Power communication network physical path automatic planning system and method

Abstract

The invention belongs to the technical field of power communication network systems, and aims to solve the problems that the current power communication network physical path automatic planning has a serious blank and the existing road network automatic planning technology cannot be directly applied to the physical path planning of the power communication network. According to the electric power communication network physical path automatic planning system provided by the invention, visual features are extracted from terrain and geographic images through a convolutional neural network, spatial topological features are extracted from a road network graph through a graph convolutional neural network, and the visual features and the spatial topological features are effectively fused and enhanced through a convolutional attention mechanism; and generating an electric power communication network physical path, and continuously updating the model parameters by using the calculated image loss in the electric power communication network physical path until the optimal electric power communication network physical path is obtained. The physical path generated by the method has good planning precision and availability.

Description

Technical Field

[0001] The present invention belongs to the technical field of power communication systems, and particularly relates to a power communication network physical path automatic planning system and method. Background Art

[0002] Power communication networks are the communication backbone of the power grid, providing important guarantees for the stability, efficiency, and reliability of modern power systems. These networks support a variety of key functions, including energy management, equipment fault alarms, and video surveillance. Reasonable planning of physical paths can effectively manage and maintain power communication networks, ensuring their rapid recovery in case of emergencies. For example, when a fire, landslide, or flood damages an underground tunnel or a utility pole carrying fiber optic communication cables, it is necessary to re-plan and deploy new physical lines to repair communication interruptions and restore network services.

[0003] The planning of the physical path of a power communication network needs to comprehensively consider various factors, such as terrain changes and existing urban infrastructure (including buildings, roads, and bridges). Currently, this work mainly relies on manual operations, which require the professional knowledge and collaboration of power engineers, construction teams, and urban management experts. The entire process is not only labor-intensive and time-consuming but also difficult to meet the increasingly complex planning requirements of communication networks. Therefore, there is an urgent need for an automated and efficient physical path planning method to improve the management and maintenance efficiency of power communication networks.

[0004] However, the automated planning of power communication networks is still blank in technology. Research has found that there are automated generation technologies in the field of street networks, which are closely related to the application of power communication networks. Some studies have developed deep learning-based models. For example, StreetGAN generates new street network layouts by learning the characteristics of existing street networks; another model, DeepStreet, predicts the future expansion pattern of street networks in a specific area based on the surrounding street networks. In addition, these technologies have been extended to combine various geographical information to generate more realistic prediction results.

[0005] Although these technologies have achieved remarkable results in the field of street network generation, they usually regard street networks as image data, rely on extracting visual features from images, and regard the approximate network generation task as an Image Inpainting task. However, in practical applications, the physical path planning of power communication networks is highly correlated with the spatial layout and topological structure of street networks. Such basic network information is usually encoded in the form of graph data, and existing methods cannot directly process graph data. Therefore, these successful street network generation technologies cannot be directly applied to the physical path planning of power communication networks. Summary of the Invention

[0006] The object of the present invention is to overcome the deficiencies of the prior art and provide an automatic planning system and method for the physical path of a power communication network.

[0007] The present invention establishes a hybrid model (AutoEPCNRoute), which combines a convolutional neural network (CNN) and a graph convolutional neural network (GCN) to predict and plan the physical path of a power communication network.

[0008] The first object of the present invention is to provide an automatic planning system for the physical path of a power communication network, including: A convolutional neural network branch, which is used to extract multi-scale visual features from the input images of terrain and geographic information, and provide environmental perception information for the physical path planning of the power communication network; A graph convolutional neural network branch, which is used to extract spatial topological features from the input road network graph closely related to the power communication network, and capture the spatial correlation between the road network and the power communication network; A feature fusion and enhancement module, which is used to fuse and enhance the extracted visual features and spatial topological features, and generate the physical path of the power communication network in the area to be planned; A discriminator, which is used to evaluate the generated physical path of the power communication network, generate an evaluation result, and dynamically adjust and update the parameters of the convolutional neural network branch and the parameters of the graph convolutional neural network branch according to the image loss value obtained by the loss calculation unit until the image loss value converges, so as to obtain the optimal physical path of the power communication network; Among them, the loss calculation unit is used to compare and calculate the generated evaluation result with the real path data to obtain the image loss value.

[0009] Preferably, the convolutional neural network branch includes: An auxiliary information pre-extraction module, which is used to convert the input images of terrain and geographic information into feature maps with resolutions of 256x256, 64x64, and 16x16 through pooling layers respectively, and gradually add the features on the generated feature maps with resolutions of 64x64 and 16x16 to the generated feature map with a resolution of 256x256, and fuse all the features to obtain a multi-scale feature map rich in multi-scale information; A geographic merging module, which is used to multiply the obtained multi-scale feature map by the physical path image of the power communication network to obtain the power communication network feature map ; A pattern learning module, which is used to encode the input images of terrain and geographic information and the physical path image of the power communication network with a mask to obtain high-order features, and combine the high-order features with the power communication network feature map to obtain a rough feature map , and then for the rough feature map Decode to extract multi-scale visual features.

