Converter transformer saturation protection optimization method based on L-M algorithm neural network fitting
By using a neural network fitting method based on the LM algorithm, a nonlinear mapping relationship is established using the current characteristics of the converter transformer. This solves the problem of maloperation of traditional protection devices under no-load closing conditions, and enables accurate differentiation between inrush current and DC bias, thereby improving the reliability and speed of the protection device.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional converter transformer saturation protection may malfunction or operate prematurely under no-load closing conditions due to inrush current interference. It cannot effectively distinguish between inrush current and DC bias, leading to misjudgment and safety hazards.
A neural network fitting method based on the LM algorithm is adopted. By monitoring the three-phase current and neutral current of the converter transformer, the peak value of the fundamental component and its attenuation characteristics are extracted. The nonlinear mapping relationship from the characteristic quantity to the magnitude of DC bias is established using the LM algorithm neural network model, and the protection criteria are optimized to improve accuracy.
It enables precise differentiation between inrush current and DC bias, improves the reliability and speed of protection devices, reduces the risk of malfunction, and ensures the safe and stable operation of the system.
Smart Images
Figure CN121863304A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system relay protection technology, specifically to an optimization method for converter transformer saturation protection based on LM algorithm neural network fitting. Background Technology
[0002] High-voltage direct current (HVDC) transmission technology has become an important method for modern power grid interconnection and clean energy consumption due to its significant technical and economic advantages in long-distance, large-capacity power transmission. Compared with AC transmission, HVDC systems offer superior performance in terms of line loss, stability control, and overcoming geographical barriers. However, the operation of HVDC systems also brings new technical challenges, one of which is the DC bias problem of converter transformers (or converter transformers).
[0003] When direct current enters the transformer windings, it causes the core operating point to deviate from the linear region and enter a saturation state, leading to a series of serious consequences: Firstly, the excitation current becomes severely distorted, and the harmonic content increases sharply, potentially causing relay protection devices to malfunction. Secondly, transformer vibration intensifies, and noise increases significantly, threatening safe operation. Thirdly, localized overheating accelerates insulation aging and may even cause thermal failures. Finally, increased leakage flux leads to eddy current losses in metal components such as the transformer tank and may cause corrosion of surrounding metal pipes. Therefore, saturation protection devices are commonly configured for converter transformers in engineering, which typically reflect the DC bias level indirectly by monitoring the neutral point current.
[0004] While traditional saturation protection is effective in dealing with steady-state DC bias, its principle has inherent flaws, particularly its reliability is severely insufficient under no-load closing conditions of converter transformers. No-load closing generates a high-amplitude, slowly decaying inrush current containing significant DC decay and aperiodic components, leading to a significant increase in the measured neutral point current of the transformer. Protection devices cannot effectively distinguish this inrush current from a genuine DC bias fault, often misinterpreting it as a severe DC intrusion, thus causing maloperation. Several domestic engineering projects have been plagued by this problem. For example, since 2019, a converter station in a certain province has repeatedly delayed system commissioning due to maloperation of saturation protection during no-load closing operations. The root cause lies in the fact that traditional protection criteria rely solely on a simple comparison of the neutral current peak value and inverse-time characteristics, lacking a deep understanding of the fundamental differences between the inrush current decay characteristics and DC bias.
[0005] Therefore, there is an urgent need to develop a saturation protection principle that can accurately distinguish between no-load closing excitation inrush current and actual DC bias, in order to make up for the shortcomings of traditional saturation protection and ensure the safe and stable operation of UHVDC projects. Summary of the Invention
[0006] To address the problem of maloperation or premature operation of existing converter transformer saturation protection under no-load closing conditions due to inrush current interference, this invention provides an optimization method for converter transformer saturation protection based on LM algorithm neural network fitting. This method accurately distinguishes between inrush current and DC bias by quantifying the magnitude of DC bias, thereby improving the reliability of protection operation.
