Data-driven short-circuit current interval prediction method for grid-connected wind power systems

Abstract

The application provides a data-driven short-circuit current interval prediction method for a fan grid-connected power transmission system, and belongs to the technical field of power systems, and comprises the following steps: S1, obtaining operation parameters of the fan grid-connected power transmission system; S2, extracting operation characteristics of the operation parameters, inputting the extracted operation characteristics into a trained BP neural network, and outputting a short-circuit current when a short-circuit fault occurs to the current operation parameters of the fan grid-connected power transmission system by the BP neural network; and S3, inputting the short-circuit current into a short-circuit current interval prediction model, and outputting a prediction interval of the short-circuit current by the short-circuit current interval prediction model. The short-circuit current of the fan grid-connected power transmission system can be accurately predicted based on the BP neural network, and then the prediction interval of the short-circuit current is obtained through the short-circuit current interval prediction model, so that the confidence of the short-circuit current prediction result is given, and the reliability of the short-circuit current prediction result is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a data-driven short-circuit current interval prediction method for a wind turbine grid-connected power transmission system. BACKGROUND

[0002] With the large number of large-capacity generator units and power transformation equipment in use, the continuous increase of power load, and the access of new energy such as wind power and nuclear power, the power grid in China has entered a new period of rapid development. The emergence of load center large power plants and the interconnection between large power systems make the entire power grid present large-scale and complex heterogeneous characteristics, and at the same time, various problems also appear, one of which is that the level of system short-circuit current irregularly changes in size and distribution with the development of the power grid. Therefore, it is increasingly urgent to calculate and predict the short-circuit current in real time and quickly according to the increase of system power load and the change of natural conditions. However, after the high proportion of new energy is connected to the grid, the traditional short-circuit current calculation method gradually loses accuracy, and the simulation-based short-circuit current calculation method is time-consuming, and the existing short-circuit current prediction method cannot give the confidence of the prediction result, so the reliability of the prediction result is not high. SUMMARY

[0003] Therefore, the present application provides a data-driven short-circuit current interval prediction method for a wind turbine grid-connected power transmission system to give the confidence of the short-circuit current prediction result and improve the reliability of the short-circuit current prediction.

[0004] The technical solution adopted by the present application to solve its technical problems is:

[0005] A data-driven short-circuit current interval prediction method for a wind turbine grid-connected power transmission system, comprising the following steps:

[0006] S1, obtaining the operating parameters of the wind turbine grid-connected power transmission system;

[0007] S2, extracting the operating characteristics of the operating parameters, and inputting the extracted operating characteristics into a trained BP neural network, and the BP neural network outputs the short-circuit current when the wind turbine grid-connected power transmission system current operating parameters have a short-circuit fault;

[0008] S3, inputting the short-circuit current into a short-circuit current interval prediction model, and the short-circuit current interval prediction model outputs the prediction interval of the short-circuit current.

[0009] Preferably, before step S1, the method further comprises:

[0010] The wind turbine grid-connected power transmission system includes a doubly-fed wind turbine generator system equipped with a rotor Crowbar circuit, the response stage of the doubly-fed wind turbine generator system when a short-circuit fault occurs is analyzed, and a short-circuit current calculation formula of the response stage is determined;

[0011] A simulation model of a wind turbine grid-connected power transmission system is established, and different system operation modes are set to generate a large number of simulation samples;

[0012] A BP neural network is established, and the simulation samples and historical data are divided into a training set, a validation set and a test set according to a proportion;

[0013] The BP neural network is iteratively trained using the training set, and the network performance is verified using the validation set after each iteration;

[0014] When the iterative training is completed and the network performance is optimized, the accuracy of the BP neural network calculation is tested by the test set, and the trained BP neural network is obtained.

[0015] Preferably, the analysis of the response stage of the doubly-fed wind turbine power generation system when a short-circuit fault occurs includes:

[0016] The response of the doubly-fed wind turbine power generation system when a short-circuit fault occurs includes two stages, the first stage is a sub-transient operation stage, and the second stage is a Crowbar protection system triggering stage.

