A power distribution network power quality disturbance analysis method of bimodal feature fusion

Abstract

The application provides a dual-mode feature fusion power quality disturbance analysis method for a power distribution network, which comprises the following steps: 1, using a Markov transition field to perform mode transformation on a power quality time sequence signal collected and processed on site to obtain an aggregated image of dynamic transition probability; 2, using a convolutional neural network to extract features from the aggregated image of dynamic transition probability to obtain a first feature vector; 3, using a gated recurrent unit to extract features from the power quality time sequence signal collected and processed on site to obtain a second feature vector; 4, using a deep learning feature fusion-based method to fuse the first feature vector and the second feature vector to obtain a fused power quality disturbance feature of the power distribution network; and 5, classifying the power quality disturbance feature of the power distribution network through a power quality disturbance classifier.

Description

TECHNICAL FIELD

[0001] The application relates to a dual-mode feature fusion power quality disturbance analysis method for a power distribution network, and belongs to the field of power system power quality disturbance identification. BACKGROUND

[0002] The proportion of new energy power connected to the power grid is increasing, the scale of the power grid is expanding, the structure and operation mode of the power system are changing, and more randomness and nonlinearity are introduced into the stable operation of the power system; power quality disturbance is one of the main problems existing in the power system; disturbances or interferences are usually caused by many factors, such as nonlinear or fluctuating loads, power electronic devices, system faults, etc., and in some extreme cases, these factors can cause waveform distortion; if identification and control actions are not taken to appropriately prevent and mitigate these disturbances, overall interruption events of the transmission and distribution networks may occur, causing significant social impact and huge economic losses.

[0003] Actual power quality disturbances are divided into single disturbances and composite disturbances, but in a power distribution network containing photovoltaic power generation, simple single-type disturbances occur less frequently, and composite disturbances, especially double disturbances, occur more frequently; actual power quality disturbances have characteristic aliasing phenomena, and the pattern similarity between composite disturbances and single disturbances significantly increases the difficulty of accurate identification; traditional power quality disturbance identification methods are complex to calculate or require some manual intervention to extract feature quantities; in the current complex waveform changes, accurate identification or real-time calculation cannot be achieved; with the development of big data and artificial intelligence, related methods make real-time and accurate identification possible.

[0004] At present, deep learning is developing rapidly and is widely used, and the generalization of problem solving can be well applied to the field of power quality disturbance identification, and still has good self-adaptive ability under certain noise; but for complex power quality identification problems, the identification accuracy of a single deep learning model still needs to be improved. SUMMARY

[0005] The application provides a dual-mode feature fusion power quality disturbance analysis method for a power distribution network.

[0006] The technical solution of the application is a dual-mode feature fusion power quality disturbance analysis method for a power distribution network, which comprises the following steps:

[0007] 1. Using a Markov transition field to perform modal transformation on the power quality time sequence signals of the power distribution network collected and processed on site to obtain an aggregated image of dynamic transition probability;

[0008] 2. performing feature extraction on the aggregated image of dynamic transition probability by using a convolutional neural network to obtain a first feature vector;

[0009] 3. performing feature extraction on the power quality time series signal of the power distribution network collected and processed by using a gated recurrent unit to obtain a second feature vector;

[0010] 4. performing feature fusion on the first feature vector and the second feature vector by using a deep learning feature fusion method to obtain a fused power quality disturbance feature of the power distribution network;

[0011] 5. classifying the power quality disturbance feature of the power distribution network by using a power quality disturbance classifier.

[0012] Further, the modal transformation of the power quality time series signal of the power distribution network collected and processed by using the Markov transition field to obtain the aggregated image of dynamic transition probability comprises the following steps:

[0013] 1-1. Discretization of the time series signal;

[0014] 1-2. Establishment of the Markov probability transition matrix and calculation of the transition probability;

[0015] 1-3. Calculation of the Markov transition field and aggregation into a two-dimensional image.

[0016] Further, the discretization of the time series signal comprises the following steps: transmitting the power quality time series signal of the power distribution network collected and processed to a power distribution network background analysis system, and then discretizing the power quality time series signal of the power distribution network into a number of quantile units required for actual power distribution network background analysis.

[0017] Further, the establishment of the Markov probability transition matrix and the calculation of the transition probability comprise the following steps:

[0018] First, count the number of elements in the power quality time series signal of the power distribution network corresponding to the transition from the i-th quantile unit to the j-th quantile unit from the sampling time to the next sampling time;

[0019] Then, record the statistical number of each cross-quantile unit in the Markov probability transition matrix;

[0020] Finally, divide the statistical number of each cross-quantile unit in each row of the Markov probability transition matrix by the total statistical number of the cross-quantile units in the row to obtain the transition probability of each cross-quantile unit. ​​​​​​​​

[0021] Furthermore, the calculation of the Markov transformed field and its aggregation into a two-dimensional image specifically includes:

[0022] 1) Determine the probability transition matrix by referring to the Markov probability transition matrix and according to equation (1). , ..., The transition probability between them ultimately yields a dimension of Markov transition field Markov transition field Each element in Indicates the time interval as The transition probability between points;

[0023] 2) By using a fuzzy kernel For each non-overlapping The pixels in the block are averaged to obtain a two-dimensional image of the aggregated Markov transition field, and the image is saved in JPG format to obtain an aggregated image of dynamic transition probabilities.

[0024] (1);

[0025] It is the first quantile units, It is the first Quantile units.

[0026] Further, the convolutional neural network includes an input layer, a convolutional layer, a normalization layer, an activation layer, and a pooling layer; the step of using the convolutional neural network to extract features from the aggregated image of dynamic transition probabilities to obtain a first feature vector specifically includes:

[0027] 2-1. Convolutional layers are used to perform image modality convolution calculations on the aggregated image of dynamic transition probabilities; the specific calculation method for the convolution calculation is as shown in formula (2):

[0028] (2);

[0029] in, Represents the first convolutional layer layer; For the convolutional layer Extracted from the layer Feature vectors of power quality disturbances in the power distribution network for each image mode; For activation functions; For the first digit extracted by the convolution kernel Feature vectors of image modes of power quality disturbance in a power distribution network; For the first Layer convolution kernel parameters, here and the inputted first and second power quality disturbance feature vectors of the power distribution network, respectively; bias of the first layer; convolution operation;

[0030] 2-2, normalizing the power quality disturbance feature vector of the power distribution network extracted by the convolution layer from the aggregated image of the dynamic transition probability using the normalization layer, and the specific calculation method is as shown in formula (3):

[0031] (3);

[0032] wherein, represents the power quality disturbance feature vector of the power distribution network extracted by the convolution layer, represents the data after the normalization of the power quality disturbance feature vector of the power distribution network, represents the mean square deviation of the power quality disturbance feature vector of the power distribution network; represents the mean value of the power quality disturbance feature vector of the power distribution network;

[0033] 2-3, nonlinearizing the normalized power quality disturbance feature vector of the power distribution network using the excitation layer, and the specific formula is as shown in formula (4):

[0034] (4);

[0035] wherein, is the outputted nonlinearized power quality disturbance feature vector of the power distribution network, and herein is the inputted normalized power quality disturbance feature vector of the power distribution network;

[0036] 2-4, highlighting the key part of the power quality disturbance feature vector of the power distribution network using the pooling layer, and the calculation method of the maximum pooling method is as shown in formula (5):

[0037] (5);

[0038] wherein, represents the number of network layers of the pooling layer, represents the pooling kernel, represents the bias vector of the first layer, and herein represents the excitation function, represents the first feature vector obtained by feature extraction on the aggregated image of the dynamic transition probability at the first layer, represents the inputted first feature vector of the power distribution network at the first layer, represents the second feature vector obtained by feature extraction on the aggregated image of the dynamic transition probability at the second layer, represents the inputted second feature vector of the power distribution network at the second layer, represents the third feature vector obtained by feature extraction on the aggregated image of the dynamic transition probability at the third layer, represents the inputted third feature vector of the power distribution network at the third layer,​​The power quality disturbance feature vector of the power distribution network after the nonlinearization of the pooling core.

[0039] Further, the power quality time sequence signal of the power distribution network is extracted by the gated recurrent unit to obtain a second feature vector, specifically including:

[0040] 3-1, input the field collected power quality time sequence signal of the power distribution network into the update gate and reset gate of the gated recurrent unit component module;

[0041] 3-2, first control the degree of the state information of the previous time of the power quality time sequence signal of the power distribution network being brought into the current state through the update gate, so that the power quality disturbance feature of the previous time step is saved in the information of the current time step;

[0042] 3-3, then combine the power quality disturbance feature information of the current sampling time with the disturbance feature memory of the previous sampling time through the reset gate to calculate the candidate state.

[0043] Further, the first feature vector and the second feature vector are fused by the deep learning feature fusion-based method to obtain the fused power quality disturbance feature of the power distribution network, specifically including:

[0044] 4-1, the first feature vector is , and the second feature vector is ;

[0045] 4-2, when the computing resources of the power distribution network background analysis system are not sufficient, the feature fusion method of feature addition is selected, at this time, if the dimensions of the two feature vectors are the same, the elements in the first feature vector are added to the corresponding elements in the second feature vector to obtain the fused feature vector; if the dimensions of the two feature vectors are different, first make the dimensions of the two feature vectors the same through linear transformation, and then add the elements in the first feature vector to the corresponding elements in the second feature vector to obtain the fused feature vector;

[0046] 4-3, when the computing resources of the power distribution network background analysis system are sufficient, the feature fusion method of feature splicing is preferred, the elements in the first feature vector are spliced with the elements in the second feature vector to form a new feature vector.

[0047] Further, the power quality disturbance feature of the power distribution network is classified by the power quality disturbance classifier, specifically by a fully connected neural network; the fully connected neural network adopts a back propagation algorithm, and each parameter of the deep learning is updated by calculating a loss value, and the loss value calculation method is shown in formula (10):

[0048] (10) ;

[0049] wherein, represents the output power grid power quality disturbance type, is the actual power grid power quality disturbance type, here is the number of power grid power quality disturbance classification categories.

[0050] Further, the classification of the power grid power quality disturbance features by the power grid power quality disturbance classifier is specifically the classification of the finally extracted power grid power quality disturbance features by a SoftMax function, specifically as formula (11):

[0051] (11) ;

[0052] wherein, represents the probability of each type of power grid power quality disturbance output by the power grid power quality disturbance classifier, here represents the output disturbance feature vector of the fully connected layer, here is the number of SoftMax function classification categories, here refers to the first disturbance.

[0053] The beneficial effects of the present application are:

[0054] 1) The power grid power quality time sequence signal of the present application is converted into a Markov transition field, which effectively describes the transition probability of the power grid power quality time sequence signal under different time scales; compared with the single comparison of the power grid power quality time sequence signals before and after, the present application can better depict the change of the power grid power quality disturbance;

[0055] 2) The present application is aimed at the problem that the existing method is difficult to accurately identify the complex power grid power quality disturbance, the mode transformation of the power grid power quality time sequence signal is carried out by using the Markov transition field to obtain the aggregated image of dynamic transition probability, which enhances the input data and constitutes sufficient data to describe the change of the power grid power quality; the convolutional neural network suitable for extracting image features is used to extract the features of the aggregated image of dynamic transition probability, the gated recurrent unit which can efficiently extract the time sequence signal is used to extract the features of the power grid power quality time sequence signal, and the disturbance change features of the input power grid power quality across time scales are fully extracted;

[0056] 3) This invention uses a deep learning feature fusion method to fuse features extracted from two different modalities of power quality data in distribution networks. Then, a power quality disturbance classifier performs multi-label classification, which can make full use of the power quality disturbance signal features of the two different modalities of data in distribution networks. This can improve the accuracy of power quality disturbance identification in complex distribution networks and maintain a high recognition rate even in high-noise environments. Furthermore, the convolutional neural network and gated recurrent unit are computed in parallel, which speeds up the calculation.

[0057] 4) This invention enhances the data input by generating an aggregated image of the dynamic transition probability of the power quality time correlation of the distribution network and combining it with the original time series signal, thereby describing the characteristics of the power quality disturbance signal of the distribution network from multiple perspectives. Through feature extraction and fusion of two deep learning networks, the algorithm performance can be effectively improved and the robustness of the model in high-noise environments can be enhanced.

[0058] 5) This invention is particularly suitable for real-time and accurate identification and classification of power quality disturbances in distribution networks containing photovoltaic power generation. Attached Figure Description

[0059] Appendix Figure 1 This is a schematic diagram of the method flow of the present invention.

[0060] Appendix Figure 2 This is a schematic diagram of a Markov transformation field.

[0061] Appendix Figure 3 This is a schematic diagram of the gated loop unit structure.

[0062] Appendix Figure 4 This is a schematic diagram of the power quality time sequence signal of the distribution network after on-site acquisition and processing.

[0063] Appendix Figure 5 This is a schematic diagram of the training results of the power quality disturbance analysis method for distribution networks based on dual-modal feature fusion of the present invention after training as a whole model. Detailed Implementation

[0064] As attached Figure 1 As shown, a dual-modal feature fusion method for power quality disturbance analysis in distribution networks is presented. This method includes the following steps:

[0065] 1. Using Markov transform fields, the power quality time-series signal of the distribution network acquired and processed on-site is subjected to mode transformation to obtain an aggregated image of dynamic transition probabilities;

[0066] 2. A convolutional neural network is used to extract features from the aggregated image of dynamic transition probabilities to obtain the first feature vector;

[0067] 3. The power quality time sequence signal of the power distribution network collected and processed on site is extracted by using a gated recurrent unit to obtain a second feature vector;

[0068] 4. The first feature vector and the second feature vector are fused by using a deep learning feature fusion method to obtain a fused power quality disturbance feature of the power distribution network;

[0069] 5. The power quality disturbance feature of the power distribution network is classified by using a power quality disturbance classifier of the power distribution network.

[0070] In the present application, steps 2 and 3 are parallel computing methods, the convolutional neural network in step 2 and the gated recurrent neural network in step 3 are simultaneously parallel computing; steps 2 and 3 do not have an overall sequence in the execution sequence; the parallel computing of steps 2 and 3 can fully utilize the GPU (Graphics Processing Unit, graphics processor) for accelerated operation.

[0071] The common power quality disturbances of the power distribution network include voltage temporary rise, voltage sag, voltage flicker, transient oscillation, voltage sag + harmonic, voltage short interruption + harmonic, voltage short interruption, voltage temporary rise + harmonic, transient impulse, voltage flicker + harmonic, etc.

[0072] The dual-modal feature fusion power quality disturbance analysis method of the power distribution network is a method used in a background analysis system of the power distribution network, mainly for identifying and classifying complex power quality disturbances; the present application is especially suitable for real-time and accurate identification and classification of power quality disturbances of the power distribution network containing photovoltaic power generation; the present application considers the volatility and randomness of photovoltaic power sources, and can accurately identify and classify the power quality disturbances of the power distribution network containing photovoltaic power generation in real time.

[0073] When the application is used for real-time and accurate identification and classification of power quality disturbances in a power distribution network containing photovoltaic power generation, the power distribution network refers to a power distribution network containing photovoltaic power generation, in which power generation and power consumption coexist, and power quality disturbances occur frequently. Compared with wind power generation, nuclear power generation and other mature new energy generation technologies with fewer disturbance factors, power quality disturbances in a power distribution network containing photovoltaic power generation are mainly caused by photovoltaic power generation abnormalities caused by various factors, including natural factors (such as light, rain, snow, storms, etc.), equipment aging and failure, human factors, etc. For example, photovoltaic power generation mainly relies on photovoltaic panels, which often have a large area and are exposed to the outside. If any small part of the photovoltaic panel has a problem, it will cause power quality disturbances. Therefore, photovoltaic power generation has the characteristics of strong vulnerability. In addition, photovoltaic power generation often appears in a distributed form in the power distribution network, which is widely distributed and not concentrated. Voltage temporary rise, voltage sag, voltage flicker, transient oscillation, voltage sag + harmonics, voltage short-term interruption + harmonics, voltage short-term interruption, voltage temporary rise + harmonics, transient impulse, voltage flicker + harmonics and other power quality disturbance problems are more prominent in a power distribution network containing photovoltaic power generation.

[0074] The field collection and processing refers to real-time data collection in a region by a scanner or intelligent measurement device and transmission to a power distribution network background analysis system through wireless communication. The power distribution network background analysis system automatically corrects the error or missing data. The correction method includes two steps: 1. Search the power quality disturbance database in the power distribution network background analysis system within a specified time to find similar data, interpolate the missing data, and delete and refill the error data; 2. If the search is successful, normalize the overall modified data after modification; if the search fails, replace or complete the data that needs to be modified by summing and averaging the left and right adjacent data of the missing or error data, and then normalize the overall modified data.

[0075] When the Markov transition field is used to perform modal transformation on the power quality time series signal of the field collection and processing of the power distribution network, Python, C and other engineering practical programming languages are preferably used for modal transformation.

[0076] As shown in Figure 2 The use of Markov transition field for modal transformation of power quality time series signal of field collection and processing of power distribution network obtains an aggregated image of dynamic transition probability, which specifically includes the following steps:

[0077] 1-1, Time series signal discretization: Because there are many numerical values in the power quality time series signal of the power distribution network, time series signal discretization can facilitate subsequent data analysis and processing. The power quality time series signal of the field collection and processing of the power distribution network is discretized. (Which contains An element, here refers to a current signal or a voltage signal in an electrical quantity, to the power grid background analysis system, for the power grid power quality time series signal with a large number of elements to be discretized into the actual power grid background analysis 4 quantile units (for example: Preferably 4, which means that the Each element in the above is arranged from small to large according to the value, and is divided into 0-25%, 25-50%, 50-75%, and 75-100% four quantile units, specifically, the Each element in the above is arranged from small to large according to the value, and is divided into 0-25%, 25-50%, 50-75%, and 75-100% four quantile units, specifically, the Quantify each value of the power grid power quality time series signal, and then identify the quantile to classify the power grid power quality time series signal Each element in the above is arranged from small to large according to the value, and is divided into 0-25%, 25-50%, 50-75%, and 75-100% four quantile units, specifically, the

[0078] 1-2, establish Markov probability transition matrix and calculate transition probability: the purpose is to find out the probability of power grid power quality time series signal from the current quantile unit to other quantile units; construct a Markov probability transition matrix of dimension ( The number of quantile units, Each element in the above is ) as a selection table for each element in the Markov conversion field constructed in 1-3, specifically including the following steps:

[0079] First, count the number of elements in the power grid power quality time series signal From the sampling time to the next sampling time, the corresponding from the first Quantile unit to the first Quantile unit conversion element;

[0080] Then, record each cross-quantile unit statistics to the Markov probability transition matrix (For example, in the power grid power quality time series signal From the sampling time to the next sampling time, the corresponding from the first quantile unit to the second quantile unit has 100 elements, so the statistics is 100);

[0081] Finally, the Markov probability transition matrix Each row of the statistics of each cross quantile unit is divided by the total statistics of the cross quantile unit in the row, and is converted into the transition probability of each cross quantile unit ;

[0082] 1-3, Calculate the Markov transition field and aggregate into a two-dimensional image: generate image modal input data for enhanced power grid background analysis; against the Markov probability transition matrix described in step 1-2 and determine according to formula (1) , …, The transition probability between them (for example The first element of the first row in transitions to itself, that is, from the quantile unit where transitions to the quantile unit where , against the Markov probability transition matrix find the corresponding transition probability as the first element of the first row in ; for example The first element of the first row in transitions to , that is, from the quantile unit where transitions to the quantile unit where , against the Markov probability transition matrix find the corresponding transition probability as the first element of the first row in ), and finally obtain a Markov transition field with dimensions ; Markov transition field Each element in represents the transition probability between points with a time interval of ; for example, illustrates that there is only one interval in the transition process along the time axis; the main diagonal , which is a special case of a time interval of 0, which obtains the probability of each quantile to itself, that is, the self-transition probability; but considering that is relatively large, if the original Markov transition field is directly used to generate an image, it will make the image too large and occupy more computer storage space, which is not conducive to the rapid analysis of the power grid background analysis system and occupies too much storage resources. Therefore, by using a fuzzy kernel to each non-overlapping ​​​The pixels in the block are averaged to obtain a two-dimensional image of the aggregated Markov transformation field, and the image is saved in JPG format to obtain an aggregated image of dynamic transition probabilities. The aggregated image of dynamic transition probabilities is used as input data for a set of image modes in the power distribution network background analysis system.

[0083] (1);

[0084] It is the first quantile units, It is the first Quantile units.

[0085] The convolutional neural network includes an input layer, a convolutional layer, a normalization layer, an activation layer, and a pooling layer; the preferred method for constructing the convolutional neural network is to use the TensorFlow 2.0 functional API.

[0086] The step of using a convolutional neural network to extract features from the aggregated image of dynamic transition probabilities to obtain the first feature vector specifically includes the following steps:

[0087] 2-1. Using convolutional layers, perform convolutional calculations on the image modality of the aggregated image of the dynamic transfer probability of power quality in the distribution network obtained by the transformation in step 1-3; Under the influence of photovoltaic power generation, compared with other communication signal disturbances, gear signal disturbances, etc., the power quality disturbance of the distribution network has more complex composite disturbances. In order to accurately extract the power quality disturbance features of the image modality of the distribution network, the specific calculation method of the convolutional calculation is as shown in formula (2):

[0088] (2);

[0089] In formula (2), Represents the first convolutional layer layer; For the convolutional layer Extracted from the layer Feature vectors of power quality disturbances in the power distribution network for each image mode; For activation functions; For the first digit extracted by the convolution kernel Feature vectors of image modes of power quality disturbance in a power distribution network; For the first Layer convolution kernel parameters, here and The input number The and the first A power quality disturbance feature vector of a distribution network; For the first Layer bias; for convolution operation;

[0090] 2-2, the normalized layer is used for normalizing the power quality disturbance feature vector of the power distribution network extracted from the aggregated image of the power quality dynamic transfer probability of the power distribution network, in order to accelerate the calculation speed of the power distribution network background analysis system and improve the accuracy of feature extraction, z-score normalization is preferably adopted, and the specific calculation method is as shown in formula (3):

[0091] (3) ;

[0092] In formula (3), denotes the power quality disturbance feature vector of the power distribution network extracted by the convolution layer, denotes the data after the power quality disturbance feature vector of the power distribution network is normalized, denotes the mean square deviation of the power quality disturbance feature vector of the power distribution network; denotes the mean value of the power quality disturbance feature vector of the power distribution network;

[0093] 2-3, the excitation layer is used for nonlinearizing the normalized power quality disturbance feature vector of the power distribution network, and ReLU function is preferably adopted for feature mapping, and the specific formula (4) is as follows:

[0094] (4) ;

[0095] In formula (4), is the output nonlinearized power quality disturbance feature vector of the power distribution network, and here is the input normalized power quality disturbance feature vector of the power distribution network;

[0096] 2-4, the pooling layer is used for highlighting the key part of the power quality disturbance feature vector of the power distribution network; the pooling layer preferably adopts the maximum pooling method to highlight the key part of the power quality disturbance feature vector of the power distribution network, and the calculation method of the maximum pooling method is as shown in formula (5):

[0097] (5) ;

[0098] In formula (5), denotes the number of network layers of the pooling layer, denotes the pooling kernel, denotes the bias vector of the i-th layer, and here denotes the excitation function, denotes the first feature vector obtained by extracting features from the aggregated image of the dynamic transfer probability in the i-th layer, denotes the input in the i-th layer of the i-th layer, denotes the excitation function, denotes the first feature vector obtained by extracting features from the aggregated image of the dynamic transfer probability in the i-th layer, denotes the input in the i-th layer of the i-th layer, The power quality disturbance feature vector of the power distribution network after the nonlinearization of the pooling core.

[0099] As shown in the accompanying Figure 3 The specific construction method of the gating recurrent unit preferably uses the TensorFlow2.0 functional API to construct the gating recurrent unit, which is calculated simultaneously with the convolutional neural network to form parallel calculation, and the calculation formula of the gating recurrent unit is as formula (6)-formula (9):

[0100] (6) ;

[0101] (7) ;

[0102] (8) ;

[0103] (9) ;

[0104] In formula (6)-formula (9): is the update gate; is the reset gate; is the power quality time series signal of the power distribution network input at the current time; is the current hidden unit candidate state; is the state of the hidden unit at the current time; is the state of the hidden unit at the previous time; , , are all trainable weight coefficients; , , are all bias matrices; and Tanh respectively represent the Sigmoid activation function and the Tanh activation function; represents the Hadamard product operation.

[0105] As shown in the accompanying Figure 3 The feature extraction of the power quality time series signal of the power distribution network using the gating recurrent unit obtains a second feature vector, and specifically includes the following steps:

[0106] 3-1, input the power quality time series signal collected on site into the component modules of the gating recurrent unit, namely the update gate and the reset gate;

[0107] 3-2, first control the degree of the state information of the power quality time series signal at the previous time step being brought into the current state through the update gate, so that the power quality disturbance feature at the previous time step is memorized into the information at the current time step;

[0108] 3-3, further inputting the power quality disturbance characteristic information of the current sampling time into the power distribution network and combining the disturbance characteristic memory of the previous sampling time to calculate the candidate state; the role is to forget the power quality time sequence signal characteristic information of the power distribution network with less historical influence, and output the characteristics that can efficiently express the power quality disturbance of the power distribution network.

[0109] The first feature vector and the second feature vector are fused by the method based on deep learning feature fusion to obtain the fused power quality disturbance feature of the power distribution network; according to the actual power distribution network power quality disturbance identification demand and the hardware configuration of the power distribution network background analysis system, the first feature vector and the second feature vector are preferably fused in the form of feature splicing or feature addition, which specifically includes the following steps:

[0110] 4-1, the first feature vector is , and the second feature vector is ;

[0111] 4-2, when the computing resources of the power distribution network background analysis system are not sufficient, the feature fusion method of feature addition is preferred, at this time, if the dimensions of the two feature vectors are the same, the elements in the first feature vector are added to the corresponding elements in the second feature vector to obtain the fused feature vector; if the dimensions of the two feature vectors are different, first make the dimensions of the two feature vectors the same through linear transformation, and then add the elements in the first feature vector to the corresponding elements in the second feature vector to obtain the fused feature vector; the feature fusion method of feature addition has the advantages that the dimension of the fused feature is the same as that of the initial two feature vectors, and less computing resources are occupied; when the computing resources of the power distribution network background analysis system are sufficient, the feature fusion method of feature splicing is preferred, the elements in the first feature vector are spliced with the elements in the second feature vector to form a new feature vector, which has the advantages that the fused feature is more complete and has less feature loss, but the dimension of the feature vector increases and more computing resources are needed.

[0112] The specific implementation scheme of the feature splicing method for fusing the first feature vector and the second feature vector preferably uses python language to realize feature addition or feature splicing, so as to realize deep learning parallel computing.

[0113] The parallel computing refers to using GPU to make multiple deep learning models calculate simultaneously.

[0114] The power quality disturbance feature of the power distribution network is classified by the power quality disturbance classifier of the power distribution network, and the power quality disturbance feature of the power distribution network is preferably classified by the full connection neural network.

[0115] The fully connected neural network is preferably constructed using TensorFlow 2.0 functional API.

[0116] The fully connected neural network adopts a back propagation algorithm, the main purpose of which is to compare and calculate the actual power quality disturbance of the power distribution network with the power quality disturbance of the power distribution network discriminated by the method of the application, and update the parameters of the deep learning in the method of the application through the calculation of the loss value, where the deep learning includes convolutional neural network, gated recurrent unit, feature fusion, fully connected neural network, power distribution network power quality disturbance classifier and the like, and the loss value calculation method is as shown in formula (10):

[0117] (10) ;

[0118] In formula (10), represents the output power quality disturbance type of the power distribution network, is the actual power quality disturbance type of the power distribution network, where is the number of power quality disturbance classification categories of the power distribution network.

[0119] The power quality disturbance features of the power distribution network are further classified by the power quality disturbance classifier of the power distribution network, and 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 final feature value extracted, and is not directly output as a classification or as the probability of each classification, and this step of converting the final feature value into the probability of each classification is called a classifier, the classifier converts the finally extracted features into the probability of each classification through the SoftMax function, and then selects the maximum probability as the output discrimination result), the SoftMax function can effectively classify the recognized disturbance features, and the specific formula is as shown in formula (11):

[0120] (11) ;

[0121] In formula (11), represents the probability of each type of power quality disturbance of the power distribution network output by the power quality disturbance classifier of the power distribution network, where represents the output disturbance feature vector of the fully connected layer, where is the number of SoftMax function classification categories, where refers to the th disturbance.

[0122] The main body of the power distribution network power quality disturbance analysis method of the dual-modal feature fusion is a deep learning feature fusion method, which is preferably implemented in a TensorFlow 2.0 environment based on Python; the power distribution network power quality disturbance analysis method of the dual-modal feature fusion of the present application is used as a whole model, and the whole model can be trained before the whole model is actually used; when the method of the present application is used as a whole model for training, the power distribution network power quality disturbance time series signal samples are obtained by historical field collection or background time domain simulation, and the training steps specifically include: 1. initializing all weight parameters of the model; 2. inputting the processed power distribution network power quality disturbance time series signal samples into the whole model; 3. calculating the loss function according to the disturbance classification result output by the model; 4. updating the weight parameters of the model according to the loss function; 5. repeating the iteration for a specified number of times, and selecting the optimal weight parameter model obtained by training; in actual application, the processed power distribution network power quality time series signals collected in the field are input into the whole model which has been trained, so as to obtain the power distribution network power quality disturbance recognition classification result.

[0123] The Figure 4 The power distribution network power quality time series signals collected in the field after processing are shown, including voltage temporary rise, voltage sag, voltage flicker, transient oscillation, voltage sag + harmonic, voltage short-time interruption + harmonic, voltage short-time interruption, voltage temporary rise + harmonic, transient impulse, voltage flicker + harmonic, etc.; by comparing these disturbance curves, it can be found that different disturbances have different obvious characteristics: such as the sawtooth characteristic of harmonic disturbance, the sharp drop of waveform amplitude in a period of time of sag and short-time interruption, etc.; such distinguishable characteristics are the key extraction objects of the disturbance characteristics of the present application, and the final output result also refers to the recognition classification result calculated according to these disturbance characteristics.

[0124] As shown in the Figure 5 The training result of the power distribution network power quality disturbance analysis method of the dual-modal feature fusion of the present application as a whole model after training is shown, with the increase of the number of training, the power distribution network power quality disturbance recognition accuracy is continuously rising, at the same time, the loss value of the built model is continuously falling, and finally stabilizes to a better value, reaching excellent performance; through high noise and noiseless test, the method of the present application can accurately distinguish different single or composite disturbances of power distribution network power quality, and has better performance compared with single convolutional neural network or gated recurrent unit.

[0125] The application applies deep learning to analyze power quality disturbance of a distribution network to improve the intelligent level of a power system under the trend of developing a new energy power generation new-type power system; meanwhile, the power quality data of the distribution network in an image mode and a time sequence mode are taken as input, the characteristics of the strong image feature extraction ability of a convolutional neural network and the accurate time sequence signal feature extraction ability of a gating recurrent unit are combined, the deep learning feature fusion method is used to identify and analyze the power quality disturbance of the distribution network, and finally the identification result is obtained; this method can better identify the features of the power quality disturbance signals of the distribution network, improve the intelligent level of the distribution network, and has better anti-disturbance ability.

Claims

1. A method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion, characterized by: include: 1) Using Markov transform fields to perform mode transformation on the power quality time-series signals of the power distribution network acquired and processed on-site to obtain an aggregated image of dynamic transition probabilities; 2) The first feature vector is obtained by using a convolutional neural network to extract features from the aggregated image of dynamic transition probabilities; 3) A gated loop unit is used to extract features from the power quality time-series signal of the power distribution network acquired and processed on-site to obtain the second feature vector; 4) A deep learning-based feature fusion method is used to fuse the first feature vector and the second feature vector to obtain the fused power quality disturbance characteristics of the distribution network; 5) Classify the power quality disturbance characteristics of the distribution network using a power quality disturbance classifier; The convolutional neural network includes an input layer, a convolutional layer, a normalization layer, an activation layer, and a pooling layer; the step of using the convolutional neural network to extract features from the aggregated image of dynamic transition probabilities to obtain a first feature vector specifically includes: 2-1) Convolution calculation of image modalities is performed on the aggregated image of dynamic transition probabilities using convolutional layers; the specific calculation method of the convolution calculation is as shown in formula (2): （2）； in, Represents the first convolutional layer layer; For the convolutional layer The first extracted from the layer Feature vectors of power quality disturbances in the power distribution network for each image mode; For activation functions; For the first digit extracted by the convolution kernel Feature vectors of image modes of power quality disturbance in a power distribution network; For the first Layer convolution kernel parameters, here and Each refers to the input number The and the first A power quality disturbance feature vector of a distribution network; For the first Layer bias; This is a convolution operation; 2-2) The distribution network power quality disturbance feature vector extracted from the aggregated image of dynamic transition probability by the convolutional layer is normalized using a normalization layer. The specific calculation method is shown in formula (3): （3）； in, This represents the power quality disturbance feature vector of the distribution network extracted by the convolutional layer. This represents the data after normalizing the power quality disturbance eigenvector of the distribution network. The mean square error represents the characteristic vector of power quality disturbances in the distribution network. This represents the mean of the characteristic vector of power quality disturbances in the distribution network; 2-3) The normalized power quality disturbance characteristic vector of the distribution network is nonlinearized using the excitation layer, as shown in formula (4): （4）； in, Here, the output is the nonlinearized power quality disturbance characteristic vector of the distribution network. The input is the normalized power quality disturbance feature vector of the distribution network; 2-4) The pooling layer uses the max pooling method to highlight the key parts of the power quality disturbance characteristic vector of the distribution network. The calculation method of the max pooling method is as follows: (5) （5）； in, This indicates the number of network layers in the pooling layer. Indicates the pooling kernel, Indicates the first The layer's bias vector, here Represents the activation function. Indicates the first The first feature vector is obtained by extracting features from the aggregated image of the dynamic transition probability layer. This indicates the input at the 1st position. Layer The power quality disturbance feature vector of the distribution network after nonlinearization of the pooled kernel; The second feature vector is obtained by using a gated cyclic unit to extract features from the power quality time-series signal of the distribution network. Specifically, this includes: 3-1) Input the power quality timing signal of the distribution network collected on site into the update gate and reset gate of the gated loop unit; 3-2) First, control the extent to which the previous state information of the power quality timing signal of the distribution network is brought into the current state by the update gate, so that the power quality disturbance characteristics of the distribution network in the previous time step are stored in the information of the current time step. 3-3) Then, the power quality disturbance characteristic information of the power distribution network input at the current sampling time is combined with the disturbance characteristic memory of the previous sampling time through the reset gate to calculate the candidate state.

2. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion as described in claim 1, characterized in that: The method of using Markov transformed fields to perform mode transformation on the power quality time-series signal of the distribution network acquired and processed on-site to obtain an aggregated image of dynamic transition probabilities specifically includes the following steps: 1-1) Discretization of time-series signals; 1-2) Establish the Markov probability transition matrix and calculate the transition probabilities; 1-3) Calculate the Markov transformed field and aggregate it into a two-dimensional image.

3. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion as described in claim 2, characterized in that: The discretization of the time-series signal specifically includes: processing the power quality time-series signal of the distribution network acquired and processed on-site. The data is transmitted to the distribution network backend analysis system, and then the power quality time-series signal of the distribution network is processed. Discretization is required for actual distribution network backend analysis. Quantile units.

4. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion according to claim 2, characterized in that: The establishment of the Markov probability transition matrix and the calculation of the transition probabilities specifically include: First, statistical analysis of power quality time-series signals in the distribution network. From each sampling time to the next sampling time, the corresponding time is from the first sampling time. The quantile unit is converted to the quantile unit. The number of elements in each quantile unit; Then, the statistics for each quantile unit are recorded into the Markov probability transition matrix. middle; Finally, the Markov probability transition matrix The statistic for each quantile span in each row is divided by the sum of the statistics for that row's quantile spans, transforming it into the transition probability for each quantile span. .

5. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion according to claim 2, characterized in that: The calculation of the Markov transformed field and its aggregation into a two-dimensional image specifically includes: 1-3-1) Determine the probability transition matrix by referring to the Markov probability transition matrix and according to equation (1). , ..., The transition probability between them ultimately yields a dimension of Markov transition field Markov transition field Each element in Indicates the time interval as The transition probability between points; 1-3-2) By using a fuzzy kernel For each non-overlapping The pixels in the block are averaged to obtain a two-dimensional image of the aggregated Markov transition field, and the image is saved in JPG format to obtain an aggregated image of dynamic transition probabilities. （1）； It is the first quantile units, It is the first Quantile units.

6. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion according to claim 1, characterized in that: The method of using deep learning-based feature fusion to fuse the first feature vector and the second feature vector to obtain the fused power quality disturbance features of the distribution network specifically includes: 4-1) Let the first eigenvector be... The second feature vector is ; 4-2) When the computing resources of the distribution network background analysis system are insufficient, the feature fusion method of feature addition is selected. In this case, if the two feature vectors have the same dimension, the elements in the first feature vector are added to the corresponding elements in the second feature vector. This yields the fused feature vector. If the two feature vectors have different dimensions, a linear transformation is first applied to make their dimensions the same, and then each element of the first feature vector is added to the corresponding element of the second feature vector. , thus obtaining the fused feature vector; 4-3) When the power distribution network backend analysis system has sufficient computing resources, select the feature fusion method of feature splicing, and splice the elements in the first feature vector with the elements in the second feature vector. This forms a new feature vector.

7. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion according to claim 1, characterized in that: The classification of power quality disturbance features in the distribution network using a power quality disturbance classifier specifically involves classifying the power quality disturbance features in the distribution network using a fully connected neural network. The fully connected neural network employs a backpropagation algorithm, and updates the parameters of the deep learning process by calculating the loss value. The loss value is calculated using formula (10). （10）； in, This indicates the type of power quality disturbance in the output distribution network. This refers to the actual power quality disturbance type in the distribution network. This represents the number of categories of power quality disturbances in the distribution network.

8. The method for analyzing power quality disturbances in distribution networks based on dual-modal feature fusion according to claim 1, characterized in that: The classification of power quality disturbance features in the distribution network using the power quality disturbance classifier specifically involves classifying the finally extracted power quality disturbance features using the SoftMax function, as shown in formula (11): （11）； in, This represents the probability of various power quality disturbances in the distribution network output by the power quality disturbance classifier. This represents the output perturbation feature vector of the fully connected layer, here. The number of categories for the SoftMax function is shown here. Refers to the first Such disturbances.

Images