Fault recovery method and system for data-driven DC power distribution network
Through the data-driven fault detection method, DC distribution network data is collected in real time, the improved adaboost algorithm model is used to identify fault points and types, and power supply is restored through the grid load scheduling system, which solves the problems of insufficient fault detection sensitivity and high error rate in the existing technology, and achieves high-precision fault positioning and system stability.
Patent Information
- Application Number
- CN202510020656.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-07
- Publication Date
- 2025-06-03
AI Technical Summary
The existing DC distribution network fault detection methods have problems such as insufficient sensitivity, high misjudgment rate, complex calculations and susceptible to noise interference, especially in the case of high resistance and complex system changes.
Using a data-driven method, the current and voltage data of the DC distribution network are collected in real time, and fault characteristic samples are preprocessed, and fault points and types are identified using the improved adaboost algorithm model combined with the BP neural network, and fault areas are isolated through the grid load scheduling system to restore power supply.
It improves the sensitivity and anti-interference ability of fault detection, and can achieve high-precision fault positioning and type selection within 2.5 milliseconds, ensuring the system's detection accuracy and stability under complex fault conditions, with an accuracy rate of 99.57%.
Smart Images
Figure CN120090149A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of distribution network fault detection, and particularly relates to a data-driven fault recovery method and system for a DC distribution network. Background Art
[0002] With the rapid development of distributed generation, electric vehicles, and energy storage, there are more and more DC scenarios on the distribution network. There are many problems with AC distribution networks in terms of energy efficiency, power conversion, and power source flexibility. In contrast, DC distribution networks based on voltage source converters (VSCs) have advantages in terms of power supply radius, efficiency, and quality, and are regarded as an important development direction. However, one of the main challenges in the practical application of DC distribution networks is that DC fault currents increase rapidly with high amplitudes, seriously endangering the safety of DC distribution networks. Since flexible converters cannot withstand excessive fault currents, the fault lines in DC distribution networks are mainly isolated by DC circuit breakers after receiving fault identification signals. DC faults must be identified within a short time before the rapidly developing fault current damages the converter equipment. Therefore, rapid fault detection is crucial for preventing fault hazards in DC distribution networks.
[0003] The rapid fault identification of DC distribution networks mainly prevents the sudden change of fault characteristics. Most existing studies can be divided into three categories according to the sudden change of fault characteristics. They are DC fault detection methods based on traveling waves, methods based on the amplitude of DC line fault characteristics, and methods based on boundary characteristics provided by current-limiting reactors.
[0004] The disadvantages of the above several methods are as follows:
[0005] (1) The principle of DC system fault identification based on differential voltage traveling waves. The traveling wave-based fault detection method is sensitive, but usually requires a high sampling frequency, such as 500 kHz, to record traveling waves. Obviously, the traveling wave-based fault detection method is more suitable for long lines and high sampling frequencies in DC transmission systems.
[0006] (2) The method based on the amplitude of DC line fault characteristics is prone to misjudging faults on adjacent DC lines. Although several fault detection methods with communication enhance the effectiveness of DC fault judgment, they are uneconomical.
[0007] (3) The method based on boundary characteristics provided by current-limiting reactors is complex in calculation and vulnerable to noise interference in DC fault detection. In addition, the above methods are applicable to specific network topologies based on boundary conditions, and they lack the ability to adapt to changes in topology and voltage levels.
[0008] These research methods involve complex feature extraction and training processes, lack a general rectification method, and do not have sufficient adaptability due to the complexity and variability of DC distribution systems.
[0009] In summary, the existing DC fault detection scheme based on the amplitude of fault characteristics lacks the ability to withstand high resistance, and the training process of the existing DC fault detection scheme is complex and cannot adapt to the changes of the system. Summary of the Invention
[0010] The object of the present invention is to provide a fault recovery method and system for a data-driven DC distribution network in view of the above problems existing in the prior art.
