CN112345888A - Power distribution network fault detection method and system under multi-hop D2D networking - Google Patents

Abstract

The invention relates to a method and a system for detecting a power distribution network fault under multi-hop D2D networking, which comprise the following steps: the distribution network information collector is used for transmitting data streams of the distribution network in a given time period to the edge of the communication network by using multi-hop D2D networking, calculating corresponding characteristic vectors and analyzing related data in real time by using calculation and storage resources in the edge server; an abnormal feature monitor for classifying the training data into abnormal or normal features; the fault detector receives the abnormal characteristic data transmitted from the abnormal characteristic monitor, classifies the abnormal characteristic data according to the time identification and identifies which type of power distribution network fault the abnormal characteristic belongs to. The method can not only overcome the problem of generalization capability of machine learning in the process of processing high-dimensional data, but also accurately identify which type of power distribution network fault the abnormal characteristic belongs to on the premise of lacking fault information.

Description

Power distribution network fault detection method and system under multi-hop D2D networking

Technical Field

The invention relates to the technical field of big data analysis, in particular to a method and a system for detecting a power distribution network fault under multi-hop D2D networking.

Background

The rapid development of current economy and technology has made power consumers place higher demands on the power efficiency and power reliability of the power sector. In the power network, the power distribution network is used as a key ring for directly supplying power to users, and the normal operation of the power distribution network directly influences the power supply efficiency of a power department and the user experience of power utilization. Therefore, when a power distribution network fails, the realization of rapid and accurate fault detection of the power distribution network is a key problem to be solved currently.

The current more and more intelligent electricity meters and the upgraded electricity utilization information acquisition system enable the data volume generated by the power distribution network to show exponential growth, and the power department is provided with more opportunities for deeply knowing the operation state of the power distribution network and the use row and row of users, and the information can be used for fault detection of the power distribution network.

With the exponential increase of the data volume on the power distribution network side, in order to realize the rapid judgment and detection of possible faults in the power distribution network, the relevant data of the power distribution network can be transmitted to the edge of the communication network, and the relevant data is analyzed in real time by utilizing the calculation and storage resources in the edge server to find out the possible fault conditions in the power distribution network.

Disclosure of Invention

In view of this, the present invention provides a method and a system for detecting a power distribution network fault under a multi-hop D2D networking, which can not only overcome the problem of generalization capability of machine learning in processing high dimensional data, but also accurately identify which type of power distribution network fault an abnormal feature belongs to on the premise of lacking fault information.

In order to achieve the purpose, the invention adopts the following technical scheme:

a power distribution network fault detection method under multi-hop D2D networking comprises the following steps:

step S1, acquiring data flow of the power distribution network in a given time period;

step S2, transmitting the obtained data flow to the edge of the communication network, calculating the corresponding characteristic vector, and analyzing the relevant data in real time by using the calculation and storage resources in the edge server to obtain the training data;

s3, constructing a DBN model, and training based on training data to obtain a trained DBN model;

step S4, dividing the training data into abnormal features and normal features according to the trained DBN model;

and step S5, classifying the obtained abnormal characteristic data according to the time identification, and identifying which type of power distribution network fault the abnormal characteristic belongs to by adopting an LSTM network.

Further, the step S2 employs multi-hop D2D networking to transmit the obtained data stream to the edge of the communication network.

Furthermore, the DBN consists of a plurality of RBMs, the electricity utilization data collected by the intelligent electric meters of each region are input into an input layer of the DBN, after training of a first RBM, data of the input layer are mapped to a first hidden layer, and the output of the hidden layer can be used as the input of the next hidden layer; the output of the last layer of RBM is the output of the DBN, and the data after sign extraction by the DBN is transmitted to a fault detector.

Further, the RBM is an undirected neural network model consisting of a layer of visible neurons and a layer of hidden neurons; wherein there is no connection between the neurons of the visible layer and the hidden layer.

Further, in the training process of the RBM, after a piece of data x is assigned to the visible layer, the RBM determines to open or close a hidden layer neuron according to the weight, specifically: the excitation value of each hidden layer neuron is calculated according to the excitation function, then the excitation value is normalized by a sigmoid function to become a probability value in an on state, and then the value is compared with a random value extracted from a [0,1] uniform distribution to decide whether to turn on or off the corresponding hidden layer neuron.

Further, the sigmoidal function uses a Logistic function as the sigmoidal function to normalize the excitation value:

where x represents the excitation value of the hidden layer neurons calculated from the excitation function.

