A deep learning-based method for generating power grid design schemes

Abstract

The present invention discloses a method for generating a power grid design scheme based on deep learning, which belongs to the field of power grid planning technology. The method comprises the following steps: S1, obtaining power grid description text and power grid line text of the power grid design scheme; S2, performing word embedding representation on the texts to obtain a word vector matrix; S3, extracting features from the word vector matrix using two parallel convolutional networks that introduce an attention mechanism to obtain a text feature vector; S4, performing feature fusion on the text feature vector using a bidirectional LSTM network to obtain a comprehensive feature representation; S5, learning the comprehensive feature representations separately through a collaborative filtering module to obtain a prediction vector representation; and S6, fusing the prediction vector representations to obtain a generated power grid design scheme. The method can establish a deep learning model to learn power grid features by utilizing power grid data, including implicit feedback data and auxiliary features, thereby achieving accurate, efficient, and personalized recommendations for power grid design schemes.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid planning, and specifically relates to a method for generating a power grid design scheme based on deep learning. Background Art

[0002] As power grids expand in size and complexity, traditional grid planning and design methods are no longer able to meet the growing data processing demands. To improve the efficiency and quality of grid planning and design, design recommendation methods are beginning to incorporate deep learning technology. Deep learning can automatically learn complex patterns and relationships from massive amounts of data, providing more accurate predictions and analysis for grid design. Design recommendation methods, combined with deep learning technology, aim to accurately describe specific grid design solutions by analyzing the lines used and detailed information of grid solutions, thereby providing effective support for power system solution selection. This approach not only improves design efficiency but also ensures the quality of grid planning and the stability of grid operation.

[0003] The recent development of deep neural network technology has brought new insights to power grid design. Collaborative filtering (CF) algorithms, particularly matrix factorization (MF) algorithms, have been widely used in the construction of recommendation systems due to their unique ability to leverage collective intelligence and experience. Matrix factorization assumes that user-item relationships are linked by hidden factors. By mapping users and items into the same representation space until they can be directly compared, the similarity between them is used to measure the quality of the match. In this case, the model learns low-dimensional vectors for users and items and ultimately uses the dot product as the matching equation to calculate the score. Deep neural networks (DNNs) have excellent performance in representation learning and are widely used. In this way, we apply this approach to power grid design. By leveraging the powerful learning capabilities of deep learning, we learn the connections between the two and recommend suitable power grid design solutions to users, thereby improving the efficiency and quality of power grid planning and design. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the deep learning-based power grid design scheme generation method provided by the present invention solves the inefficient design situation in the traditional power grid design scheme recommendation process that relies on expert experience and manual analysis, as well as the limitations of traditional data processing methods in extracting valuable information, which in turn affects the accuracy of power grid design scheme recommendations.

[0005] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a method for generating a power grid design scheme based on deep learning, comprising the following steps:

[0006] S1. Obtaining a grid description text and grid line text of a grid design plan;

[0007] S2. Perform word embedding on the power grid description text and the power grid line text to obtain the corresponding power grid description word vector matrix and power grid line word vector matrix;

[0008] S3. Using two parallel convolutional networks that introduce an attention mechanism to perform feature extraction on the power grid description word vector matrix and the power grid line word vector matrix, respectively, to obtain corresponding power grid description text feature vectors and power grid line text feature vectors;

[0009] S4. Using a bidirectional LSTM network to perform feature fusion on the power grid description text feature vector and the power grid line text feature vector to obtain a comprehensive feature representation;

[0010] S5. Sequentially learning the comprehensive feature representation through a collaborative filtering module based on representation learning and a collaborative filtering module based on matching function learning to obtain a corresponding prediction vector representation;

[0011] S6. Fusing the prediction vector representations obtained based on representation learning and matching function learning to obtain a generated power grid design solution.

