CN111401783A - A feature selection method for power system operation data integration - Google Patents

Abstract

The invention discloses a feature selection method for power system operation data integration, comprising the following steps: S1, extracting K training subsets; S2, correlation analysis; S3, sorting the features on each training subset according to weights, Different results are obtained; S4, the results are aggregated to obtain the optimal feature subset. The invention proposes an integrated feature selection framework for power plant operation data on the basis of random sampling and RReliefF feature selection algorithm, improves the stability of the algorithm, removes redundant data, and improves time efficiency.

Description

A feature selection method for power system operation data integration

technical field

The invention relates to a feature selection method, in particular to a feature selection method for power system operation data integration.

Background technique

At present, due to the continuous expansion of smart grid, the amount of historical data is huge and has many dimensions. When analyzing and modeling historical data, if all attributes are used as the input of the model, it will not only increase the computational difficulty, but also lead to dimensional disaster, resulting in overfitting of the model and reducing the generalization ability. Therefore, before modeling, dimensionality reduction processing is required.

The commonly used methods of dimensionality reduction technology include feature extraction and feature selection. Feature extraction generally maps data from high-dimensional space to low-dimensional feature space through data methods (such as projection). Typical methods include principal component analysis (PCA), kernel principal component analysis (KPCA), canonical correlation analysis (CCA) and so on. However, the physical meaning of the new features extracted by features is far from the original features, or even completely different, and the interpretability of the extracted features is weak, which is unacceptable in many problems. Feature selection is to select the smallest subset of features that maximizes the evaluation criteria from the original feature space. The features selected by feature selection have the same physical meaning, strong interpretability and obvious advantages. In recent years, they have been widely used in bioinformatics, computer vision, target recognition, etc. In practical applications, not only the feature selection algorithm is required to have a good ability to select features, but also the stability of feature selection is required. Feature selection stability means that the feature selection method has a certain robustness to the small perturbations of the training samples. A stable feature selection method should generate the same or similar feature subsets when the training samples have small perturbations. Improving the stability of feature selection can discover relevant features, enhance the confidence of domain experts in the results, and further reduce the complexity and time consumption of data acquisition.

The sources of the instability of feature selection are as follows: (1) The instability of the algorithm itself: Most of the existing feature selection methods only aim to select the smallest subset of features and ignore the stability of the selection. (2) Features are highly redundant: If a feature selection algorithm obtains the same learning accuracy on different feature subsets of the dataset, the selected attributes are unstable. (3) High-dimensional small sample problem: In some practical problems such as genetic testing, there are usually only a few hundred samples, but thousands of features. In a high-dimensional small sample space, small changes in the sample data set will affect the distribution of data, and changes in data distribution will affect the selection results, resulting in differences in feature selection results.

The RReliefF algorithm can deal with nonlinear data, and the algorithm is more efficient, and has become a well-known feature selection method. However, the RReliefF algorithm also has certain defects: First, it is unstable. The algorithm itself does not consider the stability of feature selection. Multiple results may produce different optimal subsets, which is caused by the characteristics of the data itself and the algorithm. Second, the redundant features cannot be removed, and the features that play a positive role in the predicted value have a large weight and will be retained. The algorithm itself does not take into account the correlation between the features, and the power plant operation data not only has a large number of redundant features, but also features There is high coupling and strong correlation between them.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a method for selecting integrated features of power system operation data.

The technical scheme adopted in the present invention is: a method for selecting integrated features of power system operation data, comprising the following steps:

S1, extract K training subsets;

S2, correlation analysis;

S3, sort the features by weight on each training subset to get different results;

S4, summarize the results to obtain the optimal feature subset.

Further, the step S1 is specifically:

。Extract K training subsets from the original dataset D via Bootstrap random sampling method

.

Further, the step S2 is specifically:

，通过这步来去除掉冗余数据。The correlation analysis algorithm Pearson is used for correlation analysis on each training subset. If the absolute value of the correlation coefficient between two features is greater than a certain threshold, a feature is randomly deleted to obtain K training subsets.

, through this step to remove redundant data.

Further, the step S3 is specifically:

On each training subset, the features are sorted by weight by the RReliefF algorithm, and K different results are obtained.

Advantages of the present invention:

The invention proposes an integrated feature selection framework for power plant operation data on the basis of random sampling and RReliefF feature selection algorithm, improves the stability of the algorithm, removes redundant data, and improves time efficiency.

In addition to the objects, features and advantages described above, the present invention has other objects, features and advantages. The present invention will be described in further detail below with reference to the drawings.

Description of drawings

The accompanying drawings constituting a part of the present application are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

FIG. 1 is a flowchart of a method for selecting integrated features of power system operation data according to an embodiment of the present invention;

2 is an integrated feature selection framework diagram of an embodiment of the present invention;

Fig. 3 is the experimental result stability comparison diagram of the embodiment of the present invention;

FIG. 4 is a time efficiency comparison diagram of an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Referring to FIG. 1, as shown in FIG. 1, a method for selecting integrated features of power system operation data includes the following steps:

S1, extract K training subsets;

S2, correlation analysis;

S3, sort the features by weight on each training subset to get different results;

S4, summarize the results to obtain the optimal feature subset.

The invention proposes an integrated feature selection framework for power plant operation data on the basis of random sampling and RReliefF feature selection algorithm, improves the stability of the algorithm, removes redundant data, and improves time efficiency.

The step S1 is specifically:

。Extract K training subsets from the original dataset D via Bootstrap random sampling method

.

Bootstrap is a random sampling method with replacement. Its specific steps are to extract n samples from the original data each time, select m dimensions to form a sub-data set, repeat K times, and set n, m, and k by yourself. determined.

The step S2 is specifically:

，通过这步来去除掉冗余数据。The correlation analysis algorithm Pearson is used for correlation analysis on each training subset. If the absolute value of the correlation coefficient between two features is greater than a certain threshold, a feature is randomly deleted to obtain K training subsets.

, through this step to remove redundant data.

The step S3 is specifically:

On each training subset, the features are sorted by weight by the RReliefF algorithm, and K different results are obtained.

RReliefF algorithm introduction:

The Relief algorithm was first proposed by Kira et al. in 1992. It is an efficient feature weight algorithm for data reduction and is suitable for binary classification problems. Later, some scholars proposed the ReliefF algorithm to solve the multi-classification problem. The main idea of the algorithm is to weight the features according to their ability to distinguish samples of different categories in the neighborhood. Good features can make the samples of the same category close, and the samples of different categories are far away from each other. According to the set weight threshold, the features with weights less than the threshold will be removed, and finally the optimal feature subset will be obtained.

The ReliefF algorithm provides initial values for each feature in the dataset. Iteratively update the weight W[A] through the following formula, and iterate k times to finally obtain the result.

（3）

(3)

表示特征A的权重。

是从训练样本中随机选择的样本，ReliefF通过算法找到其最近的实例。

代表距离

最近的并且相同类别的样本。

代表距离

最近的并且不同类别的样本。如果样本

和

具有不同的特征A，说明A特征会分离两个类别相同的样本，则降低特征A的权重。如果样本

和

具有不同的特征A，说明A特征会分类两个不同类别的样本，则增加特征A的权重。函数

的计算公式为：In the above formula,

represents the weight of feature A.

is a randomly selected sample from the training samples, and ReliefF algorithmically finds its nearest instance.

representative distance

recent samples of the same class.

representative distance

recent samples of different classes. If the sample

and

If there are different features A, it means that the A feature will separate two samples of the same category, and the weight of feature A will be reduced. If the sample

and

With different feature A, it means that feature A will classify samples of two different categories, then increase the weight of feature A. function

The calculation formula is:

For numeric properties:

（4）

(4)

For discrete attributes:

（5）

(5)

On the basis of ReliefF, Kononenko et al. proposed the RReliefF algorithm for solving regression problems. Since the predicted value of the regression problem is continuous, there is no category distinction. In order to solve this problem, the probability of two different instances is introduced to judge whether the two instances belong to the same class. This probability can simulate the relative distance between the predicted instances.

（6）

(6)

（7）

(7)

（8）

(8)

The integrated feature selection framework BPRR (Bootstrap-Pearson-RReliefF) of the present invention is an improved algorithm.

Stability metrics:

The present invention selects an index of extensions of Kuncheva similarity measure to evaluate the stability of the algorithm. The Kuncheva similarity measure is one of the stability metrics of the subset method. It is extended from the Kuncheva similarity measure and can be used to measure the similarity of feature subsets with different numbers of features. . Calculated as follows:

（9）

(9)

和

表示两个特征子集，

表示特征子集的基，

为两个子集的特征总数。扩展昆彻瓦相似度度量指标的取值在[-1, 1]之间，扩展昆彻瓦相似度度量指标取值越大，两个特征子集相似度越高。in,

and

represents two feature subsets,

represents a basis for a subset of features,

is the total number of features for the two subsets. The value of the extended Kunchewa similarity measure is between [-1, 1], and the larger the value of the extended Kunchewa similarity measure, the higher the similarity of the two feature subsets.

Experimental verification:

In the experiment of the present invention, the subsets are respectively divided into 10, 20, . , and then average the results to obtain the stability of each algorithm on different numbers of subsets.

The above algorithm is used to measure the stability of RReliefF and the integrated feature selection framework, and the results are shown in Figure 3 and Figure 4;

in conclusion:

The following two main aspects can be seen from the figure: First, the stability of the two algorithms is compared on the stability comparison chart. The stability of RReliefF is poor, and the performance is not very good in the number of subsets from 10 to 90, while the integrated feature selection framework shows better stability. The results show that the ensemble feature selection framework is effective in improving the stability of RReliefF.

Second, the time efficiency of the two algorithms is compared on the time efficiency graph. It can be seen that the running time of RReliefF is much slower than that of the integrated feature selection framework. For example, when the number of subsets reaches 90, RreliefF is 3 times slower than the integrated feature selection framework.

The experimental results show that compared with the RReliefF algorithm, the stability of feature selection is improved, redundant features are eliminated and the efficiency is improved. It is proved that the method of the present invention can be applied to the feature selection step in the data preprocessing before the power big data modeling and prediction, screen out the attributes related to the target parameters from a large number of power parameters, and can be used when the data has a small disturbance. The same results are obtained in the same situation, and the confidence of feature selection is improved.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (4)

1. The method for selecting the running data integration characteristics of the power system is characterized by comprising the following steps

The method comprises the following steps:

s1, extracting K training subsets;

s2, correlation analysis;

s3, sorting the features on each training subset according to the weight to obtain different results;

and S4, summarizing the results to obtain the optimal feature subset.

2. The power system operational data integration feature selection method of claim 1, characterized in that

Characterized in that, the step S1 specifically comprises:

extracting K training subsets from an original data set D by a Bootstrap random sampling method

。

3. The power system operational data integration feature selection method of claim 1, characterized in that

Characterized in that, the step S2 specifically comprises:

performing correlation analysis on each training subset by using a correlation analysis algorithm Pearson, and if the absolute value of a correlation coefficient between certain two features is larger than a certain threshold, deleting one feature randomly to obtain K training subsets

By this step, redundant data is removed.

4. The power system operational data integration feature selection method of claim 1, characterized in that

Characterized in that, the step S3 specifically comprises:

and sequencing the features according to the weights on each training subset through an RReliefF algorithm to obtain K different results.

Images