[0017] Preferably, the short-circuit current calculation formula of the response stage includes:

[0018] The first stage is the sub-transient operation stage, and the doubly-fed wind turbine power generation system is regarded as an induction generator operating, and the model of the induction generator is represented as a transient voltage source followed by a series of stator transient reactance, and its specific expression is as follows:

[0019]

[0020] In the formula, the subscripts {s, r} represent the electrical quantities of the generator stator winding and the rotor winding, respectively, and the subscripts {d, q} represent the d-axis electrical quantity and the q-axis electrical quantity in the dq synchronous coordinate system, respectively, L ls and L lr are the stator leakage inductance and the rotor leakage inductance, respectively, and L m is the mutual inductance of the stator and the rotor;

[0021] The stator short-circuit current of the induction generator can be obtained by the following formula:

[0022]

[0023] In the formula:

[0024]

[0025] The second stage is the Crowbar protection system triggering stage, and if a short-circuit occurs at the terminals of the doubly-fed wind turbine power generation system, the maximum short-circuit current is about:

[0026]

[0027] Where R cb is the resistance value of the Crowbar resistor.

[0028] Preferably, different system operation modes are set to generate a large number of simulation samples, including:

[0029] Select the wind turbine grid-connected power transmission system balancing synchronous machine, select the random switching line, set the basic operation mode of the wind turbine grid-connected power transmission system and set the total number of generated simulation samples N, where the active output of the synchronous machine is P Gi,base (i=1,2,…,m), the load consumes active power P loadj,base (j=1,2,…,n), the inductive reactive power absorbed by the load is Q loadj,base (j=1,2,…,n);

[0030] Randomly generate the active output of the kth group of synchronous machines and the load power consumption, where k≤N;

[0031] Randomly generate the kth group of wind speeds and set the response wind speed scenario;

[0032] Randomly generate the kth group of line switching conditions;

[0033] Set the kth group of operation modes and lines in the simulation system and initialize the power flow;

[0034] Set the fan to be off and calculate the short-circuit current I through simulation. f0 ;

[0035] Set the fan input and calculate the short-circuit current I through simulation f ;

[0036] Repeat the above steps until N groups of simulation samples are obtained.

[0037] Preferably, the kth group of operating modes is set in the simulation system as follows:

[0038] P Gi,k =k Gi,k ·P Gi ,i=1,2,...,n

[0039]

[0040] Where, P Gi,k is the active power output of the i-th synchronous machine under the k-th operation, P loadj,k is the active power consumed by the jth load under the kth operation, Q loadj,k is the inductive reactive power absorbed by the jth load under the kth operation; k Gi,k 、k PLj,k 、kQLj,k are independent random quantities in [0.8, 1.2], and the distribution is uniform, so that each type of operation mode has a probability and the same possibility of occurrence.

[0041] Preferably, the accuracy of the BP neural network calculation is tested by a test set, including:

[0042] The accuracy of the BP neural network calculation is tested by the mean absolute percentage error (MAPE), which is defined as:

[0043]

[0044] where y i is the actual result in the i-th sample, i.e. the short-circuit current value obtained by simulation; is the output value predicted by the machine learning method at the i-th input, i.e. the short-circuit current prediction value; N is the total number of test samples, i.e. the total number of samples used to test the machine learning model.

[0045] Preferably, the construction of the short-circuit current interval prediction model includes:

[0046] The upper and lower limits of the BPNN parameterized short-circuit current interval prediction model are used to train the ensemble learner to construct the short-circuit current interval prediction model.

[0047] Preferably, the construction of the short-circuit current interval prediction model is represented by the following formula:

[0048]

[0049] where β is the weight of the ensemble learner, c is the significance level, 100x(1-c)% is the confidence level, PICP is the prediction interval coverage probability (PI coverage probability), PINAW is the prediction interval normalized average width (PI normalized average width), and AWD is the accumulated width deviation (Accumulated width deviation).

[0050] Preferably, PICP is calculated using the following formula:

[0051]

[0052] where κ i is a Boolean variable;

[0053] PINAW is calculated using the following formula:

[0054]

[0055] wherein: Z is a normalization factor;

[0056] AWD is calculated using the following formula:

[0057]

[0058] wherein: Z is a normalization factor.

[0059] Compared with the prior art, the present application has the beneficial effects that:

[0060] The present application can accurately predict the short-circuit current of a wind turbine grid-connected power system based on a BP neural network, and then obtain a prediction interval of the short-circuit current through a short-circuit current interval prediction model, thereby giving a confidence level of the short-circuit current prediction result and improving the reliability of the short-circuit current prediction result. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 Fig. 1 is a flowchart of a data-driven wind turbine grid-connected power system short-circuit current interval prediction method of the present application.

[0062] Figure 2 Fig. 2 is a schematic diagram of a data-driven short-circuit current prediction application framework.

