CN115659254A - Power quality disturbance analysis method for power distribution network with bimodal feature fusion - Google Patents
Power quality disturbance analysis method for power distribution network with bimodal feature fusion Download PDFInfo
- Publication number
- CN115659254A CN115659254A CN202211134779.7A CN202211134779A CN115659254A CN 115659254 A CN115659254 A CN 115659254A CN 202211134779 A CN202211134779 A CN 202211134779A CN 115659254 A CN115659254 A CN 115659254A
- Authority
- CN
- China
- Prior art keywords
- distribution network
- power distribution
- power quality
- power
- quality disturbance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention provides a power quality disturbance analysis method of a power distribution network with fusion of bimodal characteristics, which comprises the following steps: 1. carrying out modal transformation on the power distribution network power quality time sequence signals acquired and processed on site by using a Markov conversion field to obtain a polymerization image of dynamic transition probability; 2. performing feature extraction on the aggregation image of the dynamic transition probability by adopting a convolutional neural network to obtain a first feature vector; 3. performing feature extraction on the power quality time sequence signals of the power distribution network acquired and processed on site by using a gate control circulation unit to obtain a second feature vector; 4. performing feature fusion on the first feature vector and the second feature vector by adopting a method based on deep learning feature fusion to obtain fused power quality disturbance features of the power distribution network; 5. and classifying the power quality disturbance characteristics of the power distribution network through a power quality disturbance classifier of the power distribution network.
Description
Technical Field
The invention relates to a power quality disturbance analysis method of a power distribution network with fusion of bimodal characteristics, and belongs to the field of power quality disturbance identification of a power system.
Background
The proportion of new energy power accessed to a power grid is higher and higher, the scale of the power grid is continuously enlarged, the structural form and the operation mode of a power system are changed, and more randomness and nonlinear characteristics are introduced when the power system stably operates; power quality disturbances are one of the major problems in power systems; disturbances or interference are typically caused by many factors, such as nonlinear or fluctuating loads, power electronics, system faults, etc., which can, in extreme cases, cause distortion of the waveform; without taking identification and control actions to properly prevent and mitigate these disturbances, it is possible to generate an overall outage event of the transmission and distribution network, causing significant social impact and huge economic losses.
The actual power quality disturbance is divided into single disturbance and composite disturbance, but in a power distribution network containing photovoltaic power generation, the simple single type of disturbance is less, the composite disturbance, especially the double disturbance, is more, the actual power quality disturbance has a characteristic aliasing phenomenon, and the difficulty of accurate identification is obviously increased due to the mode similarity between the composite disturbance and the single disturbance; the traditional electric energy quality disturbance identification method is complex in calculation or requires certain manual intervention to extract characteristic quantities; under the condition of complex waveform change at present, accurate identification cannot be realized or real-time calculation cannot be realized; with the development of big data and artificial intelligence, the adoption of related methods makes it possible to accurately identify the big data in real time.
At present, deep learning is developed rapidly and widely applied, the generalization of the problem solving method can be well applied to the field of electric energy quality disturbance identification, and better self-adaptive capability can be still kept under certain noise; however, in order to solve the problem of complex power quality recognition, the recognition accuracy of a single deep learning model still needs to be improved.
Disclosure of Invention
The invention provides a bimodal feature fused power quality disturbance analysis method for a power distribution network, and aims to provide a method capable of identifying and classifying power quality disturbance of the power distribution network.
The technical scheme of the invention is as follows: a method for analyzing power quality disturbance of a power distribution network with fusion of bimodal characteristics comprises the following steps:
1. carrying out modal transformation on the power quality time sequence signals of the power distribution network acquired and processed on site by using a Markov conversion field to obtain a polymerization image of dynamic transfer probability;
2. performing feature extraction on the aggregation image of the dynamic transition probability by adopting a convolutional neural network to obtain a first feature vector;
3. performing feature extraction on the power quality time sequence signals of the power distribution network acquired and processed on site by adopting a gate control circulation unit to obtain a second feature vector;
4. performing feature fusion on the first feature vector and the second feature vector by adopting a method based on deep learning feature fusion to obtain fused power quality disturbance features of the power distribution network;
5. and classifying the power quality disturbance characteristics of the power distribution network through a power quality disturbance classifier of the power distribution network.
Further, the method for obtaining the aggregation image of the dynamic transition probability by carrying out modal transformation on the power quality time sequence signals of the power distribution network acquired and processed on site by using the Markov conversion field specifically comprises the following steps:
1-1, discretizing a time sequence signal;
1-2, establishing a Markov probability transition matrix and calculating transition probability;
1-3, calculating a Markov conversion field and aggregating the Markov conversion field into a two-dimensional image.
Further, the discretization of the timing signal specifically includes: power quality time sequence signal of power distribution network acquired and processed on site
Transmitting the power quality timing sequence signal to a background analysis system of the power distribution network
Discretization required for background analysis of actual distribution network
A unit of quantile.
