU model. If the fault is at the busbar, perform the corresponding busbar protection; if the fault is at the DC line, perform the corresponding DC line protection; if the fault is outside the zone, no protection action is performed; after the protection is completed, return to S3.1.
[0081] The present invention provides a flexible DC transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU, and the technical effects are as follows:
[0082] 1) Step 1 of the present invention utilizes empirical mode decomposition (EMD) to decompose the power at the bus and the DC line to obtain the intrinsic mode components (IMFs) representing the high-frequency components, and calculates their energy as the fault characteristic quantity. EMD does not require pre-setting of basis functions, so it can adaptively process nonlinear and non-stationary signals. Fault signals in flexible DC transmission systems often exhibit nonlinear and non-stationary characteristics, and EMD can effectively decompose the intrinsic mode components (IMFs) of these signals, which makes it very suitable for such complex signal processing tasks. EMD can decompose complex signals into multiple intrinsic mode components, among which the IMF1 component with the highest frequency usually contains the main high-frequency characteristics of the fault signal. By calculating the energy of these high-frequency components, fault signals and normal signals can be effectively distinguished. This method not only improves the sensitivity of fault detection, but also can significantly improve the detection capability of high-resistance faults.
[0083] 2) Step 2 of the present invention utilizes the multi-strategy improved dung beetle optimization algorithm (MIDBO) to perform parameter optimization on the BiGRU model. The traditional dung beetle optimization algorithm (DBO) is prone to fall into local optimality when dealing with high-dimensional optimization problems. In order to solve this problem, the present invention introduces infinite folding iterative chaotic mapping (ICMIC), a spiral search strategy and a random differential mutation strategy to improve the dung beetle optimization algorithm. ICMIC can generate a more uniform initial population distribution, the spiral search strategy enhances the global search capability, and the random differential mutation strategy increases the diversity of the population. These improvements enable MIDBO to more effectively explore the global optimal solution, avoid falling into the local optimal solution, and thus improve the training effect of the BiGRU model.
[0084] The learning rate and number of neurons are key hyperparameters that influence BiGRU model performance. Optimizing these parameters using MIDBO ensures rapid convergence to the global optimal solution during training, reducing training time and improving final model accuracy. Compared to traditional grid search or random search methods, MIDBO offers greater efficiency and global search capabilities, enabling better optimization of complex neural network models.
[0085] 3) The datasets generated in step 3 of the present invention using the PSCAD / EMTDC simulation platform cover a wide range of fault types, locations, and operating conditions, making them highly representative. These datasets not only provide ample training samples for the MIDBO-BiGRU model but also ensure the model's robustness and accuracy under a variety of complex operating conditions.
[0086] By proportionally dividing the dataset into training and validation sets, model parameters can be continuously optimized during training, and the model structure can be adjusted based on the test results of the validation set until optimal accuracy is achieved. This training process ensures that the model has strong generalization capabilities when facing real-world conditions and avoids overfitting.
[0087] The trained and validated MIDBO-BiGRU model was ultimately applied to online fault diagnosis in flexible HVDC transmission systems. It can detect busbar and DC line faults in real time, maintaining high-precision detection performance, especially under high-resistance fault conditions. This provides a reliable technical guarantee for the safe operation of flexible HVDC transmission systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] The present invention will be further described below with reference to the accompanying drawings and examples:
[0089] Figure 1 It is the BiGRU model structure.
[0090] Figure 2 This is the four-terminal flexible DC transmission topology diagram.
[0091] Figure 3 This is the system online monitoring and diagnosis flow chart.
[0092] Figure 4 This is the confusion matrix diagram of the validation set.
[0093] Figure 5 This is the voltage waveform under AC fault.
[0094] Figure 6 The figure below is a comparison chart of the fitness convergence speed of the four algorithms.
[0095] Figure 7 Schematic diagram of the diagnostic process of the present invention. DETAILED DESCRIPTION
[0096] The present invention will be further described in detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0097] A fault diagnosis method for flexible HVDC transmission lines based on traveling wave characteristics and MIDBO-BiGRU was proposed. First, this method constructs a fault signature based on traveling wave characteristics. By utilizing the blocking effect of boundary elements on high-frequency current and voltage signals and measuring the power at the empirical mode decomposition (EMD) measurement point, the energy of IMF1 is calculated as the fault signature. A dataset is obtained through multiple simulations and used to train a BiGRU model. Second, this method improves the Dung Beetle Optimization (DBO) algorithm using infinite folding iterative chaotic mapping, a spiral search strategy, and a randomized differential mutation strategy. This accelerates the Dung Beetle Optimization (DBO) algorithm's convergence speed in optimizing the BiGRU model parameters and enhances its global search capability. Finally, a dataset is obtained using the PSCAD / EMTDC simulation platform. The MIDBO-BiGRU model is trained using the training set and tested using the validation set until the model accuracy reaches the expected level. The trained MIDBO-BiGRU model is then used for fault diagnosis in flexible HVDC transmission systems. Finally, a simulation model was built in PSCAD / EMTDC. The results showed that this method can not only detect busbar faults and line faults, but also solve the problem of high-resistance fault protection failure while meeting the protection reliability and speed.