[0063] Figure 3 Fig. 3 is a schematic diagram of a doubly-fed wind turbine power generation system equipped with a rotor Crowbar circuit.

[0064] Figure 4 Fig. 4 is a flowchart of simulation sample generation.

[0065] Figure 5 Fig. 5 is a schematic diagram of a topology structure of an improved New England 39-node system.

[0066] Figure 6 Fig. 6 is a performance of a BPNN, a SVM model and an EL method on the first 10 samples in a test set.

[0067] Figure 7 Fig. 7 is an envelope curve of PIs based on an EBPNN, PIs based on RL and PIs based on a SVM. DETAILED DESCRIPTION

[0068] The technical solutions and technical effects of the embodiments of the present application are further described in detail below in combination with the accompanying drawings of the present application.

[0069] Please refer to Figure 1 A data-driven wind turbine grid-connected power system short-circuit current interval prediction method, comprising the following steps:

[0070] S1, obtaining operation parameters of a wind turbine grid-connected power system;

[0071] S2, extract the operating characteristics of the operating parameters, input the extracted operating characteristics to the trained BP neural network, and the BP neural network outputs the short-circuit current when the current operating parameters of the fan grid-connected power transmission system occur short-circuit fault;

[0072] S3, input the short-circuit current into a short-circuit current interval prediction model, and the short-circuit current interval prediction model outputs the prediction interval of the short-circuit current.

[0073] The BP neural network (BPNN, Back Propagation Neural Network) learns and simulates complex electrical behavior patterns by training on simulation samples and historical data. Based on the BP neural network, the short-circuit current of the fan grid-connected power transmission system can be accurately predicted. Then, through the short-circuit current interval prediction model, the prediction interval of the short-circuit current is obtained, so as to give the confidence of the short-circuit current prediction result and improve the reliability of the short-circuit current prediction result.

[0074] Further, before step S1, the method further comprises:

[0075] The fan grid-connected power transmission system includes a doubly-fed fan power generation system equipped with a rotor Crowbar circuit. The response stage of the doubly-fed fan power generation system when a short-circuit fault occurs is analyzed, and a short-circuit current calculation formula of the response stage is determined.

[0076] A simulation model of the fan grid-connected power transmission system is established, and a large number of simulation samples are generated by setting different system operating modes.

[0077] A BP neural network is established, and the simulation samples and historical data are divided into a training set, a validation set and a test set in proportion.

[0078] The BP neural network is iteratively trained using the training set, and the network performance is verified using the validation set after each iteration.

[0079] When the iterative training is completed and the network performance is optimized, the accuracy of the BP neural network calculation is tested by the test set, and the trained BP neural network is obtained.

[0080] The short-circuit current level and the power system unit combination, operating mode and topological structure have a mapping relationship. The prediction of the short-circuit current of the power system belongs to the regression problem. Artificial neural networks have been widely used to solve such complex regression problems and are suitable for fast prediction of short-circuit current. Among them, using BP neural network to predict the short-circuit current in the power system has many advantages. BP neural networks are very effective in dealing with the complex nonlinear relationships inherent in power system dynamics. They can learn and simulate complex electrical behavior patterns by training on historical data, thereby accurately predicting short-circuit current. And the adaptability of BPNN ensures that they can continuously improve the prediction when more data is obtained.

[0081] Please refer to Figure 2 , the present application is based on BP neural network, using the offline learning + online application of short-circuit current prediction framework as shown in Figure 2 . Among them, the offline training stage, BPNN trains neural network parameters according to simulation samples and historical data as training data. After inputting the training data, the training data is divided into training set, validation set and test set according to the proportion. The training set is the training sample for training the calculation model. The validation set is used to verify whether the performance of the network is improving. The network performance is verified after each iteration, and if the network performance has not been optimized for several consecutive iterations, the training stops. The role of the validation set is to prevent the neural network from overfitting. The test set is used to test the accuracy of the neural network calculation. Then, the data is normalized, that is, all data is mapped to a unified scale to prevent the training result from being dominated by one or several sample features. The structure, parameters and training method of the network will have a great influence on the training result, so it is necessary to select appropriate neural network structure and parameters according to the characteristics of the sample, and try to use multiple training methods to train and test the effect. In the online application stage, the trained BP neural network quickly predicts the short-circuit current according to the operating parameters of the wind turbine grid-connected power system.

