Method for evaluating severity of grid-connected fault of network-forming energy storage system

Abstract

The invention discloses a method for evaluating the severity of a grid-connected fault of a network construction type energy storage system, and the method comprises the following steps: setting a feature vector according to the characteristics of operation monitoring data obtained by a grid-connected metering device of the network construction type energy storage system, and setting a scoring function related to the fault degree; carrying out probability and score conversion on the feature vectors by adopting a sigmoid function; according to the converted feature vector, constructing a ranked learning RankNet model of the grid-connected fault of the grid-forming type energy storage system; and taking the feature vectors as samples, training a ranking learning RankNet model by adopting a full-connection neural network, and then ranking and grading the samples according to the fault degree. The method has the beneficial effects that the feature vectors are extracted from the operation monitoring data, the scoring function associated with the fault degree is set, and the fault severity is scored and sequenced by using the neural network model, so that efficient evaluation of the grid-connected fault severity of the network-building energy storage system is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of grid-connected fault assessment, and particularly to a method for assessing the severity of grid-connected faults of a network-forming energy storage system. Background Art

[0002] With the rapid development of renewable energy, especially the wide application of intermittent energy sources such as wind energy and solar energy, the stability and reliability of the power system are facing unprecedented challenges. In this context, network-forming energy storage systems, as a key technology for regulating power supply fluctuations, balancing loads, and promoting optimal energy scheduling, have received extensive attention. These energy storage systems can effectively balance power supply and demand and ensure the stable operation of the power grid. However, in actual operation, network-forming energy storage systems may be affected by various factors, resulting in grid-connected faults. These faults not only affect the safety of the power system but may also cause serious consequences such as damage and outage of grid equipment. Therefore, timely identification and accurate assessment of the severity of grid-connected faults are crucial for ensuring the normal operation of the power system and improving the efficiency of fault response.

[0003] Traditional grid-connected fault assessment methods usually rely on rule-based judgment, physical model analysis, or simplified mathematical calculations. These methods may not fully reflect the actual situation of faults when dealing with complex power system faults, especially when involving large-scale energy storage systems, it is difficult to meet the requirements of fault assessment accuracy and efficiency. Therefore, how to accurately assess grid-connected faults based on the real-time operation data of energy storage systems using intelligent technologies and rank the severity of faults according to the assessment results is an important issue in current power system research. Summary of the Invention

[0004] Aiming at the scientific problem of sorting non-emergency fault handling and abnormal maintenance work in the grid-connected operation of network-forming energy storage systems, the present invention proposes a method for assessing the severity of grid-connected faults of network-forming energy storage systems, aiming to assist maintenance personnel in achieving efficient maintenance of grid-connected faults of network-forming energy storage systems.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A method for assessing the severity of grid-connected faults of a network-forming energy storage system, comprising the following steps:

[0007] Set a feature vector according to the characteristics of the operation monitoring data obtained by the grid-connected metering device of the network-forming energy storage system, and set a scoring function associated with the fault degree;

[0008] Use the sigmoid function to convert the probability and score of the feature vector;

[0009] Construct a ranking learning RankNet model for grid connection faults of a grid-connected energy storage system based on the converted feature vector;

[0010] Use the feature vector as a sample, train the ranking learning RankNet model using a fully connected neural network, and then rank and classify the samples according to the degree of fault.

[0011] In some embodiments, the process of constructing the feature vector includes:

[0012] Construct a feature parameter field according to the characteristics of the operation monitoring data, and the feature parameter field includes at least voltage, current, and phase-separated power;

[0013] Reduce the dimension of the feature parameter field to obtain the feature vector.

[0014] In some embodiments, the scoring function includes:

[0015]

[0016] where s i is the score of the i-th sample, the coefficients included in F are the normalized fault ratios corresponding to the feature parameter fields, and X i is the normalized feature vector of the i-th sample.

[0017] In some embodiments, the conversion of the feature vector using the sigmoid function for probability and scoring includes:

[0018]

[0019] where P represents the sigmoid probability function, i and j respectively represent the i-th and j-th samples, rank(i)>rank(j) indicates that the i-th sample is a fault sample ranked before the j-th sample, and s i and s j are the scores of the i-th and j-th samples respectively.

[0020] In some embodiments, the process of constructing the ranking learning RankNet model includes:

[0021] Define that the training objective of the ranking learning RankNet model is to minimize a loss function, and the loss function takes the negative logarithm of the sigmoid probability function, that is:

[0022]

[0023] where J ij is the loss function.