Further, the establishing a markov probability transition matrix and calculating a transition probability specifically includes:
firstly, counting the power quality time sequence signal of the power distribution network
From the sampling instant to the next sampling instant corresponds to the first
Conversion of a fractional unit to
The number of elements of the quantile unit;
then, the statistics of each cross quantile unit are recorded to a Markov probability transition matrix
The preparation method comprises the following steps of (1) performing;
finally, the Markov probability transition matrix is used
The statistics of each span quantile unit in each line is divided by the total statistics of the span quantile units in the line, and the transition probability of each span quantile unit is converted into the transition probability of each span quantile unit
。
Further, the calculating the markov transform field and aggregating into the two-dimensional image specifically includes:
1) Cross reference Markov probability transition matrix and cross reference determination according to equation (1)
，…，
Transition probability between, finally obtaining the dimension of
Markov transition field
(ii) a Markov transition field
Each element in (1)
Representing a time interval of
A transition probability between points of (a);
2) By using fuzzy kernels
For each non-overlap
Pixels in the block are averaged toObtaining a two-dimensional image of a polymerized Markov conversion field, and storing the image in a JPG format to obtain a polymerized image of dynamic transition probability;
Further, the convolutional neural network comprises an input layer, a convolutional layer, a normalization layer, an excitation layer and a pooling layer; the method for extracting the features of the aggregated image of the dynamic transition probability by adopting the convolutional neural network to obtain a first feature vector specifically comprises the following steps:
2-1, carrying out convolution calculation of an image mode on the aggregation image of the dynamic transition probability by utilizing a convolution layer; the specific calculation method of the convolution calculation is as formula (2):
wherein the content of the first and second substances,
indicating the second in a convolutional layer
A layer;
is a convolution layer of
Extracted in a layer
The power quality disturbance characteristic vector of the power distribution network in each image mode;
is an activation function;
as extracted by a convolution kernel
Characteristic vectors of power quality disturbance image modalities of the power distribution network;
is as follows
Parameters of the layer convolution kernel, here
And
respectively indicate the input of
Is first and second
The power quality disturbance characteristic vector of each power distribution network;
is as follows
Biasing of the layers;
is a convolution operation;
2-2, normalizing the power quality disturbance characteristic vector of the power distribution network, extracted from the dynamic transition probability aggregation image by the convolution layer, by using a normalization layer, wherein the specific calculation method is as shown in a formula (3):
wherein the content of the first and second substances,
representing the power quality disturbance characteristic vector of the distribution network extracted by the convolution layer,
representing the data after the characteristic vector of the power quality disturbance of the normalized distribution network,
representing the mean square error of the power quality disturbance characteristic vector of the power distribution network;
representing the mean value of the power quality disturbance characteristic vector of the power distribution network;
2-3, nonlinearizing the normalized power quality disturbance characteristic vector of the power distribution network by utilizing the excitation layer, specifically as a formula (4):
wherein the content of the first and second substances,
is the output nonlinear power quality disturbance characteristic vector of the distribution network, here
The input normalized power quality disturbance characteristic vector of the power distribution network is obtained;
2-4, highlighting the key part of the power quality disturbance characteristic vector of the power distribution network by using a pooling layer, wherein the calculation method of the maximum pooling method is specifically shown as a formula (5):
wherein, the first and the second end of the pipe are connected with each other,
indicates the number of network layers of the pooling layer,
the number of pooled kernels is shown,
is shown as
Offset vector of layer, here
The function of the excitation is represented by,
is shown in
The layer carries out feature extraction on the aggregation image of the dynamic transition probability to obtain a first feature vector,
indicating input in the second place
Layer one
And (4) carrying out nonlinear power quality disturbance characteristic vector on the power distribution network after the individual pooling kernel.
Further, adopt the gate control circulation unit to carry out the feature extraction to distribution network electric energy quality sequential signal and obtain the second eigenvector, specifically include:
3-1, inputting a power quality time sequence signal of the power distribution network acquired on site into a component module updating gate and a resetting gate of a gating circulation unit;
3-2, controlling the degree of state information of the power distribution network power quality time sequence signal at the previous moment brought into the current state by updating a door, so that the power distribution network power quality disturbance characteristics at the previous time step are memorized and stored in the information of the current time step;
and 3-3, combining the power quality disturbance characteristic information input into the power distribution network at the current sampling moment with the disturbance characteristic memory at the previous sampling moment through a reset door to calculate a candidate state.
Further, the method for performing feature fusion on the first feature vector and the second feature vector based on the deep learning feature fusion to obtain the fused power quality disturbance feature of the power distribution network specifically includes:
4-1, setting the first characteristic vector as
The second feature vector is
；
4-2, when the background analysis system of the power distribution network has insufficient computing resources, selecting a feature fusion method for feature addition, and adding each element in the first feature vector and each corresponding element in the second feature vector if the two feature vectors have the same dimension
Obtaining a fused feature vector; if the two feature vector dimensions are different, then the two feature vectors are switched on firstThe two eigenvectors have the same dimension through linear transformation, and then each element in the first eigenvector is added with each corresponding element in the second eigenvector
Obtaining a fused feature vector;
4-3, when the background analysis system of the power distribution network has sufficient computing resources, preferably selecting a feature fusion method for feature splicing, and carrying out vector splicing on elements in the first feature vector and elements in the second feature vector
And forming a new feature vector.
Further, the classification of the power quality disturbance characteristics of the power distribution network by the power quality disturbance classifier of the power distribution network is specifically to classify the power quality disturbance characteristics of the power distribution network by a fully-connected neural network; the fully-connected neural network adopts a back propagation algorithm, and updates each parameter of deep learning by calculating a loss value, wherein the loss value calculating method is as shown in a formula (10):
wherein the content of the first and second substances,
the output power quality disturbance type of the power distribution network is represented,
of the actual power quality disturbance type of the distribution network, here
And classifying the number of the types of the power quality disturbance of the power distribution network.
Further, the classifying the power quality disturbance characteristics of the power distribution network by the power quality disturbance classifier of the power distribution network specifically includes classifying the finally extracted power quality disturbance characteristics of the power distribution network by a SoftMax function, which is specifically shown in formula (11):
wherein, the first and the second end of the pipe are connected with each other,
the probability of various disturbances of the power quality of the power distribution network output by the power quality disturbance classifier of the power distribution network is shown, here
Represents the output perturbation feature vector of the fully-connected layer, here
Sort the number of classes for the SoftMax function, here
First finger
And (4) disturbance.