[0012] Furthermore, in step S3, the convolutional network with the attention mechanism is introduced to the word vector matrix A of any power grid description / power grid line u. u The method for feature extraction and obtaining corresponding text features is as follows:

[0013] S31. Determine the word vector matrix A u Each word vector w k u The attention weight a k ;

[0014] S32, the attention weights of all word vectors are calculated based on the word vector matrix A u The word order is spliced ​​to obtain the word vector matrix A u The corresponding attention weight A(u)=(a1,a2,…,a n ), k = 1, 2, 3, ..., n, k represents the index of the word vector element;

[0015] S33, the attention weight A(u) and the word vector matrix A u Perform corresponding multiplication to obtain the updated word vector matrix

[0016] S34, normalize the updated word vector matrix and extract its features through a set of convolutional layers to generate the corresponding feature map C i ;

[0017] S35, feature map C iPerform maximum pooling to obtain a simplified feature representation;

[0018] S36. The simplified feature representation is subjected to low-dimensional feature mapping through a fully connected layer to obtain text features corresponding to the power grid description / power grid line.

[0019] Furthermore, in step S31, the attention weight a k Expressed as:

[0020]

[0021] Where g k represents the result of linear transformation of the kth element in the input sequence, ReLU(·) represents the ReLU activation function, A1 represents the weight parameter matrix, b1 represents the bias term when the first element in the input sequence is linearly transformed, and g j Represents the result of linear transformation of the jth element in the input sequence, j represents the element index, and n represents the sequence length;

[0022] In step S34, the method for extracting features through the convolutional layer is:

[0023] Let the convolution kernel W of each convolution layer ij After updating the word vector matrix Slide up, perform weighted summation and add bias terms, and then generate feature map C through activation function i , which is expressed as:

[0024]

[0025] Where σ(·) represents the activation function, q i represents the bias term corresponding to the i-th feature map, Represents the updated word vector matrix The j-th element in , the subscript i represents the index of the feature map, and m represents the number of feature maps;

[0026] In step S36, the grid description / grid line text feature F is represented as:

[0027] F=σ(W f *P f +z f )

[0028] Where σ(·) represents the activation function, W f represents the weight, z f Represents the bias term corresponding to the network neuron f, P f Represents the feature representation of the pooling layer output.

[0029] Furthermore, in step S4, the output y of the bidirectional LSTM network at the last time step T is converted to T As the comprehensive fusion representation of the power grid description text feature vector and the power grid line text feature vector, output = y T .

[0030] Furthermore, in step S5, the collaborative filtering network based on learning representation uses a multi-layer perceptron to learn the potential vector representation p of the power grid description for the comprehensive feature representation. j and the latent vector representation p of the power grid lines u , get the prediction vector representation based on representation learning ⊙ represents the Hadamard product operation.

[0031] Furthermore, the collaborative filtering network based on learning representation uses a multi-layer perceptron to learn the corresponding latent vector representation of the comprehensive feature representation as follows:

[0032]

[0033] In the formula, ReLU(·) is the ReLU activation function, c0 represents the hidden state of the input sequence, c1 represents the output of the first hidden layer, c2 represents the output after the next hidden layer processes the state of c1, and c x-1 Representative in c x-2 The output after the next hidden layer processing, c x Representative in c x-1 The output after the next hidden layer processing, W0, W1, ..., W x Represent the weight matrices of different intermediate layers, d1, d2, ..., d x Represents the bias items corresponding to the first hidden layer, the second hidden layer to the xth hidden layer, y u Represents input data.

[0034] Furthermore, in step S5, the collaborative filtering network based on matching function learning first uses a linear embedding layer to learn the potential vector representation p of the power grid description. j and the latent vector representation p of the power grid lines u , and then process it using a multi-layer perception network to obtain the corresponding prediction vector representation based on the matching function learning.

[0035] Furthermore, the corresponding prediction vector representation is learned based on the matching function The process is expressed as:

[0036]

[0037] Where P and Q represent the weight matrices of the linear embedding layer, which are used to transform l u and l j The original features are mapped to the vector representation in the latent space, l u The characteristics of the lines representing the power grid scheme, l j Represents the features of the detailed situation of the power grid scheme, v0 represents the potential vector representation p j and the latent vector representation p u The concatenated vector, v1 represents the output vector after the first hidden layer, v Y Represents the final output vector of the multilayer perceptron, v Y-1 Represents the output vector of the second-to-last layer of the multilayer perceptron, W y1 、W Y They represent the weight matrices of the first hidden layer and the last hidden layer in the multilayer perceptron, k1 represents the bias term corresponding to the first layer in the multilayer perceptron, and k Y represents the bias term corresponding to the y-th layer in the multilayer perceptron, and ReLU(·) represents the ReLU activation function.