Further, in the training process of the RBM, a method for updating the weight value is as follows:

W←W+λP(h⁽⁰⁾＝1|v⁽⁰⁾)v^(0)T-P(h⁽¹⁾＝1|v⁽¹⁾)v^(1)T

wherein the parameter lambda is used to adjust the update speed, h⁽⁰⁾To use the visible layer vector v⁽⁰⁾Calculated probability of hidden layer neuron being turned on, using h⁽⁰⁾Reconstructing a visible layer vector v⁽⁰⁾To obtain v⁽¹⁾，h⁽¹⁾To use reconstructed visible layer vector v⁽¹⁾And calculating the probability of the hidden layer neuron being opened.

Further, the LSTM network specifically includes:

c_t＝f_t⊙c_t-1+i_t⊙g_t

h_t＝o_t⊙tanh(c_t)

wherein i_tTo the input gate, f_tTo forget the door o_tTo output gate, g_tDetermining how much of the input is saved into the cell state, c_tIs a memory cell, h_t-1For the hidden state, the vector elements are multiplied one by one, and σ denotes the activation function sigmoid.

A power distribution network fault detection system under multi-hop D2D networking comprises a power distribution network information collector, an abnormal characteristic monitor and a fault detector which are connected in sequence;

the distribution network information collector is used for transmitting data streams of the distribution network in a given time period to the edge of the communication network by using multi-hop D2D networking, calculating corresponding characteristic vectors and analyzing related data in real time by using calculation and storage resources in the edge server;

the abnormal feature monitor is used for dividing the training data into abnormal or normal features;

and the fault detector receives the abnormal characteristic data transmitted from the abnormal characteristic monitor, classifies the abnormal characteristic data according to the time identification and identifies which type of power distribution network fault the abnormal characteristic belongs to.

Compared with the prior art, the invention has the following beneficial effects: a

The method can not only overcome the problem of generalization capability of machine learning in the process of processing high-dimensional data, but also accurately identify which type of power distribution network fault the abnormal characteristic belongs to on the premise of lacking fault information.

Drawings

FIG. 1 is a system architecture diagram of the present invention;

FIG. 2 is a schematic diagram of a deep learning-based fault detection method according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a long-short term memory cycle network model according to an embodiment of the present invention;

FIG. 4 is a flowchart illustrating a deep belief network model training method according to an embodiment of the present invention;

fig. 5 is a flowchart of a power distribution network fault detection system based on deep learning in a multi-hop D2D networking according to an embodiment of the present invention.

Detailed Description

The invention is further explained below with reference to the drawings and the embodiments.

Referring to fig. 1, the present invention provides a power distribution network fault detection system under a multi-hop D2D networking, including a power distribution network information collector, an abnormal characteristic monitor and a fault detector, which are connected in sequence;

the distribution network information collector mainly has the functions of transmitting data streams of a distribution network in a given time period to the edge of a communication network by using multi-hop D2D networking, calculating corresponding characteristic vectors and analyzing related data in real time by using calculation and storage resources in an edge server;

the abnormal feature monitor mainly aims to divide training data into abnormal or normal features, the initial classification speed is as fast as possible, and therefore the unit selects a Deep Belief Network (DBN) for the abnormal feature monitoring process.

And the fault detector is mainly used for receiving the abnormal characteristic data transmitted from the abnormal characteristic monitor, classifying the abnormal characteristic data according to the time identification and identifying the type of the power distribution network fault to which the abnormal characteristic belongs. The unit selects a Long and short term memory network (LSTM) capable of processing a complex characteristic mode to determine the specific type of the power distribution network fault.

Preferably, in the embodiment, in order to analyze and process the measurement data generated by the smart meters in each region in detail within a given time period, the distribution network information collector establishes a multi-hop D2D networking between the distribution network and the edge server with strong computing and storing capacity on the base station side of the communication network as a bridge for information exchange

As shown in fig. 2, this embodiment further provides a method for detecting a fault of a power distribution network under a multi-hop D2D networking, including the following steps:

step S1, acquiring data flow of the power distribution network in a given time period;

step S2, transmitting the obtained data flow to the edge of the communication network, calculating the corresponding characteristic vector, and analyzing the relevant data in real time by using the calculation and storage resources in the edge server to obtain the training data;

s3, constructing a DBN model, and training based on training data to obtain a trained DBN model;

step S4, dividing the training data into abnormal features and normal features according to the trained DBN model;

and step S5, classifying the obtained abnormal characteristic data according to the time identification, and identifying which type of power distribution network fault the abnormal characteristic belongs to by adopting an LSTM network.