[0082] Further, the wind turbine grid-connected power system includes a doubly-fed wind turbine generator system equipped with a rotor Crowbar circuit. In the doubly-fed wind turbine generator system, due to the magnetic coupling between the stator and the rotor, when a short-circuit fault occurs in the system, a large current will flow through the rotor circuit and pass through the generator-side converter. Therefore, it is necessary to protect the generator-side converter during short-circuit. The protection system can be based on an external resistance connected to the rotor circuit, which is called Crowbar resistance. The goal of this protection system is to let the current pass through the Crowbar resistance instead of the converter. The doubly-fed wind turbine generator system equipped with a rotor Crowbar circuit is shown in Figure 3 .

[0083] Further, when a short-circuit fault occurs, the response of the doubly-fed wind turbine generator system equipped with a rotor Crowbar circuit can be divided into two stages: the first stage is the sub-transient operation stage (about 2-3 cycles), at this time the doubly-fed wind turbine generator system can be regarded as an induction generator operation; the second stage is that the Crowbar protection system is triggered, the transient of the rotor flux starts to decay, and the rotor current is limited. The doubly-fed wind turbine model in the two stages is briefly introduced below.

[0084] The first stage is the sub-transient operation stage, and the doubly-fed wind turbine generator system is regarded as an induction generator operation. The model of the induction generator is represented as a transient voltage source followed by a series stator transient reactance, and its specific expression is as follows:

[0085]

[0086] where the subscripts {s, r} refer to the stator and rotor electrical quantities, respectively, and the subscripts {d, q} refer to the d-axis and q-axis electrical quantities in the dq synchronous reference frame, L ls and L lr are the stator and rotor leakage inductances, respectively, and L m is the mutual inductance between the stator and rotor;

[0087] The stator short-circuit current of an induction generator can be obtained from the following equation:

[0088]

[0089] where:

[0090]

[0091] where T s ' is the stator transient time constant, T r ' is the rotor transient time constant, L s ' is the stator transient inductance, and L r ' is the rotor transient inductance. Although the current vector does not reach its maximum value exactly at t = T / 2, the current at half a cycle is a good approximation to the maximum current. Therefore, the current at t = T / 2 can be used in place of the maximum current. If the purpose of the analysis is to calculate the maximum value of the root mean square (rms) value of the alternating current component, the stator voltage u s (0 - ) and the stator transient reactance X s ' prior to the fault are used. However, if the purpose of the analysis is to estimate the peak value of the fault current, X s ' / C(T / 2) must be used instead of X s .

[0092] The second stage is the Crowbar protection system triggering stage. If a short circuit occurs at the terminals of the doubly-fed wind generator system, the maximum short circuit current is approximately:

[0093]

[0094] where R cb is the Crowbar resistance value.

[0095] Thus, the maximum short circuit current of a doubly-fed wind generator can be well calculated by simply using the transient reactance of the corresponding induction motor.

[0096] Further, referring to Figure 4 , a large number of simulation samples are generated by setting different system operating modes using, for example, Figure 4The flowchart shows that, in particular, a selected wind turbine is connected to a grid-connected power system to balance a synchronous machine, a selected random switching line is set, a basic operation mode of the wind turbine grid-connected power system is set, and a total number N of simulation samples is generated, wherein the active power of the synchronous machine is P Gi,base (i = 1, 2, …, m), the active power consumed by the load is P loadj,base (j = 1, 2, …, n), the inductive reactive power absorbed by the load is Q loadj,base (j = 1, 2, …, n); randomly generate the kth set of synchronous machine active power and load power consumption, wherein k ≤ N; randomly generate the kth set of wind speed, set the response wind speed scene; randomly generate the kth set of line switching; set the kth set of operation mode and line in the simulation system, and initialize the power flow; set the wind turbine not to be put into operation, and calculate the short-circuit current I f0 by simulation; set the wind turbine to be put into operation, and calculate the short-circuit current I f by simulation; repeat the above steps until N sets of simulation samples are obtained.

[0097] Before generating the samples, the sample labels and sample characteristics need to be determined first. In a power system, the main factors affecting the size of the short-circuit current are the topology of the power grid, the short-circuit parameters of the power equipment, and the operation mode of the power system. Among them, the topology of the power system is relatively fixed after construction, and the equipment updating cycle is long, so in the topology and equipment parameters of the power system, the invention only considers the switching of part of the lines in the case of determining the topology and fixed equipment. The operation mode of the power system will be adjusted according to the changes of user demand and primary energy supply and other factors. Especially after the grid connection of new energy, due to the randomness of the primary energy, the operation mode of the entire power system also has a certain randomness, so the invention mainly considers obtaining training samples by changing the operation mode of the power system.