[0011] To achieve the above object, the technical solution of the present invention is as follows:
[0012] In a first aspect, the present invention proposes a fault recovery method for a data-driven DC distribution network, including:
[0013] S1. Real-time collect the current and voltage data of the DC distribution network at time t, and determine whether the distribution network has a fault according to the collected current and voltage data. If so, enter S2;
[0014] S2. Obtain the fault feature samples and preprocess the fault feature samples;
[0015] S3. Based on the preprocessed fault feature samples, use the improved adaboost algorithm model to identify the fault point and the fault type;
[0016] S4. The grid load dispatching system isolates the fault area according to the fault point and the fault type, and restores power supply to the non-fault area by adjusting the load of the grid;
[0017] S5. Judge whether the fault is recovered. If not, return to S1 to collect the current and voltage data of the DC distribution network at time t+1 and enter the next cycle.
[0018] In S2, the preprocessing includes:
[0019] S21. Calculate the fault response sensitivity of each fault feature according to the following formula, and remove the fault features with sensitivity less than the set threshold:
[0020]
[0021] In the above formula, J P is the fault response sensitivity of the fault feature X p , t 0 is the fault occurrence time, Δt is the sampling interval, n is the number of fault feature samples, and X p (t 0 ) is the fault feature X 0 at time t p ;
[0022] S21. Calculate the comprehensive correlation coefficients of each fault feature after being processed by S21 respectively, and screen out the fault features whose comprehensive correlation coefficients are greater than the set threshold:
[0023]
[0024] In the above formula, d p is the comprehensive correlation coefficient of the p-th fault feature X p , c p,q is the Pearson correlation coefficient between the fault feature X p and X q . d is the number of fault features under each fault type, and b is the number of fault types.
[0025] The said S3 includes:
[0026] S31. Use the improved curve smoothing algorithm based on Apollo to smooth the preprocessed fault feature samples. Among them, the objective function of the improved curve smoothing algorithm based on Apollo is:
[0027] B = α 1 B 1 + α 2 B 2 + α 3 B 3
[0028] B 1 = ∑|2x p (k) - x p (k - 1) - x p (k + 1)|
[0029] B 2 = ∑|x p (k + 1) - x p (k)|
[0030] B 3 = ∑x p (k) - x p (k) ref |
[0031] In the above formula, B is the objective function, B 1 , B 2 , B 3 are the improved smoothness cost function, length cost function and deviation cost function respectively, α 1 , α 2 , α 3 are the weight coefficients of B 1 , B 2 , B 3 respectively, and x p (k) is the k-th sample value, xp (k) ref is the optimal solution for the k-th sample;
[0032] S32. Normalize the fault feature samples after smoothing;
[0033] S33. Input the normalized fault feature samples into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type. Among them, the improved adaboost algorithm model uses a BP neural network as a weak classifier.
[0034] The training process of the improved adaboost algorithm model includes:
[0035] S331. Initialize the parameters. Among them, the weights of each training sample are initialized to 1 / M, where M is the number of training samples;
[0036] S332. Start the first round of iteration. First, train a weak classifier based on the current sample weights, and then use the training efficiency determined by the following formula as the importance weight of the weak classifier, and update the weights of each training sample:
[0037]
[0038] In the above formula, β t is the training efficiency of the t-th round of iteration, η is the learning rate, and φ is the convergence target;
[0039] S333. Repeat S332 in a loop until the training of the target number of weak classifiers is completed;
[0040] S334. Combine each weak classifier into a strong classifier to obtain the output of the model. Among them, the output function H(X) is as follows:
[0041]
[0042] In the above formula, G t (X), W t are the output and weight of the t-th weak separator respectively, and N is the number of weak classifiers.
[0043] The calculation method of the said W t is:
[0044] Expand and take the derivative of the following objective function, and take W when the derivative is zero t :
[0045]
[0046] In the above formula, y i is the i-th training sample, H i-1(X) is the output of the (i - 1)-th training sample obtained through training.
[0047] In step S2, before preprocessing the fault feature samples, data expansion of the fault feature samples is performed by linear interpolation according to the following formula:
[0048]
[0049] In the above formula, is the p-th fault data of the d-type fault occurring at position c at time k, which is a row vector, and e is a number based on the fault impedance. is at and the g-th group of the G groups of fault data expanded between, is adjacent fault data.