In the present embodiment, the DBN is composed of a plurality of Restricted Boltzmann Machine (RBM) stacks. And inputting the electricity utilization data collected by the intelligent electric meters in each region into an input layer of the DBN, mapping the data of the input layer to a first hidden layer after first RBM training, and taking the output of the hidden layer as the input of the next hidden layer. The output of the last layer of RBM is the output of the DBN, and the data after sign extraction by the DBN is transmitted to a fault detector.

Referring to fig. 4, in this embodiment, the deep belief network model (DBN) training specifically includes:

step 401: training data X collected from D2D networking and measured by smart meters in each distribution network region_m×n；

Step 402: assigning the training data to a first RBM in the DBN model and fully training the RBM;

and step 3: using the state of the hidden layer neuron of the previous RBM as an input vector of the current RBM;

step 404: fully training the RBM and stacking the RBM above a previous RBM;

step 405: if the training times exceed the total number of RBMs in the DBN model, skipping to 406, otherwise skipping to 403 to restart the training of a new layer of RBMs;

step 406: adjusting and optimizing the parameters from top to top by using a BP neural network;

step 407: performing multiple Gibbs samples in the top RBM;

step 408: transmitting the sampling result downwards to obtain the state of each layer of RBM;

step 409: outputting the hidden layer of the last layer of RBM as feature data after dimension reduction;

step 410: and sending the characteristic data to a fault detection module to judge the specific type of the power distribution network fault.

Preferably, the RBM is an undirected neural network model consisting of a layer of visible neurons and a layer of hidden neurons. Wherein there is no connection between the neurons of the visible layer and the hidden layer. Assuming that the number of samples collected from each area smart meter is m, and the dimension of each sample is n, the training data set of each area can be represented as X_m×nAfter multi-layer RBM training, a data set X with the sample number m and the dimension d can be obtained_m×dD n, so the DBN can achieve the purposes of feature extraction and data dimension reduction.

Assuming that there is already a trained RBM, the weights for each visible layer neuron and hidden layer neuron are represented using the matrix W:

wherein w_i,jRepresenting the weight from the ith visible layer neuron to the jth hidden layer neuron, M representing the number of visible layer neurons, and N generationsNumber of table hidden layer neurons.

Preferably, in the RBM training process, after a piece of data x is assigned to the visible layer, the RBM determines to turn on or off hidden layer neurons according to the weight.

The excitation value of each hidden layer neuron is first calculated according to the excitation function, then normalized by the sigmoid function to become the probability value that they are in the on state, and then compared with a random value extracted from the [0,1] uniform distribution to decide whether to turn on or off the corresponding hidden layer neuron.

Normalizing the excitation value by adopting a Logistic function as a sigmoid function:

where x represents the excitation value of the hidden layer neurons calculated from the excitation function.

Preferably, in this embodiment, in the RBM training process, the method for updating the weight value is as follows:

W←W+λP(h⁽⁰⁾＝1|v⁽⁰⁾)v^(0)T-P(h⁽¹⁾＝1|v⁽¹⁾)v^(1)T

wherein the parameter lambda is used to adjust the update speed, h⁽⁰⁾To use the visible layer vector v⁽⁰⁾Calculated probability of hidden layer neuron being turned on, using h⁽⁰⁾Reconstructing a visible layer vector v⁽⁰⁾To obtain v⁽¹⁾，h⁽¹⁾To use reconstructed visible layer vector v⁽¹⁾And calculating the probability of the hidden layer neuron being opened.

After training one RBM, the next RBM is stacked on the upper layer of the RBM, and the hidden layer unit of the previous RBM is used as the visible layer vector of the next RBM for training. The DBN is formed by stacking a plurality of RBMs, each RBM can be used as a feature extractor of different layers, and as the number of layers of the RBMs is increased, features extracted by the model are more and more abstract, so that the DBN can represent complex data structures in high-dimensional data.

In this embodiment, the LSTM includes three main stages:

a forgetting stage: the forgetting stage has the main function of selectively forgetting input data transmitted by the previous node;

selecting a memory stage: the function of the selective memorization phase is to selectively 'memorize' the input data;

an output stage: the main function of the output stage is to determine which data will be output as the current state.

