CN108802565B - Medium-voltage power distribution network disconnection ungrounded fault detection method based on machine learning - Google Patents

Abstract

The invention relates to a method for detecting disconnection and non-grounding faults of a medium-voltage distribution network based on machine learning, comprising the following steps: S1, extracting daily electricity consumption data from the distribution transformer level; S2, preprocessing the extracted data, Including detrending and periodic processing of the time series; S3, using the random forest algorithm based on feature selection to classify the preprocessed data to obtain the fault detection result. Compared with the prior art, the present invention proposes a set of detection methods for intelligently detecting the disconnection of power distribution network based on the data of the power distribution information system and the machine learning algorithm with a pure data-driven method. The accuracy is high, and the parameters of the main standard for judging the disconnection of the distribution network can be selected.

Description

A method for detection of disconnected and ungrounded faults in medium-voltage distribution networks based on machine learning

technical field

The invention relates to a fault diagnosis technology for a distribution network, in particular to a method for detecting disconnection and non-grounding faults of a medium-voltage distribution network based on machine learning.

Background technique

The power system is the core part of the urban infrastructure, and the medium voltage grid is the center of the load and an important part of the power system distribution network. It has a wide coverage and can directly provide the required power resources for the city and its regions. As an important part of urban power supply system, the importance of medium voltage distribution network is self-evident. Since the medium-voltage distribution network substation often adopts the neutral point ungrounded method, when the wires on both sides of the fracture are not grounded or the non-power side wires are grounded after the single-phase line is broken, there is no obvious fault feature. And the fault cannot be detected by the existing relay protection device in the substation.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for detecting disconnection and non-grounding faults of a medium voltage distribution network based on machine learning in order to overcome the above-mentioned defects of the prior art.

The object of the present invention can be realized through the following technical solutions:

A method for detecting disconnected and ungrounded faults in a medium-voltage distribution network based on machine learning, comprising the following steps:

S1. Extract daily electricity consumption data from the distribution transformer level;

S2. Preprocess the extracted data, including detrending and periodic processing of the time series;

S3. Use a random forest algorithm based on feature selection to classify the preprocessed data to obtain a fault detection result.

Preferably, in the step S1, after the daily electricity consumption data is extracted, the electricity consumption data that does not meet the preprocessing requirements are screened out.

Preferably, the step S2 specifically includes:

S21, extracting the time series of the corresponding time period in the first three weeks of each transformer;

S22. The average value of each window corresponding to the time points in the time series of the previous three weeks is used as a template;

S23, subtracting the template from the original time series to obtain a preprocessed time series;

S24, performing a stationarity test.

Preferably, the random forest algorithm based on feature selection specifically includes:

Use the variable importance measurement results of random forest to arrange the features in descending order, take the first p% as the feature subset, and then randomly divide the training set and test set, and calculate the average and standard deviation of the accuracy, AUC value after multiple averages.

Preferably, the ratio of the randomly divided training set and test set is: 70% as the training set and 30% as the test set.

Preferably, the daily electricity consumption data includes: voltage, current, active power and reactive power of a distribution transformer.

Compared with the prior art, the present invention proposes a set of detection methods for intelligently detecting the disconnection of power distribution network based on the data of the power distribution information system and the machine learning algorithm with a pure data-driven method. The accuracy is high, and the parameters of the main standard for judging the disconnection of the distribution network can be selected.

Description of drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation manner and a specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

As shown in Figure 1, a method for detecting disconnection and non-grounding faults in a medium-voltage distribution network based on machine learning includes the following steps:

S1. Extract daily electricity consumption data from the distribution transformer level;

S2. Preprocess the extracted data, including detrending and periodic processing of the time series;

S3. Use a random forest algorithm based on feature selection to classify the preprocessed data to obtain a fault detection result.

In step S1, after extracting the daily electricity consumption data, the electricity consumption data that does not meet the preprocessing requirements are screened out.

Step S2 specifically includes:

S21, extracting the time series of the corresponding time period in the first three weeks of each transformer;

S22. The average value of each window corresponding to the time points in the time series of the previous three weeks is used as a template;

S23, subtracting the template from the original time series to obtain a preprocessed time series;

S24, performing a stationarity test.