[0038] Furthermore, in step S6, the prediction vector representation is fused to form:

[0039]

[0040] Where, represents the generated power grid design scheme, σ(·) represents the activation function, W out Represents the weight matrix of the output layer.

[0041] The beneficial effects of the present invention are:

[0042] (1) By studying relevant knowledge in the field of power grid planning and design and combining the characteristics and shortcomings of traditional power grid design solution recommendation methods, this paper proposes an innovative deep learning recommendation network, GridDeepCF, to effectively assist power engineers in identifying and selecting the optimal power grid design solution. In GridDeepCF, the traditional recommendation system is improved by introducing two key network modules. Through in-depth feature extraction and fusion of power grid data and an enhanced focus on the key elements of power grid solutions, the quality and relevance of the recommended solutions are improved.

[0043] (2) Through extensive experiments on multiple real-world power grid datasets, we validate the effectiveness of our proposed recommendation network, GridDeepCF. Compared with other state-of-the-art methods, our proposed method achieves significant improvements in accuracy, efficiency, and personalization of power grid design recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the method for generating a power grid design solution based on deep learning in the present invention. DETAILED DESCRIPTION

[0045] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0046] The embodiment of the present invention provides a method for generating a power grid design scheme based on deep learning, such as Figure 1 As shown, the following steps are included:

[0047] S1. Obtaining a grid description text and grid line text of a grid design plan;

[0048] S2. Perform word embedding on the power grid description text and the power grid line text to obtain the corresponding power grid description word vector matrix and power grid line word vector matrix;

[0049] S3. Using two parallel convolutional networks that introduce an attention mechanism to perform feature extraction on the power grid description word vector matrix and the power grid line word vector matrix, respectively, to obtain corresponding power grid description text feature vectors and power grid line text feature vectors;

[0050] S4. Using a bidirectional LSTM network to perform feature fusion on the power grid description text feature vector and the power grid line text feature vector to obtain a comprehensive feature representation;

[0051] S5. Sequentially learning the comprehensive feature representation through a collaborative filtering module based on representation learning and a collaborative filtering module based on matching function learning to obtain a corresponding prediction vector representation;

[0052] S6. Fusing the prediction vector representations obtained based on representation learning and matching function learning to obtain a generated power grid design solution.

[0053] In an embodiment of the present invention, the convolutional network, bidirectional LSTM network and collaborative filtering module that introduce the attention mechanism in the above method are formed into a network framework GridDeepCF for generating power grid design schemes. The input of the network framework is the lines used in the power grid scheme and a detailed description of the power grid scheme. The generated power grid design scheme is obtained through the processing output of the above network framework.

[0054] In step S3 of the embodiment of the present invention, in order to enhance the importance of text features, an attention mechanism is introduced into the convolutional network CNN, which can assign different weights to each word in the input text, thereby highlighting the words that have the greatest impact on the text description.

[0055] In step S3 of this embodiment, the convolutional network with the attention mechanism is used to generate the word vector matrix A of any power grid description / power grid line u. u The method for feature extraction and obtaining corresponding text features is as follows:

[0056] S31. Determine the word vector matrix A u Each word vector in The attention weight a k ;

[0057] S32, the attention weights of all word vectors are calculated based on the word vector matrix A u The word order is spliced ​​to obtain the word vector matrix A u The corresponding attention weight A(u)=(a1,a2,…,a n ), k = 1, 2, 3, ..., n, k represents the index of the word vector element;

[0058] S33, the attention weight A(u) and the word vector matrix A u Perform corresponding multiplication to obtain the updated word vector matrix

[0059] S34, normalize the updated word vector matrix and extract its features through a set of convolutional layers to generate the corresponding feature map C i ;

[0060] S35, feature map C i Perform maximum pooling to obtain a simplified feature representation;

[0061] S36. The simplified feature representation is subjected to low-dimensional feature mapping through a fully connected layer to obtain text features corresponding to the power grid description / power grid line.