The LSTM network model is described as:

c_t＝f_t⊙c_t-1+i_t⊙g_t

h_t＝o_t⊙tanh(c_t)

wherein i_tTo the input gate, f_tTo forget the door o_tTo output gate, g_tDetermining how much of the input is saved into the cell state, c_tIs a memory cell, h_t-1For the hidden state, the vector elements are multiplied one by one, and σ denotes the activation function sigmoid.

The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.

Claims (9)

1. A power distribution network fault detection method under multi-hop D2D networking is characterized by comprising the following steps:

step S1, acquiring data flow of the power distribution network in a given time period;

step S2, transmitting the obtained data flow to the edge of the communication network, calculating the corresponding characteristic vector, and analyzing the relevant data in real time by using the calculation and storage resources in the edge server to obtain the training data;

s3, constructing a DBN model, and training based on training data to obtain a trained DBN model;

step S4, dividing the training data into abnormal features and normal features according to the trained DBN model;

and step S5, classifying the obtained abnormal characteristic data according to the time identification, and identifying which type of power distribution network fault the abnormal characteristic belongs to by adopting an LSTM network.

2. The method for detecting the fault of the power distribution network under the multi-hop D2D networking according to claim 1, wherein the step S2 adopts the multi-hop D2D networking to transmit the obtained data stream to the edge of the communication network.

3. The method as claimed in claim 1, wherein the DBN is composed of a plurality of RBMs, the electricity consumption data collected by the smart meters of each station are input to the input layer of the DBN, after training of the first RBM, the data of the input layer are mapped to the first hidden layer, and the output of the hidden layer is used as the input of the next hidden layer; the output of the last layer of RBM is the output of the DBN, and the data after sign extraction by the DBN is transmitted to a fault detector.

4. The method for detecting the fault of the power distribution network under the multi-hop D2D networking as claimed in claim 3, wherein the RBM is a non-directional neural network model consisting of a layer of visible neurons and a layer of hidden neurons; wherein there is no connection between the neurons of the visible layer and the hidden layer.

5. The method as claimed in claim 3, wherein in the training process of the RBM, after assigning a piece of data x to a visible layer, the RBM determines to turn on or off a hidden layer neuron according to a weight value, specifically: the excitation value of each hidden layer neuron is calculated according to the excitation function, then the excitation value is normalized by a sigmoid function to become a probability value in an on state, and then the value is compared with a random value extracted from a [0,1] uniform distribution to decide whether to turn on or off the corresponding hidden layer neuron.

6. The method of claim 5, wherein the sigmoid function normalizes the excitation values by using a Logistic function as the sigmoid function:

where x represents the excitation value of the hidden layer neurons calculated from the excitation function.

7. The method as claimed in claim 3, wherein in the training process of the RBM, the method for updating the weight values is as follows:

W←W+λP(h⁽⁰⁾＝1|v⁽⁰⁾)v^(0)T-P(h⁽¹⁾＝1|v⁽¹⁾)v^(1)T

wherein the parameter lambda is used to adjust the update speed, h⁽⁰⁾To use the visible layer vector v⁽⁰⁾Calculated probability of hidden layer neuron being turned on, using h⁽⁰⁾Reconstructing a visible layer vector v⁽⁰⁾To obtain v⁽¹⁾，h⁽¹⁾To use reconstructed visible layer vector v⁽¹⁾And calculating the probability of the hidden layer neuron being opened.

8. The method for detecting the fault of the power distribution network under the multi-hop D2D networking according to claim 1, wherein the LSTM network specifically comprises:

c_t＝f_t⊙c_t-1+i_t⊙g_t

h_t＝o_t⊙tanh(c_t)

wherein i_tTo the input gate, f_tTo forget the door o_tTo output gate, g_tDetermining how much of the input is saved into the cell state, c_tIs a memory cell, h_t-1For the hidden state, the vector elements are multiplied one by one, and σ denotes the activation function sigmoid.

9. A power distribution network fault detection system under multi-hop D2D networking is characterized by comprising a power distribution network information collector, an abnormal characteristic monitor and a fault detector which are sequentially connected;

the distribution network information collector is used for transmitting data streams of the distribution network in a given time period to the edge of the communication network by using multi-hop D2D networking, calculating corresponding characteristic vectors and analyzing related data in real time by using calculation and storage resources in the edge server;

the abnormal feature monitor is used for dividing the training data into abnormal or normal features;

and the fault detector receives the abnormal characteristic data transmitted from the abnormal characteristic monitor, classifies the abnormal characteristic data according to the time identification and identifies which type of power distribution network fault the abnormal characteristic belongs to.

Images