Power system state evaluation model tendency correction method based on spatial distribution

Abstract

The invention relates to a power system state evaluation model tendency correction method based on spatial distribution. According to the method, spatial distribution information of training samples is quantized according to a support vector machine (SVM), the training samples are divided into subsets of different grades, then the unbalance degree between two types of samples in the subsets of the same grade is quantized, and the evaluation tendency correction method of the power system stability evaluation model is formulated based on the unbalance degree. Compared with an existing method which only considers the unbalance of the number of the samples, the method comprehensively considers the unbalance distribution condition of the number and the space of the samples, the calculation of the mean ratio of the fitting degree is more accurate, the evaluation tendency problem of the stability evaluation model of the power system is obviously improved, and the overall prediction accuracy of the evaluation model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a method for correcting the tendency of a power system state assessment model based on spatial distribution. Background Art

[0002] As an important field of power system analysis, power system stability analysis is crucial to ensure the safe and stable operation of power systems. In recent years, with the rapid development of artificial intelligence technology, through deep mining and analysis of massive data resources, the rapid and accurate judgment of the stable state of the power system can be achieved, which can effectively avoid the problems of low accuracy and efficiency of traditional power system stability analysis methods. However, these methods are based on data-driven, and sufficient and balanced training samples can ensure that the artificial intelligence algorithm has excellent generalization ability. However, due to the robustness of the power system, the system is in a stable state most of the time, resulting in a very scarce number of unstable samples. This leads to a certain evaluation tendency of the trained power system stability assessment model, and the prediction ability of unstable samples is reduced. In the actual operation of the power system, the consequences of missed judgment of system instability are more serious than the misjudgment of system stable operation. Therefore, it is necessary to design a method that can correct the evaluation tendency of the power system stability assessment model to improve the prediction accuracy of unstable samples and even improve the overall accuracy of the assessment model. Summary of the invention

[0003] The purpose of the present invention is to provide a method for correcting the tendency of a power system state assessment model based on spatial distribution, using SVM to quantify the spatial distribution information of samples, dividing them into subsets of different levels according to the spatial distribution information of samples, and quantifying the imbalance between two types of samples in the same level subsets, and then correcting the SSAE model through the cost-sensitive method, and then fine-tuning to obtain the final power system stability assessment model. The assessment tendency of the model has been improved, and the overall prediction accuracy of the model has also been improved.

[0004] To achieve the above purpose, the technical solution of the present invention is: a method for correcting the tendency of a power system state assessment model based on spatial distribution, characterized in that the training samples are divided into subsets of different levels according to the spatial distribution information of the SVM quantified training samples, and then the imbalance degree between two types of samples in the subsets of the same level is quantified, based on which the assessment tendency correction of the power system stability assessment model is realized. The method is implemented as follows:

[0005] (1) A large number of training samples are obtained through a large number of offline simulation calculations, and the power angle and angular velocity deviations of each generator set before, at, and after the system fault is cleared are selected as sample feature quantities;

[0006] (2) Input the sample set, obtain the classification hyperplane through pre-training SVM, and calculate the Euclidean distance between each sample and the classification hyperplane, eliminate the influence of the differential distribution of marginal samples, obtain the spatial distribution information of the samples, and then eliminate the marginal samples on this basis, and divide the training samples into multiple level subsets according to the normal distribution;

[0007] (3) Input the sample set, pre-train the stacked sparse autoencoder SSAE, and calculate the imbalance degree between the two types of samples in the same level subset;

[0008] (4) Determine the cost-sensitive correction coefficient according to the degree of imbalance between the two types of samples in each subset, correct the loss function of SSAE, and finally train SSAE to obtain the final power system stability assessment model.

[0009] In one embodiment of the present invention, the method is specifically implemented as follows:

[0010] 1) Obtain a large number of training samples for the simulation model through different fault information; obtain a large number of training samples through the permutations and combinations of different operating conditions, transmission lines, fault locations, and fault durations; select the power angle and angular velocity deviations of each generator set in the system before, at, and after the fault is cleared as the feature quantities of the samples, and perform z-score normalization on the sample features to construct a sample set S, in which the stable sample label is 1 and the unstable sample label is 0;

[0011] 2) Based on the sample set S, the pre-trained SVM obtains the classification hyperplane, and the Euclidean distances from unstable samples and stable samples to the classification hyperplane are calculated respectively, as shown in formula (1) and formula (2):

[0012]

[0013]