[0024] In some embodiments, the parameters of the loss function are set to θ k , and the gradient descent method is used to adjust the ranking learning RankNet model, and then the loss function J ij derives with respect to the parameter θ k , including:

[0025]

[0026] In the formula,

[0027] In some embodiments, the overall gradient calculation of the loss function includes:

[0028]

[0029] For a single sample, the gradient of the loss function is:

[0030]

[0031] The beneficial effects of the present invention are as follows: By extracting feature vectors from operation monitoring data and setting a scoring function associated with the fault degree, a neural network model is used to perform fault severity scoring and ranking, so as to realize the efficient evaluation of the grid-connected fault severity of the network-forming energy storage system. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic flow chart of the method for evaluating the grid-connected fault severity of the network-forming energy storage system disclosed in the embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of the simulation topology model of the grid-connected fault of the network-forming energy storage system;

[0034] Figure 3 is the score distribution corresponding to the test set samples. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0036] It should be noted that the terms "including" and "having" in the embodiments of the present invention and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0037] The present invention proposes a method for evaluating the severity of grid connection faults in a network-forming energy storage system, which can be applied to the Figure 2 topological structure described above. A three-phase power metering device operation monitoring data array including 4 fields and 9 features is designed to achieve dimensionality reduction of the data; the negative logarithm of the Sigmoid function is used for the conversion of sorting probability and scoring, and a fully connected neural network (9 neurons in the input layer corresponding to 9 features) is used to train the RankNet model to achieve the classification of the fault degree of the metering device. The fault samples are expanded by a random combination method to realize the training of the network model in the case of insufficient fault samples. The proposed scoring model can achieve accurate sorting of faults of different degrees, thereby realizing the efficient evaluation of the severity of metering device faults.

[0038] This embodiment proposes a method for evaluating the severity of grid connection faults in a network-forming energy storage system, as Figure 1 shown, including the following steps:

[0039] Step 1, set a feature vector according to the characteristics of the operation monitoring data obtained by the grid connection metering device of the network-forming energy storage system, and set a scoring function associated with the fault degree.

[0040] In the above step 1, a feature parameter field is constructed according to the characteristics of the operation monitoring data. The feature parameter field includes at least more than a dozen physical quantities such as voltage, current, and phase power. Each physical quantity is a time-domain sequence with a sampling time interval of 0.2 s. According to the requirements of grid connection fault discrimination of the network-forming energy storage system, a feature parameter field table related to the operation monitoring data of the grid connection metering device is designed, as shown in Table 1, which reflects the magnitude, fluctuation, and waveform characteristics of each index and serves as the data input for the sorting learning RankNet model in step 3.

[0041] Table 1 Feature parameter field table

[0042]

[0043] The time-domain sequence of each type field of the grid-connected metering device operation monitoring parameters is an independent one-dimensional array. Therefore, all the monitoring field parameters form a complex high-dimensional array, which cannot be used as the input of the classification algorithm. The feature parameter fields are dimensionally reduced to obtain feature vectors. According to the simulation operation monitoring data of the grid-forming energy storage system grid-connected metering device, 9 grid-connected metering device operation monitoring feature vectors are designed. In one example, the present invention is processed by a feature extraction method, and the feature vector includes:

[0044] X = (R AB , R AC , R BC , S, E A , E B , E C , N Udc , N Idc )

[0045] Feature 1: R AB Represents the Euclidean distance between the effective values of the currents of phases AB,

[0046] Feature 2: R AC Represents the Euclidean distance between the effective values of the currents of phases AC,

[0047] Feature 3: R BC Represents the Euclidean distance between the effective values of the currents of phases BC,

[0048] Feature 4: S represents the maximum value of the approximate entropy of the three-phase voltage, which can characterize the complexity of the voltage curve in the three-phase voltage, S = max{ApEn(U A ), ApEn(U B ), ApEn(U C )}.

[0049] Feature 5: E A Represents the effective value of the voltage of phase A.

[0050] Feature 6: E B Represents the effective value of the voltage of phase B.

[0051] Feature 7: E C Represents the effective value of the voltage of phase C.

[0052] Feature 8: N Udc Represents the absolute value of the difference between the DC bus voltage and the normal voltage.

[0053] Feature 9: N Idc Represents the absolute value of the difference between the DC bus current and the normal current.