[0010] Preferably, the pattern learning module includes: An encoding and decoding module based on gated convolution, which encodes the input images of terrain and geographic information and the physical path image of the power communication network with a mask to obtain high-order features, and decodes the rough feature map to extract multi-scale visual features; The dilated gated convolution module is used to combine the obtained high-order features with the power communication network feature map to obtain a rough feature map .

[0011] Preferably, the graph convolutional neural network branch includes: A graph construction module, which constructs a local graph of the road network within the image sample area according to the input road network graph closely related to the power communication network for each image sample; The ROI construction and initialization module is used to expand each road node in the constructed local graph into a region of interest pixels and initialize the features of the pixel region to obtain an initialized local graph; The spatial feature extraction and update module is used to update and extract the features in the obtained initialized local graph to obtain spatial topological features.

[0012] Preferably, the spatial feature extraction and update module includes: The graph convolutional network module is used to update and extract the features in the obtained initialized local graph according to the graph convolutional neural network to obtain updated node information; The feature mapping module is used to map the updated node information back to the rough feature map according to the position coordinates to obtain spatial topological features.

[0013] Preferably, the feature fusion and enhancement module includes: The Unet-style learning module is used to learn the visual features through the encoder to obtain feature maps at different scales, and in the decoding stage, the decoder uses skip connections to splice the feature maps of the corresponding sizes to enhance the feature information and obtain the visual feature mapping map; The convolutional attention module is used to fuse and enhance the obtained visual feature mapping map, the power communication network feature map and the spatial topological features to generate the physical path of the power communication network in the area to be planned.

[0014] Preferably, the convolutional attention module includes: A channel attention module, which is used to adaptively allocate weights for different channels, and enhance relevant features and suppress irrelevant or redundant features according to the importance of each channel in the visual feature map, the power communication network feature map and the spatial topology feature; A spatial attention module, which is used to adaptively capture significant regions of the visual feature map, the power communication network feature map and the spatial topology feature in the spatial dimension. By calculating the weight distribution in the spatial dimension, it enhances the perception ability of key positions, optimizes feature expression, fuses each feature, and generates a physical path of the power communication network in the area to be planned.

[0015] The second object of the present invention is to provide a method for automatically planning a physical path of a power communication network, including the following steps: S1. Extract multi-scale visual features from the input images of terrain and geographical information to provide environmental perception information for the physical path planning of the power communication network; S2. Extract spatial topology features from the input road network diagram closely related to the power communication network to capture the spatial correlation between the road network and the power communication network; S3. Fuse and enhance the extracted visual features and spatial topology features to generate a physical path of the power communication network in the area to be planned; S4. Evaluate the generated physical path of the power communication network to generate an evaluation result; S5. Compare and calculate the generated evaluation result with the real path data to obtain an image loss value; S6. According to the obtained image loss value, dynamically adjust and update the parameters of the convolutional neural network branch and the parameters of the graph convolutional neural network branch until the image loss value converges, and obtain the optimal physical path of the power communication network.

[0016] Preferably, step S1 includes the following steps: S11. Convert the input images of terrain and geographical information into feature maps with resolutions of 256x256, 64x64, and 16x16 through a pooling layer respectively, and gradually add the features on the generated feature maps with resolutions of 64x64 and 16x16 to the generated feature map with a resolution of 256x256 for fusion to obtain a multi-scale feature map rich in multi-scale information; S12. Multiply the obtained multi-scale feature map by the physical path image of the power communication network to obtain a power communication network feature map ; S13. Encode the input images of terrain and geographical information and the physical path image of the power communication network with a mask to obtain high-order features, and combine the high-order features with the power communication network feature map Combine to obtain a rough feature map , and then decode the rough feature map to extract multi-scale visual features.

[0017] Preferably, step S2 includes the following steps: S21. For each image sample, construct a local map of the road network within the image sample area according to the input road network diagram closely related to the power communication network; S22. Expand each road node in the constructed local map into a pixel area of interest, and initialize the features of the pixel area to obtain an initialized local map; S23. Update and extract the features in the obtained initialized local map to obtain spatial topological features.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the combined design of the convolutional neural network branch and the graph convolutional neural network branch, the present invention can not only extract visual features, but also synchronously extract spatial topological features. That is, through the design of the above two branch structures, the present invention can simultaneously process heterogeneous data of two modalities, namely images and graphs, so that the extracted data can be highly relevant to the power communication network. Moreover, the present invention fuses and enhances the extracted visual features and spatial topological features through the feature fusion and enhancement module, and continuously updates the network parameters by using the calculated image loss value, so that the finally generated physical path of the power communication network can highly coincide with the real path, and the simulation test data verifies the correctness of the path planning method of the present invention.