[0098] The specific steps include:
[0099] Step 1: First, use empirical mode decomposition (EMD) to decompose the power at the busbar and DC line to obtain the intrinsic mode component (IMF) representing the high-frequency component, and calculate its energy as the fault characteristic;
[0100] Step 2: Secondly, the multi-strategy improved dungbeetle optimizer (MIDBO) algorithm is used to optimize the parameters of the BiGRU model;
[0101] Step 3: Build a simulation platform based on PSCAD / EMTDC Figure 2 A four-terminal flexible DC transmission system is constructed and a data set is obtained. A MIDBO-BiGRU model is trained using the training set and tested using the validation set until the accuracy of the MIDBO-BiGRU model reaches the expected level. The trained MIDBO-BiGRU model is then used for fault diagnosis of the flexible DC transmission system.
[0102] Step 4: Finally, a simulation model was built in PSCAD / EMTDC. The results showed that this method can not only detect busbar faults and line faults, but also solve the problem of high-resistance fault protection failure while ensuring protection reliability and speed.
[0103] The present invention utilizes the blocking effect of the current limiting reactors on both sides of the line on high-frequency signals, uses EMD to decompose the original data into multiple intrinsic modal components, and extracts the energy of the first intrinsic modal component with the highest frequency as the feature quantity. 12 A large number of simulations are carried out as an example to verify the feasibility of this solution. The sampling frequency is 20kHz, and fault data under different conditions are collected. The fault location includes lines and buses inside and outside the area; the fault distance is 0%, 25%, 50%, 75% and 100% of each line; the transition resistance is 0.01Ω~700Ω; the fault type includes bipolar short circuit fault and positive and negative short circuit fault. The present invention selects the power data of measurement points b and l within 1.5ms after the starting element is started. There are 1070 groups of fault data in total, which are divided into training set and verification set at 4:1. After offline training of the KOA-CNN-BiGRU-AM model, the confusion matrix of the verification set is as follows Figure 4 As shown. Figure 4 As can be seen, the proposed method divides the dataset into a training set and a validation set at a ratio of 4:1, and continuously adjusts the network parameters based on the accuracy of the validation set. After multiple adjustments, the accuracy of the validation set reached 100%, indicating that the proposed model can accurately identify the seven fault types under different fault conditions after learning from the training set.
[0104] After adjusting the model parameters using the training set and the validation set, the DC line OHL 12 2 / 5, 3 / 5, 4 / 5, OHL 13 1 / 5, OHL 24 Three fault conditions are set at 1 / 5 and Bus1 respectively, with transition resistance ranging from 0.01 to 700Ω, and 30 sets of resistors are randomly selected. The protection status of the system is shown in Table 1:
[0105] Table 1 System protection status
[0106]
[0107] As can be seen from Table 1, the model can accurately identify various faults with high accuracy.
[0108] The protection area of the present invention is the busbar and DC line, so AC faults are considered out-of-area faults. Figure 2 Taking the Fs three-phase AC fault as an example, the DC voltage waveform under the three-phase fault with a transition resistance of 0.01Ω is obtained. The 5th second is the time when the fault occurs, as shown in Figure 5 As shown. Figure 5 It can be seen that the DC voltage fluctuation is small, and its maximum Teager energy value is 1161, which does not reach the threshold value, so the protection does not operate. This shows that the method proposed by the present invention is not affected by AC faults.
[0109] The proposed solution is a protection strategy based on high-frequency energy. The size of the reactor and the line length determine the weakening and blocking effects of high frequencies, so it is necessary to discuss the generalizability of the proposed solution. In the simulation model of the present invention, the line reactor size is 0.2H and the line length is 207km. The fault diagnosis accuracy of the proposed solution is discussed for reactor sizes of 0.1H, 0.15H, and 0.25H, and line lengths of 150km, 250km, and 300km. Twenty groups of faults were simulated and tested under each line parameter, and the results are shown in Table 2. As can be seen from Table 2, the proposed solution has a certain degree of generalization performance and is equally applicable to lines and reactors within a certain range.
[0110] Table 2 Weak boundary experiment
[0111]
[0112] In order to verify the problem that the MIDBO proposed in the present invention has strong global exploration ability and is easy to escape the local optimum, MIDBO is compared with DBO, particle swarm optimization algorithm (PSO) and sparrow algorithm (SSA), and the above four algorithms are used to optimize the BiGRU model of the present invention respectively. The results are as follows: Figure 6 As shown. Figure 6 It can be seen that the MIDBO proposed in the present invention has a strong global search capability and is more likely to find the optimal solution.
