JP7240691B1 - Data drive active power distribution network abnormal state detection method and system - Google Patents

Abstract

Kind Code: A1 A data drive active power grid abnormal condition detection method and system for improving data set utilization efficiency, improving abnormal condition prediction accuracy, and reducing prediction error. The method obtains node parameter data after the occurrence of a distribution network abnormality, inputs the data into a trained distribution network abnormality prediction model, outputs a distribution network abnormality prediction result, and outputs a distribution network history. Cluster the abnormal node parameter data, find data samples with strong correlation with the grid abnormal condition type by association rule algorithm, form a sample data set for training, and determine the grid abnormal condition based on the sample data set A prediction model is trained, and a three-layer data mining structure is used to extract and obtain data samples that have a strong correlation with corresponding fault types according to data classification and association rules. [Selection drawing] Fig. 1

Description

TECHNICAL FIELD The present invention relates to the technical field of power grid abnormal condition classification and prediction, and more particularly to a data drive active power grid abnormal condition detection method and system.

This section merely provides background information related to the present invention and may not necessarily constitute prior art.

With the deepening of active power grid intelligent real-time monitoring, more and more monitoring and protection devices are installed in the active power grid. These devices collect the power grid operating status data, and from the data analysis process, the active power grid operating status is normal. determine whether there is However, such a diagnosis method has the problems of high cost due to the installation of the monitoring device and low diagnosis accuracy.

In order to avoid the above-mentioned problems of conventional fault diagnosis methods, artificial intelligence is used to diagnose abnormal conditions such as active power grid short-circuit faults, disconnection faults and overloads, relying on the advantages of self-learning and self-optimization. Common fault diagnosis methods include methods such as expert systems, fuzzy logic, artificial neural networks, Bayesian networks, and Petri networks. In the expert system-based fault diagnosis method, data such as the operating state of the active power distribution network is used to construct an expert model in combination with the expert's empirical knowledge to infer the fault type. A diagnostic method based on fuzzy theory realizes an approximate simulation of an arbitrary nonlinear continuous function by fuzzy control. Diagnostic methods based on artificial neural networks mimic the human brain and natural neural networks, model existing historical data, complete training of relevant parameters, and obtain a solution set for the problem. Although the Petri network-based diagnostic approach does not require a complex information base and can effectively handle discrete dynamic active grid faults, it suffers from low fault tolerance.

When the scale of the active distribution network is very large and there is a lot of fault data, the analysis effect on the fault data is highly dependent on the choice of algorithms. In addition, the current failure diagnosis method adopted by artificial intelligence always adopts only a single artificial intelligence algorithm or model, which has room for improvement in diagnosis accuracy and low fault tolerance, and strongly depends on the experience of experts. There are restrictions such as

In order to solve the above problems, the present invention uses three-layer data mining to classify and predict active distribution network abnormal conditions, so as to solve the problem of low accuracy and long time of abnormal condition classification and prediction correspondingly. A data drive active grid abnormal state detection method and system are extracted.

To achieve the above objectives, the following technical solutions are adopted in some embodiments.

A data drive active grid abnormal condition detection method, comprising:
Acquiring node parameter data after the occurrence of a distribution grid abnormality, inputting the data into a trained distribution grid abnormality prediction model, and outputting a distribution grid abnormality prediction result;
The training process of the grid abnormal condition prediction model includes clustering the grid historical abnormal condition node parameter data, finding data samples with a strong correlation with the grid abnormal condition type by association rule algorithm, and obtaining sample data for training. forming a set; and training the grid abnormal condition prediction model based on the sample data set.

As a further solution, the step of clustering the grid historical abnormal condition node parameter data specifically comprises:
classifying the data samples based on Euclidean distance;
determining whether the data sample clustering is complete by the discriminant function;
calculating the error sum of squares, determining the clustering number referenced by the SSE-K graph, and determining the final clustering number based on the referenced clustering number.

As a further solution, after clustering the grid historical abnormal state node parameter data,
Self-encoding the clustered data samples to form a first sample data set, wherein the elements in the first sample data set are vectors, each vector composed of TV and F _i , where TV = {N0C ₁ , N0C ₂ ,..., N0C _a } is the data after self-coding, the range of C values is [1, K], a is the node number where data collection is performed, F _i is the value of the abnormal condition type),
The self-coded data includes at least node location information where the data is located and clustering category information to which the data belongs.