[0062] In the above step S31 of this embodiment, the attention weight a k Expressed as;

[0063]

[0064] Where g k represents the result of linear transformation of the kth element in the input sequence, ReLU(·) represents the ReLU activation function, A1 represents the weight parameter matrix, b1 represents the bias term when the first element in the input sequence is linearly transformed, and g jRepresents the result of linear transformation of the jth element in the input sequence, j represents the element index, and n represents the sequence length;

[0065] In step S34, the method for extracting features through the convolutional layer is:

[0066] In the above step S34 of this embodiment, the method of performing feature extraction through the convolution layer is:

[0067] Let the convolution kernel W of each convolution layer ij After updating the word vector matrix Slide up, perform weighted summation and add bias terms, and then generate feature map C through activation function i , which is expressed as:

[0068]

[0069] Where σ(·) represents the activation function, q i represents the bias term corresponding to the i-th feature map, Represents the updated word vector matrix The j-th element in , the subscript i represents the index of the feature map, and m represents the number of feature maps;

[0070] In step S34 of this embodiment, the weighted summed word vector matrix is ​​normalized. Taking the power grid line as an example, the usage of each line is normalized to be within the range of [0, 1]. This can accelerate the network and improve the generalization ability of the model.

[0071] In step S35 of this embodiment, in order to reduce the spatial dimension of the feature, a pooling layer is introduced to select the maximum value max(C i ) to simplify feature representation.

[0072] In step S36 of this embodiment, the grid description / grid line text feature F is represented as:

[0073] F=σ(W f *P f +z f )

[0074] Where σ(·) represents the activation function, W f represents the weight, z f Represents the bias term corresponding to the network neuron f, P f Represents the feature representation of the pooling layer output.

[0075] In this embodiment, the power grid description / power grid line text feature F is obtained through the above steps, and then the corresponding feature vector is obtained, including the power grid description text feature vector output j and power grid line text feature vector output u , which is used as the input of the subsequent bidirectional LSTM network.

[0076] In this embodiment, text features play an important role in the process of realizing power grid design scheme generation. By mapping professional terms into dense vector representations, the network structure can capture the deep semantic information in the text data.

[0077] In a specific example of this embodiment, the professional terms embedded in the matrix, such as "steel core aluminum stranded wire," "110kV," "transmission line," "transformer," "GB standard," "high temperature," "industrial load," and "mountainous area," not only represent the physical properties and engineering requirements of the power grid design, but also reflect the operating characteristics of the power grid under different environmental conditions. These text features serve as input features, enabling the neural network to learn the similarities and differences between different design solutions in terms of material selection, voltage level, line type, equipment configuration, compliance standards, environmental challenges, load types served, and geographical location. When faced with a new power grid design project, such as one located in a mountainous area, the model has learned through training to identify text features related to "mountainous area." These text features carry the environmental factors and design considerations unique to mountainous areas, enabling the model to accurately recommend power grid design solutions that are adapted to the complex terrain and climatic conditions of mountainous areas, ensuring that the design solutions are both economical, efficient, safe, and reliable. In this way, the network structure transforms the word vector matrix of professional terms into a knowledge representation that deeply understands the design solutions, thereby greatly improving the accuracy and practicality of the recommendations.

[0078] In a specific example of this embodiment, the designed convolutional network includes two convolutional layers, two pooling layers, and a fully connected layer. Convolutional layer 1 uses 32 3x3 convolution kernels with a stride of 1, a border padding of 1, an activation function of ReLU, weights initialized with He, and bias items initialized to 0; followed by pooling layer 1, using a 2x2 maximum pooling window with a stride of 2 to reduce the spatial dimension of the feature map. Convolutional layer 2 further extracts features, using 64 3x3 convolution kernels with a stride of 1, a border padding of 1, and also using the ReLU activation function. The initialization of weights and bias items is the same as that of convolutional layer 1; the following pooling layer 2 also uses a 2x2 maximum pooling window with a stride of 2. Fully connected layer 1 maps the output of pooling layer 2 to a low-dimensional space, has 128 neurons, and uses ReLU as the activation function. The weights and bias items are initialized with He and 0, respectively.

[0079] In step S4 of the embodiment of the present invention, the output y of the bidirectional LSTM network at the last time step T is converted to T As the comprehensive fusion representation of the power grid description text feature vector and the power grid line text feature vector, output = y T .

