Fault identification method for sending line of double-fed wind farm based on multi-feature combination
By constructing voltage and current feature vectors for doubly-fed wind turbines and using a fully connected neural network to identify faults in the transmission lines of doubly-fed wind farms, the problem of difficulty in identifying faults using traditional methods in new power systems is solved, achieving fault identification with high accuracy and wide applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU POWER SUPPLY BUREAU GUANGDONG POWER GRID CO LTD
- Filing Date
- 2022-11-30
- Publication Date
- 2026-06-09
AI Technical Summary
In existing technologies, traditional power systems have difficulty accurately identifying faults in the transmission lines of doubly fed wind farms. In particular, in new power systems, traditional current characteristic phase selection methods are not applicable, voltage quantity phase selection sensitivity is insufficient, and the fault characteristics of deep learning algorithms are unclear and have insufficient generalization ability.
By constructing a simulation model of a double-fed induction generator (DFIG) and a synchronous generator for a double-ended transmission line, multidimensional fault feature vectors of voltage and current are extracted. A three-layer fully connected neural network is used for iterative training. By combining features such as voltage drop, voltage wavelet energy, and zero-sequence current component, the fault type is identified.
It improves the accuracy and generalization ability of fault identification, and can correctly identify faults in the doubly fed wind turbine output line in various scenarios, especially when the voltage change of high resistance faults is not obvious.
Smart Images
Figure CN115796835B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of new power technology, specifically, it relates to a method for identifying faults in the outgoing lines of doubly fed wind farms based on the combination of multiple features. Background Technology
[0002] In recent years, with the large-scale integration of power electronic equipment into the power grid and the gradual increase in the proportion of new energy sources, the fault characteristics of power systems have become increasingly complex. Taking doubly-fed induction generators (DFIGs) as an example, because DFIGs belong to weakly fed systems, their short-circuit current amplitude is not large. Furthermore, due to the influence of transformer wiring, the current at the protection installation point of the DFIG wind farm's transmission line is mainly zero-sequence, and the fault current amplitude characteristics and phase-to-phase differences differ from those of traditional synchronous generators. In addition, the wind turbine's control strategy also affects the fault current characteristics. For example, when a grid fault occurs, the wind turbine starts in low-voltage ride-through mode, and the current transmitted by the wind power also includes a frequency component determined by the current wind speed. Therefore, it is necessary to analyze and appropriately select the fault characteristics of DFIGs to accurately classify faults in the DFIG wind farm's transmission line, providing a basis for the reliable operation of protection devices and improving the stability of the new power system.
[0003] Currently, there are few methods for fault classification in new power systems. Traditional power systems typically use sudden changes in current or voltage and sequence components for phase selection. However, due to factors such as weak feedforward of new energy sources and power electronic control strategies, the current-based phase selection method is no longer applicable in traditional power systems, and voltage-based phase selection also faces the problem of insufficient sensitivity.
[0004] Currently popular deep learning algorithms can achieve feature extraction and fit complex input-output mapping relationships. However, due to their "black box" nature, the extracted fault features are not meaningful, and their generalization ability needs improvement. Therefore, fault feature analysis for specific problems, selecting prominent indicator inputs, and using neural networks as classifiers is an efficient and reliable method. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies and provide a fault identification method for doubly fed wind farm transmission lines based on multi-feature combination. This invention utilizes voltage and current information, and the extracted fault features have good discriminative power, thereby improving the classification accuracy and generalization ability of the model.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A method for fault identification of doubly-fed induction generator (DFIG) wind farm transmission lines based on multi-feature combination includes the following steps:
[0008] S1. Build a simulation model of the double-ended transmission line of the doubly-fed induction generator (DFIG) and synchronous machine, set different fault conditions and generate simulation fault data, label the fault types according to the simulation fault data, and obtain the voltage simulation fault data and current simulation fault data of the DFIG.
[0009] S2. Based on the voltage simulation fault data and current simulation fault data of the doubly fed wind turbine, extract the fault voltage feature vector and the fault current feature vector, and form a multi-dimensional fault feature vector;
[0010] S3. Construct a neural network for fault identification. Input the multidimensional fault feature vector into the neural network and iteratively train the neural network until the error between the output of the neural network and the label tends to stabilize, thus obtaining a trained neural network.
[0011] S4. Obtain real-time fault data of the doubly fed wind turbine, extract fault voltage feature vector and fault current feature vector, and form a multi-dimensional fault feature vector. Input the vector into a neural network, and the neural network outputs the fault type probability. The type with the highest probability is determined as the fault category.
[0012] Further, in step S1, the three-phase voltage, three-phase current and zero-sequence current data of the doubly-fed induction generator (DFIG) are extracted from one cycle before the fault to three cycles after the fault to obtain the voltage simulation fault data and current simulation fault data of the DFIG.
[0013] Furthermore, in step S2, the fault voltage feature vector includes the voltage drop degree, voltage wavelet energy, and voltage modulus maxima, and the fault current feature vector includes the zero-sequence component, current energy, and cosine similarity.
[0014] Furthermore, both the voltage sag and current energy are calculated using the second moment. The formula for calculating the voltage sag is:
[0015]
[0016] In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This represents the k-th sampled value of the X-phase voltage before the fault. This is the kth sampled value of the X-phase voltage after the fault.
[0017] Furthermore, the voltage wavelet energy and voltage modulus maxima are both derived through wavelet transform. The db4 wavelet is selected as the wavelet basis function, and the coefficients of the first and second layers are extracted. The second moment and maximum value of the coefficients within the first quarter cycle after the fault are calculated as part of the input features.
[0018] Furthermore, current cosine similarity can depict the distortion of fault current and is suitable for fault currents with strong randomness on the doubly-fed induction generator (DFIG) side. The formula for calculating current cosine similarity is:
[0019]
[0020] In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This is the k-th sampled value of the X-phase current before the fault. This is the kth sampled value of the X-phase current after the fault.
[0021] Furthermore, in step S3, the neural network is a three-layer fully connected neural network. The training of the neural network adopts the backpropagation algorithm. The objective function of the neural network is the error between the output value and the label. The backpropagation algorithm calculates the partial derivative of the objective function with respect to each neuron layer by layer to form a gradient. Based on this gradient, the weight values of each neuron are gradually modified until the error between the output and the label is not large or the training cycle is completed.
[0022] Furthermore, the gradient calculation formula for the connection weights is as follows:
[0023]
[0024] Where, ω lm The weights connecting neuron m in layer j and neuron l in layer j+1, The loss function of a neural network, Loss with respect to ω lm Finding the partial derivative yields the gradient information we're looking for. k-1 For the output of the (k-1)th hidden layer, net k For the output of the k-th hidden layer, net j+1 For the output of the (j+1)th hidden layer, net j For the output of the j-th hidden layer, o m is the input value of the m-th layer network;
[0025] After the gradient is calculated, update the weights according to the learning rate:
[0026]
[0027] Where η is the learning rate, representing the rate of weight gradient descent, and the learning rate η ranges from 0.0001 to 0.1.
[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0029] 1. This invention extracts key features of fault voltage and fault current, and imports them into a fully connected neural network to realize fault identification of the transmission line of a doubly fed wind farm. The fault voltage is used as the main feature, which is not affected by weak wind power supply or frequency fluctuations. The fault current is used as an auxiliary feature to ensure that the system can still be correctly identified when the high-resistance fault voltage change is not obvious.
[0030] 2. This invention uses multidimensional fault feature vectors as input to a fully connected neural network. The input has a clear physical meaning. Fully connected neural networks are simpler to train than deep neural networks and have excellent classification and recognition effects.
[0031] 3. The fault features extracted by this invention are the general features of doubly fed wind turbine faults, and are not limited to a specific model. Faults in the output line of doubly fed wind turbines can be effectively identified in various scenarios. Attached Figure Description
[0032] Figure 1 This is a flowchart illustrating the fault identification method for doubly-fed wind farm transmission lines based on multi-feature combination according to the present invention.
[0033] Figure 2 This is a schematic diagram illustrating the effects of fault voltage and its corresponding characteristics.
[0034] Figure 3 This is a schematic diagram illustrating the effects of fault current and its corresponding characteristics.
[0035] Figure 4 This is a schematic diagram of the structure of a fully connected neural network. Detailed Implementation
[0036] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the present invention's method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination.
[0037] Please see Figure 1 This invention discloses a method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination, comprising the following steps:
[0038] S1. Construct a simulation model of the double-fed induction generator (DFIG) and synchronous generator (SEG) double-ended transmission line, set different fault conditions and generate simulation fault data, label the fault types based on the simulation fault data, and obtain the voltage and current simulation fault data of the DFIG.
[0039] S2. Based on the voltage simulation fault data and current simulation fault data of the doubly fed wind turbine, extract the fault voltage feature vector and the fault current feature vector, and form a multi-dimensional fault feature vector.
[0040] S3. Construct a neural network for fault identification. Input the multidimensional fault feature vector into the neural network and iteratively train the neural network until the error between the output of the neural network and the label tends to stabilize, thus obtaining a trained neural network.
[0041] S4. Obtain real-time fault data of the doubly fed wind turbine, extract fault voltage feature vector and fault current feature vector, and form a multi-dimensional fault feature vector. Input the vector into a neural network, and the neural network outputs the fault type probability. The type with the highest probability is determined as the fault category.
[0042] Specifically, in step S1, the three-phase voltage, three-phase current and zero-sequence current data of the doubly fed induction generator (DFIG) are extracted from one cycle before the fault to three cycles after the fault to obtain the voltage simulation fault data and current simulation fault data of the DFIG.
[0043] In step S2, the fault voltage feature vector includes the voltage drop degree, voltage wavelet energy, and voltage modulus maxima, while the fault current feature vector includes the zero-sequence component, current energy, and cosine similarity.
[0044] The voltage sag and current energy are both calculated using the second moment. The formula for calculating the voltage sag is:
[0045]
[0046] In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This represents the k-th sampled value of the X-phase voltage before the fault. This is the kth sampled value of the X-phase voltage after the fault.
[0047] The voltage wavelet energy and voltage modulus maxima are both derived from wavelet transform. The db4 wavelet is selected as the wavelet basis function. The coefficients of the first and second layers are extracted, and the second moment and maximum value of the coefficients within the first quarter cycle after the fault are calculated as part of the input features.
[0048] Current cosine similarity can depict the distortion of fault current and is suitable for fault currents with strong randomness on the doubly-fed induction generator (DFIG) side. The formula for calculating current cosine similarity is:
[0049]
[0050] In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This is the k-th sampled value of the X-phase current before the fault. This is the kth sampled value of the X-phase current after the fault.
[0051] like Figure 2 As shown, a 15-dimensional fault voltage feature vector, including voltage drop degree, voltage wavelet energy, and voltage modulus maxima, is extracted from three-phase voltage simulation fault data; for example... Figure 3 As shown, a total of 22-dimensional fault feature vectors are formed by extracting the zero-sequence component, post-fault current energy, and cosine similarity from the three-phase current simulation fault data.
[0052] In step S3, the neural network is a three-layer fully connected neural network. The training of the neural network adopts the backpropagation algorithm. The objective function of the neural network is the error between the output value and the label. The backpropagation algorithm calculates the partial derivative of the objective function with respect to each neuron layer by layer to form a gradient. Based on this gradient, the weight values of each neuron are gradually modified until the error between the output and the label is small or the training cycle is completed.
[0053] The gradient calculation formula for the connection weights is:
[0054]
[0055] Where, ω lm The weights connecting neuron m in layer j and neuron l in layer j+1, The loss function of a neural network, Loss with respect to ω lm Finding the partial derivative yields the gradient information we're looking for. k-1 For the output of the (k-1)th hidden layer, net k For the output of the k-th hidden layer, net j+1 For the output of the (j+1)th hidden layer, net j For the output of the j-th hidden layer, o m is the input value of the m-th layer network.
[0056] After the gradient is calculated, update the weights according to the learning rate:
[0057]
[0058] Where η is the learning rate, representing the rate of weight gradient descent, and the learning rate η ranges from 0.0001 to 0.1.
[0059] A three-layer fully connected neural network was constructed as the classifier, with 64, 32, and 10 neurons in each layer. The fully connected neural network was trained using a 22-dimensional fault feature vector with a learning rate of 0.005 and 40 training epochs. The input x to the fully connected neural network was a 22-dimensional feature vector, and the output label was... This represents the probability value of the predicted fault type out of 10 categories.
[0060] In step S4, the trained fully connected neural network can be used to identify ten fault types in the transmission lines of a doubly-fed wind farm. The three-phase voltage and current data measured for the fault are used to form a 22-dimensional fault feature vector according to step S2, and this vector is input into the fully connected neural network. The class with the highest probability in the output y of the fully connected neural network is the fault category. For example, as... Figure 4As shown, if the label output by the fully connected neural network is y = [0.01, 0.05, 0.7, 0.02, 0.03, 0.01, 0.02, 0.0, 0.1, 0.06], and the maximum probability is 0.7, then the fault is determined to be a C-phase ground fault corresponding to the maximum probability value.
[0061] In summary, the present invention has the following advantages and beneficial effects:
[0062] 1. This invention extracts key features of fault voltage and fault current, and imports them into a fully connected neural network to realize fault identification of the transmission line of a doubly fed wind farm. The fault voltage is used as the main feature, which is not affected by weak wind power supply or frequency fluctuations. The fault current is used as an auxiliary feature to ensure that the system can still be correctly identified when the high-resistance fault voltage change is not obvious.
[0063] 2. This invention uses multidimensional fault feature vectors as input to a fully connected neural network. The input has a clear physical meaning. Fully connected neural networks are simpler to train than deep neural networks and have excellent classification and recognition effects.
[0064] 3. The fault features extracted by this invention are the general features of doubly fed wind turbine faults, and are not limited to a specific model. Faults in the output line of doubly fed wind turbines can be effectively identified in various scenarios.
[0065] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit disclosed in the present invention should fall within the patent scope covered by the present invention.
Claims
1. A method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination, characterized in that, Includes the following steps: S1. Build a simulation model of the double-ended transmission line of the doubly-fed induction generator (DFIG) and synchronous machine, set different fault conditions and generate simulation fault data, label the fault types according to the simulation fault data, and obtain the voltage simulation fault data and current simulation fault data of the DFIG. S2. Based on the voltage simulation fault data and current simulation fault data of the doubly fed wind turbine, extract the fault voltage feature vector and the fault current feature vector, and form a multi-dimensional fault feature vector; S3. Construct a neural network for fault identification. Input the multidimensional fault feature vector into the neural network and iteratively train the neural network until the error between the output of the neural network and the label tends to stabilize, and obtain a trained neural network. S4. Obtain real-time fault data of the doubly fed wind turbine, extract fault voltage feature vector and fault current feature vector, and form a multi-dimensional fault feature vector. Input the vector into a neural network, and the neural network outputs the fault type probability. The type with the highest probability is determined as the fault category. In step S2, the fault voltage feature vector includes voltage drop degree, voltage wavelet energy, and voltage modulus maxima, and the fault current feature vector includes zero-sequence component, current energy, and cosine similarity. The voltage sag and current energy are both calculated using the second moment. The formula for calculating the voltage sag is: ； In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This represents the k-th sampled value of the X-phase voltage before the fault. This is the kth sampled value of the X-phase voltage after the fault.
2. The method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination as described in claim 1, characterized in that, In step S1, the three-phase voltage, three-phase current and zero-sequence current data of the doubly-fed induction generator (DFIG) are extracted from one cycle before the fault to three cycles after the fault to obtain the voltage simulation fault data and current simulation fault data of the DFIG.
3. The method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination as described in claim 1, characterized in that, The voltage wavelet energy and voltage modulus maxima are both derived from wavelet transform. The db4 wavelet is selected as the wavelet basis function. The coefficients of the first and second layers are extracted, and the second moment and maximum value of the coefficients within the first quarter cycle after the fault are calculated as part of the input features.
4. The method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination as described in claim 1, characterized in that, Current cosine similarity can depict the distortion of fault current and is suitable for fault currents with strong randomness on the doubly-fed wind farm side. The formula for calculating current cosine similarity is: ； In the formula, n is the number of sampling points in one cycle, and X represents the three phases A, B, and C. This represents the k-th sampled value of the X-phase current before the fault. This is the kth sampled value of the X-phase current after the fault.
5. The method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination according to claim 1, characterized in that, In step S3, the neural network is a three-layer fully connected neural network. The training of the neural network adopts the backpropagation algorithm. The objective function of the neural network is the error between the output value and the label. The backpropagation algorithm calculates the partial derivative of the objective function with respect to each neuron layer by layer to form a gradient. Based on this gradient, the weight values of each neuron are gradually modified until the error between the output and the label is small or the training cycle is completed.
6. The method for fault identification of doubly-fed wind farm transmission lines based on multi-feature combination according to claim 5, characterized in that, The gradient calculation formula for the connection weights is: ； in, The weights connecting neuron m in layer j and neuron l in layer j+1, The loss function of a neural network is represented by the loss pair. Finding the partial derivative yields the gradient information we seek. This is the output of the (k-1)th hidden layer. This is the output of the k-th hidden layer. This is the output of the (j+1)th hidden layer. This is the output of the j-th hidden layer. is the input value of the m-th layer network; After the gradient is calculated, update the weights according to the learning rate: ； in, The learning rate represents the rate at which the weight gradient descent occurs. Take a value between 0.0001 and 0.1.