The invention has the beneficial effects that:
1) According to the invention, the power quality time sequence signals of the power distribution network are converted into Markov conversion fields, so that the power quality time sequence signal conversion probability of the power distribution network under different time scales is effectively described; compared with the power quality time sequence signals of the power distribution network at the moments before and after single comparison, the method can better depict the condition of power quality disturbance change of the power distribution network;
2) Aiming at the problem that the power quality disturbance of a complex power distribution network is difficult to accurately identify by the existing method, a Markov conversion field is utilized to carry out modal transformation on a power quality time sequence signal of the power distribution network to obtain a polymerization image of dynamic transition probability, so that input data are enhanced, and sufficient data for describing the power quality change of the power distribution network are formed; performing feature extraction on the aggregation image of the dynamic transition probability by using a convolutional neural network suitable for extracting image features, performing feature extraction on a power distribution network power quality time sequence signal by using a gate control cycle unit capable of efficiently extracting the time sequence signal, and fully extracting the input disturbance change features of the power distribution network power quality across time scales;
3) According to the method, the features extracted from the power quality data of the power distribution network in two different modes are fused through a deep learning feature fusion method, and then the multi-label classification is carried out through the power quality disturbance classifier of the power distribution network, so that the power quality disturbance signal features of the power distribution network in two different modes can be more fully utilized, the accuracy of the power quality disturbance identification of the complex power distribution network can be improved, and a high identification rate can be still kept in a high-noise environment; the convolutional neural network and the gated circulation unit are calculated in parallel, so that the calculation speed is increased;
4) According to the method, the aggregated image of the dynamic transition probability of the power quality time correlation of the power distribution network is generated and is combined with the original time sequence signal, so that the data input can be enhanced, and the characteristics of the power quality disturbance signal of the power distribution network can be described from multiple angles; through the feature extraction and fusion of two deep learning networks, the algorithm performance can be effectively improved, and the robustness of the model in a high-noise environment can be enhanced.
5) The method is particularly suitable for real-time accurate identification and classification of the power quality disturbance of the power distribution network containing photovoltaic power generation.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
Figure 2 is a diagram of a markov transition field.
FIG. 3 is a schematic diagram of a gated loop unit.
Fig. 4 is a schematic diagram of a power quality time sequence signal of the power distribution network after field acquisition and processing.
FIG. 5 is a schematic diagram of a training result of a power distribution network power quality disturbance analysis method with the bimodal feature fusion as an integral model after training.
Detailed Description
As shown in fig. 1, a method for analyzing power quality disturbance of a power distribution network with fusion of bimodal features includes the following steps:
1. carrying out modal transformation on the power quality time sequence signals of the power distribution network acquired and processed on site by using a Markov conversion field to obtain a polymerization image of dynamic transfer probability;
2. performing feature extraction on the aggregation image of the dynamic transition probability by adopting a convolutional neural network to obtain a first feature vector;
3. performing feature extraction on the power quality time sequence signals of the power distribution network acquired and processed on site by using a gate control circulation unit to obtain a second feature vector;
4. performing feature fusion on the first feature vector and the second feature vector by adopting a method based on deep learning feature fusion to obtain fused power quality disturbance features of the power distribution network;
5. and classifying the power quality disturbance characteristics of the power distribution network through a power quality disturbance classifier of the power distribution network.
In the invention, a parallel computing mode is adopted between the step 2 and the step 3, and the convolution neural network in the step 2 and the gated cyclic neural network in the step 3 are simultaneously computed in parallel; the step 2 and the step 3 have no integral sequence in the execution sequence; the parallel computation of the step 2 and the step 3 can fully utilize a Graphics Processing Unit (GPU) to carry out accelerated operation.
Common power quality disturbances of the distribution network include: voltage sag, voltage flicker, transient oscillation, voltage sag + harmonic, voltage short interruption, voltage sag + harmonic, transient impulse, voltage flicker + harmonic, and the like.
The analysis method for the power quality disturbance of the power distribution network with the bimodal feature fusion is a method used in a background analysis system of the power distribution network, and is mainly used for identifying and classifying complex power quality disturbance; the method is particularly suitable for accurately identifying and classifying the power quality disturbance of the power distribution network containing photovoltaic power generation in real time; the method considers the volatility and randomness of the photovoltaic power supply, and can accurately identify and classify the energy quality disturbance of the power distribution network containing photovoltaic power generation in real time.
When the method is used for accurately identifying and classifying the energy quality disturbance of the power distribution network containing photovoltaic power generation in real time, the power distribution network refers to the power distribution network containing photovoltaic power generation, power generation and power utilization coexist in the power distribution network, the power quality disturbance of the power distribution network is frequent, compared with new energy power generation with mature technologies and few disturbance factors, such as wind power generation and nuclear power generation, the power quality disturbance of the power distribution network containing photovoltaic power generation is mainly caused by photovoltaic power generation abnormity caused by various factors, and the various factors comprise: natural factors (such as light, rain, snow, storm, etc.), equipment aging and malfunction, human causes, etc., such as: the photovoltaic power generation mainly depends on a photovoltaic panel, the photovoltaic panel usually has a large area and is exposed outside, and if any small part of the photovoltaic panel has a problem, the problem of electric energy quality disturbance can be caused, so the photovoltaic power generation has the characteristic of strong vulnerability; and photovoltaic power generation in a power distribution network usually occurs in a distributed mode, the distribution is wide and not concentrated, and the problems of electric energy quality disturbance such as voltage temporary rise, voltage temporary fall, voltage flicker, transient oscillation, voltage temporary fall + harmonic wave, voltage short-time interruption, voltage temporary rise + harmonic wave, transient impact, voltage flicker + harmonic wave and the like are more prominent in the power distribution network containing the photovoltaic power generation.
The field acquisition processing means that a scanner or intelligent measuring equipment is adopted to perform regional real-time data acquisition and transmit the data to a power distribution network background analysis system through wireless communication, the power distribution network background analysis system automatically corrects error or missing data, and the correction method comprises the following two steps: 1. searching a power quality disturbance database in a power distribution network background analysis system within a specified time, searching data similar to the situation, performing interpolation processing on missing data, and performing deletion and refill processing on error data; 2. if the search is successful, the overall modified data is normalized after modification; if the search fails, the left and right adjacent data of the missing or error data are summed and averaged, the data needing to be modified is replaced or supplemented, and the overall modified data is normalized.