[0098] Among them, there are many characteristics to describe the operation mode of the power system, including the power transmitted and power loss of each line, the voltage amplitude and phase angle of each node, the output of the generator, and the load power. For a large power system, the large number of characteristics describing the operation mode is not conducive to artificial intelligence methods, and not all operation characteristics have a great impact on the short-circuit current level, so it is not necessary to use all characteristics as input of the machine learning model. Considering that the invention involves a system with sufficient reactive power, the generator voltage can be maintained stable under different operation modes, so the load power and generator output are used as sample characteristics. In addition, since the calculation of the short-circuit current of the power system containing only rotating power sources can be accurately and quickly calculated by the traditional method, the invention considers using the system short-circuit current when the double-fed wind turbine is not in operation as a sample characteristic, which reflects the operation mode of the system to a certain extent, greatly improving the training efficiency and accuracy of the machine learning method.

[0099] Based on the above considerations, for a power system with (m+1) synchronous machines (of which 1 is a balanced node), n loads (constant PQ model) and l wind farms, a system operating mode is set as a basic operating mode, in which the generator output is P Gi,base (i=1, 2,..., m), the active power of the load is P loadj,base (j=1, 2,..., n), the reactive power of the load is Q loadj,base (j=1, 2,..., n). On the basis of the basic operating mode, the values of the synchronous machine output of the unbalanced unit, the wind speed and the load consumption power are adjusted, the synchronous machine output, the active power and the reactive power of the load consumption are limited to fluctuate within 0.8-1.2 times of the reference value, and the wind speed V wind is also randomly fluctuated. In the simulation system, the kth operating mode of the system is set as:

[0100] P Gi,k =k Gi,k ·P Gi ,i=1,2,...,n

[0101]

[0102] In the formula, P Gi,k is the active output of the ith synchronous machine in the kth operation, P loadj,k is the active power consumed by the jth load in the kth operation, Q loadj,k is the inductive reactive power absorbed by the jth load in the kth operation; k Gi,k , k PLj,k , k QLj,k are independent random quantities in [0.8, 1.2], and the probability distribution is uniform distribution, so that each operating mode has a probability and the same possibility of occurrence.

[0103] At the same time, in the wind turbine grid-connected power transmission system containing double-fed wind turbines, the most common cause of operating mode change is the fluctuation of wind speed, so it is necessary to make the wind speed also have randomness. According to the statistics, the probability distribution of wind speed is generally normal distribution, so in the present application, the normal distribution model is used to describe the wind speed. In each operation, the wind speed is fixed as V wind , which is normally distributed with a mean of 12 and a variance of 4 (if the wind speed is less than 0, the wind speed is set to 0).

[0104] In addition to changing the system's operating mode, a simple change in the system topology is considered, that is, controlling the switching of some lines. Some lines (a total of l branches) are selected as switchable lines. The selection is based on the fact that the normal operation of the system will not be affected after the line is cut off, and there will be no islands in the system after all lines are cut off, and no generators or loads will be disconnected from the grid. Set the line state variable matrix Line = [L1, L2, ..., L l ], an element in the matrix of 0 indicates that the corresponding line is disconnected, and conversely, an element in the matrix of 1 indicates that the line is in operation. The elements in the state variable matrix Line are either 0 or 1, and their probability follows a 0-1 distribution, with a probability of 1 of 0.7.

[0105] Through Figure 4 The flowchart shown can obtain the simulation samples required by the BP neural network. Each sample contains complete sample features and sample labels. The specific composition of each sample is shown in the following table:

[0106] Table 1 Sample composition

[0107]

[0108] Among them, when the BP neural network is iteratively trained using the training set, the input of the BP neural network is the characteristics of each sample, and the output is the sample label, so that in the application process, the operating parameters of the wind turbine grid-connected transmission system are obtained; the operating characteristics of the operating parameters are extracted, including the active output of the synchronous machine, wind speed, active power consumed by the load, inductive reactive power absorbed by the load, line switching state vector and short-circuit current when the wind turbine is not put into use; the extracted operating characteristics are input into the trained BP neural network, and the short-circuit current when a short-circuit fault occurs in the current operating parameters of the wind turbine grid-connected transmission system output by the BP neural network can be obtained.

[0109] Furthermore, the accuracy of BP neural network calculation is tested through the test set, including:

[0110] The accuracy of BP neural network calculation is tested by the mean absolute percentage error (MAPE), which is defined as:

[0111]

[0112] Among them, y i is the actual result in the i-th sample, that is, the short-circuit current value obtained by simulation; is the output value predicted by the machine learning method for the i-th input group, i.e., the predicted short-circuit current value; N is the total number of test samples, i.e., the total number of samples used to test the machine learning model. The smaller the MAPE, the more accurate the short-circuit current prediction using the BP neural network.

[0113] Further, the prediction interval (PI) technique shows the best interpretability due to its interpretable loss function, and the PIs also have the advantages of excellent scalability and easy implementation. Therefore, the short-circuit current prediction interval is obtained by the short-circuit current interval prediction model to improve the reliability of the short-circuit current prediction based on the BP neural network.

[0114] Since the information accuracy provided by the PIs constructed by a single learner is limited, the robustness and quality of the PIs are improved by using ensemble learning (EL) in the application, and a PI construction method based on ensemble backpropagation neural networks (EBPNN) is provided. Specifically, the EBPNN first parameterizes the upper and lower limits of the PIs by using the BPNN, and then trains the EL. The main goal of the EBPNN is to construct the PIs with the optimal quality under the given c. Therefore, the PI construction is theoretically a multi-objective optimization problem, in which the decision variable is the weight of the ensemble learner, defined as β, (wherein, 2M represents the total sum of the base DBNs participating in the PI construction). Therefore, from the optimization point of view, the PI construction can be represented by the following formula:

[0115]

[0116] In the formula, c is the significance level, 100x(1-c)% is the confidence level, PICP is the PI coverage probability, PINAW is the PI normalized average width, and AWD is the accumulated width deviation.

[0117] In order to simplify the solving process, the application introduces an importance factor v i c i = 1, 2, 3. The above formula is further converted into a single-objective optimization problem, as shown in the following formula:

[0118]

[0119] The particle swarm optimization (PSO) is used to solve the above formula. Based on this method, the short-circuit current interval prediction model based on ensemble learning can be constructed.

[0120] In some embodiments, a typical PI refers to a predicted interval range in which the actual target value will fall with a certain probability (i.e. confidence level). A PI consists of a lower bound an upper bound and a certain probability 100x(1-c)%, where c is the significance level and 100x(1-c)% is the confidence level. Suppose there is a training sample set containing n t samples then the mathematical definition of the PI trained on Ψ t is shown as follows:

[0121]

[0122] Further, the PICP index indicates the probability that the target of the training sample set is covered by the predicted interval, which is calculated using the following formula:

[0123]

[0124] wherein k i is a Boolean variable; its definition is shown as follows. The higher the PICP value, the higher the quality of the PI.

[0125]

[0126] wherein and represent the i-th sample vector of the input feature set and the target feature set, respectively. For the convenience of description, the present application only considers a single target in the target set, i.e.

[0127] The PICP and AWD values can be optimal when the upper bound of the predicted interval is positive infinity and the lower bound is negative infinity. However, such a PI cannot provide any useful information. Therefore, in order to construct a PI with a high degree of information, PINAW needs to be introduced. The smaller the PINAW value, the higher the quality of the PI. The PINAW is calculated using the following formula:

[0128]

[0129] wherein Z is a normalization factor;

[0130] The AWD is used to quantify the degree to which the training target deviates from the upper or lower bound of the PI, and can be calculated using the following formula:

[0131]

[0132] wherein Z is a normalization factor.

[0133] The lower the AWD, the higher the quality of the PI.

[0134] Embodiments

[0135] The data-driven wind generator grid-connected power system short-circuit current interval prediction method of the present application is verified based on the improved New England 39-bus system. Compared with the standard 39-bus system, the grid-connected synchronous motor at bus 38 is replaced by a grid-connected wind generator with the same capacity. The topology structure of the improved New England 39-bus system is shown in Figure 5 .

[0136] Firstly, the effectiveness of the BP neural network, i.e., the short-circuit current prediction model based on BPNN, is verified. The fault is set as a three-phase short-circuit fault occurring at bus 2, with a fault duration of 0.1 s and a grounding impedance of 0.01 Ω. Three switchable lines are set, including the connection line of nodes 26-29, the connection line of nodes 22-23, and the connection line of nodes 4-14. The basic operating mode of the system is the initial power flow operating mode. According to the above-mentioned operating mode setting method, 1500 operating modes are generated, and 1500 simulation samples are constructed as a training set through electromagnetic transient simulation. In addition, 100 simulation samples are reserved as a test set and do not participate in model training. The BPNN contains two hidden layers, each containing 24 neurons, and the training function is based on the Powell / Beale restart conjugate gradient algorithm. The comparison algorithm is the Support Vector Machines (SVM) model and the Ensemble Learning (EL) method, with the SVM based on a quadratic kernel function and the EL method being a random forest. The performance of each method on the first 10 samples in the test set is shown in Figure 6 .

[0137] As can be seen from Figure 6 , the MAPE of BPNN is less than 0.5% except for sample #3, with good prediction effect, and the other two machine learning algorithms have a larger gap compared with BPNN. The average effect and calculation speed of the three methods on the test set are shown in Table 2. In order to reflect the calculation efficiency, the Time Domain Simulation (TDS) method is introduced for comparison.

[0138] Table 2 Comparison of short-circuit current prediction effect and calculation efficiency of different methods

[0139]

[0140] As shown in Table 2, although BPNN requires longer training time and computation time than RL and SVM, its average percentage error is only 27.24% and 22.16% of the other two methods, respectively. Compared with TDS, the BPNN method improves the average computation efficiency by more than 200 times, reaching the millisecond level, which can meet the real-time computing requirements.

[0141] Next, the effectiveness of the short-circuit current interval prediction model based on EBPNN is verified. In addition to the PIs based on EBPNN, PIs based on RL and SVM are also used for comparative experiments. In particular, to ensure unbiased testing, the learning methods of the four PIs will be trained and tested based on the same training sample set and test sample set. The training algorithm of each PI is executed 10 times, and the average value of the PI quality index is taken to evaluate the performance of each algorithm. The results after testing on the test sample set are shown in Table 3. Figure 7 Figure 7 As can be seen from Table 3, the PIs based on EBPNN are more accurate than the PIs based on RL and SVM.

[0142] The quantitative evaluation indicators of each PI are shown in Table 3. It should be noted that in Table 3, in addition to the AWD and PINAW indicators used by the present application, the average coverage error (ACE) indicator is also used to evaluate the PI quality, and its calculation formula is as follows:

[0143]

[0144] Table 3 Quantitative evaluation indicators of PIs constructed by each method

[0145] Method SVM RL EBPNN ACE 6.7e-3 5.1e-3 3.2e-3 PINAW 0.6823 0.6476 0.6191 AWD 0.0128 0.0084 0.0093

[0146] As can be seen from Table 3, the performance of PIs based on SVM is the worst, with ACE, PINAW and AWD values of 6.7e-3, 0.6823 and 0.0128, respectively. The PIs based on EBPNN driven by the present application have the best performance, with ACE, PINAW and AWD values of 3.2e-3, 0.6191 and 0.0093, respectively.

[0147] The above only discloses the preferred embodiments of the present application, and of course cannot limit the scope of the rights of the present application. Those skilled in the art can understand that the above-mentioned embodiments can be implemented in whole or in part, and equivalent changes made in accordance with the claims of the present application still fall within the scope of the present application.​

Claims

1. A data-driven short-circuit current interval prediction method for a wind turbine grid-connected power transmission system, characterized in that: The following steps are involved: S1. Obtaining the operating parameters of the wind turbine grid-connected power transmission system; S2. Extracting the operating characteristics of the operating parameters, inputting the extracted operating characteristics into a trained BP neural network, and the BP neural network outputting the short-circuit current when a short-circuit fault occurs in the current operating parameters of the wind turbine grid-connected power transmission system; S3, inputting the short-circuit current into a short-circuit current interval prediction model, and the short-circuit current interval prediction model outputting a prediction interval of the short-circuit current; Before step S1, the method further includes: The wind turbine grid-connected transmission system includes a doubly-fed wind turbine generator system equipped with a rotor crowbar circuit. The response phases of the doubly-fed wind turbine generator system when a short-circuit fault occurs are analyzed, and the short-circuit current calculation formula during the response phase is determined. The response of the doubly-fed wind turbine generator system when a short-circuit fault occurs consists of two phases: the first phase is the subtransient operation phase, and the second phase is the triggering phase of the crowbar protection system. Establish a simulation model of the wind turbine grid-connected power transmission system, set different system operation modes to generate a large number of simulation samples; establish a BP neural network, and divide the simulation samples and historical data into training sets, validation sets, and test sets in proportion; Iteratively training the BP neural network using the training set, and verifying the network performance using the validation set after each iteration; When the iterative training is completed and the network performance is optimized, the accuracy of the BP neural network calculation is tested through the test set to obtain the trained BP neural network; The calculation formula for determining the short-circuit current in the response stage includes: The first stage is the sub-transient operation stage. The doubly fed wind turbine generator system is regarded as an induction generator. The model of the induction generator is represented by a transient voltage source followed by a stator transient reactance X in series. s ', its specific expression is as follows: Where, the subscripts {s, r} represent the electrical quantities of the stator winding and rotor winding of the generator, respectively; the subscripts {d, q} represent the d-axis electrical quantities and q-axis electrical quantities in the dq synchronous coordinate system, respectively; L ls and L lr are stator leakage inductance and rotor leakage inductance respectively, L m is the mutual inductance between stator and rotor; The stator short-circuit current of an induction generator can be obtained by the following formula: Where: Among them, T s ' is the stator transient time constant, T r ' is the rotor transient time constant, L' s is the stator transient inductance, L' r is the rotor transient inductance; The second stage is the Crowbar protection system triggering stage. If a short circuit occurs at the terminals of the doubly fed wind turbine generator system, the maximum short-circuit current is approximately: Where R cb is the resistance value of the Crowbar resistor.

2. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 1 is characterized in that: Set up different system operation modes to generate a large number of simulation samples, including: Select the wind turbine grid-connected power transmission system balancing synchronous machine, select the random switching line, set the basic operation mode of the wind turbine grid-connected power transmission system and set the total number of generated simulation samples N, where the active output of the synchronous machine is P Gi,base (i=1,2,…,m), the load consumes active power P loadj,base (j=1,2,…,n), the inductive reactive power absorbed by the load is Q loadj,base (j=1,2,…,n); Randomly generate the active output of the kth group of synchronous machines and the load power consumption, where k≤N; Randomly generate the kth group of wind speeds and set the response wind speed scenario; Randomly generate the kth group of line switching conditions; Set the kth group of operation modes and lines in the simulation system and initialize the power flow; Set the fan to be off and calculate the short-circuit current I through simulation. f0 ; Set the fan input and calculate the short-circuit current I through simulation f ; Repeat the above steps until N groups of simulation samples are obtained.

3. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 2 is characterized in that: The operation mode of the kth group is set in the simulation system as follows: P Gi,k =k Gi,k ·P Gi ,i=1,2,...,n Where, P Gi,k is the active power output of the i-th synchronous machine under the k-th operation, P loadj,k is the active power consumed by the jth load under the kth operation, Q loadj,k is the inductive reactive power absorbed by the jth load under the kth operation; k Gi,k 、k PLj,k 、k QLj,k They are all independent random quantities in the range of [0.8, 1.2], and the general distribution is uniform, so that all types of operation modes have probabilities and the possibility of occurrence is the same.

4. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 1, characterized in that: The accuracy of BP neural network calculation is tested through the test set, including: The accuracy of BP neural network calculation is tested by the mean absolute percentage error (MAPE), which is defined as: Among them, y i is the actual result in the i-th sample, that is, the short-circuit current value obtained by simulation; is the output value predicted by the machine learning method when the i-th group of input is input, that is, the short-circuit current prediction value; N is the total number of test samples, that is, the total number of samples used to test the machine learning model.

5. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 1, characterized in that: The construction of the short-circuit current interval prediction model includes: The upper and lower limits of the short-circuit current interval prediction model are parameterized using BPNN, and the ensemble learner is trained to construct the short-circuit current interval prediction model.

6. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 5, characterized in that: The short-circuit current interval prediction model is constructed as follows: Where β is the weight of the ensemble learner, c is the significance level, 100×(1-c)% is the confidence level, PICP is the prediction interval coverage probability (PI coverage probability), PINAW is the prediction interval normalized average width (PI normalized average width), and AWD is the accumulated width deviation (Accumulated width deviation).

7. The data-driven short-circuit current interval prediction method for wind turbine grid-connected power transmission system according to claim 6, characterized in that: PICP is calculated using the following formula: Where: κ i is a Boolean variable; PINAW is calculated using the following formula: Where: Z is the normalization factor, is the i-th group of sample vectors of the input feature set; AWD is calculated using the following formula: Where: Z is the normalization factor.

Images