[0014] Where: Dist u (x u,i ) is the distance between the unstable sample and the classification hyperplane; Dist s (x s,i ) is the distance between the stable sample and the classification hyperplane; x u,i is the i-th unstable sample in the sample set S; x s,i is the i-th stable sample in the sample set S; w is the normal vector of the classification hyperplane of SVM; b is the intercept;

[0015] 3) Normalize the calculated Dist u (x u,i ) and Dist s (x s,i), and obtain the fault severity of each unstable sample u (x u,i ) and the stability level of each stable sample s (x s,i ):

[0016]

[0017]

[0018] Where: N u is the number of unstable samples, N s To stabilize the number of samples;

[0019] 4) Eliminate samples that are on the edge of the stable or unstable class, that is, eliminate edge samples whose normalized results of formulas (3) and (4) are between 0.9 and 1;

[0020] Assume that the fault severity distribution of the sample is normal distribution, and assume that the normal distribution parameter μ = 0. Calculate the normal distribution parameter σ of the stable sample and unstable sample respectively through equations (5) and (6): s , σ u ; Refer to the 3σ principle of normal distribution and set Severity u (x u,i ) and Severity s (x s,i ) is divided into three intervals; accordingly, the sample set S is divided into three level subsets:

[0021]

[0022]

[0023] 5) Input the sample set S, pre-train the stacked sparse autoencoder SSAE, and after the number of iterations reaches 1000, calculate the fitting degree of SSAE to the sample, as shown in formula (7):

[0024] G(x i )=u(x i )ln(v(x i ))(v(x i )-u(x i )) 2 +(1-u(x i ))ln(1-v(x i ))(v(x i )-u(x i )) 2 (7)

[0025] Where: u(x i ) is the sample x i The true label of v(x i ) is the sample x i The predicted probability of (v(x i )-u(x i )) 2 For sample x i The square value of the deviation; ln(v(x i )) is when the sample x i is the loss value of the stable sample; ln(1-v(x i )) is the sample x i is the loss value of unstable samples; G(x i ) represents SSAE for sample x i The degree of fit;

[0026] 6) Calculate the mean ratio D of the degree of fit of stable samples and unstable samples in the three subsets respectively imb(k) ;

[0027]

[0028] Where: k = 1, 2, 3; N s(k) is the number of stable samples in the k-th level subset; N u(k) is the number of unstable samples in the k-th level subset; x u,k,r is the rth unstable sample in the kth level subset; x s,k,z is the zth stable sample in the kth level subset;

[0029] 7) According to the D of each level subset imb(k) Assign a corresponding cost coefficient to each sample and use the cost-sensitive method to modify the loss function L of SSAE SSAE :

[0030]

[0031] Where: N k is the number of samples in the k-th level subset; x k,n is the nth sample in the kth level subset;

[0032] 8) After the loss function of SSAE is corrected, the sample set S is re-input and the trained SSAE is fine-tuned to obtain the final power system stability assessment model.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention obtains the spatial distribution of samples by pre-training SVM, thereby obtaining the fault severity of each sample, and then quantifies the sample imbalance information in spatial distribution through the pre-trained SSAE model, while avoiding the influence of the differential distribution of edge samples on the result quantification, thereby formulating an evaluation tendency correction method for the power system stability assessment model. Compared with the existing methods that usually only consider the imbalance in the number of samples, the present invention comprehensively considers the imbalance distribution of samples in quantity and space, and the calculation of the mean ratio of the degree of fitting is more accurate, which significantly improves the evaluation tendency problem of the power system stability assessment model and improves the overall prediction accuracy of the assessment model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The present invention provides a specific flow chart of the offline training of the power system stability assessment model. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0037] The present invention provides a method for correcting the tendency of a power system state assessment model based on spatial distribution, which is characterized in that the training samples are divided into subsets of different levels according to the spatial distribution information of the SVM quantified training samples, and then the imbalance degree between two types of samples in the subsets of the same level is quantified, based on which the assessment tendency correction of the power system stability assessment model is realized. The method is implemented as follows:

[0038] (1) A large number of training samples are obtained through a large number of offline simulation calculations, and the power angle and angular velocity deviations of each generator set before, at, and after the system fault is cleared are selected as sample feature quantities;

[0039] (2) Input the sample set, obtain the classification hyperplane through pre-training SVM, and calculate the Euclidean distance between each sample and the classification hyperplane, eliminate the influence of the differential distribution of marginal samples, obtain the spatial distribution information of the samples, and then eliminate the marginal samples on this basis, and divide the training samples into multiple level subsets according to the normal distribution;

[0040] (3) Input the sample set, pre-train the stacked sparse autoencoder SSAE, and calculate the imbalance degree between the two types of samples in the same level subset;

[0041] (4) Determine the cost-sensitive correction coefficient according to the degree of imbalance between the two types of samples in each subset, correct the loss function of SSAE, and finally train SSAE to obtain the final power system stability assessment model.

[0042] The following are specific implementation examples of the present invention.

[0043] The simulation software uses PSD-BPA, and the test system uses an IEEE-39 node system with a speed regulator and a voltage regulator. The operating state of the power system is set to five load levels: 90%, 95%, 100%, 105% and 110%, and permanent three-phase short-circuit faults are set at 10%, 50% and 90% of the distance from the head end of each fault line. The fault removal time is set to 5 to 20 cycles after the fault occurs, the maximum simulation time is 4s, and the system operation information is sampled 120 times per second. A total of 15012 samples are generated on the IEEE-39 node system, including 7191 stable samples and 7821 unstable samples. Randomly extract sample data from stable samples and unstable samples to form training sets and test sets. The number of training sample sets is 4000, the number of test samples is 2000, the ratio of unstable samples to stable samples in the training set is 1:3, and the ratio of unstable samples to stable samples in the test set is 1:1.

[0044] The calculation example uses SSAE to establish a power system stability assessment model. The parameters of SSAE are as follows: the number of neurons in the hidden layer is 100, 50, and 25 respectively; pre-training is 100 times; fine-tuning is 400 times; α is 0.0005, and β is 0.00005. The sample grouping principle is as follows: M is set to 2.

[0045] Step 1: Input the training set to pre-train SVM and obtain the classification hyperplane. Then calculate the Euclidean distance between each sample and the hyperplane and normalize it. Based on this, divide the training set into two levels of subsets according to the normal distribution. The specific grouping is as follows;

[0046] The number of unstable samples in the first-level subset is: N u(1) =744

[0047] The number of stable samples in the first-level subset is: N s(1) =2234

[0048] The number of unstable samples in the second-level subset is: N u(2) =256

[0049] The number of stable samples in the second-level subset is: N s(2) =766

[0050] Step 2: Calculate the mean ratio D of the degree of fit between the two types of samples in the same level subset imb(k) ;

[0051] The specific results of the mean degree of fit between the two types of samples in each level subset are as follows:

[0052] The mean ratio of the degree of fit between the two types of samples in the first level subset: D imb(1) =5.949824

[0053] The mean ratio of the degree of fit between the two types of samples in the second-level subset: D imb(2) =5.099570

[0054] Step 4: According to D imb(1) and D imb(2) Each sample is assigned a corresponding cost-sensitive coefficient, and the loss function of SSAE is modified. The training set is input to train the power system stability assessment model based on the modified SSAE. The model is applied to the real-time assessment of power system stability.

[0055] This example is based on the SSAE classifier and compares several commonly used data imbalance processing methods, including oversampling methods (ADASYN and SMOTE), random undersampling methods, traditional cost-sensitive methods, and the cost-sensitive method introduced in the present invention; among them, the oversampling methods (ADASYN and SMOTE) generate system unstable samples until the number is consistent with the system stable samples, and then input the samples into the SSAE for training; the random undersampling method randomly extracts samples with the same number of system unstable samples from the system stable samples, and then inputs the extracted samples into the SSAE model for training; the traditional cost-sensitive method uses the ratio of the number of stable samples to the number of unstable samples of 3:1 as the cost coefficient of the unstable samples to correct the loss function of the model, and trains to obtain the corrected SSAE model.

[0056] The stable sample prediction error rate is defined as follows.

[0057]

[0058] In the formula, F s is the stable number of samples with incorrect predictions, T s is the stable number of samples for correct prediction.

[0059] The unstable sample prediction error rate is defined as follows.

[0060]

[0061] In the formula, F u is the number of unstable samples with incorrect predictions, T u is the number of unstable samples predicted correctly.

[0062] The global accuracy of the evaluated model is shown below.

[0063]

[0064] The experimental results of various methods are compared in Table 1:

[0065] Table 1 Comparison of fault diagnosis results of different unbalance correction methods

[0066]

[0067] It can be seen from Table 1 that the correction effect of the method of the present invention is the best and the overall accuracy of the model is the highest.

[0068] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A method for correcting the tendency of a power system state assessment model based on spatial distribution, characterized in that: According to the spatial distribution information of training samples quantified by SVM, the training samples are divided into subsets of different levels, and then the imbalance degree between two types of samples in the subsets of the same level is quantified, based on which the assessment tendency correction of the power system stability assessment model is realized.

2. The method for correcting the tendency of the power system state assessment model based on spatial distribution according to claim 1 is characterized in that: The method is implemented as follows: (1) Obtain training samples through offline simulation calculation, and select the power angle and angular velocity deviation of each generator set before, at, and after the system fault is cleared as sample feature quantities; (2) Input the sample set, obtain the classification hyperplane through pre-training SVM, and calculate the Euclidean distance between each sample and the classification hyperplane, eliminate the influence of the differential distribution of marginal samples, obtain the spatial distribution information of the samples, and then eliminate the marginal samples on this basis, and divide the training samples into multiple level subsets according to the normal distribution; (3) Input the sample set, pre-train the stacked sparse autoencoder SSAE, and calculate the imbalance degree between the two types of samples in the same level subset; (4) Determine the cost-sensitive correction coefficient according to the degree of imbalance between the two types of samples in each subset, correct the loss function of SSAE, and finally train SSAE to obtain the final power system stability assessment model.

3. The method for correcting the tendency of the power system state assessment model based on spatial distribution according to claim 1 or 2, characterized in that: The specific implementation of this method is as follows: 1) Obtain a large number of training samples for the simulation model through different fault information; obtain a large number of training samples through the arrangement and combination of different operating conditions, transmission lines, fault locations, and fault durations; The power angle and angular velocity deviations of each generator set in the system before, at, and after fault clearing are selected as the feature quantities of the samples, and the sample features are normalized by z-score to construct a sample set S, where the stable sample label is 1 and the unstable sample label is 0; 2) Based on the sample set S, the pre-trained SVM obtains the classification hyperplane, and the Euclidean distances from unstable samples and stable samples to the classification hyperplane are calculated respectively, as shown in formula (1) and formula (2): Where: Dist u (x u,i ) is the distance between the unstable sample and the classification hyperplane; Dist s (x s,i ) is the distance between the stable sample and the classification hyperplane; x u,i is the i-th unstable sample in the sample set S; x s,i is the i-th stable sample in the sample set S; w is the normal vector of the classification hyperplane of SVM; b is the intercept; 3) Normalize the calculated Dist u (x u,i ) and Dist s (x s,i ), and obtain the fault severity of each unstable sample u (x u,i ) and the stability level of each stable sample s (x s,i ): Where: N u is the number of unstable samples, N s To stabilize the number of samples; 4) Eliminate samples that are on the edge of the stable or unstable class, that is, eliminate edge samples whose normalized results of formulas (3) and (4) are between 0.9 and 1; Assume that the fault severity distribution of the sample is normal distribution, and assume that the normal distribution parameter μ = 0. Calculate the normal distribution parameter σ of the stable sample and the unstable sample respectively through equations (5) and (6): s , σ u ; Refer to the 3σ principle of normal distribution and set Severity u (x u,i ) and Severity s (x s,i ) is divided into three intervals; accordingly, the sample set S is divided into three level subsets: 5) Input the sample set S, pre-train the stacked sparse autoencoder SSAE, and after the number of iterations reaches 1000, calculate the fitting degree of SSAE to the sample, as shown in formula (7): G(x i )=u(x i )ln(v(x i ))(v(x i )-u(x i )) 2 +(1-u(x i ))ln(1-v(x i ))(v(x i )-u(x i )) 2 (7) Where: u(x i ) is the sample x i The true label of v(x i ) is the sample x i The predicted probability of (v(x i )-u(x i )) 2 For sample x i The square value of the deviation; ln(v(x i )) is when the sample x i is the loss value of the stable sample; ln(1-v(x i )) is the loss value of sample xi as an unstable sample; G(x i ) represents SSAE for sample x i The degree of fit; 6) Calculate the mean ratio D of the degree of fit of stable samples and unstable samples in the three subsets respectively imb(k) ; Where: k = 1, 2, 3; N s(k) is the number of stable samples in the k-th level subset; N u(k) is the number of unstable samples in the k-th level subset; x u,k,r is the rth unstable sample in the kth level subset; x s,k,z is the zth stable sample in the kth level subset; 7) According to the D of each level subset imb(k) Assign a corresponding cost coefficient to each sample and use the cost-sensitive method to modify the loss function L of SSAE SSAE : Where: N k is the number of samples in the k-th level subset; x k,n is the nth sample in the kth level subset; 8) After the loss function of SSAE is corrected, the sample set S is re-input and the trained SSAE is fine-tuned to obtain the final power system stability assessment model.

Images