[0080] In this embodiment, a bidirectional LSTM network structure is used to simultaneously capture the forward and backward information of text features. At each time step, the forward LSTM layer processes the information from the beginning of the text to the current point, while the backward LSTM layer processes the information from the end of the text to the current point. The outputs of these two layers of LSTM are merged at each time step, thereby forming a comprehensive feature vector at each moment, which integrates the forward-looking and retrospective context of the text. Subsequently, these merged feature vectors are sent to the fully connected layer, where deeper processing and abstraction are performed to extract richer and more abstract feature representations. When all time steps are processed, the output of the bidirectional LSTM at the last time step is taken as the final fusion representation of the two text features. This representation fully reflects the global semantic information of the two text features and provides strong feature support for subsequent tasks.

[0081] In step S5 of the embodiment of the present invention, the collaborative filtering network based on the learning representation uses a multi-layer perceptron to learn the potential vector representation p of the power grid description for the comprehensive feature representation. j and the latent vector representation p of the power grid lines u , get the prediction vector representation v based on representation learning Y rl =p u ⊙p j , ⊙ represents the Hadamard product operation.

[0082] In this embodiment, the collaborative filtering network based on learning representation uses a multi-layer perceptron to learn the corresponding latent vector representation of the comprehensive feature representation as follows:

[0083]

[0084] In the formula, ReLU(·) is the ReLU activation function, c0 represents the hidden state of the input sequence, c1 represents the output of the first hidden layer, c2 represents the output after the next hidden layer processes the state of c1, and c x-1 Representative in c x-2 The output after the next hidden layer processing, c x Representative in c x-1 The output after the next hidden layer processing, W0, W1, ..., W x Represent the weight matrices of different intermediate layers, d1, d2, ..., dx Represents the bias items corresponding to the first hidden layer, the second hidden layer to the xth hidden layer, y u Represents input data.

[0085] In step S5 of the embodiment of the present invention, the collaborative filtering network based on matching function learning first uses a linear embedding layer to learn the potential vector representation p of the power grid description. j and the latent vector representation p of the power grid lines u , and then process it using a multi-layer perception network to obtain the corresponding prediction vector representation based on the matching function learning.

[0086] In this embodiment, the corresponding prediction vector representation is learned based on the matching function The process is expressed as:

[0087]

[0088] Where P and Q represent the weight matrices of the linear embedding layer, which are used to transform l u and l j The original features are mapped to the vector representation in the latent space, l u The characteristics of the lines representing the power grid scheme, l j Represents the features of the detailed situation of the power grid scheme, v0 represents the potential vector representation p j and the latent vector representation p u The concatenated vector, v1 represents the output vector after the first hidden layer, v Y Represents the final output vector of the multilayer perceptron, v Y-1 Represents the output vector of the second-to-last layer of the multilayer perceptron, W y1 、W Y They represent the weight matrices of the first hidden layer and the last hidden layer in the multilayer perceptron, k1 represents the bias term corresponding to the first layer in the multilayer perceptron, and k Y represents the bias term corresponding to the y-th layer in the multilayer perceptron, and ReLU(·) represents the ReLU activation function.

[0089] In step S6 of the embodiment of the present invention, the prediction vector representations obtained from the two parts are finally fused to obtain the final output of the entire model; based on this, the prediction vector representation is fused as follows:

[0090]

[0091] Where, represents the generated power grid design scheme, σ(·) represents the activation function, W out Represents the weight matrix of the output layer.

[0092] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0093] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A method for generating a power grid design scheme based on deep learning, characterized in that: The following steps are involved: S1. Obtaining a grid description text and grid line text of a grid design plan; S2. Perform word embedding on the power grid description text and the power grid line text to obtain the corresponding power grid description word vector matrix and power grid line word vector matrix; S3. Using two parallel convolutional networks that introduce an attention mechanism to perform feature extraction on the power grid description word vector matrix and the power grid line word vector matrix, respectively, to obtain corresponding power grid description text feature vectors and power grid line text feature vectors; S4. Using a bidirectional LSTM network to perform feature fusion on the power grid description text feature vector and the power grid line text feature vector to obtain a comprehensive feature representation; S5. Sequentially learning the comprehensive feature representation through a collaborative filtering module based on representation learning and a collaborative filtering module based on matching function learning to obtain a corresponding prediction vector representation; S6. Fusing the prediction vector representations obtained based on representation learning and matching function learning to obtain a generated power grid design solution; In step S5, the collaborative filtering network based on the learning representation uses a multi-layer perceptron to learn the potential vector representation of the power grid description for the comprehensive feature representation. and the latent vector representation of the power grid lines , get the prediction vector representation based on representation learning , represents the Hadamard product operation; The collaborative filtering network based on learning representation uses a multi-layer perceptron to learn the corresponding latent vector representation of comprehensive feature representation as follows: Where, for activation function, Represents the hidden state of the input sequence, represents the output of the first hidden layer, Representatives in The output after the next hidden layer processing, Representatives in The output after processing by the next hidden layer is Representatives in The output after processing by the next hidden layer is 、 、…、 Represent the weight matrices of different intermediate layers, 、 、…、 Represent the bias items corresponding to the first hidden layer, the second hidden layer to the xth hidden layer, respectively. Represents input data; In step S5, the collaborative filtering network based on matching function learning first uses a linear embedding layer to learn the potential vector representation of the power grid description. and the latent vector representation of the power grid lines , and then process it using a multi-layer perception network to obtain the corresponding prediction vector representation based on the matching function learning. .