[0007] The technical solution adopted in this invention is as follows: The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting includes the following steps: Step 1: After the protection device is activated, initialize the system parameters; Step 2: Monitor the circuit breaker position status in real time. When a closing signal is detected, collect the three-phase current on the valve side of the converter transformer. , , and instantaneous value of neutral line current ; Step 3: Perform digital filtering on the collected three-phase current data to extract the fundamental component; Then, the peak curves of the fundamental components of each phase are calculated to identify the phase with the fastest decay. Record its fundamental frequency component peak value And the time required for the peak to decay to 5%. ; Step 4: Include , The feature dataset is used as input. As output, based on the inclusion , The dataset was used to train the neural network model of the LM algorithm; Step 5: Based on the input feature dataset, obtain the output value through a neural network. Query the inverse time-limit characteristic curve to determine the protection action after the corresponding running time delay.
[0008] It also includes step 6: For certain special closing conditions, which cause the peak values of the three-phase inrush current to rise, the neural network calculation link is directly bypassed, and the traditional inverse time criterion based on the measured peak value of the neutral line current is adopted to ensure the speed and accuracy of the protection action.
[0009] In step 1, the system parameters are initialized, including: setting the sampling frequency to be no less than 4kHz, loading the neural network model, and recording the inverse time-limit characteristic curve. The inverse-time characteristic curve is a standard DC current-allowable operating time curve provided by the converter transformer manufacturer. For example... Figure 2As shown, the core of the inverse-time characteristic of the allowable operating time of the converter transformer under DC bias is that the larger the DC current, the shorter the allowable time. When the DC current is close to 500A, the allowable time of the converter transformer is as short as a second, and the protection needs to act quickly; when the DC current is between 100-200A, the allowable time is tens of seconds, and the protection needs to balance fault clearing efficiency and system selectivity; when the current is less than or equal to 100A, the allowable time is hundreds of seconds or more, and the protection can act slowly or monitor to reduce unnecessary tripping. The inverse-time characteristic curve provides operating criteria for the setting of power grid relay protection and the planning of operating logic.
[0010] In step 2, the three-phase current on the valve side of the converter transformer is collected using a current transformer. , , and neutral line current Instantaneous values are extracted, and peak current is continuously recorded for at least 20 cycles. Figure 3 and Figure 4 The figures show the three-phase inrush current curve and neutral current peak curve under a simulation model of a ±800kV converter station, with the operating condition being the A-phase closing angle. The three-phase residual magnetism of transformer T1 is 0.7pu, 0pu, and -0.7pu, while that of transformer T2 is -0.7pu, 0pu, and 0.7pu. The DC bias is 300A. Figure 3 It can be seen that after no-load closing, the three-phase inrush current reaches its peak value, and then rapidly decays over time. Figure 4 In the diagram, the black curve represents the peak value of the neutral line current. The red curve represents the inverse time characteristic curve. Under DC bias of 300A, according to... Figure 2 The inverse-time characteristic curve shown indicates a theoretical operating time of 28.5 seconds. However, the actual measured operating time under this condition is 21.2 seconds, indicating that the saturation protection device will activate prematurely.
[0011] In step 3, the collected three-phase current data undergoes digital filtering, and a 50Hz bandpass filter is used to extract the fundamental component. Then, the peak value curves of the fundamental component for each phase are calculated to identify the phase with the fastest attenuation. The phase with the fastest attenuation refers to the phase in the three-phase excitation current where the fundamental component decays the fastest during no-load closing. The peak value of its fundamental component is recorded. Time required for the peak value to decay to 5% .exist Figure 3 In the three-phase inrush current curve, phase A decays the fastest, and its fundamental component peak value is recorded. and the time required for the peak to decay to 5%. ,like Figure 5 As shown in the figure, the peak value of the fundamental current in phase A can be obtained. The current is 2.199 kA, and the decay time is... It takes 7.56 seconds.