As a further solution, the step of finding data samples with strong correlations with grid abnormal condition types by an association rule algorithm to form a sample data set for training specifically comprises:
a step of merging the two elements TV and F _i of each vector in the first sample data set as one element to form a new data set M: {Z ₁ , Z ₂ , . . . , Z _i }; , Z _i represents a new vector consisting of one post-self-coding data and the corresponding abnormal condition type, i is the number of new vectors),
Quantifying the connections between two or more elements in the new vectors using the support and confidence of frequent itemsets, extracting the non-zero vectors that satisfy the support and confidence requirements, and distributing power forming a second sample data set for training a network abnormal condition prediction model.

As a further solution, the step of training the grid abnormal condition prediction model based on the sample data set specifically comprises:
Regression training is performed using a stochastic gradient descent algorithm based on the obtained sample data set to obtain an optimal parametric solution of the grid abnormal condition prediction model.

である
（ただし、ｗは予測モデル関数の重みベクトルであり、ｂは予測モデル関数のインターセプトを表し、ｗ^Ｔ（ＣＶ_ｊ）はベクトルｗとベクトルＣＶ_ｊのスカラー積を表し、ベクトルＣＶ_ｊは訓練サンプルである）。 As a further solution, the distribution network abnormal state prediction model specifically:

where w is the weight vector of the predictive model function, b represents the intercept of the predictive model function, w ^T (CV _j ) represents the scalar product of vector w and vector CV _j , and vector CV _j is the training sample).

である
（ただし、Ｌ（Ｆ_ｉ，ｆ（ＣＶ_ｊ））はロス関数を表し、αはハイパーパラメータで設定値であり、Ｒ（ｗ）は正則化項であり、ｎはサンプルデータセット総量であり、Ｆ_ｉは異常状態タイプの値である）。 A further solution further comprises measuring the fit of the grid abnormal condition prediction model by optimizing an objective function, wherein the optimization of the objective function is specifically:

where L(F _i , f(CV _j )) represents the loss function, α is the hyperparameter and set value, R(w) is the regularization term, and n is the total amount of the sample data set. Yes and F _i is the value of the abnormal condition type).

In some other embodiments, the following technical solutions are adopted.

An active grid abnormal condition sensing system for a data drive, comprising:
a data acquisition module for acquiring node parameter data after a grid abnormal condition occurs;
an abnormal condition prediction module that inputs the data to a trained distribution network abnormal condition prediction model and outputs a distribution network abnormal condition prediction result;
The training step of the grid short circuit abnormal condition model includes clustering the grid historical fault node parameter data, finding data samples with a strong correlation with the grid short circuit abnormal condition by association rule algorithm, and obtaining a sample data set for training. and training the grid abnormal condition prediction model based on the sample data set.

The beneficial effects of the present invention compared to the prior art are as follows.
(1) The distribution network abnormal state prediction including distributed power sources of the present invention uses a three-layer data mining structure to extract and obtain data samples that have a strong correlation with the corresponding fault type through data classification and association rules. However, it can improve the utilization efficiency of the data set, and the three-layer data mining structure can improve the prediction accuracy of the abnormal state, shorten the calculation time, and reduce the prediction error.
(2) The present invention extracts a data format simplification method by self-encoding and sets different abnormal state type labels to effectively avoid the problem of unclear and unclear data relevance. , it can accurately extract data with strong correlations with abnormal condition types, enhancing the effectiveness of datasets.

FIG. 1 is a flow chart of the first layer data mining process in an embodiment of the present invention. FIG. 2 is a flow chart of the third layer data mining process in an embodiment of the present invention. FIG. 3 is a curve diagram of error sum of squares SSE and reference K value in an embodiment of the present invention.

It should be noted that the following detailed description is exemplary and is intended to provide further explanation for the present application. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments of the present application. As used herein, unless the context clearly indicates otherwise, the singular forms also include the plural forms and, as used herein, it should be understood that the terms "include" and/or "include" , indicates the presence of features, steps, operations, devices, components, and/or combinations thereof.
Example 1

In one or more embodiments, a data drive active grid fault condition detection method is disclosed that includes the following steps.
(1) obtaining node parameter data after occurrence of a grid abnormal condition, the node parameter data including, but not limited to, node voltage data or node current data;
(2) Inputting the data to a trained distribution network abnormality prediction model, and outputting a distribution network abnormality prediction result.

The training process of the grid abnormal condition prediction model includes clustering the grid historical abnormal condition node parameter data, finding data samples with a strong correlation with the grid abnormal condition type by association rule algorithm, and obtaining sample data for training. forming a set; and training a grid abnormal condition prediction model based on the obtained sample data set.

The power distribution network abnormal state prediction of this embodiment mainly includes four types of active power distribution network short-circuit faults: single-phase short circuit, two-phase short circuit, two-phase short circuit, and three-phase short circuit, load abnormality and disconnection. It is against failure.