[0113] To address the complex manual calculation and setting of fault detection thresholds and the poor fault resistance tolerance of current four-terminal flexible DC transmission lines, this paper proposes a fault diagnosis method based on EMD and MIDBO-BiGRU, based on analysis of fault traveling wave transmission characteristics. Results demonstrate that this fault diagnosis method, based on a combined EMD and MIDBO-BiGRU model, can effectively identify various faults.
Claims
1. A flexible DC transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU is characterized by The following steps are involved: Step 1: Use EMD to decompose the power at the bus and DC line to obtain the intrinsic mode components representing the high-frequency components, and calculate their energy as the fault characteristic quantity; Step 2: Use the multi-strategy improved dung beetle optimization algorithm to optimize the parameters of the BiGRU model; Step 3: Get the data set; The MIDBO-BiGRU model is trained with the training set and tested with the validation set until the accuracy of the MIDBO-BiGRU model reaches the expected level. The trained MIDBO-BiGRU model is then used for fault diagnosis of the flexible DC transmission system.
2. The flexible DC transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU according to claim 1 is characterized in that: In step 1, the empirical mode decomposition (EMD) is selected to process the signal, and the process of the empirical mode decomposition (EMD) decomposition includes the following steps: S1.1: Find the extreme point of the original signal s(t), t = 1, 2, ..., T, where T is the length of the original data s(t). If it does not exist, the decomposition ends; The maximum value formula is as follows: s max (t)>s(t-1)&s max (t)>s(t+1)(1); In formula (1), s max (t) is the maximum value in the original signal s(t); s(t-1) is the t-1th data in the original signal s(t); s(t+1) is the t+1th data in the original signal s(t); The minimum value formula is as follows: s min (t)<s(t-1)&s min (t)<s(t+1)(2); In formula (2), s min (t) is the minimum value in the original signal s(t); S1.2: Use cubic spline function to fit the extreme points, and fit the maximum points (t, s max (t)), and the minimum point (t,s min (t)) perform upper envelope fitting and lower envelope fitting respectively to obtain the upper envelope u(t) and lower envelope l(t); S1.3: Calculate the mean of the envelope; In formula (5), m(t) is the mean value of the envelope; S1.4: Calculate the difference between the original signal and the mean, i.e. the intermediate signal; h(t)=s(t)-m(t) (6) In formula (6), h(t) is the intermediate signal; S1.5: Determine whether the intermediate signal meets the IMF component condition. If so, it is used as the IMF component. If not, return to S1.
1. The determination conditions are as follows: In formula (7), n zero 、n max and n min are the number of zero crossing points, the number of maximum points and the number of minimum points of the intermediate signal respectively; s h (t) and l h (t) represent the upper and lower envelopes of the intermediate signal h(t), respectively; S1.6: Remove the intermediate signal that meets the IMF component condition from the original signal and use it as the original signal, and repeat S1.1 to S1.5 to continue decomposition; s new (t)=s(t)-c j (t) (8) In formula (8), s new (t) is the new original signal; c j (t) is the IMF component that satisfies the IMF component conditions obtained through the above steps.
3. The flexible DC transmission line fault diagnosis method based on traveling wave characteristics and MIDBO-BiGRU according to claim 1 is characterized in that: In step 1, the original signal s(t) is the IMF component c of each order j (t) is the superposition of the residual function r(t), that is, In formula (9), j is the order of IMF; c j (t) characterizing the signal from high frequency to low frequency, each component corresponding to a local oscillation mode in the signal; r(t) represents the average trend of the signal, that is, the direction of change of the fault signal; n is the maximum number of IMF components that can be decomposed from the original signal s(t); When a fault occurs, the high-frequency components of the positive and negative lines in different sections are different. The empirical mode decomposition (EMD) is used to analyze the positive power P P-bus , Negative power P N-bus And the positive power P at the DC line measurement point P-line , Negative power P N-line To decompose; The highest frequency IMF1 component calculates the high frequency energy E of the positive pole of the bus P-bus , bus negative pole high frequency energy E N-bus , DC line positive high frequency energy E P-line , DC line negative pole high frequency energy E N-line The calculation formula is shown in formula (10); In formula (10), and are the elements in IMF1 of the bus positive pole, bus negative pole, line positive pole, and line negative pole power, respectively, and N is the number of elements; The energy value obtained by formula (10) is input into the MIDBO-BiGRU model neural network as the fault feature; The Teager energy operator (TEO) is used as the start-up criterion; the energy change in the signal is highlighted by applying the Teager energy operator (TEO) to the DC voltage signal; The Teager energy operator of a discrete signal x(n) is defined as: In formula (11): is the instantaneous energy value; x(n), x(n+1) and x(n-1) are the signals at sampling points n, n+1 and n-1 respectively; when Exceeding the threshold After that, protection starts.
4. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 1 is characterized in that: In step 2, the learning rate and neuron parameters of the BiGRU model are optimized using the multi-strategy improved dung beetle optimization algorithm; specifically, as follows: The infinite folding iterative chaotic map ICMIC, spiral search strategy and random difference mutation strategy are introduced to improve the dung beetle optimization algorithm DBO. The initial population of the Dung Beetle Optimization Algorithm (DBO) is randomly generated. ICMIC is introduced to generate the initial population. The expression of ICMIC is as follows: In formula (12), n is the population individual; x n+1 is the nth mapping individual; a is a control parameter, ranging from 0 to positive infinity; the dung beetle optimization algorithm DBO only considers the worst case when simulating the rolling ball behavior, and uses the spiral search strategy to replace the position update method of the rolling ball behavior. The spiral search strategy formula is shown in formula (13); x(t+1)=e bk cos(2πk)*|x * (t)-x(t)|+x * (t) (13); In formula (13), x(t+1) is the new position; b is a constant representing the spiral shape; k is an arbitrary number from 0 to 1; x * (t) is the current best position; x(t) is the current position; e bk Indicates the extension amplitude of the spiral search; In the stealing stage, the position of the dung beetle is updated according to the individual's current optimal value, which causes the algorithm to enter a local optimal solution; therefore, a random difference mutation strategy is introduced to enhance population diversity. The formula is as follows: P(t+1)=r1·(P * -P(t))+r2·(P′(t)-P(t)) (14); In formula (14), P(t+1) is the position of the new individual obtained by random mutation; r1 and r2 are random values between 0 and 1; P * is the current best individual position; P′(t) is the position of an individual randomly selected from the population; P(t) is the current individual position.
5. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 4 is characterized in that: In step 2, the BiGRU model introduces a reverse loop, which better captures the complex relationship in the sequence data by considering both forward and backward information at each time step. The model is composed of two layers of gated recurrent units GRU. The forward GRU unit starts from the beginning of the sequence and processes each time step of the input sequence in sequence. The forward hidden state of each time step depends on the state of the previous time step and the current input. In contrast to the forward GRU unit, the backward GRU unit starts from the end of the sequence and processes the input sequence in reverse order. The backward hidden state of each time step also depends on the state of the next time step and the current input. The BiGRU model processes the sequence from the forward and reverse directions respectively, and can fully capture the timing information, and concatenate the hidden vectors of the hidden neurons as the final output result.
6. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 5 is characterized in that: The expression of the BiGRU model is: Where: and They are the forward and backward output at time t respectively; and are the forward and backward outputs at time t-1 respectively; the GRU function represents the nonlinear transformation of the input; y t is the output at time t; x t Input at time t; is bitwise addition.
7. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 6 is characterized in that: Learning rate and number of neurons are key hyperparameters in neural network training; optimizing these two parameters using the multi-strategy improved dung beetle optimization algorithm MIDBO can effectively improve the final accuracy and generalization ability of the BiGRU model.
8. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 1 is characterized in that: In step 3, a four-terminal flexible DC transmission system is built based on the PSCAD / EMTDC simulation platform, and a data set is obtained. The MIDBO-BiGRU model is trained using the training set, and the model is tested using the validation set until the model accuracy reaches the expected level. The trained MIDBO-BiGRU model is used for fault diagnosis of the flexible DC transmission system.
9. The method for fault diagnosis of flexible DC transmission lines based on traveling wave characteristics and MIDBO-BiGRU according to claim 8 is characterized in that: When collecting data sets, the sampling frequency is 20kHz, and fault data under different conditions are collected; the fault locations include lines and busbars inside and outside the area; the fault distances are 0%, 25%, 50%, 75% and 100% of each line; the transition resistance is 0.01Ω~1000Ω; the fault types include bipolar short circuit faults and positive and negative short circuit faults; the collected data sets are divided into training sets and validation sets at a ratio of 4:1; the trained MIDBO-BiGRU model is used for online detection of faults in flexible DC transmission systems; online monitoring of faults includes the following steps: S3.1: The system continuously detects the DC voltage fluctuation; when the DC voltage fluctuation caused by the system fault reaches the threshold, the data acquisition device at the measurement point collects power data; S3.2: Decompose the power data using EMD, calculate the energy value of IMF1 obtained by decomposition, and input the energy value into the MIDBO-BiGRU model; S3.3: Perform corresponding protection according to the diagnosis result of the MIDBO-BiGRU model. If the fault is at the busbar, perform corresponding busbar protection; if the fault is at the DC line, perform corresponding DC line protection; If it is an out-of-zone fault, no protection action will be performed; after the protection is completed, return to S3.1.