When the Markov conversion field is used for carrying out modal conversion on the power quality time sequence signals of the power distribution network collected and processed on site, the programming languages of practical engineering applications such as Python and C are preferably selected for carrying out modal conversion.
As shown in fig. 2, the obtaining of the aggregate image of the dynamic transition probability by performing modal transformation on the power quality time sequence signal of the power distribution network acquired and processed on site by using the markov conversion field specifically includes the following steps:
1-1, discretization of time sequence signals: because the power quality time sequence signals of the power distribution network have more complex numerical values, the discretization of the time sequence signals can facilitate the subsequent data analysis and processing; power quality time sequence signal of power distribution network acquired and processed on site
(which comprises
The element refers to a current signal or a voltage signal in the electric quantity,
data volume collected at a certain sampling frequency within a specified time) is transmitted to a power distribution network background analysis system, so as to obtain a power distribution network power quality time sequence signal with a large number of elements
Discretization required for background analysis of actual distribution network
Quantile cells (e.g.:
preferably 4, then this means that
The elements in the Chinese character are divided into four quantile units of 0-25%, 25-50%, 50-75% and 75-100% according to the size of the numerical value after being arranged from small to large, specifically the four quantile units are
The elements in the Chinese character are divided into 4 steps after being arranged from small to large according to the size of the numerical value, if the size of the element is in the smallest step, the element is divided into 0% -25%, if the size of the element is in the second smallest step, the element is divided into 25% -50%, if the size of the element is in the third smallest step, the element is divided into 50% -75%, if the size of the element is in the third largest step, the element is divided into 75% -100%), and the quantiles are firstly used for quantiles
Quantizing each value of the power quality time sequence signal of the power distribution network, and identifying quantiles to obtain the power quality time sequence signal of the power distribution network
Each element in the group is classified into a corresponding quantile unit area;
1-2, establishing a Markov probability transition matrix and calculating transition probability: the method aims to find out the probability that the power quality time sequence signal of the power distribution network is transferred from the current quantile unit to other quantile units; construct a
Markov probability transition matrix of dimension
（
The number of the quantile unit is the number of the quantile unit,
each element in (1) is
) The method for constructing the selection comparison table of each element in the Markov conversion field by using the selection comparison table as 1-3 specifically comprises the following steps:
firstly, counting the power quality time sequence signal of the power distribution network
From the sampling instant to the next sampling instant correspondingly from the first
Conversion of unit of quantile to
The number of elements of the quantile unit;
then, the statistics of each cross quantile unit are recorded to a Markov probability transition matrix
In (e.g. distribution network power quality timing signals)
When the total number of elements from the sampling time to the next sampling time is 100 corresponding to the situation of transferring from the 1 st quantile unit to the 2 nd quantile unit, the statistic number is 100);
finally, the Markov probability transition matrix is used
The statistics of each trans-quantile unit in each line is divided by the sum of the statistics of the trans-quantile unit in the line and converted into the transition probability of each trans-quantile unit
；
1-3, calculating a Markov conversion field and aggregating the Markov conversion field into a two-dimensional image: generating image modal input data for enhancing background analysis of the power distribution network; markov probability transition matrix against steps 1-2
And determined by comparison according to formula (1)
，…，
Transition probability (e.g. between)
The first row of the first element in (1) refers to
Transition to
By itself, i.e. from
Transition to place quantile cell
Place quantile unit, against Markov probability transition matrix
Finding corresponding transition probabilities
As
First element of the first row of (1); also for example
First row of (3)
An element means
Transition to
I.e. from
Transition to place quantile cell
Place quantile unit, against Markov probability transition matrix
Finding corresponding transition probabilities
As
First row of (1)
Element) to finally obtain a dimension of
Markov transition field
(ii) a Markov transition field
Each element in (1)
Representing a time interval of
A transition probability between points of (a); for example, in the case of a liquid,
illustrating that there is only one interval in the transition along the time axis; main diagonal line
This is a special case of a time interval of 0, which is obtained every timeThe probability of each fractional digit to itself, namely the self-transition probability; but consider if
Is larger if the original Markov transition field is used
The direct generation of the image may result in an oversized image, occupy more computer storage space, is not conducive to the rapid analysis of the background analysis system of the power distribution network, and occupies too many storage resources, by using the fuzzy kernel
For each non-overlap
Averaging pixels in the blocks to obtain a two-dimensional image of a polymerized Markov conversion field, storing the image in a JPG format to obtain a polymerized image of dynamic transition probability, and taking the polymerized image of the dynamic transition probability as input data of a group of image modalities of a power distribution network background analysis system;
The convolutional neural network comprises an input layer, a convolutional layer, a normalization layer, an excitation layer and a pooling layer; the specific building method of the convolutional neural network preferably uses TensorFlow2.0 functional API to build the convolutional neural network.