In the present embodiment, one prediction model is constructed for each abnormal state, and four types of short-circuit faults, load abnormalities, and disconnection faults are used as examples, and prediction models are constructed for each of these abnormal states, modeling The method and idea are the same, and the whole prediction system can include these abnormal states, input the data after the occurrence of the abnormal state, predict and calculate the model, and finally obtain a more reliable prediction result. to get

In this embodiment, the three-layer data mining method is used to realize the training of the grid abnormal condition prediction model.

The first layer data mining uses a K-means clustering algorithm to classify the raw anomaly data, and then uses self-encoding rules to simplify the format of the raw anomaly data.

In the 2nd layer data mining, on the classified abnormal state data obtained in the 1st layer, data that has a small impact on the abnormal state prediction result is culled by the association rule, and the data has a strong correlation with the distribution network abnormal state type. as the data set for predictive model parameter regression training.

A data sample that has a strong correlation with the grid abnormal condition type is a data sample that satisfies the set minimum support and satisfies the set minimum confidence.

The third layer data mining uses the data obtained in the second layer data mining as a training set to obtain prediction models for various types of abnormal conditions by means of a stochastic gradient descent algorithm.

Each of the three layers of data mining steps is described in detail below.
1. First layer data mining

First, we collect the node voltage data after the occurrence of an abnormal condition, and then classify these data by a K-means clustering algorithm. In this embodiment, we encode the classified data and extract self-encoding rules to simplify the formatting of the data. The K-means clustering algorithm together with auto-encoding constitutes the first layer data mining process, and the data generated by the first layer data mining process constitutes the first sample data set.

Referring to FIG. 1, the K-means clustering algorithm is mainly that the Euclidean distance is used for data sample classification, the discriminant function is used to determine whether the sample clustering is complete, and the elbow method is It includes three aspects: identifying the referenced clustering number K value, which is then used to identify the final K value based on the referenced clustering number K value.

であり、
ただし、Ｐ＝（ｐ_１，ｐ_２，．．．，ｐ_ｎ）はｎ次元空間内の１つのデータサンプルを表し、例えば、三機九ノードシステムにおいて、ｎ＝９となると、Ｐ＝（ｐ_１，ｐ_２，．．．，ｐ_９）はある時刻で９つのノードの電圧データを表す。 Specifically, the K-means clustering algorithm determines that the data sample belongs to a clustering based on the Euclidean distance between the sample point and each class central sample point. When the Euclidean distance between a data sample and a kind of central sample point is minimal, the data sample belongs to the clustering where this central sample point is located. The formula for Euclidean distance is

and
where P=(p ₁ , p ₂ , . . . , p _n ) represents one data sample in n-dimensional space, e.g. ₁ , p ₂ ,..., p ₉ ) represent the voltage data of the 9 nodes at a certain time.

はｊ番目のクラスタリングの中心サンプル点を表し、Ｐ、Ｑ^ｊはいずれもベクトルである。最初に、ユークリッド距離を反復算出する工程において、各クラスタリングの中心サンプル点を任意に選択することができる。

represents the central sample point of the j-th clustering, and both P and Q ^j are vectors. First, in the step of iteratively calculating the Euclidean distance, a central sample point for each clustering can be chosen arbitrarily.

であり、
ただし、Ｋは分類の数を表し、

はｊ番目のクラスタリングのサンプル中心を表し、Ｐ^ｊはｊ番目のクラスタリングのうちのいずれも１つのサンプルデータを表し、

はいずれもｎ次元のサンプルデータのうちの１つの元素を表す。 If the classification is not complete, we need to update the center points for each clustering. The average value of all sample data in each clustering is taken as the new center point for the next iteration, and at the same time the discriminant function is used to determine whether the iterative update has stopped, i.e. whether the classification is complete. . The discriminant function here represents the minimum sum of squares of the difference between the sample and the center of the sample among all clusterings, and its calculation formula is

and
where K represents the number of classifications,

represents the sample center of the j-th clustering, P ^j represents one sample data for any of the j-th clustering,

each represent one element in the n-dimensional sample data.

When the discriminant function represented by equation (2) converges to a minimum, the sample center of each clustering does not change obviously, in which case the iteration is stopped, the classification process is completed, and all sample points are classified as class K. be done.

One of the drawbacks of the K-means clustering algorithm is that the number of classifications cannot be known in advance, so the clustering process is difficult to perform smoothly and the clustering quality is not high at the same time. Hence, the referenced clustering number K is the key point of this clustering algorithm.