2. The method for generating a power grid design scheme based on deep learning according to claim 1, characterized in that: In step S3, the convolutional network with the attention mechanism is introduced to any power grid description / power grid line The word vector matrix The method for feature extraction and obtaining corresponding text features is as follows: S31. Determine the word vector matrix Each word vector in The attention weight ; S32, the attention weights of all word vectors are calculated based on the word vector matrix The word order is spliced ​​to obtain the word vector matrix The corresponding attention weight , k =1,2,3,…, n , k Represents the index of the word vector element; S33, the attention weight and word vector matrix Perform corresponding multiplication to obtain the updated word vector matrix ; S34, normalize the updated word vector matrix and extract its features through a set of convolutional layers to generate the corresponding feature map ; S35. Feature map Perform maximum pooling to obtain a simplified feature representation; S36. The simplified feature representation is subjected to low-dimensional feature mapping through a fully connected layer to obtain text features corresponding to the power grid description / power grid line.

3. The method for generating a power grid design scheme based on deep learning according to claim 2, characterized in that: In step S31, the attention weight Expressed as: Where, Indicates the first k The result after linear transformation of elements is express activation function, represents the weight parameter matrix, Represents the bias term when the first element in the input sequence is linearly transformed, Indicates the first j The result after linear transformation of elements is j Represents the element index, n Indicates the sequence length; In step S34, the method for extracting features through the convolutional layer is: Make the convolution kernel of each convolution layer After updating the word vector matrix Slide up, perform weighted summation and add bias terms, and then generate feature maps through activation functions , which is expressed as: Where, represents the activation function, Indicates the i The bias term corresponding to the feature map, Represents the updated word vector matrix The j elements, subscripts i represents the index of the feature map, m Indicates the number of feature maps; In step S36, the grid description / grid line text features Expressed as: Where, represents the activation function, represents the weight, represents the bias term corresponding to the network neuron f, Represents the feature representation of the pooling layer output.

4. The method for generating a power grid design scheme based on deep learning according to claim 1, characterized in that: In step S4, the output of the bidirectional LSTM network at the last time step T is converted to As a comprehensive fusion representation of the power grid description text feature vector and the power grid line text feature vector .

5. The method for generating a power grid design scheme based on deep learning according to claim 1, characterized in that: Learning the corresponding prediction vector representation based on the matching function The process is expressed as: Where P and Q represent the weight matrices of the linear embedding layer, which are used to transform and The original features are mapped to the vector representation in the latent space, Characteristic of the lines representing the grid scheme, Characteristics that represent the details of the grid scheme, Representation of latent vector and latent vector representation The concatenated vector, represents the output vector after the first hidden layer, represents the final output vector of the multilayer perceptron, represents the output vector of the second-to-last layer of the multi-layer perceptron, 、 Represent the weight matrices of the first hidden layer and the last hidden layer in the multilayer perceptron, Represents the bias term corresponding to the first layer in the multi-layer perceptron, Represents the bias term corresponding to the y-th layer in the multilayer perceptron, express Activation function.

6. The method for generating a power grid design scheme based on deep learning according to claim 1, characterized in that: In step S6, the prediction vector representation is fused as follows: Where, represents the generated power grid design scheme, represents the activation function, Represents the weight matrix of the output layer.

Images