[0050] In a second aspect, the present invention proposes a fault recovery system for a data-driven DC distribution network, including a data acquisition and fault judgment module, a sample preprocessing module, a fault identification module, and a power grid load scheduling system;
[0051] The data acquisition and fault judgment module is used for:
[0052] Real-time collecting the current and voltage data of the DC distribution network at time t, and determining whether a fault occurs in the distribution network according to the collected current and voltage data. If so, the sample preprocessing module is activated;
[0053] After power supply is restored in the non-fault area, it is judged whether the fault is restored. If not, the current and voltage data of the DC distribution network at time t + 1 are collected and enter the next cycle;
[0054] The sample preprocessing module is used to obtain fault feature samples and preprocess the fault feature samples;
[0055] The fault identification module is used to identify the fault point and the fault type based on the preprocessed fault feature samples by using an improved adaboost algorithm model;
[0056] The power grid load scheduling system is used to isolate the fault area according to the fault point and the fault type, and restore power supply to the non-fault area by adjusting the load of the power grid.
[0057] The preprocessing includes:
[0058] A1. Calculate the sensitivity of each fault feature according to the following formula, and remove the fault features with sensitivity less than the set threshold:
[0059]
[0060] In the above formula, J P is the fault response sensitivity of the fault feature X p , t 0 is the fault occurrence time, Δt is the sampling interval, n is the number of fault feature samples, and X p (t 0 ) is the fault feature X 0 at time t p .
[0061] A2. Calculate the comprehensive correlation coefficients of each fault feature after being processed by A1 respectively, and screen out the fault features whose comprehensive correlation coefficients are greater than the set threshold:
[0062]
[0063] In the above formula, d p is the comprehensive correlation coefficient of the p-th fault feature X p , c p,q is the Pearson correlation coefficient between the fault feature X p and X q , d is the number of fault features under each fault type, and b is the number of fault types.
[0064] The fault identification module identifies the fault point and the fault type based on the following steps:
[0065] B1. Use the improved curve smoothing algorithm based on Apollo to smooth the preprocessed fault feature samples. Among them, the objective function of the improved curve smoothing algorithm based on Apollo is:
[0066] B = α 1 B 1 + α 2 B 2 + α 3 B 3
[0067] B 1 = ∑|2x p (k) - x p (k - 1) - x p (k + 1)|
[0068] B 2 = ∑|x p (k + 1) - x p (k)|
[0069] B 3 = ∑x p (k) - x p (k) ref |
[0070] In the above formula, B is the objective function, B1 , B 2 , B 3 are respectively the improved smoothness cost function, length cost function, and deviation cost function. α 1 , α 2 , α 3 are respectively the weight coefficients of B 1 , B 2 , B 3 . x p (k) is the k-th sample value, and x p (k) ref is the optimal solution of the k-th sample;
[0071] B2. Normalize the fault feature samples after smoothing;
[0072] B3. Input the normalized fault feature samples into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type. Among them, the improved adaboost algorithm model uses a BP neural network as a weak classifier, and its training process includes:
[0073] B31. Initialize the parameters. Among them, the weights of each training sample are initialized to 1 / M, and M is the number of training samples;
[0074] B32. Start the first round of iteration. First, train a weak classifier based on the current sample weights, and then use the training efficiency determined by the following formula as the importance weight of this weak classifier, and update the weights of each training sample:
[0075]
[0076] In the above formula, β t is the training efficiency of the t-th round of iteration, η is the learning rate, and φ is the convergence target;
[0077] B33. Repeat B32 in a loop until the training of the target number of weak classifiers is completed;
[0078] B34. Combine each weak classifier into a strong classifier to obtain the output of the model. Among them, the output function H(X) is as follows:
[0079]
[0080] In the above formula, G t (X), W t are respectively the output and weight of the t-th weak separator, and N is the number of weak classifiers;
[0081] The calculation method of the said W t is:
[0082] Expand and take the derivative of the following objective function, and find W when the derivative is zero. t :
[0083]
[0084] In the above formula, y i is the i-th training sample, and H i-1 (X) is the output of the (i - 1)-th training sample obtained through training.
[0085] The system further includes a data expansion module, which is used to expand the fault feature samples by linear interpolation according to the following formula before preprocessing the fault feature samples:
[0086]
[0087] In the above formula, is the p-th fault data of the d-type fault occurring at position c at time k, which is a row vector, and e is a number based on the fault impedance. is at and The g-th group of the G groups of fault data expanded between, is Adjacent fault data.