This embodiment uses the elbow method to calculate the sum of squared errors (SSE) to obtain the referenced K value. As K gets larger, the raw data is classified more and more finely, while each clustering gets more and more compact and the sum of the squared errors gets smaller and smaller. Therefore, for a certain value of K, the sum of squared errors drops off sharply, reflected in the SSE-K graph, where the slope of the SSE curve suddenly slows down, the rate of slope decrease being the fastest, and as K increases, The SSE will change slowly. This method is called the elbow method because this tendency of change resembles a curved elbow. The K value corresponding to the elbow site is the referenced clustering number.

である。 The formula for calculating the sum of squared errors is

is.

In the process of iteratively calculating the SSE, K is an artificially chosen value, often increasing from 1 to a suitable upper limit (eg 15, to better observe the trend of the curve), and finally the SSE- A K value referenced by the K curve is determined. C _i represents the ith clustering, p represents any one data sample of C _i , and m _i represents the average value of all sample data of C _i , ie its sample center.

Based on the referenced K value, the data sample capacity and complexity of the data format are determined, and the final K value is greater than or equal to the reference K value.

After the K-means clustering algorithm, each clustering data sample has a certain similarity, but in this case the data format is still complicated and not suitable for further processing. This embodiment adopts the method of self-encoding to simplify the classified data. To preserve the key information of the raw data, such as the node where the data is located and the clustering where the data is located, self-encode according to the encoding rule N0C(4) below.

Here, N represents the node position where the data is located, C represents the clustering to which the data belongs, and 0 in the middle is for identification and has no practical significance. Taking the data format as 406 after some data is self-encoded, 406 indicates that the voltage data of node 4 at a certain time belongs to clustering 6 .

The above self-encoding rule is for node data only, and the abnormal condition type data label is set to person, for example, the abnormal condition type values F _i ε {110,120,130, 100, 140, 150}
, if F _i =110, the fault _condition is a single phase-to-ground fault; if F _i =120, the fault condition is a two-phase-to-phase fault; is a two-to-ground fault, when F _i =100, the fault condition is a three-phase short circuit fault, when F _i =140, the fault condition is a load fault, and when F _i =150, , the abnormal state is a disconnection fault.

The initial data set {NV, F _i } constitutes the first sample data set after the K-means clustering and autoencoding process described above, the elements in the first sample data set being vectors, each vector being TV and F _i , TV={N0C ₁ , N0C ₂ , . . . , N0C _a } is the data after self-encoding, the range of C values is [1, K], a is the node number where data collection is performed, but no faulty nodes are included, and F _i is the abnormal State type value.
2. Second layer data mining

In the active power distribution network, there is a certain connection between the abnormal condition type and the node voltage, and when the abnormal condition occurs, the node voltage waveform will also have a corresponding change. is used to find data samples that have a strong correlation with the corresponding abnormal condition types, and use these data as a training set to train an abnormal condition prediction model, thereby improving the accuracy of the model.

The Apriori algorithm is an association rule algorithm that mines frequent itemsets. First, the two elements of TV and F _i in the first sample data set are merged into one element, that is, the corresponding abnormal state type and node voltage data are formed into one vector, and a new data set M: Form _{ Z _{1 ,} Z ₂ , .

Quantifying the connections between two or more elements in the new vectors using the support and confidence of frequent itemsets, extracting the non-zero vectors that satisfy the support and confidence requirements, and distributing power A second sample data set is formed for training the network anomaly prediction model.

A frequent itemset association rule is a connection between two or more elements in a vector, and such a connection is quantified using the frequent itemset support and confidence, Measures the degree of association between elements. Non-zero vectors that satisfy the support and confidence requirements are extracted to constitute the second sample data set.

であり、
ただし、Ｚ_ｘ、Ｚ_ｙはそれぞれデータセットＭのうちの２つの非０ベクトルである。 Let Z _x , Z _y be two non-zero vectors in the data set M, respectively, the formulas for calculating support and confidence are described below.
The degree of support represents the probability that Z _x and Z _y appear at the same time, and the calculation formula is

and
where Z _x , Z _y are two non-zero vectors in dataset M, respectively.

である。 Reliability represents the probability that _Zy appears at the same time when _Zx appears.

is.

Of the data set M, the non-zero vectors that satisfy the above support and confidence requirements are extracted, then each element of the data set M is divided into [CV, Fi] format to form a second sample data set, CV representing the extracted sample data and corresponding to the abnormal condition type values F _i .

The differences between the raw data set, the first sample data set and the second sample data set are:
The raw data set and the first sample data set have the same dimension and number of samples, and the first sample data set is generated by subjecting the raw data set to K-means clustering and auto-encoding processing. , the second sample data set is generated by subjecting the first sample data set to the association rule algorithm processing, and includes only data samples that have a strong correlation with the corresponding abnormal condition type, and the sample quantity is the first Much less than the one-sample data set.
3. Third layer data mining

After the previous two-layer data mining, the raw data was classified and the data samples with strong correlations with active grid fault condition types were mined to constitute a second sample data set.