[0012] In step 4, for those containing , The dataset was augmented to obtain a large amount of data; The sample augmentation method employs linear interpolation to expand the dataset under typical operating conditions, including combinations of different closing angles, residual magnetism, and DC bias conditions. The specific sample augmentation method is as follows: First, 100 sets of data are simulated under typical initial conditions, considering CT saturation and DC bias currents of 0A, 150A, 300A, and 450A. For each operating condition with the same initial closing conditions and DC magnitudes of 0A, 150A, 300A, and 450A, linear interpolation is performed, along with linear interpolation of the decay time. This is done in 28 intervals of 5A DC. Since the influence of DC bias gradually increases, its effect on the peak value of the three-phase current fundamental component is not significant; the peak value of the phase current fundamental component is taken as the average of the peak values under the corresponding two boundary DC conditions.
[0013] The large amount of data is optimized to obtain an effective dataset; the effective dataset is then used to train the LM algorithm neural network model to obtain a mathematical model.
[0014] LM algorithm neural network model: The LM algorithm for neural network fitting is a method that uses the LM algorithm to optimize the parameter vector of a neural network. The supervised learning process. The LM algorithm is an intelligent fusion of gradient descent and Gauss-Newton methods, and its core objective is to adjust... This minimizes the sum of squared errors between the fitted output and the true output of the neural network for all training samples. The objective function of the LM algorithm is: ; in: These are observed values. P is the model function, and P is the parameter vector to be optimized.
[0015] The update rule of the LM algorithm can also be expressed by a mathematical formula, that is... ; in: Let be the parameter vector of the k-th iteration, J be the Jacobian matrix, where represents the partial derivatives of the objective function with respect to the parameters, and r be the residual vector. It is a damping factor that enables a smooth transition between the Gauss-Newton method and the gradient descent method.
[0016] A large amount of data was obtained through sample augmentation. The LM algorithm neural network model was used to train the data and perform fitting. Figure 6(a) shows the training results, and Figure 6(b) shows the test results. It can be seen that the coefficient of determination (R²) for both training and testing reached 99.9%, and the root mean square error (RMSE) for training and testing were 3.29 and 3.72, respectively. Compared with the actual value, the error of 100A DC is relatively small. The relative error of the test set is as follows: Figure 7 As can be seen, the relative error of 95% of the data points is within 5%, indicating high accuracy.
[0017] In step 5, different output values are obtained through a neural network based on different input feature datasets. The inverse time characteristic curve is queried, and the protection action is determined after the corresponding operating time delay. Table 1 presents different no-load closing conditions of the converter transformer, and under different DC biases, the DC current magnitude is fitted by inputting the feature values of the LM algorithm neural network. For the above... Figure 5 Under the corresponding operating conditions, the fundamental component peak value of phase A, which has the fastest decay rate, is 2.199 kA, and the decay time is 7.56 s. Using the LM algorithm neural network fitting model, its DC bias magnitude is calculated to be 302.64 A. Figure 2 The relationship between DC current and protection action time shows that a delay of 28.82s results in saturation protection action, which is close to the theoretical action time of 29s. This effectively prevents premature protection action and improves protection reliability. The action signal curve is shown below. Figure 8 As shown.
[0018] In step 6, for certain special closing conditions, such as those with optimized closing angles, very low residual magnetism, and rising peak values of all three-phase inrush currents, such as... Figure 9 As shown. The closing angle of phase A is... The residual magnetism of T1 is -0.6 pu, -0.6 pu, and 0.9 pu, and the residual magnetism of T2 is -0.6 pu, -0.6 pu, and 0.9 pu. The applied DC bias is 400A, and the inrush current of all three phases shows an increasing trend. The curve is similar to the curve when DC bias occurs during normal operation. The curve uses the traditional inverse-time criterion based on the measured peak value of the neutral current, such as... Figure 10 As shown. Under this operating condition, the original criteria can be used directly for judgment. The theoretical action time is 26.5s, and the protection action time is 26.36s, which is close to the theoretical action time, ensuring the speed and accuracy of the protection action.