The method for extracting the features of the dynamic transition probability aggregated image by adopting the convolutional neural network to obtain the first feature vector specifically comprises the following steps:
2-1, performing image mode convolution calculation on the aggregation image of the power quality dynamic transfer probability of the power distribution network obtained through conversion in the step 1-3 by utilizing a convolution layer; under the condition of being influenced by photovoltaic power generation, compared with other communication signal disturbance, gear signal disturbance and the like, the power quality disturbance of the power distribution network has more complex composite disturbance, and in order to accurately extract the power quality disturbance characteristics of the power distribution network in an image mode, a specific calculation method of the convolution calculation is as shown in a formula (2):
in the formula (2), the first and second groups of the compound,
indicating the second in a convolutional layer
A layer;
is a convolution layer of
Extracted in a layer
The power quality disturbance characteristic vector of the power distribution network in each image mode;
is an activation function;
is extracted by a convolution kernel
Characteristic vectors of the power quality disturbance image modalities of the individual power distribution networks;
is as follows
Parameters of the layer convolution kernel, here
And
first of finger input
A first and a second
The power quality disturbance characteristic vector of each power distribution network;
is as follows
Biasing of the layers;
is a convolution operation;
2-2, normalizing the power quality disturbance characteristic vector of the power distribution network, which is extracted by the convolution layer from the aggregation image of the power quality dynamic transfer probability of the power distribution network, by using a normalization layer, preferably adopting z-score normalization for accelerating the calculation speed of a background analysis system of the power distribution network and improving the accuracy of characteristic extraction, wherein the specific calculation method is as shown in a formula (3):
in the formula (3), the first and second groups,
representing the power quality disturbance characteristic vector of the distribution network extracted by the convolution layer,
representing the data after normalizing the power quality disturbance characteristic vector of the power distribution network,
representing the mean square error of the power quality disturbance characteristic vector of the power distribution network;
representing the mean value of the power quality disturbance characteristic vector of the power distribution network;
2-3, nonlinearizing the normalized power quality disturbance characteristic vector of the power distribution network by using an excitation layer, preferably performing characteristic mapping by using a ReLU function, specifically as a formula (4):
in the formula (4), the first and second groups,
is the output nonlinear power quality disturbance characteristic vector of the distribution network, here
The input normalized power quality disturbance characteristic vector of the power distribution network is obtained;
2-4, wherein the pooling layer is used for highlighting the key part of the power quality disturbance characteristic vector of the power distribution network; the pooling layer preferably adopts a maximum pooling method to highlight the key part of the power quality disturbance characteristic vector of the power distribution network, and the calculation method of the maximum pooling method is specifically as shown in a formula (5):
in the formula (5), the first and second groups,
indicates the number of network layers of the pooling layer,
the representation of the pooled kernel is shown,
is shown as
Offset vector of layer, here
The function of the excitation is represented by,
is shown in
The layer carries out feature extraction on the aggregation image of the dynamic transition probability to obtain a first feature vector,
indicating input in the second place
Layer one
And (4) power quality disturbance characteristic vectors of the power distribution network after the individual pooling cores are subjected to nonlinear transformation.
As shown in fig. 3, in the specific construction method of the gated cyclic unit, preferably, a tensrflow2.0 functional API is used to construct the gated cyclic unit, which is calculated simultaneously with the convolutional neural network to form parallel calculation, and the calculation formula of the gated cyclic unit is as formula (6) -formula (9):
equation (6) -equation (9):
to update the door;
to reset the gate;
the power quality time sequence signal of the power distribution network is input at the current moment;
is a current hidden unit candidate state;
hiding the state of the unit for the current moment;
hiding the state of the cell at the previous time;
、
、
are trainable weight coefficients;
、
、
are all bias matrices;
tanh and Sigmoid activation functions are referred to as a Sigmoid activation function and a Tanh activation function, respectively;
representing a hadamard product operation.
As shown in fig. 3, the performing feature extraction on the power quality timing signal of the power distribution network by using the gate control cycle unit to obtain a second feature vector specifically includes the following steps:
3-1, inputting the power quality time sequence signals of the power distribution network collected on site into a gate control cycle unit to form a module updating gate and a reset gate;
3-2, controlling the degree of bringing the state information of the power distribution network power quality time sequence signal at the previous moment into the current state by updating a door, so that the power distribution network power quality disturbance characteristics at the previous time step are memorized and stored in the information of the current time step;
3-3, combining the power quality disturbance characteristic information input into the power distribution network at the current sampling moment with the disturbance characteristic memory at the previous sampling moment through a reset door to calculate a candidate state; the method has the effects that the characteristic information quantity of the power quality time sequence signals of the power distribution network with small historical influence is forgotten, and the characteristic that the power quality of the power distribution network is disturbed can be efficiently expressed by outputting.
Performing feature fusion on the first feature vector and the second feature vector by using a deep learning feature fusion-based method to obtain fused power quality disturbance features of the power distribution network; according to the actual power quality disturbance identification requirement of the power distribution network and the hardware configuration of the background analysis system of the power distribution network, feature fusion is preferably performed on the first feature vector and the second feature vector in a feature splicing or feature adding mode, and the method specifically comprises the following steps:
4-1, setting the first characteristic vector as
The second feature vector is
；
4-2, when the background analysis system of the power distribution network has insufficient computing resources, preferably selecting a feature fusion method of feature addition, and adding each element in the first feature vector and each corresponding element in the second feature vector if the two feature vectors have the same dimension at the moment
Obtaining a fused feature vector; if the two eigenvectors have different dimensions, the dimensions of the two eigenvectors are made to be the same through linear transformation, and then each element in the first eigenvector is added with each corresponding element in the second eigenvector
Obtaining a fused feature vector; the feature fusion method of feature addition has the advantages that the fused feature dimension is the same as the dimensions of the initial two feature vectors, and less computing resources are occupied; when the background analysis system of the power distribution network has sufficient computing resources, the feature fusion method of feature splicing is optimized, and elements in the first feature vector and elements in the second feature vector are subjected to vector splicing
And forming a new feature vector, wherein the fused features are more complete, the feature loss is less, but the feature vector dimension is increased, and more computing resources are needed.
The specific implementation scheme of performing feature fusion on the first feature vector and the second feature vector in a feature splicing manner preferably utilizes python language to realize feature addition or feature splicing, so as to realize deep learning parallel computation.
The parallel computation refers to the simultaneous parallel computation of multiple deep learning models by using a graphics card GPU.
The power distribution network power quality disturbance characteristics are classified through the power distribution network power quality disturbance classifier, and preferably, the power distribution network power quality disturbance characteristics are classified through a full-connection neural network.
The fully-connected neural network is preferably constructed using a TensorFlow2.0 functional API.