As a specific example, there were 40,000 sets of data in the raw data set, and 25,000-30,000 sets of data in the second sample data set after processing, greatly reducing the amount of data.

Referring to FIG. 2, the third layer data mining takes the second sample data set as the training set and adopts the stochastic gradient descent (SGD) algorithm to obtain the optimal parameters of the abnormal condition prediction model through regression training.

The stochastic gradient descent algorithm is an iterative optimization algorithm and is always used to solve model parameter optimization problems in machine learning. The stochastic gradient descent algorithm is an improved form of the gradient descent algorithm and has been successfully applied to large-scale sparse machine learning problems such as text classification, natural language processing. Gradients are for taking the parametric partial derivatives of a multivariate function and constructing them into a vector. The optimal solution for the model parameters is obtained when all partial derivatives on the gradients are all zero. The stochastic gradient descent algorithm randomly uses only one sample data in each iteration, and when the total number of samples is large, only some samples are used for iterative computation, thus shortening the training time of the model.

ただし、ｗは予測モデル関数の重みベクトルであり、ｂは予測モデル関数のインターセプトを表し、ｗ^Ｔ（ＣＶ_ｊ）はベクトルｗとベクトルＣＶ_ｊとのスカラー積を表す。 This embodiment takes the second sample data set as the training set, and assumes that the weight of the voltage data of each node is linear, thereby constructing a linear model function, as follows:

where w is the weight vector of the prediction model function, b represents the intercept of the prediction model function, and w ^T (CV _j ) represents the scalar product of vector w and vector CV _j .

である。 The loss function is for measuring the difference between the actual abnormal condition type value F _i and the model prediction value f(CV _j ), denoted by L(F _i , f(CV _j )), the present invention adopts the logistic regression loss function, and as a calculation formula,

is.

である。 The risk function is the desired value of the loss function, is represented by Er, and is calculated as follows:

is.

であり、
ただし、αは設定されたハイパーパラメータ、例えば０．１であり、αを設定することでパラメータの範囲を小さくして、モデルを簡略化する目的を達成するとともに、モデルによい一般化能力を具備させる。正則化項Ｒ（ｗ）は、ロス関数の複雑さを測定するためのもので、ロス関数パラメータを規制する役割を果たす。本実施例では、Ｌ２正則化を採用し、即ち、

である。 As can be seen from this, the objective function minimizes the risk function, but since there are many historical data and the function is complicated, it easily leads to overfitting of the prediction result. To avoid this situation, the present invention introduces a structural risk function, denoted by Sr,

and
However, α is a set hyperparameter, such as 0.1, and setting α reduces the range of parameters to achieve the purpose of simplifying the model, and the model has good generalization ability. Let The regularization term R(w) is a measure of the complexity of the loss function and serves to regulate the loss function parameters. In this example, we employ L2 regularization, i.e.

is.

であり、
ｗは予測モデル関数の重みベクトルであり、ｂは予測モデル関数のインターセプトを表し、ｎは第２サンプルデータセット総量であり、Ｆ_ｉは異常状態タイプの値である。 In this embodiment, the degree of fit of the prediction model is measured by the optimization objective function, and the smaller the risk function and the structural risk function, the higher the degree of model fit. Finally, as the optimization objective function,

and
w is the weight vector of the prediction model function, b represents the intercept of the prediction model function, n is the total amount of the second sample data set, and _Fi is the value of the abnormal condition type.

ただし、ηは、経時的な学習率であり、算出式として、

であり、
ただし、ｔはタイムステップであり、ｔ_０は始まる時間である。 In the stochastic gradient descent algorithm, iteratively computes with some test set each time, and at each iteration, the model parameters are updated by

However, η is the learning rate over time, and the calculation formula is

and
where t is the time step and _t0 is the starting time.

In this embodiment, from the three-layer data mining structure, the raw data is processed by clustering algorithm, association rule and stochastic gradient descent in order to train a prediction model of active grid abnormal conditions, abnormal condition classification and prediction. Problems with low accuracy and long time can be solved correspondingly, which contributes to taking timely countermeasures so as to shorten the duration of abnormal conditions and reduce power interruption losses.

Other abnormal state sensing prediction methods are similar to the above steps.

In order to verify the effect of the method of this embodiment, this embodiment takes the IEEE three-machine nine-node system as an example, and debugs four common short-circuit faults, load abnormalities and disconnection faults on this model. and collect node voltage data, a short circuit fault sets the fault node to node 8, a load fault installation is load drop at node 8, an open fault installation is line 7-8 disconnection, node Collect the voltage data of the other nodes except 8.
(1) First layer data mining

In the K-mean value clustering algorithm, first, an appropriate K value is specified as the clustering number by the elbow method, and when four types of short-circuit faults occur, the short-circuit voltage width value data of node 2 is extracted and acquired by the elbow method. A curve of error sum of squares SSE and K value is shown in FIG. As can be seen from FIG. 3, when K=5, the rate of decrease in slope is the fastest, and when K>5, the change in SSE is slow, so the reference K value corresponding to the elbow site is 5, and the final The K value may be 5 or more (if the final K value is less than 5, the classification is not fine enough and the data aggregation is not high enough). In this way, the range of K values is narrowed, which contributes to obtaining a more suitable K value quickly.

If the final K value is 5 or more, similarity features between data are not lost, but if the value is too large, the number of clusterings becomes too large again. In accordance with the data sample capacity of this embodiment, in the subsequent self-encoding process, the data format does not frequently overlap because the clustering number is small, and the data format does not become complicated if the clustering number is too large. In order for the two-layer data mining to proceed smoothly, in this experiment, on top of K = 5, a test was conducted in which the K value was sequentially increased and selected, and finally K = 8 was determined, and all data samples are divided into 8 classes, and the data of each clustering have greater similarity.

After the data is classified, the data format is simplified according to the self-encoding rule, taking the voltage data of node 2 after the occurrence of a three-phase short circuit fault as an example, the voltage value is in the interval [-16.01 kV, 15.85 kV] , and when divided into 8 types, each type is [-16.01 kV, -12.03 kV], [-12.03 kV, -8.04 kV], [-8.04 kV, -4.06 kV], [ -4.06kV, -0.08kV], [-0.08kV, 3.90kV], [3.90kV, 7.88kV], [7.88kV, 11.87kV], [11.87kV, 15.85kV ], and when the voltage value at a certain time is 5.36 kV, the data format becomes 206 according to the self-encoding rule, and the voltage value is in the section [-196.582 kV, 206.754 kV], and divided into eight classes, each class is [-196.582 kV, -146.1648 kV], [-146.164 kV, -95.747 kV], [-95 .747 kV, −45.330 kV], [−45.330 kV, 5.086 kV], [5.086 kV, 55.503 kV], [55.503 kV, 105.920 kV], [105.920 kV, 156.337 kV], [156.337 kV, 206.754 kV], and when the voltage value at a certain time is 20.048 kV, the data format becomes 405 according to the self-encoding rule, and node 1 voltage data after the occurrence of a disconnection fault is an example. , the voltage value is in the section [-14.454 kV, 14.468 kV], and when divided into 8 classes, each class is [-14.454 kV, -10.839 kV], [-10.839 kV, -7 .224 kV], [−7.224 kV, −3.608 kV], [−3.608 kV, 0.006 kV], [0.006 kV, 3.621 kV], [3.621 kV, 7.237 kV], [7. 237 kV, 10.852 kV], [10.852 kV, 14.468 kV], and when the voltage value at a certain time is 9.581 kV, the data format becomes 107 according to the self-encoding rule. The voltage data for the remaining nodes are similarly processed and the final results are shown in Table 1.

（２）第２層データマイニング Table 1 Database I Some self-encoding data

(2) Second layer data mining

After the first round of data mining, the data in database I are input into the Apriori algorithm for strong correlation data mining. First, the minimum support is 0.2, the minimum reliability is 0.8, the data sample that satisfies the minimum support in database I is a frequent itemset, and the data that satisfies the minimum reliability among the frequent itemsets is forced. Set as correlation data. Table 2 shows the results of the second round of data mining.

Table 2 Result of second data mining

Taking the association rule {205, 407, 603}→{130} of number 8 as an example, the voltage of node 2 is in the fifth clustering, the voltage of node 4 is in the seventh clustering, and the voltage of node 6 is in the third clustering. , it is very likely that the abnormal condition type is a two-to-ground short circuit. , it is expressed that the fault condition type is very likely to be load fault, and the voltage at node 1 is the fifth When the voltage at node 4 is at the fifth clustering and the voltage at node 6 is at the third clustering, the fault condition type is most likely open fault. Table 3 shows the database II obtained by the association rule algorithm.

（３）３回目のデータマイニング Table 3 Data of database II part

(3) Third data mining

In the third data mining, the hyperparameter α is set to 0.1 and the number of iterations is set to 500. The model parameters trained by inputting the Database II data samples into the stochastic gradient descent algorithm are shown in Table 4.

（４）結果分析 Table 4 Model parameter data

(4) Result analysis

To measure the accuracy of the above classification and prediction models, we use 10000 data samples as the test set. The accuracy of abnormal state classification and prediction is high, with an accuracy rate of 85.30% for the three-phase short-circuit fault prediction model, an accuracy rate of 74.80% for the single-to-ground short-circuit fault model, and an accuracy rate of 74.80% for the two-phase short-circuit fault model. is 78.20%, the accuracy rate of the two-to-ground short circuit prediction model is 87%, the accuracy rate of the load abnormality prediction model is 93.29%, and the accuracy rate of the disconnection failure prediction model is 89.25%.

In this method, from three-layer data mining, firstly by clustering and self-encoding, all data sample formats are simplified, and there are only fixed K kinds of formats in the data of a node, which may generate anomalous data and effectively eliminate the difference between missing dots and normal data (e.g. sudden changes in width values or missing data), and at the same time sufficiently mine the similarity between data, e.g. Among the data, 15 abnormal data samples occur, after clustering and self-encoding, these abnormal data points all belong to the 8th clustering, i.e. their data formats are all 408, in which case, If there is a large amount of data in the eighth clustering, self-encoding will assimilate these abnormal data points into normal data to reduce the effects of abnormal data and missing dots, and if there are only these 15 data samples in the eighth clustering. For example, these data can be reduced by extracting with the second layer strong correlation.

Through the second layer data mining, according to the two quantification criteria of support and confidence, extract the data samples with strong correlation of abnormal conditions, improve the fault tolerance of the algorithm training the model, and improve the accuracy of the model training. To increase.

Compared with the fault tolerance of traditional methods, this method does not directly consider the situation of raw data, but processes the data first, improving fault tolerance, reducing the number of data samples, and shortening the training time. , the time for the model to train is about 350 seconds and the prediction time is within 20 seconds.

Comparing the abnormal state detection method of the present embodiment with the conventional method of directly using the support vector machine to detect abnormal states, Tables 5, 6 and 7 show the prediction accuracy, prediction time and Give comparative results in terms of fault tolerance.

Table 5 Comparison of prediction accuracy

Table 6 Prediction time comparison

Table 7 Fault tolerance comparison (quantification by prediction accuracy)

From the above comparison results, it can be seen that the method of this embodiment clearly outperforms the conventional support vector machine method in terms of prediction time, prediction accuracy and fault tolerance.
Example 2

In one or more embodiments,
a data acquisition module for acquiring node parameter data after a grid abnormal condition occurs;
a failure prediction module that inputs the data to a trained distribution grid abnormal state prediction model and outputs a distribution grid short circuit abnormal state prediction result;
The training process of the grid abnormal condition prediction model includes clustering the grid historical abnormal condition node parameter data, finding data samples with a strong correlation with the grid abnormal condition type by association rule algorithm, and obtaining sample data for training. A data-driven active grid fault detection system is disclosed that includes forming a set and training a grid fault prediction model based on the sample data set.

It should be noted that the specific implementation methods of the above modules in this embodiment have already been described in detail in Embodiment 1, and will not be repeated here.

Although the specific embodiments of the present invention have been described with reference to the above drawings, they are not intended to limit the protection scope of the present invention. It should be understood that various modifications or variations that can be made without significant effort still fall within the protection scope of the present invention.

Claims (3)

A method for active grid abnormal condition detection of a data drive by an active grid abnormal condition sensing system , wherein a data acquisition module acquires node voltage data of a plurality of nodes as node parameter data after a grid abnormal condition occurs, an abnormal condition prediction module inputting the data into a trained distribution network abnormal condition prediction model and outputting a distribution network abnormal condition prediction result;
The active grid abnormal condition sensing system training the grid abnormal condition prediction model comprises:
The node voltages obtained by referring to the node voltage data of the plurality of nodes obtained after occurrence of the distribution network history abnormal state in a plurality of voltage intervals for classifying the node voltage data predetermined for each node. The number C of the class to which the data corresponds is obtained for each node, thereby clustering the node voltage data of each node and self-encoding the data sample after clustering, the node number N and the class of the node voltage data for each node A step of _forming a first sample data set composed of N0C in which the numbers C of the , the elements in the first sample data set are vectors, each vector consisting of TV and F _i , TV={N0C ₁ , N0C ₂ , . . . , N0C _a }, where N represents the position of each node for which data was collected , the range of C values is [1, K], K is the clustering number, and a is the node number for which data was collected. , F _i are values indicating a currently occurring abnormal state type among a plurality of predetermined abnormal state types;
Finding data samples with a strong correlation with the grid abnormal condition type by an association rule algorithm to form a sample data set for training, specifically the TV of each vector in the first sample data set and The two elements _of F _i are merged as one element to form a new data set M _: {Z ₁ , Z ₂ , . where i is the number of new vectors, and between two or more elements of said new vectors using the support and confidence of frequent itemsets quantifying connections and extracting non-zero vectors that satisfy support and confidence requirements to form a second sample data set, a training sample data set;
A step of training the distribution grid abnormal condition prediction model based on the training sample data set,
Specifically, the distribution network abnormal state prediction model is:

and
where w is the weight vector of the predictive model function, b represents the intercept of the predictive model function, w ^T (CV _j ) represents the scalar product of vector w and vector CV _j , and vector CV _j is the training sample. a step;
The step of measuring the fit of the grid abnormal condition prediction model by optimizing an objective function, wherein the optimization of the objective function specifically includes:

and
where L(F _i , f(CV _j )) represents the loss function, α is the hyperparameter and set value, R(w) is the regularization term, n is the sample data set total amount, and F a step where _i is the value of the abnormal condition type;
Based on the obtained sample data set, perform regression training using a stochastic gradient descent algorithm to obtain the optimal parametric solution of the grid abnormal condition prediction model, the stochastic gradient descent algorithm each time Iteratively calculate using a partial test set, and when iterating, the model parameters are updated by the following formula,

However, η is the learning rate over time, and the calculation formula is

and
where t is the time step and t ₀ is the time to begin
A data drive active power distribution network abnormal state detection method characterized by: Specifically, the step of clustering the distribution network history abnormal state node parameter data includes:
classifying the data samples based on Euclidean distance;
determining whether the data sample clustering is complete by the discriminant function;
calculating the sum of error squares, identifying the clustering number referenced by the SSE-K graph, and identifying the final clustering number based on the referenced clustering number;
The active power grid abnormal state detection method for a data drive as claimed in claim 1 . An active grid abnormal condition sensing system for a data drive, comprising:
a data acquisition module for acquiring node voltage data as node parameter data after occurrence of a grid abnormal condition;
an abnormal condition prediction module that inputs the data to a trained distribution network abnormal condition prediction model and outputs a distribution network abnormal condition prediction result;
The power grid abnormal condition model training step includes:
By referring to the node voltage data acquired after the occurrence of the distribution network history abnormal state in a plurality of voltage intervals for classifying the node voltage data predetermined for each node, the class number to which the acquired node voltage data corresponds. N0C in which the node voltage data of each node is clustered and the clustered data samples are self-encoded , and the node number N for each node and the node voltage data class number C are arranged; forming a first sample data set composed of values F _i indicative of a currently occurring abnormal condition type of a plurality of predetermined abnormal condition types, wherein the elements in the first sample data set are vectors; , where each vector consists of TV and F _i , TV={N0C ₁ , N0C ₂ , . . . , N0C _a }, where N represents the position of each node for which data was collected , the range of C values is [1, K], K is the clustering number, and a is the node number for which data was collected. , F _i are values indicating a currently occurring abnormal state type among a plurality of predetermined abnormal state types;
Finding data samples with a strong correlation with the grid abnormal condition type by an association rule algorithm to form a sample data set for training, specifically the TV of each vector in the first sample data set and The two elements _of F _i are merged as one element to form a new data set M _: {Z ₁ , Z ₂ , . where i is the number of new vectors, and between two or more elements of said new vectors using the support and confidence of frequent itemsets quantifying connections and extracting non-zero vectors that satisfy support and confidence requirements to form a second sample data set, a training sample data set;
A step of training the distribution grid abnormal condition prediction model based on the training sample data set,
Specifically, the distribution network abnormal state prediction model is:

and
where w is the weight vector of the predictive model function, b represents the intercept of the predictive model function, w ^T (CV _j ) represents the scalar product of vector w and vector CV _j , and vector CV _j is the training sample. a step;
The step of measuring the fit of the grid abnormal condition prediction model by optimizing an objective function, wherein the optimization of the objective function specifically includes:

and
where L(F _i , f(CV _j )) represents the loss function, α is the hyperparameter and set value, R(w) is the regularization term, n is the sample data set total amount, and F a step where _i is the value of the abnormal condition type;
Based on the obtained sample data set, perform regression training using a stochastic gradient descent algorithm to obtain the optimal parametric solution of the grid abnormal condition prediction model, the stochastic gradient descent algorithm each time Iteratively calculate using a partial test set, and when iterating, the model parameters are updated by the following formula,

However, η is the learning rate over time, and the calculation formula is

and
where t is the time step and t ₀ is the time to begin
An active power grid abnormal state sensing system for a data drive, characterized by:

Images