[0088] Compared with the prior art, the beneficial effects of the present invention are:
[0089] 1. A fault recovery method for a data-driven DC distribution network according to the present invention first collects the current and voltage data of the DC distribution network in real time, and determines whether a fault occurs in the distribution network according to the collected current and voltage data. If so, obtain the fault feature samples, and preprocess the fault feature samples. Then, based on the preprocessed fault feature samples, use an improved adaboost algorithm model to identify the fault point and the fault type. Finally, the grid load dispatching system isolates the fault area according to the fault point and the fault type, and restores power supply to the non-fault area by adjusting the load of the grid. This method improves the sensitivity and anti-interference ability of fault detection by optimizing feature extraction and algorithm models, can achieve high-precision fault location and pole selection within 2.5 milliseconds, ensures the detection accuracy and stability of the system under complex fault conditions, and the accuracy rate can reach 99.57%.
[0090] 2. In the improved AdaBoost algorithm model of the data-driven DC distribution network fault recovery method of the present invention, on the one hand, the training efficiency is proposed as the importance weight of the weak classifier, making the weight dynamic adjustment more flexible, adapting to different data sets or fault characteristics, and at the same time, it can also improve the convergence speed of the model in the initial stage of training, providing a more flexible adaptation ability in complex problems such as DC distribution network fault recovery; on the other hand, the algorithm introduces an objective function, which can be used as a constraint condition to improve the training efficiency, and the optimal weight is obtained by derivation, which can effectively avoid model overfitting. Description of the Drawings
[0091] Figure 1 It is the DC distribution topology adopted in Embodiment 1.
[0092] Figure 2 It is the flowchart of the method described in Embodiment 1.
[0093] Figure 3 It is the simulation model of the DC distribution network.
[0094] Figure 4 It is the structure diagram of the system described in Embodiment 2.
[0095] Figure 5 It is the structure diagram of the system described in Embodiment 3.
[0096] Figure 6 It is the fault diagnosis result of Embodiment 1. Detailed Embodiments
[0097] The present invention will be further described in detail below in conjunction with the description of the drawings and the detailed embodiments.
[0098] Embodiment 1:
[0099] This embodiment takes Figure 1 The shown DC distribution as the research object. In the topology of this DC distribution, an internal DC fault occurs on Line 1, including positive ground fault (PGF, F1), negative ground fault (NGF, F2) and short circuit fault (SCF, F3). The internal fault path occurs between the positive poles or between the positive / negative poles and the ground, and they have low / high resistance characteristics. Since the DC line fault is affected by the fault circuit parameters and control strategies, the DC line fault characteristics are diverse and complex. In addition, the external faults are external AC fault (EACF, F4) and external DC fault (EDCF, F5). The EACF occurs on the valve side of the AC transformer beside VSC 1, including single-phase ground fault (1GF), two-phase ground fault (2GF) and three-phase short circuit fault (3SF). The EDCF occurs on Line 2 and Line 4.
[0100] A fault recovery method for a data-driven DC distribution network is as follows: Figure 2 as shown below:
[0101] 1. Collect the current and voltage data of the DC distribution network in real time at time t. Determine that the distribution network has a fault based on the collected current and voltage data, and obtain the fault feature samples.
[0102] The fault data comes from the monitoring points (MP) on Line 1 near VSC 1, including internal faults (F1, F2, and F3) and external faults (F4 and F5), as Figure 3 shown below:
[0103] The parameter configuration of the fault data shows that there are 21 fault locations for internal faults. Since EACF includes 1GF (A-phase, B-phase, and C-phase), 2GF (AB-phase, BC-phase, and AC-phase), and 3SF (ABC-phase), EACF has seven fault locations. EDCF occurs on Line 2 and Line 4. EDCF includes external PGF, external NGF, and external SCF. Therefore, EDCF has 6 fault locations (2 lines * 3 fault types).
[0104] EACF occurs on the valve side of the AC Transformer next to VSC 1, including single-phase ground fault (1GF), two-phase ground fault (2GF), and three-phase short circuit fault (3SF). EDCF occurs on Line 2 and Line 4.
[0105] 2. Adopt a data expansion algorithm and perform data expansion on the fault feature samples through linear interpolation:
[0106]
[0107] In the above formula, is the p-th fault data of the d-type fault occurring at position c, which is a row vector, e is a number based on the fault impedance, is the g-th group of the G groups of fault data expanded between and , is 's adjacent fault data.