The fully-connected neural network adopts a back propagation algorithm, and the method mainly aims to compare and calculate the real power quality disturbance condition of the power distribution network with the power quality disturbance condition of the power distribution network judged by the method, and update each parameter of deep learning in the method by calculating a loss value, wherein the deep learning comprises the links of a convolutional neural network, a gated cycle unit, feature fusion, a fully-connected neural network, a power quality disturbance classifier of the power distribution network and the like, and the loss value calculating method is as shown in a formula (10):
in the formula (10), the first and second groups of the chemical reaction are shown in the formula,
the output power quality disturbance type of the power distribution network is represented,
of the actual type of power quality disturbance of the distribution network, here
And classifying the number of the types of the power quality disturbance of the power distribution network.
The power quality disturbance features of the power distribution network are classified by the power quality disturbance classifier of the power distribution network, and further preferably classified by a SoftMax function, wherein the finally extracted power quality disturbance features of the power distribution network are classified by the SoftMax function (the value output by the last layer of the fully-connected neural network is essentially the extracted final feature value, and is not the probability of outputting classification conditions or classifying the probability directly, the step of converting the final feature value into the probability of each classification is separated and called a classifier, the classifier converts the finally extracted features into the probability of each classification by the SoftMax function, and then the maximum probability is selected as the output judgment result), and the SoftMax function can effectively classify the identified disturbance features, specifically as a formula (11):
in the formula (11), the first and second groups,
representing the probability of various disturbances of the power quality of the distribution network output by the power quality disturbance classifier of the distribution network, where
Represents the output perturbation feature vector of the fully-connected layer, here
Sort the number of classes for the SoftMax function, here
Mean the first
And (4) disturbance.
The main body of the bimodal feature fusion power distribution network power quality disturbance analysis method is a deep learning feature fusion method which is preferably realized in a TensorFlow2.0 environment based on Python; the power quality disturbance analysis method of the power distribution network with the bimodal feature fusion is used as an integral model, and the integral model can be trained before the integral model is actually used; when the method is used as an integral model for training, a power distribution network power quality disturbance time sequence signal sample is acquired by using historical field acquisition or background time domain simulation, and the training step specifically comprises the following steps: 1. initializing all weight parameters of the model; 2. inputting the processed power quality disturbance time sequence signal samples of the power distribution network into an integral model; 3. calculating a loss function according to a disturbance classification result output by the model; 4. updating the model weight parameters according to the loss function; 5. repeating iteration for a specified number of times, and selecting an optimal weight parameter model obtained by training; during actual application, the power quality time sequence signals of the power distribution network after field acquisition and processing are input into the trained integral model, so that a power quality disturbance identification classification result of the power distribution network is obtained.
Fig. 4 shows a power quality timing signal of a power distribution network after field acquisition and processing, which includes voltage sag, voltage flicker, transient oscillation, voltage sag + harmonic, voltage short-time interruption, voltage sag + harmonic, transient impact, voltage flicker + harmonic, and the like; comparing these perturbation curves, it can be found that different perturbations have different distinct characteristics: if harmonic disturbance has sawtooth characteristics, the waveform amplitude in a period of time drops suddenly with short-time interruption; the distinguishable characteristics are key extraction objects of the disturbance characteristics of the invention, and the final output result also refers to the recognition and classification result obtained by calculation according to the disturbance characteristics.
As shown in fig. 5, the bimodal feature fused power quality disturbance analysis method of the power distribution network, as a training result of an integral model after training, has the advantages that as the number of times of training increases, the power quality disturbance identification accuracy of the power distribution network continuously increases, meanwhile, the loss value of a constructed model also continuously decreases, and finally, the model is stabilized to a good value, so that excellent performance is achieved; through high-noise and noiseless tests, different single or composite disturbances of the power quality of the power distribution network can be accurately judged, and compared with a single convolution neural network or a gated circulation unit, the method provided by the invention has better performance.
According to the method, under the trend of developing a novel power system for new energy power generation, the power quality disturbance of a power distribution network is analyzed by applying deep learning, so that the intelligent level of the power system is improved; meanwhile, the power quality data of the power distribution network in an image mode and a time sequence mode are used as input, the characteristic of strong capability of extracting image characteristics of a convolutional neural network is combined with the characteristic of accurate characteristics of extracting time sequence signals of a gate control circulation unit, the power quality disturbance of the power distribution network is identified and analyzed by a deep learning characteristic fusion method, and finally an identification result is obtained; the method can better identify the characteristics of the power quality disturbance signals of the power distribution network, improves the intelligent level of the power distribution network, and has better disturbance rejection capability.
Claims (10)
1. A power quality disturbance analysis method of a power distribution network with bimodal feature fusion is characterized by comprising the following steps:
1. carrying out modal transformation on the power distribution network power quality time sequence signals acquired and processed on site by using a Markov conversion field to obtain a polymerization image of dynamic transition probability;
2. performing feature extraction on the aggregation image of the dynamic transition probability by adopting a convolutional neural network to obtain a first feature vector;
3. performing feature extraction on the power quality time sequence signals of the power distribution network acquired and processed on site by using a gate control circulation unit to obtain a second feature vector;
4. performing feature fusion on the first feature vector and the second feature vector by adopting a method based on deep learning feature fusion to obtain fused power quality disturbance features of the power distribution network;
5. and classifying the power quality disturbance characteristics of the power distribution network through a power quality disturbance classifier of the power distribution network.
2. The method for analyzing the power quality disturbance of the power distribution network with the fusion of the bimodal features as claimed in claim 1, wherein the aggregation image of the dynamic transition probability is obtained by performing modal transformation on the power quality time sequence signal of the power distribution network acquired and processed on site by using a markov conversion field, and specifically comprises the following steps:
1-1, discretizing a time sequence signal;
1-2, establishing a Markov probability transition matrix and calculating transition probability;
and 1-3, calculating a Markov conversion field and aggregating the Markov conversion field into a two-dimensional image.
3. The method for analyzing the power quality disturbance of the power distribution network with the bimodal feature fusion as claimed in claim 2, wherein the discretization of the time sequence signal specifically comprises: power quality time sequence signal of power distribution network acquired and processed on site
Transmitting the power quality timing sequence signal to a background analysis system of the power distribution network
Discretization is required for background analysis of actual distribution network
A unit of quantile.