[0019] This invention discloses an optimization method for converter transformer saturation protection based on LM algorithm neural network fitting, with the following technical advantages: 1) This invention can accurately quantify the DC bias component superimposed on a strong inrush current. By extracting the peak value of the fundamental component of the fastest decaying phase and its decay time as features, and utilizing the powerful nonlinear fitting capability of the LM algorithm, the neural network can establish an accurate mapping relationship from inrush current characteristics to the magnitude of DC bias. Simulation results show that the model has a fitting determination coefficient R² as high as 99.9%, the root mean square error (RMSE) of the test set is only 3.72, and the relative fitting error under most operating conditions is controlled within 5%, achieving high-precision identification of the DC component.
[0020] 2) Traditional protection methods are prone to failure under complex and changing field conditions, while the method of this invention exhibits excellent adaptability. It can effectively adapt to different initial closing phase angles and transformer residual magnetism.
[0021] 3) This invention is more practical for engineering applications and easier to integrate and deploy. Essentially, this invention is an algorithm upgrade, requiring only modification or addition of corresponding logic programs to existing protection devices. It eliminates the need for new sensors or hardware, reducing modification costs and complexity. This method constitutes a complete, layered protection strategy. For special operating conditions where inrush flow is not significant, it can automatically switch to traditional criteria, ensuring comprehensive and reliable protection.
[0022] 4) This invention exhibits excellent anti-interference capability and robustness. Simulation verification shows that even under harsh conditions of CT saturation, although the accuracy decreases slightly, it can still maintain the correct motion trend, ensuring the safety and reliability of the system. Attached Figure Description
[0023] The present invention will be further described below with reference to the accompanying drawings and examples; Figure 1 A flowchart illustrating the method for preventing maloperation under saturation protection of converter transformers.
[0024] Figure 2 The present invention relates to the DC bias current of the converter transformer in a ±800kV converter station. A diagram showing the correspondence between allowed runtime and time.
[0025] Figure 3 The inrush current curves of a 300A DC converter under no-load switching are shown.
[0026] Figure 4 This is the peak current curve of the neutral line under no-load closing.
[0027] Figure 5 for Figure 3 Peak curve of phase current with the fastest fundamental frequency decay.
[0028] Figure 6(a) shows the training results obtained using the LM algorithm after sample augmentation; Figure 6(b) shows the test results obtained using the LM algorithm after sample augmentation.
[0029] Figure 7 This is the relative error histogram for waveform fitting in this invention.
[0030] Figure 8 The peak value curve of the fundamental component of the phase current and the operating signal curve of the present invention at 300A DC are shown. Figure 9 The inrush current curve of a 400A DC converter transformer under a certain operating condition. Figure 10 This is a comparison chart of the peak current curve and the inverse time characteristic curve of phase C. Detailed Implementation
[0031] An optimization method for converter transformer saturation protection based on LM algorithm neural network fitting is applied to the saturation protection under no-load closing conditions of converter transformers. This method aims to address abnormal operating behavior of converter transformer saturation protection and improve the reliability of converter transformer operation. The core principle of this method lies in utilizing the difference in attenuation characteristics of the inrush current under different DC bias conditions during no-load closing, and accurately quantifying the DC component through an intelligent algorithm. Specifically: First, the three-phase excitation current signal of the converter transformer is acquired in real time, and the peak value of the fundamental component of the excitation current of the phase with the fastest decay is extracted. and the time required for it to decay to 5% of its peak value. As a key feature quantity; Subsequently, the feature quantity is input into a neural network model trained based on the LM algorithm. This model can establish a nonlinear mapping relationship from the feature quantity to the magnitude of DC bias, thereby achieving accurate fitting and quantification of it. Finally, based on the fitted... The protection delay is determined by querying the inverse time characteristic curve provided by the manufacturer, and this serves as the basis for the saturation protection action. Compared to traditional protection criteria based solely on the neutral current peak value, and other correction methods based on waveform prediction, this invention is the first to apply LM algorithm neural network fitting technology to this challenging problem. By mining and utilizing the deep characteristics of the inrush current decay process, it achieves more accurate and reliable protection, providing effective assurance for the safe commissioning of converter transformers.