[0054] Meanwhile, based on the statistical analysis of the proportion of grid-connected metering device fault categories and the monitoring field data in the operation monitoring data, a scoring function corresponding to the grid-connected fault degree of the grid-forming energy storage system is designed.

[0055] In one example, the scoring function includes:

[0056]

[0057] In the formula, s i is the score of the i-th sample, the coefficients included in F are the normalized fault proportions corresponding to the characteristic parameter fields, and X i is the normalized characteristic vector of the i-th sample. According to actual maintenance experience, F can be set as (2, 1, 1, 0.5, 0.5, 0.5, 2, 3, 3).

[0058] Considering the data annotation format of the sample pairs, samples are randomly selected from the dataset, and different types of faults are evaluated according to expert experience to distinguish the severity of different fault types. Then, attribute labels are set to provide a basis for the training of the ranking learning RankNet model in step 4. There are three types of labels, and (+1, 0, -1) represent the three degrees of emergency, severity, and general respectively. If f(s i ) - f(s j ) ≥ 0.05, then the label y ij = +1; if f(s i ) - f(s j ) ≤ -0.05, then the label y ij = -1; if -0.05 ≤ f(s i ) - f(s j ) ≤ 0.05, then the label y ij = 0. Among them, s i is the score of the i-th sample, and s j is the score of the j-th sample. The proportions of each label are shown in Table 2.

[0059] Table 2 Proportions of Each Label

[0060]

[0061] Step 2, use the sigmoid function to convert the characteristic vector for probability and scoring.

[0062] Using the sigmoid function to convert the characteristic vector for probability and scoring includes:

[0063]

[0064] Wherein, P represents the sigmoid probability function, i and j respectively represent the i-th and j-th samples, rank(i) > rank(j) indicates that the i-th sample is a fault sample ranked before the j-th sample, and s i and sj are the scores of the i-th and j-th samples respectively.

[0065] Step 3: Construct a ranking learning RankNet model for the grid-connection fault of the grid-connected energy storage system according to the transformed feature vectors.

[0066] The construction process of the ranking learning RankNet model includes:

[0067] Define that the training objective of the ranking learning RankNet model is to minimize a loss function, which measures the difference between the predicted fault sample ranking and the actual fault sample ranking. The loss function takes the negative logarithm of the sigmoid probability function, that is:

[0068]

[0069] Wherein, J ij is the loss function. When s i -s j ≥0, the value of J ij is almost equal to 0, that is, when the i-th sample is ranked before the j-th sample, the score of the i-th sample should also be higher than that of the j-th sample, and at this time the corresponding loss function value is very small. On the contrary, the value of the loss function increases as the score difference increases.

[0070] Set the parameter of the loss function to θ k , adjust the ranking learning RankNet model using the gradient descent method, and then the loss function J ij derives with respect to the parameter θ k , including:

[0071]

[0072] Wherein, Since λ ij is always negative, so for the score s j of the fault sample j ranked behind, it will decrease, while the score s i of the fault sample i ranked in front will increase, and the increase amplitude depends on the absolute value of λ ij .

[0073] Perform the overall gradient calculation of the loss function, including:

[0074]

[0075] For a single sample, the gradient of the loss function is:

[0076]

[0077] Equations (5) and (6) show that among all the fault sample pairs in the data set, for any fault sample l, the scores of the fault samples in front of it decrease, while the scores of the fault samples behind it increase. The difference in scores determines the range of change. Based on this, the grid-forming energy storage system grid connection fault severity scoring model can score the fault samples, and then obtain the sorting result.

[0078] Step 4: Use the feature vector as a sample, and train the sorting learning RankNet model with a fully connected neural network, and then sort and grade the samples according to the fault degree.

[0079] In Step 4, a two-part four-layer fully connected neural network is used to train the scoring model. The input is the corresponding feature vector set according to the operation monitoring data, and the parameter of the loss function is the difference between the two outputs of the neural network. The parameters of the fully connected neural network used are as follows: There are 9 neurons in the input layer (corresponding to a feature vector dimension of 9), and there are three hidden layers in total. The first layer includes 128 neurons, the second layer includes 64 neurons, and the third layer includes 32 neurons. The fault samples in the training set and the test set all come from the simulated fault samples. The test results of the grid-forming energy storage system grid connection fault severity scoring model obtained by training on the training set using the neural network and RankNet algorithm established by the present invention on the test set and the training set are shown in Table 3. Among them, the correct number represents that the discrimination result of the grid-forming energy storage system grid connection fault severity scoring model is the same as the sample annotation. If the two are different, it is regarded as a discrimination error. For the 0-label sample pairs, they are all counted as discrimination errors.