[0108] After calculation, the data expanded using this method is very close to the actual simulated fault data, with an average error of 0.013%.
[0109] 3. Filter out the fault features with low fault response sensitivity, including:
[0110] Calculate the sensitivity of each fault feature according to the following formula, and then remove the fault features with sensitivity less than the set threshold:
[0111]
[0112] In the above formula, J P is the fault response sensitivity of the fault feature X p , t 0 is the fault occurrence time, Δt is the sampling interval, n is the number of fault feature samples, and X p (t 0 ) is the fault feature X 0 at time t p .
[0113] 4. To reduce the computational complexity of fault detection and remove potential redundancy between multi-dimensional fault features, a comprehensive correlation coefficient is introduced. Calculate the comprehensive correlation coefficient of each fault feature respectively, and screen out the fault features whose comprehensive correlation coefficient is greater than the set threshold:
[0114]
[0115] In the above formula, d p is the comprehensive correlation coefficient of the p-th fault feature X p , c p,q is the Pearson correlation coefficient between the fault feature X p and X q , d is the number of fault features under each fault type, and b is the number of fault types.
[0116] In this embodiment, X11-X14 are screened out through steps 3 and 4. These fault features are sensitive, weakly coupled, and have strong fault detection capabilities. Therefore, X11-X14 are selected to construct the fault feature vector X:
[0117] X = [X11 X12 X13 X14].
[0118] 5. Use the improved curve smoothing algorithm based on Apollo to smooth the constructed fault feature vector X. Among them, the objective function of the improved curve smoothing algorithm based on Apollo is:
[0119] B = α 1 B 1 + α 2 B 2 + α 3 B 3
[0120] B 1 = ∑|2x p (k)-x p (k - 1)-x p (k + 1)|
[0121] B 2 = ∑|x p(k + 1)-x p (k)|
[0122] B 3 = ∑x p (k)-x p (k) ref |
[0123] In the above formula, B is the objective function, and B 1 , B 2 , B 3 are respectively the improved smoothness cost function, length cost function, and deviation cost function. α 1 , α 2 , α 3 are respectively the weight coefficients of B 1 , B 2 , B 3 . x p (k) is the k-th sample value, and x p (k) ref is the optimal solution of the k-th sample.
[0124] After the smoothing process of this step, the maximum and average errors of the current curve are 8.12% and 1.64% respectively.
[0125] 6. To improve the characterization ability of fault features and eliminate the influence of the order of magnitude, the smoothed fault feature vector X is normalized.
[0126] 7. The normalized fault feature samples are input into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type. Among them, the improved adaboost algorithm model uses a BP neural network as a weak classifier, and its training process includes:
[0127] 7.1. Initialize the parameters. Among them, the weights of each training sample are initialized to 1 / M, and M is the number of training samples.
[0128] 7.2. Start the first round of iteration. First, train a weak classifier based on the current sample weights, and then use the training efficiency determined by the following formula as the importance weight of this weak classifier, and update the weights of each training sample:
[0129]
[0130] In the above formula, β t is the training efficiency of the t-th round of iteration, η is the learning rate, and φ is the convergence target, which is set to 0.01 in this embodiment.
[0131] 7.3. Repeat 7.2 in a loop until the training of the target number of weak classifiers is completed.
[0132] 7.4. Combine each weak classifier into a strong classifier to obtain the output of the model. Among them, the output function H(X) is as follows:
[0133]
[0134] In the above formula, G t (X) and W t are respectively the output and weight of the t-th weak separator, and N is the number of weak classifiers.
[0135] Example 2:
[0136] The steps are the same as those in the example, and the differences are as follows:
[0137] In step 7.4, W t is calculated by the following method:
[0138] Expand and take the derivative of the following objective function, and take W when the derivative is zero t :
[0139]
[0140] In the above formula, y i is the i-th training sample, and H i-1 (X) is the output of the (i - 1)-th training sample obtained through training.
[0141] Example 3:
[0142] A fault recovery system for a data-driven DC distribution network, as Figure 4 shown, includes a data acquisition and fault judgment module, a sample preprocessing module, a fault identification module, and a power grid load dispatching system;
[0143] The data acquisition and fault judgment module is used for:
[0144] Real-time collect the current and voltage data of the DC distribution network at time t, and determine whether the distribution network has a fault according to the collected current and voltage data. If so, activate the sample preprocessing module;
[0145] After restoring power supply in the non-fault area, judge whether the fault has been restored. If not, collect the current and voltage data of the DC distribution network at time t + 1 and enter the next cycle.
[0146] The sample preprocessing module is used to obtain fault feature samples and preprocess the fault feature samples. The preprocessing includes:
[0147] A1. Calculate the sensitivity of each fault feature according to the following formula, and remove the fault features with sensitivity less than the set threshold:
[0148]
[0149] In the above formula, J P is the fault response sensitivity of the fault feature X p , t 0 is the fault occurrence time, Δt is the sampling interval, n is the number of fault feature samples, and X p (t 0 ) is the fault feature X 0 at time t p .
[0150] A2. Calculate the comprehensive correlation coefficients of each fault feature after being processed by A1 respectively, and filter out the fault features whose comprehensive correlation coefficients are greater than the set threshold:
[0151]
[0152] In the above formula, d p is the comprehensive correlation coefficient of the p-th fault feature X p , c p,q is the Pearson correlation coefficient between the fault feature X p and X q , d is the number of fault features under each fault type, and b is the number of fault types.
[0153] The fault identification module is used to identify the fault point and fault type based on the preprocessed fault feature samples by using an improved adaboost algorithm model, and its operation process includes:
[0154] B1. Smooth the preprocessed fault feature samples by using an improved curve smoothing algorithm based on Apollo. Among them, the objective function of the improved curve smoothing algorithm based on Apollo is:
[0155] B = α 1 B 1 + α 2 B 2 + α 3 B 3
[0156] B 1 = ∑|2x p (k) - x p (k - 1) - x p (k + 1)|
[0157] B 2 = ∑|x p (k + 1) - x p (k)|
[0158] B 3 = ∑|x p(k)-x p (k) ref |
[0159] In the above formula, B is the objective function, B 1 、B 2 、B 3 are respectively the improved smoothness cost function, length cost function and deviation cost function, α 1 、α 2 、α 3 are respectively the weight coefficients of B 1 、B 2 、B 3 , x p (k) is the k-th sample value, x p (k) ref is the optimal solution of the k-th sample;
[0160] B2. Normalize the fault feature samples after smoothing;
[0161] B3. Input the normalized fault feature samples into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type. Among them, the improved adaboost algorithm model uses a BP neural network as a weak classifier, and its training process includes:
[0162] B31. Initialize the parameters. Among them, the weights of each training sample are initialized to 1 / M, and M is the number of training samples;
[0163] B32. Start the first round of iteration. First, train a weak classifier based on the current sample weights, and then use the training efficiency determined by the following formula as the importance weight of this weak classifier, and update the weights of each training sample:
[0164]
[0165] In the above formula, β t is the training efficiency of the t-th round of iteration, η is the learning rate, and φ is the convergence target;
[0166] B33. Repeat B32 in a loop until the training of the target number of weak classifiers is completed;
[0167] B34. Combine each weak classifier into a strong classifier to obtain the output of the model. Among them, the output function H(X) is as follows:
[0168]
[0169] In the above formula, G t (X), W t are respectively the output and weight of the t-th weak separator, and N is the number of weak classifiers.
[0170] The power grid load dispatching system is used to isolate the fault area according to the fault point and fault type, and restore power supply to the non-fault area by adjusting the load of the power grid.
[0171] Embodiment 4:
[0172] The structure is the same as that of Embodiment 3, except that:
[0173] As Figure 5 shown, the system further includes a data expansion module, and the data expansion module is used to perform data expansion on the fault feature samples by linear interpolation according to the following formula before preprocessing the fault feature samples:
[0174]
[0175] In the above formula, is the p-th fault data of the d-type fault occurring at position c at time k, which is a row vector, and e is a number based on the fault impedance. is between and is adjacent fault data.
[0176] The calculation method of the described W t is:
[0177] Expand and take the derivative of the following objective function, and take W t when the derivative is zero:
[0178]
[0179] In the above formula, y i is the i-th training sample, and H i-1 (X) is the output of the (i - 1)-th training sample obtained by training.
[0180] The fault recognition result of the method of the present invention is as Figure 6 shown. It can be seen that the detection rate of the method for internal DC faults reaches 100%. Although some external fault samples are not correctly identified as the corresponding external fault types, they are still determined to be external faults. Therefore, the method has good effects on distinguishing internal DC faults from external faults and selecting fault types.
Claims
1. A data driven DC distribution network fault recovery method, characterized in that: The method comprises: S1, real-time collection of current and voltage data of the DC distribution network during period t, and judging whether the distribution network has a fault based on the collected current and voltage data, if so, proceed to S2; S2. Obtain fault feature samples and preprocess the fault feature samples; S3, based on the preprocessed fault feature samples, the improved adaboost algorithm model is used to identify the fault point and fault type; S4. The power grid load dispatching system isolates the fault area according to the fault point and fault type, and restores power supply to the non-fault area by adjusting the load of the power grid; S5. Determine whether the fault is restored. If not, return to S1 to collect the current and voltage data of the DC distribution network during the period t+1 and enter the next cycle.
2. A data driven DC distribution network fault recovery method according to claim 1, characterized in that: In S2, the preprocessing includes: S21. Calculate the fault response sensitivity of each fault feature according to the following formula, and remove the fault features with a sensitivity less than a set threshold: In the above formula, J P is the fault feature X p The fault response sensitivity of X is: t0 is the time when the fault occurs, Δt is the sampling interval, n is the number of fault feature samples, and X is the fault response sensitivity of X. p (t0) is the fault feature X at time t0 p ; S21. Calculate the comprehensive correlation coefficient of each fault feature processed by S21 respectively, and select the fault features whose comprehensive correlation coefficient is greater than the set threshold: In the above formula, d p is the pth fault feature X p The comprehensive correlation coefficient, c p,q is the fault feature X p With X q The Pearson correlation coefficient of , d is the number of fault features under each fault type, and b is the number of fault types.
3. A data driven DC distribution network fault recovery method according to claim 1 or 2, characterized in that: The S3 includes: S31, using an improved curve smoothing algorithm based on Apollo to smooth the preprocessed fault feature samples, wherein the objective function of the improved curve smoothing algorithm based on Apollo is: B=α1B1+α2B2+α3B3 B1=∑|2x p (k)-x p (k-1)-x p (k+1)| B2=∑|x p (k+1)-x p (k)| B3=∑|x p (k)-x p (k) ref | In the above formula, B is the objective function, B1, B2, B3 are the improved smoothness cost function, length cost function and deviation cost function respectively, α1, α2, α3 are the weight coefficients of B1, B2, B3 respectively, and x p (k) is the kth sample value, x p (k) ref is the optimal solution for the kth sample; S32, normalizing the smoothed fault feature samples; S33, input the normalized fault feature samples into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type, wherein the improved adaboost algorithm model uses BP neural network as a weak classifier.
4. A data driven DC distribution network fault recovery method according to claim 3, characterized in that: The training process of the improved adaboost algorithm model includes: S331, initializing parameters, wherein the weight of each training sample is initialized to 1 / M, where M is the number of training samples; S332, start the first round of iteration, first train a weak classifier based on the current sample weight, then use the training efficiency determined by the following formula as the importance weight of the weak classifier, and update the weight of each training sample: In the above formula, β t is the training efficiency of the tth iteration, η is the learning rate, and φ is the convergence target; S333, repeat S332 in a loop until the training of the target number of weak classifiers is completed; S334, combining each weak classifier into a strong classifier, thereby obtaining the output of the model, wherein the output function H(X) is as follows: In the above formula, G t (X), W t are the output and weight of the tth weak separator respectively, and N is the number of weak classifiers.
5. A data driven DC distribution network fault recovery method according to claim 4, characterized in that: The W t The calculation method is: Expand the following objective function and take the derivative W when it is zero. t : In the above formula, y i is the i-th training sample, H i-1 (X) is the output of the i-1th training sample obtained during training.
6. A data driven DC distribution network fault recovery method according to claim 1 or 2, characterized in that: In S2, before preprocessing the fault feature samples, data expansion is performed on the fault feature samples by linear interpolation according to the following formula: In the above formula, is the pth fault data of a d-type fault occurring at position c at time k, which is a row vector, e is a number based on the fault impedance, For and The gth group of G group fault data extended between for Adjacent fault data.
7. A data driven DC distribution network fault recovery system, characterized in that: The system includes a data acquisition and fault judgment module, a sample preprocessing module, a fault identification module, and a power grid load dispatching system; The data collection and fault diagnosis module is used for: Collect the current and voltage data of the DC distribution network in time period t in real time, and determine whether the distribution network is faulty based on the collected current and voltage data. If so, activate the sample preprocessing module; After power supply is restored in the non-fault area, it is determined whether the fault has been restored. If not, the current and voltage data of the DC distribution network in the period t+1 are collected to enter the next cycle. The sample preprocessing module is used to obtain fault feature samples and preprocess the fault feature samples; The fault identification module is used to identify the fault point and fault type based on the preprocessed fault feature samples using an improved adaboost algorithm model; The power grid load dispatching system is used to isolate the fault area according to the fault point and the fault type, and restore power supply to the non-fault area by adjusting the load of the power grid.
8. A data driven DC distribution network fault recovery system according to claim 7, characterized in that: The pre-processing comprises: A1. Calculate the sensitivity of each fault feature according to the following formula, and remove the fault features with a sensitivity less than the set threshold: In the above formula, J P is the fault feature X p The fault response sensitivity of X is: t0 is the time when the fault occurs, Δt is the sampling interval, n is the number of fault feature samples, and X is the fault response sensitivity of X. p (t0) is the fault feature X at time t0 p . A2. Calculate the comprehensive correlation coefficient of each fault feature after processing by A1, and select the fault features whose comprehensive correlation coefficient is greater than the set threshold: In the above formula, d p is the pth fault feature X p The comprehensive correlation coefficient, c p,q is the fault feature X p With X q The Pearson correlation coefficient of , d is the number of fault features under each fault type, and b is the number of fault types.
9. A data driven DC distribution network fault recovery system according to claim 7, characterized in that: The fault identification module identifies the fault point and the fault type based on the following steps: B1. The improved curve smoothing algorithm based on Apollo is used to smooth the preprocessed fault feature samples. The objective function of the improved curve smoothing algorithm based on Apollo is: B=α1B1+α2B2+α3B3 B1=∑|2x p (k)-x p (k-1)-x p (k+1)| B2=∑|x p (k+1)-x p (k)| B3=∑x p (k)-x p (k) ref In the above formula, B is the objective function, B1, B2, B3 are the improved smoothness cost function, length cost function and deviation cost function respectively, α1, α2, α3 are the weight coefficients of B1, B2, B3 respectively, and x p (k) is the kth sample value, x p (k) ref is the optimal solution for the kth sample; B2, normalize the fault feature samples after smoothing; B3. Input the normalized fault feature samples into the improved adaboost algorithm model for training and prediction, so as to identify the fault point and fault type. The improved adaboost algorithm model uses BP neural network as a weak classifier, and its training process includes: B31, initialization parameters, where the weight of each training sample is initialized to 1 / M, where M is the number of training samples; B32. Start the first round of iteration. First, train a weak classifier based on the current sample weight. Then use the training efficiency determined by the following formula as the importance weight of the weak classifier and update the weight of each training sample: In the above formula, β t is the training efficiency of the tth iteration, η is the learning rate, and φ is the convergence target; B33, repeat B32 cyclically until the training of the target number of weak classifiers is completed; B34. Combine the weak classifiers into a strong classifier to obtain the output of the model, where the output function H(X) is as follows: In the above formula, G t (X), W t are the output and weight of the tth weak separator, respectively, and N is the number of weak classifiers; The W t The calculation method is: Expand the following objective function and take the derivative W when it is zero. t : In the above formula, y i is the i-th training sample, H i-1 (X) is the output of the i-1th training sample obtained during training.
10. A data driven DC distribution network fault recovery system according to claim 7 or 8, characterized in that: The system further includes a data expansion module, which is used to perform data expansion on the fault feature samples by linear interpolation according to the following formula before preprocessing the fault feature samples: In the above formula, is the pth fault data of a d-type fault occurring at position c at time k, which is a row vector, e is a number based on the fault impedance, For and The gth group of G group fault data extended between for Adjacent fault data.