4. The method for analyzing the power quality disturbance of the power distribution network with the bimodal feature fusion as claimed in claim 2, wherein the establishing of the markov probability transition matrix and the calculating of the transition probability specifically comprise:
firstly, counting the power quality time sequence signals of the power distribution network
From the sampling instant to the next sampling instant corresponds to the first
Conversion of unit of quantile to
Number of elements of unit of quantile；
Then, the statistics of each cross quantile unit are recorded to a Markov probability transition matrix
Performing the following steps;
finally, the Markov probability transition matrix is used
The statistics of each span quantile unit in each line is divided by the total statistics of the span quantile units in the line, and the transition probability of each span quantile unit is converted into the transition probability of each span quantile unit
。
5. The method according to claim 2, wherein the computing of the markov transform field and the aggregation thereof into a two-dimensional image comprises:
1) Cross reference to Markov probability transition matrix and cross reference determination according to equation (1)
，…，
Transition probability therebetween, finally obtaining the dimension of
Markov transition field
(ii) a Markov transition field
Each element in (1)
Representing a time interval of
A transition probability between points of (a);
2) By using fuzzy kernels
For each non-overlap
Averaging pixels in the block to obtain a two-dimensional image of a polymerized Markov conversion field, and storing the image in a JPG format to obtain a polymerized image of dynamic transition probability;
6. The analysis method for the power quality disturbance of the power distribution network with the bimodal feature fusion as claimed in claim 1, wherein the convolutional neural network comprises an input layer, a convolutional layer, a normalization layer, an excitation layer and a pooling layer; the method for extracting the features of the aggregated image of the dynamic transition probability by adopting the convolutional neural network to obtain a first feature vector specifically comprises the following steps:
2-1, performing convolution calculation of an image mode on the aggregation image of the dynamic transition probability by using a convolution layer; the specific calculation method of the convolution calculation is as formula (2):
wherein the content of the first and second substances,
indicating the second in a convolutional layer
A layer;
is a convolution layer of
Extracted from the layer
The power quality disturbance characteristic vector of the power distribution network in each image mode;
is an activation function;
as extracted by a convolution kernel
Characteristic vectors of power quality disturbance image modalities of the power distribution network;
is as follows
Parameters of the layer convolution kernel, here
And
respectively mean the input of
A first and a second
The power quality disturbance characteristic vector of each power distribution network;
is a first
Biasing of the layers;
is a convolution operation;
2-2, normalizing the power quality disturbance characteristic vector of the power distribution network, extracted from the dynamic transition probability aggregation image by the convolution layer, by using a normalization layer, wherein the specific calculation method is as shown in a formula (3):
wherein, the first and the second end of the pipe are connected with each other,
representing the power quality disturbance characteristic vector of the power distribution network extracted by the convolution layer,
representing the data after normalizing the power quality disturbance characteristic vector of the power distribution network,
representing the mean square error of the power quality disturbance characteristic vector of the power distribution network;
representing the mean value of the power quality disturbance characteristic vectors of the power distribution network;
2-3, nonlinearizing the normalized power quality disturbance characteristic vector of the power distribution network by utilizing the excitation layer, specifically as a formula (4):
wherein the content of the first and second substances,
is the output nonlinear power quality disturbance characteristic vector of the distribution network, here
The input normalized power quality disturbance characteristic vector of the power distribution network is obtained;
2-4, highlighting the key part of the power quality disturbance characteristic vector of the power distribution network by utilizing a pooling layer, wherein the calculation method of the maximum pooling method is specifically shown as a formula (5):
wherein the content of the first and second substances,
indicating the number of network layers of the pooling layer,
the representation of the pooled kernel is shown,
is shown as
Offset vector of layer, here
The function of the excitation is represented by,
is shown in
The layer carries out feature extraction on the aggregation image of the dynamic transition probability to obtain a first feature vector,
indicating input in the second place
Layer one
And (4) carrying out nonlinear power quality disturbance characteristic vector on the power distribution network after the individual pooling kernel.
7. The method for analyzing the power quality disturbance of the power distribution network based on the bimodal feature fusion as claimed in claim 1, wherein the feature extraction of the power quality time sequence signal of the power distribution network by using the gate control cycle unit to obtain the second feature vector specifically comprises:
3-1, inputting a power quality time sequence signal of the power distribution network acquired on site into a component module updating gate and a resetting gate of a gating circulation unit;
3-2, controlling the degree of state information of the power distribution network power quality time sequence signal at the previous moment brought into the current state by updating a door, so that the power distribution network power quality disturbance characteristics at the previous time step are memorized and stored in the information of the current time step;
and 3-3, combining the power quality disturbance characteristic information input into the power distribution network at the current sampling moment with the disturbance characteristic memory at the previous sampling moment through a reset gate to calculate the candidate state.
8. The method for analyzing the power quality disturbance of the power distribution network based on the bimodal feature fusion as claimed in claim 1, wherein the method for performing the feature fusion on the first feature vector and the second feature vector based on the deep learning feature fusion is adopted to obtain the fused power quality disturbance feature of the power distribution network, and specifically comprises:
4-1, setting the first characteristic vector as
The second feature vector is
；
4-2, when the background analysis system of the power distribution network has insufficient computing resources, selecting a feature fusion method for feature addition, and adding each element in the first feature vector and each corresponding element in the second feature vector if the two feature vectors have the same dimension
Obtaining a fused feature vector; if the two eigenvectors have different dimensions, the dimensions of the two eigenvectors are made to be the same through linear transformation, and then each element in the first eigenvector and each corresponding element in the second eigenvector are added
Obtaining a fused feature vector;
4-3, when the background analysis system of the power distribution network has sufficient computing resources, preferably selecting a feature fusion method for feature splicing, and enabling elements in the first feature vector and the second feature vector to be in a proper directionVector stitching of elements in a quantity
And forming a new feature vector.