[0080] Table 3 Accuracy of the scoring model based on the RankNet algorithm

[0081]

[0082] As shown in Table 3, if the 0-label sample pairs are not considered, the discrimination accuracy of the grid-forming energy storage system grid connection fault severity scoring model based on the RankNet algorithm is almost close to 100% on the training set and the test set. When using the scoring function f(s i ) for scoring, the score distribution corresponding to 1000 test set samples is as Figure 3 shown.

[0083] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Improving the efficiency of fault repair. The present invention uses the RankNet model based on sorting learning to score and sort faults, and can quickly identify the most serious fault among various fault samples. This can help maintenance personnel prioritize the most urgent and serious faults, avoid the mixing of low-priority faults and high-priority faults, and significantly improve the efficiency and accuracy of the maintenance work; (2) Utilizing the accuracy of simulation data. In the invention, the simulation operation monitoring data of the network-forming energy storage system is used as input, and by designing feature vectors and extracting key parameters (such as voltage, current, phase-separated power, etc.), these data can accurately reflect the actual operation state of the energy storage system, avoid the interference of measurement errors in actual operation, and improve the reliability of the model prediction results; (3) Innovative application of the RankNet sorting model. By using the RankNet model to sort fault samples, the model can consider the relative severity of faults and assign a score to each fault sample. This method not only realizes the classification of faults, but also can give an accurate sorting according to the relationship between faults, enabling the entire maintenance work to be reasonably arranged according to the severity of faults, and avoiding the problem that the traditional simple fault classification method cannot accurately reflect the fault priority; (4) The efficient learning ability of the neural network. The present invention uses a fully connected neural network to train the fault severity scoring model. This neural network has a strong non-linear mapping ability and can continuously adjust the network parameters by optimizing the loss function, so that the model can more accurately identify complex fault patterns and operating states. Especially by using the gradient descent method to adjust the model, it ensures the good adaptability of the algorithm in large-scale datasets.

[0084] The above embodiments are only for illustrating the technical concept and characteristics of the present invention, and the purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the essence of the content of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for evaluating the severity of grid-connected faults of a grid-connected energy storage system, characterized in that: The following steps are involved: Setting a feature vector according to the characteristics of the operation monitoring data obtained by the grid-connected metering device of the grid-connected energy storage system, and setting a scoring function associated with the degree of fault; The sigmoid function is used to convert the feature vector into probability and score; According to the converted feature vector, a ranking learning RankNet model for grid-connected faults of a grid-connected energy storage system is constructed; The feature vector is used as a sample, and a fully connected neural network is used to train the ranking learning RankNet model, and then the samples are ranked and graded according to the degree of fault.

2. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 1, characterized in that: The process of constructing the feature vector includes: Constructing a characteristic parameter field according to the characteristics of the operation monitoring data, wherein the characteristic parameter field at least includes voltage, current and phase power; The feature parameter field is reduced in dimension to obtain the feature vector.

3. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 2, characterized in that: The scoring function includes: In the formula, s i is the score of the i-th sample, the coefficient contained in F is the normalized fault proportion corresponding to the feature parameter field, and X i is the normalized feature vector of the i-th sample.

4. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 3, characterized in that: The method of converting the probability and score of the feature vector using the sigmoid function includes: Where P represents the sigmoid probability function, i and j represent the i-th and j-th samples respectively, rank(i)>rank(j) means that the i-th sample is the fault sample before the j-th sample, s i and j are the scores of the i-th and j-th samples respectively.

5. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 4, characterized in that: The construction process of the ranking learning RankNet model includes: The training objective of the ranking learning RankNet model is defined as minimizing a loss function, which takes the negative logarithm of the sigmoid probability function, namely: In the formula, J ij is the loss function.

6. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 5, characterized in that: The parameter of the loss function is set to θ k , the gradient descent method is used to adjust the ranking learning RankNet model, and then the loss function J ij For parameter θ k Derivation, including: In the formula, 7. The method for evaluating the severity of grid-connected faults of a grid-connected energy storage system according to claim 6, characterized in that: The loss function performs an overall gradient calculation, including: For a single sample, the gradient of the loss function is:

Images