The random forest algorithm based on feature selection specifically includes: using the variable importance measurement results of random forest to arrange the features in descending order, taking the first p% as the feature subset, and then randomly dividing the training set and the test set, 70% as the training set, 30 % is used as the test set to calculate the accuracy rate, the mean and standard deviation of the AUC values after multiple averaging, and the specific process is as follows:

Input: original dataset S;

Output: classification accuracy, mean and standard deviation of AUC value and importance Rank of corresponding feature set;

Step1: Initialize, set repeated training times N ₁ and N ₂ , Fori=1:N ₁ ;

Step2: In the initial data set S, use bootstrap to randomly extract K training subsets with replacement, and then construct K independent classification and regression trees, and the complement of each training subset is recorded as out-of-bag ( OOB);

Step3: For m features, randomly extract m _try features from each node of each tree, generate feature subsets, calculate the Gini index of each feature, and select the feature with the smallest Gini index in the feature subset as the splitting feature;

Step4: Each tree grows to the maximum without pruning;

Step5: average the importance of each feature, arrange in descending order to obtain Rank, take the top p% as the feature subset, Forj=1:N ₂ ;

Step6: After randomly shuffling the data set S, separate 70% of the training set and 30% of the test set, and repeat Step2-4;

Step7: The generated K trees are formed into a random forest, and the classification result is determined by the vote of the tree classifier;

Step8: Calculate the importance of each feature, and the accuracy and AUC value of the model in the test set;

Step9: Calculate the mean and standard deviation of the accuracy and AUC values.

Among them, m _try =floor(log ₂ (m)+1 is usually taken.

Daily electricity consumption data includes: voltage, current, active power and reactive power of distribution transformers. The effective efficiency of active and reactive power data is low and the 10kV feeder current generally only collects one phase, so the relevant characteristic quantities of active and reactive power and their rate of change are not considered.

Example

This embodiment uses the data of the power distribution and consumption information system in a certain region in East China from March 2016 to April 2017 to diagnose the disconnection and non-grounding fault of the branch line of the medium-voltage distribution network. Extract current, power, voltage A, B, C three-phase power data from distribution transformer level. The reason for this is to accurately locate the branches of the distribution transformer. After screening out the electricity consumption data that did not meet the preprocessing requirements, 32 disconnection records were obtained, corresponding to the three-phase electricity consumption data of 604 distribution transformers. At the same time, the three-phase power consumption data of the above-mentioned 604 distribution transformers one week before the disconnection were also extracted as the comparison of normal power consumption.

For data preprocessing, it can be seen from the time series of three-phase power consumption data that there are certain trends and periodicities. To remove these two effects, the time series is detrended and periodic. After preprocessing, a stationary time series is obtained.

The parameters in the random forest algorithm based on feature selection are taken as ntree=500 and mtry=7 respectively. After many experiments, it is found that the average importance of features is less sensitive to the values of ntree and mtry, and the influence of parameter changes on the model is weak.

The result of feature selection shows that the voltage characteristic has better disconnection detection ability. Therefore, Logistic Regression, SVC, and the aforementioned Random Forest algorithm are used below to compare the classification results.

First, the AUC, accuracy (ACC), sensitivity (TPR) and specificity (TNR) of the three classifier results were investigated.

Randomly select 70% of the data from the sample set as the training set and 30% of the data as the test set, and repeat 1000 times to obtain the average of the prediction results. The classification results were measured from the three classification AUC values, accuracy (ACC), sensitivity (TPR) and specificity (TNR). Randomly select 70% of the data from the sample set as the training set and 30% of the data as the test set, and repeat 1000 times to obtain the average of the prediction results.

的分类效果最好。After experiments, in the SVC model, select the "RBF" kernel function

The classification effect is the best.

Claims (4)

1. A medium voltage distribution network disconnection ungrounded fault detection method based on machine learning is characterized by comprising the following steps:

s1, extracting daily electric quantity data from the distribution transformer level;

s2, preprocessing the extracted data, including performing a trend removing item and periodic processing on the time sequence;

s3, classifying the preprocessed data by using a random forest algorithm based on feature selection to obtain a fault detection result;

the step S2 specifically includes:

s21, extracting a time sequence of time periods corresponding to three weeks before each transformer;

s22, taking the average value of each window corresponding to the time points in the previous three-week time sequence as a template;

s23, subtracting the template from the original time sequence to obtain a preprocessed time sequence;

s24, carrying out stability test;

the random forest algorithm based on feature selection specifically comprises the following steps:

and (3) arranging the features in a descending order by using the variable importance measurement result of the random forest, taking the top p% as a feature subset, randomly classifying a training set and a testing set, and calculating the accuracy after multiple averaging, and the mean value and the standard deviation of the AUC value.

2. The machine learning-based disconnection and non-grounding fault detection method for the medium-voltage distribution network, according to claim 1, is characterized in that after daily electricity consumption data are extracted in the step S1, electricity consumption data which do not meet preprocessing requirements are screened out.

3. The machine learning-based method for detecting the disconnection and non-grounding fault of the medium voltage distribution network according to claim 1, wherein the proportion of the randomly-divided training set to the test set is as follows: 70% as training set and 30% as test set.

4. The machine learning-based disconnection and non-grounding fault detection method for the medium-voltage distribution network, according to claim 1, is characterized in that the daily electricity consumption data comprises: voltage, current, active power and reactive power of the distribution transformer.

Images