[0032] Example: Figure 2 The following is a diagram showing the DC current-allowable operating time relationship of a 400kV Y / Y converter transformer at the lower end of a converter station. When there is a 150A DC intrusion, the saturation protection action time is 65s; when there is a 300A DC intrusion, the saturation protection action time is 29.5s; and when there is a 450A DC intrusion, the saturation protection action time is 26s.
[0033] Figure 3 The figure shows the three-phase inrush current curve under the simulation model of a ±800kV converter station, with the operating condition being the closing angle of phase A. The three-phase residual magnetism of transformer T1 is 0.7pu, 0pu, and -0.7pu, while that of transformer T2 is -0.7pu, 0pu, and 0.7pu. The DC bias is 300A. As shown in the figure, after no-load closing, the three-phase inrush current reaches its peak value and then rapidly decays over time.
[0034] Figure 4 This is the peak value curve of the neutral line current under the converter station simulation model. The black curve represents the peak value of the neutral line current. The red curve represents the inverse time characteristic curve. Under DC bias of 300A, according to... Figure 2 The inverse-time characteristic curve shown indicates a theoretical operating time of 28.5 seconds. However, the actual measured operating time under this condition is 21.2 seconds, indicating that the saturation protection device will activate prematurely.
[0035] Figure 5 for Figure 3 The three-phase inrush current curve shows the peak current of the phase with the fastest fundamental frequency decay. The phase with the fastest decay is phase A, and the peak fundamental current of phase A is... The current is 2.199 kA, and the decay time is... It takes 7.56 seconds.
[0036] Figures 6(a) and 6(b) show the neural network fitting results based on the LM algorithm. Figure 6(a) shows the training results, and Figure 6(b) shows the testing results. It can be seen that the coefficient of determination (R²) for both training and testing reaches 99.9%, and the root mean square error (RMSE) for training and testing is 3.29 and 3.72, respectively. Compared to the actual value, the error in the 100A DC range is relatively small. The relative error of the test set is as follows: Figure 7 By fitting the relative error histogram, it can be seen that the relative error of 95% of the data points is within 5%, indicating high accuracy.
[0037]
[0038] Table 1 presents different no-load closing conditions of the converter transformer and provides the DC current magnitude fitted by inputting the feature values of the LM algorithm neural network under different DC bias conditions. Figure 5 Under the corresponding operating conditions, the fundamental component peak value of phase A, which has the fastest decay rate, is 2.199 kA, and the decay time is 7.56 s. Using the LM algorithm neural network fitting model, its DC bias magnitude is calculated to be 302.64 A. Figure 2The relationship between DC current and protection action time shows that a delay of 28.82s results in saturation protection action, which is close to the theoretical action time of 29s. This effectively prevents premature protection action and improves protection reliability. The action signal curve is shown below. Figure 8 As shown.
[0039] Figure 9 The curve shows the inrush current of the three-phase excitation system. The operating condition is that the closing angle of phase A is... The residual magnetism of T1 is -0.6 pu, -0.6 pu, and 0.9 pu, and the residual magnetism of T2 is -0.6 pu, -0.6 pu, and 0.9 pu. The applied DC bias is 400A. All three-phase inrush currents show an increasing trend. The curve is similar to the curve when DC bias occurs during normal operation. The curve uses the traditional inverse-time criterion based on the measured peak value of the neutral current, such as... Figure 10 As shown. Under this operating condition, the original criterion can be used directly for judgment. The theoretical action time is 26.5s, and the protection action time is 26.36s, which is close to the theoretical action time.
Claims
1. An optimization method for converter transformer saturation protection based on LM algorithm neural network fitting, characterized in that... Includes the following steps: Step 1: After the protection device is activated, initialize the system parameters; Step 2: Monitor the circuit breaker position status in real time. When a closing signal is detected, collect the three-phase current on the valve side of the converter transformer. , , and instantaneous value of neutral line current ; Step 3: Perform digital filtering on the collected three-phase current data to extract the fundamental component; Then, the peak curves of the fundamental components of each phase are calculated to identify the phase with the fastest decay. Record its fundamental frequency component peak value And the time required for the value to decay from the peak to a certain value. ; Step 4: Include , The feature dataset is used as input. As output, based on the inclusion , The dataset was used to train the neural network model of the LM algorithm; Step 5: Based on the input feature dataset, obtain the output value through a neural network. Query the inverse time-limit characteristic curve to determine the protection action after the corresponding running time delay.
2. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 1, characterized in that: It also includes step 6: For special closing conditions that cause the peak values of the three-phase inrush current to rise, the neural network calculation link is directly bypassed, and the traditional inverse time criterion based on the measured peak value of the neutral line current is adopted to ensure the speed and accuracy of the protection action.
3. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 1, characterized in that: In step 1, the system parameters are initialized, including setting the sampling frequency, loading the neural network model, and inputting the inverse time-limit characteristic curve.
4. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 3, characterized in that: The inverse-time characteristic curve is the standard DC current-allowable operating time correspondence curve of the converter transformer. The allowable operating time of the converter transformer under DC bias exhibits an inverse-time characteristic: the larger the DC current, the shorter the allowable time; when the DC current is close to 500A, the allowable time of the converter transformer is as short as seconds, and the protection needs to act quickly; when the DC current is between 100-200A, the allowable time is tens of seconds, and the protection needs to take into account both fault clearing efficiency and system selectivity; when the current is less than or equal to 100A, the allowable time is more than hundreds of seconds, and the protection can act slowly or monitor to reduce unnecessary tripping.
5. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 4, characterized in that: In step 2, the three-phase current on the valve side of the converter transformer is collected using a current transformer. , , and neutral line current Instantaneous values are extracted, and the peak current is continuously recorded for at least 20 cycles. After no-load closing, the three-phase inrush current reaches its peak value and then decays rapidly over time.
6. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 5, characterized in that: In step 3, the collected three-phase current data is digitally filtered, and the fundamental component is extracted using a 50Hz bandpass filter. Then, the peak curves of the fundamental components of each phase are calculated to identify the phase with the fastest decay. The phase with the fastest decay refers to the phase in the three-phase excitation current where the fundamental component decays the fastest during the no-load closing process. Record its fundamental frequency component peak value Time required for the peak value to decay to 5% .
7. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 6, characterized in that: In step 4, for those containing , The dataset was augmented to obtain a large amount of data; Sample augmentation uses linear interpolation to expand the dataset under typical operating conditions, including combinations of different closing angles, residual magnetism, and DC bias conditions; the large amount of data is optimized to obtain an effective dataset; the effective dataset is used to train the LM algorithm neural network model to obtain a mathematical model.
8. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 7, characterized in that: The LM algorithm neural network model is as follows: The core objective of the LM algorithm is to adjust This minimizes the sum of squared errors between the fitted output and the true output of the neural network for all training samples; the objective function of the LM algorithm is: ； in: These are observed values. P is the model function, and P is the parameter vector to be optimized. The update rule of the LM algorithm can also be expressed mathematically, that is... ； in: Let be the parameter vector of the k-th iteration, J be the Jacobian matrix containing the partial derivatives of the objective function with respect to the parameters, and r be the residual vector. It is the damping factor; A large amount of data was obtained through sample augmentation, and the data was trained using the LM algorithm neural network model for fitting.
9. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 8, characterized in that: In step 5, different output values are obtained through a neural network based on different input feature datasets. The inverse time characteristic curve is queried, and the protection action is performed after the corresponding running time is delayed. Under different converter transformer no-load closing conditions, the DC current magnitude is fitted by inputting the feature values of the LM algorithm neural network under different DC bias conditions.
10. The optimization method for converter transformer saturation protection based on LM algorithm neural network fitting according to claim 2, characterized in that: In step 6, for special closing conditions, such as optimized closing angle, very small residual magnetism, and rising peak values of all three phase inrush currents, the original criteria are directly used for judgment under these conditions, ensuring the speed and accuracy of protection action.