9. The method for analyzing the power quality disturbance of the power distribution network with the fusion of the bimodal features as claimed in claim 1, wherein the classifying the power quality disturbance features of the power distribution network by the power quality disturbance classifier of the power distribution network is specifically classifying the power quality disturbance features of the power distribution network by a fully-connected neural network; the fully-connected neural network adopts a back propagation algorithm, and updates each parameter of deep learning by calculating a loss value, wherein the loss value calculating method is as shown in a formula (10):
wherein, the first and the second end of the pipe are connected with each other,
the output power quality disturbance type of the power distribution network is represented,
of the actual power quality disturbance type of the distribution network, here
And classifying the types of the power quality disturbance of the power distribution network.
10. The method for analyzing the power quality disturbance of the power distribution network with the bimodal feature fusion as claimed in claim 1, wherein the classifying the power quality disturbance features of the power distribution network by the power quality disturbance classifier of the power distribution network is specifically to classify finally extracted power quality disturbance features of the power distribution network by a SoftMax function, and specifically is as in formula (11):
wherein, the first and the second end of the pipe are connected with each other,
representing the probability of various disturbances of the power quality of the distribution network output by the power quality disturbance classifier of the distribution network, where
Represents the output perturbation eigenvector of the fully-connected layer, here
Sort the number of classes for the SoftMax function, here
First finger
And (4) disturbance.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211134779.7A CN115659254A (en) | 2022-09-19 | 2022-09-19 | Power quality disturbance analysis method for power distribution network with bimodal feature fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211134779.7A CN115659254A (en) | 2022-09-19 | 2022-09-19 | Power quality disturbance analysis method for power distribution network with bimodal feature fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115659254A true CN115659254A (en) | 2023-01-31 |
Family
ID=84983742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211134779.7A Pending CN115659254A (en) | 2022-09-19 | 2022-09-19 | Power quality disturbance analysis method for power distribution network with bimodal feature fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115659254A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116093939A (en) * | 2023-04-06 | 2023-05-09 | 北京昊创瑞通电气设备股份有限公司 | Lyapunov control algorithm-based power supply secondary integration distribution network automation terminal system |
| CN116342961A (en) * | 2023-03-30 | 2023-06-27 | 重庆师范大学 | Time sequence classification deep learning system based on mixed quantum neural network |
| CN116526667A (en) * | 2023-04-27 | 2023-08-01 | 北京昊创瑞通电气设备股份有限公司 | Secondary fusion distribution network feeder terminal system based on current Internet of things feedback mechanism |
-
2022
- 2022-09-19 CN CN202211134779.7A patent/CN115659254A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116342961A (en) * | 2023-03-30 | 2023-06-27 | 重庆师范大学 | Time sequence classification deep learning system based on mixed quantum neural network |
| CN116342961B (en) * | 2023-03-30 | 2024-02-13 | 重庆师范大学 | Time sequence classification deep learning system based on mixed quantum neural network |
| CN116093939A (en) * | 2023-04-06 | 2023-05-09 | 北京昊创瑞通电气设备股份有限公司 | Lyapunov control algorithm-based power supply secondary integration distribution network automation terminal system |
| CN116526667A (en) * | 2023-04-27 | 2023-08-01 | 北京昊创瑞通电气设备股份有限公司 | Secondary fusion distribution network feeder terminal system based on current Internet of things feedback mechanism |
| CN116526667B (en) * | 2023-04-27 | 2023-12-29 | 北京昊创瑞通电气设备股份有限公司 | Secondary fusion distribution network feeder terminal system based on current Internet of things feedback mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111860982A (en) | Wind power plant short-term wind power prediction method based on VMD-FCM-GRU | |
| CN115659254A (en) | Power quality disturbance analysis method for power distribution network with bimodal feature fusion | |
| Yang et al. | Hybrid prediction method for wind speed combining ensemble empirical mode decomposition and Bayesian ridge regression | |
| CN108984745A (en) | A kind of neural network file classification method merging more knowledge mappings | |
| CN111967343A (en) | Detection method based on simple neural network and extreme gradient lifting model fusion | |
| Yang et al. | Efficient and robust multiview clustering with anchor graph regularization | |
| CN111292195A (en) | Risk account identification method and device | |
| CN114169442B (en) | Remote sensing image small sample scene classification method based on double prototype network | |
| CN112733447B (en) | Underwater sound source positioning method and system based on domain adaptive network | |
| CN116316591A (en) | Short-term photovoltaic power prediction method and system based on hybrid bidirectional gating cycle | |
| CN109471049B (en) | Satellite power supply system anomaly detection method based on improved stacked self-encoder | |
| CN114120041A (en) | Small sample classification method based on double-pair anti-variation self-encoder | |
| CN111008224A (en) | Time sequence classification and retrieval method based on deep multitask representation learning | |
| CN114022311A (en) | Comprehensive energy system data compensation method for generating countermeasure network based on time sequence condition | |
| CN114692947A (en) | VMD-Lasso-CNN-LSTM-based user electricity consumption prediction method | |
| CN111222689A (en) | LSTM load prediction method, medium, and electronic device based on multi-scale temporal features | |
| CN105426923A (en) | Semi-supervised classification method and system | |
| CN115062727A (en) | Graph node classification method and system based on multi-order hypergraph convolutional network | |
| CN113987910A (en) | Method and device for identifying load of residents by coupling neural network and dynamic time planning | |
| CN112668633B (en) | Adaptive graph migration learning method based on fine granularity field | |
| CN111783688B (en) | Remote sensing image scene classification method based on convolutional neural network | |
| CN117458480A (en) | Photovoltaic power generation power short-term prediction method and system based on improved LOF | |
| CN116681921A (en) | Target labeling method and system based on multi-feature loss function fusion | |
| CN115329821A (en) | Ship noise identification method based on pairing coding network and comparison learning | |
| CN115348182A (en) | Long-term spectrum prediction method based on depth stack self-encoder |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |