Novel power distribution network voltage control operation condition dynamic discrimination method and device, electronic equipment and storage medium

Abstract

The invention provides a novel power distribution network voltage control operation condition dynamic discrimination method and device, electronic equipment and a storage medium, and relates to the field of power system research, and the discrimination method comprises the steps: constructing a first electrical characteristic quantity and a second electrical characteristic quantity, and inputting the two types of electrical characteristic quantities into a self-attention-based neural network architecture; fusing the two types of electrical characteristic quantities by adopting a cross attention mechanism; a multi-head self-attention mechanism is adopted to extract fusion features; pooling the extracted features; inputting the pooling features into a full-connection neural network, and analyzing a mutual relation between the features; inputting the feature relationship into a classifier for unsupervised classification, and generating a feature sequence corresponding to each working condition category; and decoding the feature sequence to generate an electrical information amount corresponding to each working condition category. According to the scheme, the control potential of multi-stage stations and new energy is fully considered, various types of operation conditions are generated, and an efficient and accurate guidance decision is provided for voltage control of the power distribution network.

Description

Technical Field

[0001] This application relates to the field of power system research, specifically to a novel method, device, electronic equipment, and storage medium for dynamically identifying voltage control operating conditions in distribution networks. Background Technology

[0002] The distribution network is a crucial component of the power system, responsible for converting electrical energy transmitted through high-voltage transmission lines into low-voltage electricity suitable for industrial, commercial, and residential use. With the transformation of the energy structure and the continuous increase in electricity demand, the distribution network faces numerous challenges, including the large-scale integration of distributed renewable energy sources and the increasing diversity and uncertainty of loads. New voltage control methods for distribution networks have evolved to include interactive voltage control involving sources, grids, and loads. The classification and identification of operating conditions are the primary issues in voltage control that considers the coordinated operation of sources, grids, and loads.

[0003] In existing technologies, most power distribution network voltage control methods are based on voltage control zoning for operating condition identification and control. For example, control based on the nine-zone method involves real-time monitoring of the low-voltage bus voltage and reactive power on the high-voltage side of the transformer in the substation, using these as control state variables. Based on the corresponding logical relationship of their operating points in each operating zone, control commands for capacitors and main transformer taps are derived to identify the operating condition of the substation. Simultaneously, reactive power exchange between substations at different levels is identified, and reactive power control is applied to substations with excessive reactive power exchange, thereby achieving reactive power balance and voltage stability between substations of different voltage levels. However, this method has three drawbacks: First, When the reactive power exchanged by substations at all levels exceeds the limit, the control of reactive power increase or decrease between the grids is only achieved by switching substation capacitors, without considering the reactive power margin of each substation. If there are no capacitors available for switching, the main transformer of the distribution network substation can only be adjusted, which does not achieve the expected effect. Second, the voltage control coordination capability of new energy power plants at the same voltage level is insufficient. While maintaining the stability of the grid-connected bus of the substation, there is a lack of a sound control mechanism for reactive power control of the main grid substation. Third, under conditions of insufficient reactive power margin or abnormal operation of the main grid substation, the main transformer can no longer be adjusted, and relying solely on the coordination of distribution network substations at all levels cannot achieve voltage stability control. In other words, the nine-zone method divides the control area based on static voltage and reactive power thresholds. This fixed zoning method does not consider the differences in distribution network topology and node voltage sensitivity, which leads to rigid zoning boundaries and insufficient dynamic adaptability. Although later researchers improved the nine-zone method by further refining the boundaries to obtain the thirteen-zone and seventeen-zone methods, these control methods still suffer from drawbacks. These include failing to consider the reactive power margin of substations at all levels, insufficient voltage control coordination capabilities of energy plants leading to unsatisfactory voltage stability and precise control, and not incorporating source-load interaction characteristics into the judgment criteria for voltage control guidance. Consequently, these traditional voltage operation condition judgment and control methods cannot meet the requirements of voltage control in new distribution networks.

[0004] Furthermore, in existing distribution network voltage control systems, the technology for renewable energy power plants to participate in grid voltage control generally adopts the collaborative control of the renewable energy power plant's AVC substation and the dispatch master station. The main control methods are: voltage control curve control issued by the master station and reactive power control of tie lines, etc. However, the basic control principle is the same, which is to perform reactive power distribution control on SVG, collector lines, etc. based on reactive power compensation calculation under the operating conditions of the power plant. However, this method has two drawbacks: First, the collaboration between the renewable energy power plant's AVC substation and the grid master station mainly relies on the reactive power of the grid-connected bus or tie line as the sole condition. The main purpose is to maintain the voltage stability and economic operation of the renewable energy power plant, without considering the operating conditions of the substation connected to the distribution network. When the reactive power of the upper-level distribution network is insufficient or under abnormal operating conditions, it still provides sufficient reactive power to the renewable energy power plant, which will threaten the stable operation of the grid. Second, renewable energy power plants and loads at all levels connected to the same distribution network substation lack a collaborative mechanism. The AVC substation still operates relatively independently. There is a lack of collaborative control strategies in the same area, which causes reactive power flow between substations at all levels of the distribution network, resulting in difficulties in voltage control and affecting the economic operation of the grid. In other words, this type of control method only uses the reactive power of the grid-connected bus or tie line as a condition, without considering the operating conditions of the substation connected to the distribution network. There is a lack of coordination mechanism between the various levels of new energy power plants and loads connected to the same substation. The AVC substation still operates relatively independently. The lack of coordinated control strategies in the same area leads to control lag, voltage overruns, insufficient decoupling of control strategies, reactive power distribution conflicts, limited dynamic response capabilities, voltage instability, and affects the stable operation of the power grid.

[0005] It is evident that the identification and control of voltage operating conditions in distribution networks urgently need to incorporate the application of deep correlation analysis of multi-source heterogeneous data such as new energy power plants and controllable loads, as well as cross-modal data characteristics, into the identification and control mechanism. Therefore, it is necessary to propose a new method to solve the problem of rapid and accurate classification and dynamic identification of voltage operating conditions, thereby assisting stakeholders in making effective and accurate voltage control decisions. Summary of the Invention

[0006] To address one of the aforementioned technical deficiencies, this application provides a novel method, apparatus, electronic device, and storage medium for dynamically determining the operating conditions of voltage control in power distribution networks.

[0007] According to the first aspect of this application, a novel dynamic discrimination method for voltage control operating conditions in a distribution network is provided, comprising:

[0008] Based on the collected voltage operation data, a first electrical feature quantity and a second electrical feature quantity are constructed, and the first electrical feature quantity and the second electrical feature quantity are input into a self-attention-based neural network architecture; wherein, the first electrical feature quantity is composed of the electrical features of each new energy power station connected to the target main substation, and the second electrical feature quantity is composed of the electrical features of the target main substation and the power grid substations and user substations supplied by it;

[0009] A cross-attention mechanism is used to fuse the first electrical feature quantity and the second electrical feature quantity based on the correlation between electrical features, resulting in a fused feature vector;

[0010] A multi-head self-attention mechanism is adopted to extract features from the fused feature vector based on the correlation between the fused features, thereby obtaining the extracted feature matrix;

[0011] The extracted feature matrix is ​​pooled to obtain the pooled feature matrix.

[0012] The pooled feature matrix is ​​input into the first fully connected neural network for processing, the relationships between the features are parsed, and the feature relationship matrix is ​​obtained based on the parsing results.

[0013] The feature relationship matrix is ​​input into the classifier for unsupervised classification, generating a classification feature sequence corresponding to various voltage operating conditions.

[0014] The classification feature sequence is decoded to generate and output electrical information corresponding to various voltage operating conditions.

[0015] Preferably, the first electrical characteristic quantity X1 is: ;

[0016] In the formula: j1 is the total number of new energy power stations connected to the target main substation, x i.t The electrical characteristics of the i-th renewable energy power station at time t: ;

[0017] Where: u i.t p i.t q i.t These represent the grid-connected bus voltage, tie-line active power, and reactive power of the i-th renewable energy power station at time t; p ik.t q ik.t Let Q represent the active power and reactive power of the k-th collector line of the i-th renewable energy power station at time t, respectively, and m1 be the total number of collector lines; is.t Let be the reactive power of the s-th reactive power compensation device in the i-th renewable energy power station, and s be the total number of reactive power compensation devices.

[0018] The second electrical characteristic quantity X2 is: ;

[0019] In the formula: x t Electrical characteristics of the target main substation and its supplied power grid substations and user substations:

[0020] ;

[0021] Where: u t p t q t q B.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t; u nl.t p nl.t q nl.t q nBl.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power at time t for the nth substation supplied by the target main substation, where n is the total number of substations supplied by the target main substation; u d1c.t p d1c.t q d1c.t q d1Bc.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t for the d1th user substation supplied by the target main substation, where d1 is the total number of user substations supplied by the target main substation.

[0022] Preferably, the method of employing a cross-attention mechanism to fuse the first electrical feature quantity and the second electrical feature quantity based on the correlation between electrical features to obtain a fused feature vector specifically includes:

[0023] A linear transformation is performed on the first electrical characteristic quantity X1 and the second electrical characteristic quantity X2 to obtain the first linear characteristic quantity S1 and the second linear characteristic quantity S2 with the same dimension.

[0024] A cross-attention mechanism is used to fuse the first linear feature S1 and the second linear feature S2 based on the correlation between electrical features to obtain feature S;

[0025] The feature quantity S is added to the first linear feature quantity S1 and the second linear feature quantity S2 through residual connection, and the result is then subjected to layer normalization to obtain the feature quantity Z. l ;

[0026] The characteristic Z l The input is processed in the first feedforward neural network to obtain the feature quantity Z. l+1 ;

[0027] The feature Z is connected through residual connection. l+1 With characteristic Z lThe features are added together, and the result is then normalized to obtain the fused feature vector Z.

[0028] Preferably, the multi-head self-attention mechanism is used to extract features from the fused feature vector based on the correlation between the fused features, resulting in an extracted feature matrix, specifically including:

[0029] Based on the fused feature vector Z, the query matrix Q2, key matrix K2, and value matrix V2 of the multi-head self-attention mechanism of each encoder layer in the neural network architecture are transformed to establish the relationship between the fused feature vector Z and the query matrix Q2, key matrix K2, and value matrix V2 respectively: , , Among them: W Q2 W K2 W V2 These are the weight matrices for the query matrix Q2, key matrix K2, and value matrix V2 of each head's self-attention mechanism;

[0030] The query matrix Q2, key matrix K2, and value matrix V2 of each head self-attention mechanism are transformed by the scaling dot product attention mechanism to obtain multiple attention values ​​of the fusion feature vector Z corresponding to each head self-attention mechanism.

[0031] Multiple attention values ​​are concatenated and linearly transformed to obtain the final attention value;

[0032] The final attention value is added to the fused feature vector Z through residual connections, and the result is then normalized to obtain the feature quantity H. l ;

[0033] The characteristic quantity H l The input is processed in the second feedforward neural network to obtain the feature quantity H. l+1 ;

[0034] The feature quantity H is connected via residual link. l+1 With characteristic quantity H l The sums are then added together, and the result is subjected to layer normalization to obtain the feature quantity H.

[0035] The feature quantities H output from all layers of the encoder are integrated to obtain the extracted feature matrix Hm2, where m2 is the total number of layers of the encoder based on the self-attention neural network architecture.

[0036] Preferably, the step of inputting the pooled feature matrix into the first fully connected neural network for processing, parsing the interrelationships between the features, and obtaining the feature relationship matrix based on the parsing results specifically includes:

[0037] Pooling feature matrix HF Input into the first fully connected neural network;

[0038] The pooling feature matrix H is released through the first fully connected neural network. F The relationships between all features are analyzed, and all feature relationships are re-analyzed to obtain the feature relationship matrix Y:

[0039] ;

[0040] Among them: W A B is the weight matrix of the first fully connected neural network; A Let be the bias matrix of the first fully connected neural network.

[0041] Preferably, the step of inputting the feature relationship matrix into a classifier for unsupervised classification to generate classification feature sequences corresponding to various voltage operating condition categories specifically includes:

[0042] Input the feature relation matrix Y into the classifier, and use the softmax activation function to calculate the probability distribution of each element in the feature relation matrix Y;

[0043] Based on the calculation results, a classification result is generated; the classification result is a classification feature sequence corresponding to various voltage operating conditions.

[0044] Preferably, the step of decoding the classification feature sequence to generate and output electrical information corresponding to various voltage operating condition categories specifically includes:

[0045] The classification feature sequence output by the classifier is input into the second fully connected neural network for processing;

[0046] The feature sequence C is obtained by processing the feature sequence output by the second fully connected neural network using an occlusion multi-head self-attention mechanism.

[0047] The feature sequence C and the extracted feature matrix Hm2 output by the encoder are input into the decoder for decoding.

[0048] The decoded electrical information sequence is input into a third fully connected neural network for processing to obtain electrical information corresponding to various voltage operating conditions.

[0049] According to a second aspect of this application, a novel dynamic discrimination device for voltage control operating conditions of a distribution network is provided, including a module for implementing the novel dynamic discrimination method for voltage control operating conditions of a distribution network.

[0050] According to a third aspect of this application, an electronic device is provided, comprising:

[0051] Memory;

[0052] Processor; and

[0053] Computer programs;

[0054] The computer program is stored in the memory and configured to be executed by the processor to implement the novel dynamic discrimination method for voltage control operation of power distribution networks.

[0055] According to a fourth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon; the computer program is executed by a processor to implement the novel method for dynamically determining the operating conditions of power distribution network voltage control.

[0056] The dynamic discrimination method provided in this application first constructs two types of electrical characteristic quantities: new energy power stations at all levels and target main substations, as well as the power grid substations and user substations they supply. Then, it combines cross-attention mechanism and multi-head self-attention mechanism to perform feature fusion and feature extraction on the two types of characteristic quantities. Next, it constructs a pooling layer and a classifier to perform unsupervised classification of features and generate a refined category sequence of voltage operating conditions. Finally, it uses a decoder to decode the classified category sequence and translate the classification results from model language into electrical quantity information that dispatching and operation personnel can intuitively understand and analyze.

[0057] This application aims to solve the problem of rapid and accurate dynamic identification of complex voltage operation conditions involving multiple power stations. It innovatively considers voltage operation conditions under the interaction of new energy sources, the power grid, and loads, and incorporates new energy power stations, power grid substations at all levels, and user substations at all levels into the voltage operation condition identification criteria. It fully considers the control potential of multiple power stations and new energy sources, generating more refined categories of operation conditions. This adapts to the development trend and current engineering status of modern distribution systems, breaking away from the traditional "nine-zone" voltage operation condition identification method for the power grid. It provides more scientific, reasonable, efficient, and accurate guidance strategies and decisions for distribution network voltage control; and employs a cross-attention mechanism. The fusion processing of source, grid, and load features efficiently achieves information fusion of different types of features, enhancing the model's expressive power and accuracy. An improved encoder based on a self-attention neural network architecture is used for automatic supervised classification of voltage operating conditions, improving the global feature calculation and extraction performance of multiple types of information, increasing model training efficiency and speed, and ensuring the accuracy of classification results. A decoder is employed to decode the feature quantities of different operating conditions, accurately generating electrical information for each type of operating condition, achieving efficient representation of different voltage operating conditions, facilitating reading and analysis by grid operators, and providing reliable auxiliary decision-making for grid operation.

[0058] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of what is pointed out in the written description and the accompanying drawings. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0060] Figure 1 A flowchart of the dynamic discrimination method for the voltage control operation condition of the new distribution network provided in this application. Detailed Implementation

[0061] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0062] Addressing some problems existing in the current technology:

[0063] Firstly, this application provides a novel dynamic discrimination method for voltage control operating conditions in a distribution network. This method can be executed by a dynamic discrimination device for voltage control operating conditions in a distribution network, or by components configured within the dynamic discrimination device, such as chips or chip systems, or by logic modules or software having some or all of the functions of the dynamic discrimination device. This application does not limit the scope of this method.

[0064] For example, the dynamic discrimination method for voltage control operation conditions of this novel distribution network, such as Figure 1 As shown, the dynamic discrimination method includes:

[0065] Based on the collected voltage operation data, a first electrical feature quantity and a second electrical feature quantity are constructed, and the first electrical feature quantity and the second electrical feature quantity are input into a self-attention-based neural network architecture; wherein, the first electrical feature quantity is composed of the electrical features of each new energy power station connected to the target main substation, and the second electrical feature quantity is composed of the electrical features of the target main substation and the power grid substations and user substations supplied by it;

[0066] A cross-attention mechanism is used to fuse the first electrical feature quantity and the second electrical feature quantity based on the correlation between electrical features, resulting in a fused feature vector;

[0067] A multi-head self-attention mechanism is adopted to extract features from the fused feature vector based on the correlation between the fused features, thereby obtaining the extracted feature matrix;

[0068] The extracted feature matrix is ​​pooled to obtain the pooled feature matrix.

[0069] The pooled feature matrix is ​​input into the first fully connected neural network for processing, the relationships between the features are parsed, and the feature relationship matrix is ​​obtained based on the parsing results.

[0070] The feature relationship matrix is ​​input into the classifier for unsupervised classification, generating a classification feature sequence corresponding to various voltage operating conditions.

[0071] The classification feature sequence is decoded to generate and output electrical information corresponding to various voltage operating conditions.

[0072] In some possible implementations of the first aspect, the first electrical characteristic quantity X1 is: In the formula: j1 is the total number of new energy power stations connected to the target main substation, x i.t The electrical characteristics of the i-th renewable energy power station at time t: ;

[0073] Where: u i.t p i.t q i.t These represent the grid-connected bus voltage, tie-line active power, and reactive power of the i-th renewable energy power station at time t; p ik.t q ik.t Let Q represent the active power and reactive power of the k-th collector line of the i-th renewable energy power station at time t, respectively, and m1 be the total number of collector lines; is.t Let s be the reactive power of the s-th reactive power compensation device (or SVG) in the i-th renewable energy power station, where s is the total number of reactive power compensation devices.

[0074] The second electrical characteristic quantity X2 is: In the formula: x t Electrical characteristics of the target main substation and its supplied power grid substations and user substations:

[0075] ;

[0076] Where: u t p t q t q B.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t; u nl.t p nl.t qnl.t q nBl.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power at time t for the nth substation supplied by the target main substation, where n is the total number of substations supplied by the target main substation; u d1c.t p d1c.t q d1c.t q d1Bc.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t for the d1th user substation supplied by the target main substation, where d1 is the total number of user substations supplied by the target main substation.

[0077] If the target main substation is a 220kV substation, then: u t p t q t and q B.t The values ​​of u represent the 110kV bus voltage, active power, reactive power, and remaining compensable reactive power of the 220kV substation at time t. nl.t p nl.t q nl.t and q nBl.t The bus voltage, active power, reactive power, and remaining compensable reactive power of the 220kV substation supplied to the nth 110kV substation at time t, where n is the total number of 110kV substations; u d1c.t p d1c.t q d1c.t and q d1Bc.t The bus voltage, active power, reactive power, and remaining compensable reactive power of the d1th 35kV substation supplied by the 220kV substation at time t are given, where d1 is the total number of 35kV substations.

[0078] In some possible implementations of the first aspect, before performing a linear transformation on the first electrical feature X1 and the second electrical feature X2, it is necessary to perform position encoding on all electrical feature sequences, specifically:

[0079] The positional encoding of the electrical characteristic quantity sequence at time t is performed using the sin-cos rule, as shown in the following equation:

[0080] ;

[0081] ;

[0082] Where: pos represents the physical position in the input electrical feature sequence; d5 and i represent the computational dimension and dimension count of the hidden layer inherent in the position encoder, respectively; after the input electrical features enter the position encoder, the position encoder performs position encoding calculations on each element one by one, performing a d5-dimensional scan calculation on each element to generate the corresponding encoding vector. The position encoding of each element is composed of cosine and sine functions of different frequencies, with the wavelength gradually increasing from 2π to... .

[0083] Because the input electrical features contain complex and massive electrical information such as substation electrical features, new energy power plant collectors, and reactive power regulation devices, the voltage sequences of substations and new energy power plants exhibit similarity or consistency at certain moments within this long sequence of input electrical features. However, self-attention-based neural network models primarily rely on attention mechanisms, which are permutation-invariant and unable to perceive positional information or distinguish the order of sequences. This leads to the inability to correctly distinguish similar or identical sequences with different orders. Therefore, this application incorporates positional encoding for embedding. The greatest advantage of sin-cos regular positional encoding is that it can calculate each position phasor in the positional encoding within the range of -1 to 1, ensuring stability and controllability without increasing the length of the input feature sequence or the number of failures in feature calculation.

[0084] In some possible implementations of the first aspect, the use of a cross-attention mechanism to fuse the first and second electrical features based on the correlation between electrical features to obtain a fused feature vector specifically includes:

[0085] A linear transformation is performed on the first electrical characteristic quantity X1 and the second electrical characteristic quantity X2 to obtain the first linear characteristic quantity S1 and the second linear characteristic quantity S2 with the same dimension.

[0086] A cross-attention mechanism is used to fuse the first linear feature S1 and the second linear feature S2 based on the correlation between electrical features to obtain feature S;

[0087] The feature quantity S is added to the first linear feature quantity S1 and the second linear feature quantity S2 through residual connection, and the result is then subjected to layer normalization to obtain the feature quantity Z. l ;

[0088] The characteristic Z l The input is processed in the first feedforward neural network to obtain the feature quantity Z. l+1 ;

[0089] The feature Z is connected through residual connection. l+1 With characteristic Z lThe features are added together, and the result is then normalized to obtain the fused feature vector Z.

[0090] Specifically, after constructing the first electrical feature X1 and the second electrical feature X2, the first electrical feature X1 and the second electrical feature X2 are input into the linear layer of the neural network architecture for linear transformation:

[0091] ;

[0092] Where: S1 and S2 are the output features of the linear layer (i.e., the first and second linear features); W1 and W2 are the weight matrices of the linear transformation, and b1 and b2 are the bias matrices of the linear transformation. These weight and bias matrices are composed of the connection weights between neurons in the linear layer network, mainly realizing the linear transformation of the input feature data, mapping the features to a new space, and changing the dimension of the data. Since the cross-attention mechanism requires the input feature dimensions to be consistent, in practical applications, it is necessary to define the dimensions of W1 and W2, that is, set the dimensions as (dout, din), where dou is the output data dimension and din is the input data dimension.

[0093] After obtaining the first linear feature S1 and the second linear feature S2 with consistent dimensions, an encoding operation based on the cross-attention mechanism is performed. Specifically, the cross-attention mechanism is calculated as follows:

[0094] ;

[0095] Where: CrossAttention is the cross-attention mechanism; S is the feature quantity after processing by the cross-attention mechanism, with the same dimension as S1 and S2; the softmax function is used to normalize the feature vector and map it to the output interval; W Q1 W K1 and W V1 d1 represents the linear transformation weight matrix of the query matrix Q1, key matrix K1, and value matrix V1 in the cross-attention mechanism, respectively, and d2 represents the feature vector dimension. This is a scaling factor to ensure that the gradient is more stable during backpropagation.

[0096] After the cross-attention mechanism outputs feature S, it adds feature S to the original input features (i.e., the first linear feature S1 and the second linear feature S2) through residual connections, and then performs layer normalization on the result of the addition.

[0097] ;

[0098] Where: Z lS1 represents the layer-normalized feature values; LayerNorm is the layer normalization function; S1+S2 represents the original input feature values, with the dimension unchanged.

[0099] After the first residual connection and layer normalization process, the feature Z is... l The input is further processed in the first feedforward neural network:

[0100] ;

[0101] Where: Z l+1 represents the feature quantity after processing by the first feedforward neural network, and FeedForward is the feedforward neural network function.

[0102] The feature quantity Z after processing by the first feedforward neural network l+1 Perform residual joins and layer normalization again:

[0103] ;

[0104] Where Z is the feature vector after layer normalization. The feature vector Z obtained at this time is the final fused feature vector obtained after the electrical features of the original input are processed in terms of dimension and value using the cross attention mechanism. The elements in this fused feature vector are no longer single original electrical features, but relational features rich in contextual information and interrelationship information between electrical features.

[0105] In this embodiment, since the electrical features of the substation and the electrical features of the new energy source need to be combined as the common input of the self-attention-based neural network architecture, traditional neural network models of this type usually directly concatenate or combine vectors. This results in complex and lengthy input features, increases the computational difficulty, and exposes the model to the risk of overfitting, leading to performance degradation and reduced accuracy. Therefore, this embodiment adds a cross-attention mechanism layer for the fusion calculation of the two information, realizing the fusion of two sets of different input features and improving the shortcomings of traditional models.

[0106] Furthermore, if traditional methods such as combination and splicing are used to fuse two different sets of input features, operations such as selecting some typical or representative features for fusion are performed to consider the dimensional consistency and numerical correlation of the two sets of features. This causes the original input features to lose some features, affecting the accuracy and comprehensiveness of the final discrimination result. However, the embodiments of this application first use a cross-attention mechanism to deeply and efficiently fuse the two sets of features before multi-head self-attention training. During fusion, every piece of collected running data is fully utilized. Compared with traditional vector combination and vector splicing operations, it not only retains the features of different categories of vectors, but also shortens the input vector sequence, ensuring the efficiency and speed of model training, and improving the accuracy and comprehensiveness of the final discrimination result.

[0107] In some possible implementations of the first aspect, the multi-head self-attention mechanism is employed to extract features from the fused feature vector based on the correlation between the fused features, resulting in an extracted feature matrix, specifically including:

[0108] Based on the fused feature vector Z, the query matrix Q2, key matrix K2, and value matrix V2 of the multi-head self-attention mechanism of each encoder layer in the neural network architecture are transformed to establish the relationship between the fused feature vector Z and the query matrix Q2, key matrix K2, and value matrix V2 respectively: , , Among them: W Q2 W K2 W V2 These are the weight matrices for the query matrix Q2, key matrix K2, and value matrix V2 of each head's self-attention mechanism;

[0109] The query matrix Q2, key matrix K2, and value matrix V2 of each head self-attention mechanism are transformed by the scaling dot product attention mechanism to obtain multiple attention values ​​of the fusion feature vector Z corresponding to each head self-attention mechanism.

[0110] Multiple attention values ​​are concatenated and linearly transformed to obtain the final attention value;

[0111] The final attention value is added to the fused feature vector Z through residual connections, and the result is then normalized to obtain the feature quantity H. l ;

[0112] The characteristic quantity H l The input is processed in the second feedforward neural network to obtain the feature quantity H. l+1 ;

[0113] The feature quantity H is connected via residual link. l+1 With characteristic quantity H lThe sums are then added together, and the result is subjected to layer normalization to obtain the feature quantity H.

[0114] The feature quantities H output from all layers of the encoder are integrated to obtain the extracted feature matrix Hm2, where m2 is the total number of layers of the encoder based on the self-attention neural network architecture.

[0115] Specifically, after fusing the original input electrical features through the cross-attention mechanism and the first feedforward neural network, the neural network architecture encoder, composed of a multi-layer multi-head self-attention mechanism and a second feedforward neural network, is constructed as follows based on the fused feature vector Z: The query matrix Q2, key matrix K2, and value matrix V2 of the multi-head self-attention mechanism of each encoder layer are transformed according to the following relationship:

[0116] ;

[0117] ;

[0118] ;

[0119] Among them: W Q2 W K2 W V2 These are the weight matrices for the query matrix Q2, key matrix K2, and value matrix V2 of each head's self-attention mechanism.

[0120] The scaling dot product attention mechanism is used to transform the query matrix Q2, key matrix K2, and value matrix V2 of each head self-attention mechanism obtained after the previous transformation as follows: The attention value of the fused feature vector Z is calculated through the scaling dot product attention mechanism, and the weight coefficients are calculated by performing a dot product and SoftMax normalization on the query matrix Q2 and key matrix K2:

[0121] ;

[0122] Where d3 is the dimension of the query matrix Q2, the key matrix K2, and the value matrix V2.

[0123] Multi-head self-attention mechanisms transform feature sequences using multiple self-attention mechanisms. Each self-attention head learns information from different representation subspaces, enabling simultaneous attention to information from different representation subspaces at different locations. Therefore, the attention value obtained by the i-th self-attention mechanism after transforming the input feature sequence (i.e., the fused feature vector Z) is denoted as:

[0124] ;

[0125] Among them: W i Q2 Wi K2 W i V2 Let Q2 be the weight matrix of the query matrix, key matrix, and value matrix V2 of the i-th self-attention head.

[0126] After obtaining multiple attention values ​​of the fused feature vector Z corresponding to each head's self-attention mechanism, these multiple attention values ​​are concatenated and linearly transformed to obtain the final attention value, thereby realizing the modeling expression for different constraints:

[0127] ;

[0128] Where: Concat is the concatenation function, p1 is the total number of self-attention heads; Wp1 is the weight matrix of the linear transformation relationship of the multi-head self-attention mechanism during the concatenation process;

[0129] The final attention value is added to the fused feature vector Z through residual connections, and the result is then normalized to obtain the feature quantity H. l :

[0130] ;

[0131] The characteristic quantity H l The input is processed in the second feedforward neural network to obtain the feature quantity H. l+1 :

[0132] ;

[0133] The feature quantity H is connected via residual link. l+1 With characteristic quantity H l The sums are then performed, and the result is subjected to layer normalization to obtain the feature quantity H:

[0134] ;

[0135] The feature quantities H output from all layers of the encoder are integrated to obtain the extracted feature matrix Hm2. The processing flow of a single layer encoder can be denoted as: Then the final output of the encoder is:

[0136] ;

[0137] Where: m2 is the total number of layers in the encoder of the self-attention-based neural network architecture.

[0138] After the multi-head self-attention mechanism described above, the dimension and feature content of the extracted feature matrix Hm2 are greatly expanded and enriched based on the fused feature vector Z. The contextual information and feature association information contained in each element of the extracted feature matrix Hm2 are more in-depth and comprehensive.

[0139] Since the multi-head self-attention mechanism uses multiple self-attention heads to pick up electrical features, the resulting extracted feature matrix Hm2 is relatively large, which can affect the efficiency and speed of subsequent data processing. Therefore, in this embodiment, before inputting the extracted feature matrix Hm2 into the first fully connected layer for parsing, a pooling operation is performed to generate a fixed-length sequence representation, extracting significant features, and then compressing the entire feature matrix as the input features for the classification task. This embodiment uses average pooling, which expresses features by calculating the average value of elements and their neighborhoods. This effectively preserves the global information of the feature sequence, especially in cases of accidental data distortion in electrical acquisition information. Average pooling can ignore this effect, calculate global features, and suppress the impact of accidental abrupt changes in the data sequence.

[0140] Specifically, in this embodiment, the pooling layer size is F×F, and the pooling layer stride is w, used to extract significant features:

[0141] ;

[0142] Wherein: H F Here, k represents the pooled feature matrix (i.e., the pooling feature matrix), and k is the number of steps in the pooling process. This represents the pooling region elements applied to Hm2 in each pooling process.

[0143] In some possible implementations of the first aspect, the step of inputting the pooled feature matrix into a first fully connected neural network for processing, parsing the relationships between the features, and obtaining a feature relationship matrix based on the parsing results specifically includes:

[0144] Pooling feature matrix H F Input into the first fully connected neural network;

[0145] The pooling feature matrix H is released through the first fully connected neural network. F The relationships between all features are analyzed, and all feature relationships are re-analyzed to obtain the feature relationship matrix Y:

[0146] ;

[0147] Among them: W AB is the weight matrix of the first fully connected neural network, representing the relational parameter matrix showing the connection between each neuron in each layer of the first fully connected neural network and each neuron in the previous layer. It is automatically updated and adjusted as the first fully connected layer is trained. A Let be the bias matrix of the first fully connected neural network, indicating that each output neuron has an independent bias term used to adjust the overall level of the output.

[0148] In this embodiment, a first fully connected neural network is set up here, which can infinitely release and fully analyze all feature relationships calculated and compressed through the aforementioned series of steps, and then send them to the classifier for classification, ensuring that the final classification result is both comprehensive and accurate. If the first fully connected neural network is not set up here, and the traditional model's built-in initialization neural network is used directly for processing, the subsequent classifier will ignore some features when recognizing features, resulting in inaccurate final classification results (for example, running data that should belong to six categories is ultimately only classified into four categories).

[0149] In some possible implementations of the first aspect, the step of inputting the feature relation matrix into a classifier for unsupervised classification to generate classification feature sequences corresponding to various voltage operating condition categories specifically includes:

[0150] Input the feature relation matrix Y into the classifier, and use the softmax activation function to calculate the probability distribution of each element in the feature relation matrix Y;

[0151] Based on the calculation results, a classification result is generated; the classification result is a classification feature sequence corresponding to various voltage operating conditions.

[0152] Specifically, after obtaining the feature relation matrix Y output by the first fully connected neural network, the feature relation matrix Y is used as the input feature of the classifier, and the softmax activation function is used to calculate the probability distribution of each element in the feature relation matrix Y:

[0153] ;

[0154] Where: q is the total number of all elements in the characteristic relation matrix Y; y i p is the i-th element in the characteristic relation matrix Y; i This represents the classification result corresponding to the i-th element in the feature relation matrix Y; Let θ be the probability value of the i-th element in the feature relation matrix Y corresponding to each type of output result, where θ is the built-in parameter of the Softmax network;

[0155] Specifically, the cross-entropy loss function in the training of the Softmax classifier is shown in the following equation:

[0156] ;

[0157] Where: m3 is the total number of training samples; y ij For the input sample y i Network output at time; t ij The target output category for the i-th sample;

[0158] By calculating the probability distribution above, the calculated proportion (i.e., probability) of each input feature relative to all input features is obtained. Based on this calculation result, category classification is performed, and the final output is a classification feature sequence corresponding to various voltage operating conditions of the distribution network. The above classification process is unsupervised learning, which does not require setting categories and labels. It only requires classifying the output quantity Y calculated from the features to automatically output the calculated category.

[0159] In this embodiment, a self-attention-based neural network architecture with an original structure is used to process the complex and lengthy feature sequences of a large-scale active power distribution network (including substations and new energy power plants). A large number of sequences with the same characteristics are identified and extracted. Through continuous calculation and extraction using a multi-head self-attention mechanism, a feature vector with highly representative properties is finally obtained. Then, a classifier uses probability calculation analysis (the entropy probability of a certain type of feature relative to all features, i.e., exponential expression) to achieve the inductive output of different sequences with the same feature. Therefore, the output category of the classifier is the category of the operating condition of the power distribution network in that area. However, at this point, the category expression of different operating conditions is still the extracted feature sequence, which cannot be expressed as ordinary electrical quantity information for parsing and naming operating condition features. Therefore, this embodiment adds a decoder to decode the feature sequences of different categories output by the classifier to obtain the corresponding category of electrical quantity sequences, i.e., the expression of voltage and power.

[0160] In some possible implementations of the first aspect, the decoding of the classification feature sequence to generate and output electrical information corresponding to various voltage operating condition categories specifically includes:

[0161] The classification feature sequence output by the classifier is input into the second fully connected neural network for processing;

[0162] The feature sequence C is obtained by processing the feature sequence output by the second fully connected neural network using an occlusion multi-head self-attention mechanism.

[0163] The feature sequence C and the extracted feature matrix Hm2 output by the encoder are input into the decoder for decoding.

[0164] The decoded electrical information sequence is input into a third fully connected neural network for processing to obtain electrical information corresponding to various voltage operating conditions.

[0165] Since the multi-head self-attention mechanism and pooling operations used in this scheme are all local feature extraction calculations, adding fully connected neural networks to the front and back ends of the decoder can further integrate the local features extracted by the self-attention mechanism and pooling layers into global features, thereby improving the model's ability to learn overall data features.

[0166] Specifically, in this embodiment, a decoder is used to decode the classification feature sequence output by the classifier to generate corresponding electrical information to characterize the features of different operating conditions. The input of the decoder in this embodiment mainly includes the classification feature sequence output by the classifier and the extracted feature matrix Hm2 output by the encoder. Its structure is generally similar to that of the encoder in this embodiment. The difference is that the decoder adds fully connected neural networks to both the input and output for data dimension transformation. The calculation principle of the two fully connected neural networks is consistent with the first fully connected neural network located at the front end of the classifier. At the same time, an occlusion multi-head self-attention mechanism is added to the front end of the decoder.

[0167] ;

[0168] Where: Q3, K3, and V3 are the query matrix, key matrix, and value matrix of the decoder's occlusion multi-head self-attention mechanism, respectively; d4 is the dimension of the query matrix Q3, key matrix K3, and value matrix V3; D is the introduction matrix. By introducing matrix D to set future sequence information to zero, the label data information can be occluded in the multi-head self-attention mechanism.

[0169] After decoding, the various category feature sequences corresponding to different voltage operating conditions are translated into corresponding electrical information sequences (i.e., sequences expressed in terms of electrical quantity and power). The output data structure is consistent with the input data structure, with several sequences corresponding to several categories. Each category outputs a set of data, and each set of data includes the electrical feature matrices of each renewable energy power station connected to the target substation, the electrical feature matrices of the target substation and its supplied power grid substations, and the electrical feature matrices of the target substation and its supplied user substations. At this point, the decoder outputs no longer large-scale and complex sequences, but rather electrical quantity sequences with highly representative characteristics, the number of which equals the number of categories. Each sequence highly expresses different electrical characteristics. Thus, the translation of compressed, calculated, extracted, and encoded features into electrical quantity expressions such as voltage and power that dispatching and operation personnel can analyze is completed, achieving user-friendly interaction between humans and deep learning models and improving decision support.

[0170] From the overall scheme, it can be seen that, in terms of model structure, the self-attention-based neural network architecture in this application differs from existing neural network architectures in at least the following ways: First, a cross-attention mechanism layer is added at the front end of the encoder layer to achieve deep and efficient fusion of different feature vectors; second, a first fully connected layer and a classifier are added at the back end of the encoder layer to achieve the model's self-supervised training and recognition function. In practical applications, after inputting electrical feature quantities at any time, the dynamic discrimination of voltage operating conditions at that time can be achieved. For each discrimination result, the number of operating condition categories is uncertain because the operating conditions of different substations and the number of types of new energy power plants connected are different, resulting in different operating conditions. The reason why this application proposes this operating condition discrimination method is that the traditional nine-square grid (i.e., the nine-zone method) only focuses on substation voltage and power factor, simply dividing it into nine squares to describe nine operating conditions. First, it does not include new energy sources, and second, the standard for classifying operating conditions is singular and cannot be applied to the development trend of voltage control in new distribution networks. Therefore, the advantage of this application is that this method can achieve refined discrimination of operating conditions. The operating conditions mainly include: voltage operation characteristics of new energy power plants, reactive power characteristics of new energy power plants, reactive power characteristics of substations, voltage characteristics of substations, reactive power flow between substations, and other multi-dimensional characteristics.

[0171] Secondly, this application provides an apparatus for dynamically identifying the operating conditions of a new type of distribution network voltage control. The dynamic identification apparatus includes a module for implementing the aforementioned dynamic identification method for the operating conditions of a new type of distribution network voltage control.

[0172] Thirdly, this application provides an electronic device, which can be any device capable of realizing the aforementioned dynamic discrimination method for the voltage control operation of the new power distribution network. The device can be various terminal devices, such as desktop computers, laptops, tablets, handheld devices, etc., and can be implemented through software and / or hardware.

[0173] For example, the device includes:

[0174] Memory;

[0175] Processor; and

[0176] Computer programs;

[0177] The computer program is stored in the memory and configured to be executed by the processor to implement the novel dynamic discrimination method for voltage control operation of power distribution networks.

[0178] Fourthly, this application provides a computer-readable storage medium, which may be a ROM, RAM, disk, or optical disk, etc.

[0179] For example, the computer-readable storage medium stores a computer program; the computer program is executed by a processor to implement the novel method for dynamically determining the operating conditions of power distribution network voltage control.

[0180] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0181] This application is described with reference to flowchart illustrations of methods, apparatus, and computer program products according to embodiments of this application. It should be understood that each step in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart.

[0182] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0183] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 A process or multiple process diagrams specify the steps of a function.

[0184] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0185] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0186] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic discrimination method for voltage control operation conditions in a new type of power distribution network, characterized in that: include: Based on the collected voltage operation data, a first electrical feature quantity and a second electrical feature quantity are constructed, and the first electrical feature quantity and the second electrical feature quantity are input into a self-attention-based neural network architecture; wherein, the first electrical feature quantity is composed of the electrical features of each new energy power station connected to the target main substation, and the second electrical feature quantity is composed of the electrical features of the target main substation and the power grid substations and user substations supplied by it; A cross-attention mechanism is used to fuse the first electrical feature quantity and the second electrical feature quantity based on the correlation between electrical features, resulting in a fused feature vector; A multi-head self-attention mechanism is adopted to extract features from the fused feature vector based on the correlation between the fused features, thereby obtaining the extracted feature matrix; The extracted feature matrix is ​​pooled to obtain the pooled feature matrix. The pooled feature matrix is ​​input into the first fully connected neural network for processing, the relationships between the features are parsed, and the feature relationship matrix is ​​obtained based on the parsing results. The feature relationship matrix is ​​input into the classifier for unsupervised classification, generating a classification feature sequence corresponding to various voltage operating conditions. The classification feature sequence is decoded to generate and output electrical information corresponding to various voltage operating conditions.

2. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 1, characterized in that: The first electrical characteristic quantity X1 is: ; In the formula: j1 is the total number of new energy power stations connected to the target main substation, x i.t The electrical characteristics of the i-th renewable energy power station at time t: ; Where: u i.t p i.t q i.t These represent the grid-connected bus voltage, tie-line active power, and reactive power of the i-th renewable energy power station at time t; p ik.t q ik.t Let Q represent the active power and reactive power of the k-th collector line of the i-th renewable energy power station at time t, respectively, and m1 be the total number of collector lines; is.t Let be the reactive power of the s-th reactive power compensation device in the i-th renewable energy power station, and s be the total number of reactive power compensation devices. The second electrical characteristic quantity X2 is: ; In the formula: x t Electrical characteristics of the target main substation and its supplied power grid substations and user substations: ； Where: u t p t q t q B.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t; u nl.t p nl.t q nl.t q nBl.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power at time t for the nth substation supplied by the target main substation, where n is the total number of substations supplied by the target main substation; u d1c.t p d1c.t q d1c.t q d1Bc.t These represent the bus voltage, active power, reactive power, and remaining compensable reactive power of the target main substation at time t for the d1th user substation supplied by the target main substation, where d1 is the total number of user substations supplied by the target main substation.

3. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 2, characterized in that: The method employs a cross-attention mechanism to fuse the first and second electrical feature quantities based on the correlation between electrical features, resulting in a fused feature vector. Specifically, this includes: A linear transformation is performed on the first electrical characteristic quantity X1 and the second electrical characteristic quantity X2 to obtain the first linear characteristic quantity S1 and the second linear characteristic quantity S2 with the same dimension. A cross-attention mechanism is used to fuse the first linear feature S1 and the second linear feature S2 based on the correlation between electrical features to obtain feature S; The feature quantity S is added to the first linear feature quantity S1 and the second linear feature quantity S2 through residual connection, and the result is then subjected to layer normalization to obtain the feature quantity Z. l ; The characteristic Z l The input is processed in the first feedforward neural network to obtain the feature quantity Z. l+1 ; The feature Z is connected through residual connection. l+1 With characteristic Z l The features are added together, and the result is then normalized to obtain the fused feature vector Z.

4. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 3, characterized in that: The multi-head self-attention mechanism is employed to extract features from the fused feature vector based on the correlation between the fused features, resulting in an extracted feature matrix. Specifically, this includes: Based on the fused feature vector Z, the query matrix Q2, key matrix K2, and value matrix V2 of the multi-head self-attention mechanism of each encoder layer in the neural network architecture are transformed to establish the relationship between the fused feature vector Z and the query matrix Q2, key matrix K2, and value matrix V2 respectively: , , ;W Q2 W K2 W V2 These are the weight matrices for the query matrix Q2, key matrix K2, and value matrix V2 of each head's self-attention mechanism; The query matrix Q2, key matrix K2, and value matrix V2 of each head self-attention mechanism are transformed by the scaling dot product attention mechanism to obtain multiple attention values ​​of the fusion feature vector Z corresponding to each head self-attention mechanism. Multiple attention values ​​are concatenated and linearly transformed to obtain the final attention value; The final attention value is added to the fused feature vector Z through residual connections, and the result is then normalized to obtain the feature quantity H. l ; The characteristic quantity H l The input is processed in the second feedforward neural network to obtain the feature quantity H. l+1 ; The feature quantity H is connected through residual linking. l+1 With characteristic quantity H l The sums are then added together, and the result is subjected to layer normalization to obtain the feature quantity H. The feature quantities H output from all layers of the encoder are integrated to obtain the extracted feature matrix Hm2, where m2 is the total number of layers of the encoder based on the self-attention neural network architecture.

5. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 4, characterized in that: The process of inputting the pooled feature matrix into the first fully connected neural network for processing, parsing the relationships between the features, and obtaining the feature relationship matrix based on the parsing results specifically includes: Pooling feature matrix H F Input into the first fully connected neural network; The pooling feature matrix H is released through the first fully connected neural network. F The relationships between all features are analyzed, and all feature relationships are re-analyzed to obtain the feature relationship matrix Y: ； Among them: W A B is the weight matrix of the first fully connected neural network; A Let be the bias matrix of the first fully connected neural network.

6. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 5, characterized in that: The step of inputting the feature relationship matrix into the classifier for unsupervised classification to generate classification feature sequences corresponding to various voltage operating condition categories specifically includes: Input the feature relation matrix Y into the classifier, and use the softmax activation function to calculate the probability distribution of each element in the feature relation matrix Y; Based on the calculation results, a classification result is generated; the classification result is a classification feature sequence corresponding to various voltage operating conditions.

7. The dynamic discrimination method for the voltage control operation condition of a novel power distribution network according to claim 6, characterized in that: The process of decoding the classification feature sequence to generate and output electrical information corresponding to various voltage operating condition categories specifically includes: The classification feature sequence output by the classifier is input into the second fully connected neural network for processing; The feature sequence C is obtained by processing the feature sequence output by the second fully connected neural network using an occlusion multi-head self-attention mechanism. The feature sequence C and the extracted feature matrix Hm2 output by the encoder are input into the decoder for decoding. The decoded electrical information sequence is input into a third fully connected neural network for processing to obtain electrical information corresponding to various voltage operating conditions.

8. A new dynamic discrimination device for voltage control operation conditions of power distribution networks, characterized in that: It includes a module for implementing the dynamic discrimination method for the voltage control operation condition of the novel distribution network as described in any one of claims 1 to 7.

9. An electronic device, characterized in that: include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the novel dynamic discrimination method for voltage control operation of the power distribution network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: It stores a computer program; the computer program is executed by a processor to implement the dynamic discrimination method for the operating conditions of the new distribution network voltage control as described in any one of claims 1 to 7.

Images