[0019] Therefore, the method provided by the present invention can significantly improve the planning accuracy and practicability of the physical path. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the system architecture diagram of the automatic physical path planning system for the power communication network proposed in Embodiment 1 of the present invention; Figure 2 is the working flowchart of the function fusion and enhancement unit in Embodiment 1 of the present invention; Figure 3 is the demonstration diagram of the data set constructed by the present invention; Figure 4 is the qualitative comparison diagram of the example results generated by using different models with the data set; Figure 4 In (a)- Figure 4 in (j) are respectively samples randomly selected from the data set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will be combined with the embodiments of the present invention Figures 1 to 4, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0022] Embodiment 1: As Figure 1 shown, the embodiment of the present invention provides an automatic physical path planning system for a power communication network (AutoEPCNRoute model), which is planned based on two types of input data. The first is image data, including urban auxiliary information of seven channels: urban elevation, slope, aspect, mountain shadows of four azimuths, and urban land use data. At the same time, it also includes a power communication network image sample with a mask and the expected guiding node image in the sample. The second is graphic data, that is, the topological data obtained by encoding the nodes and edges of the urban road network. The AutoEPCNRoute model consists of multiple modules. First, the convolutional neural network branch is used to extract visual features from the image data, and the graph convolutional neural network branch is used to extract spatial topological features from the graphic data. Then, the extracted features are integrated through the feature fusion and enhancement module, and the planned power communication network result is generated within the mask area. Finally, the discriminator and loss calculation unit evaluate the difference between the generated result and the real result, and update the parameters of the convolutional neural network branch and the graph convolutional neural network branch accordingly, continuously optimizing the planning performance until the optimal physical path of the power communication network is obtained.

[0023] The automatic physical path planning system for a power communication network provided by the embodiment of the present invention specifically includes: A convolutional neural network (CNN) branch unit, which is used to extract multi-scale visual features from the input images of terrain and geographic information, and at the same time learn the pattern of the power communication network to provide environmental perception information for the physical path planning of the power communication network; In the embodiment of the present invention, the convolutional neural network branch includes: An auxiliary information pre-extraction module, which is used to convert the input images of terrain and geographic information into feature maps with resolutions of 256x256, 64x64, and 16x16 respectively through the pooling layer, and gradually add the features on the generated feature maps with resolutions of 64x64 and 16x16 to the generated feature map with a resolution of 256x256, and fuse all the features to obtain a multi-scale feature map rich in multi-scale information; images of terrain and geographic information It consists of image layers of elevation, slope, aspect, hillshade, lakes, buildings, and other urban infrastructure. The auxiliary information pre-extraction module in the embodiments of the present invention is designed based on HrNet, and uses parallel multi-resolution subnets and a multi-scale fusion mechanism for subnet connection and information exchange. Specifically, this module converts the input channels into feature maps with sizes of 256x256, 64x64, and 16x16 through pooling layers. Then, it gradually incrementally adds and fuses the features from the lower-resolution subnets to generate a final multi-scale feature map rich in multi-scale information. This structure effectively solves the problem of data dispersion caused by sampling errors and can directly guide the result generation. It retains the spatial representation of terrain and regional information while learning additional focus areas from the images of terrain and geographical information through feature fusion.

[0024] A geographical merging module, which is used to explore the correlation between the physical path of the power communication network and the auxiliary information. Specifically, it is used to multiply the obtained multi-scale feature map with the image points of the physical path of the power communication network that does not need to be repaired and has a complete physical path, to obtain a power communication network feature map that can guide the generation of the power communication network. ; A pattern learning module (power communication network pattern learning module), which is used to encode the input images of terrain and geographical information and the physical path image of the power communication network with a mask (i.e., the area where the power communication is actually interrupted and the physical path of the power communication network needs to be planned, that is, the area to be planned), to obtain high-order features, and combine the high-order features with the power communication network feature map to obtain a rough feature map , and then decode the rough feature map to extract multi-scale visual features. This module mainly uses a neural network with an Encoder-Decoder structure to capture the features of the physical path image of the power communication network and generate preliminary results. The input channels of the pattern learning module in this embodiment consist of the network context I mask processed by the mask channel M context and the key guiding points P guide within the mask area.

[0025] In the embodiments of the present invention, the pattern learning module includes: An encoding and decoding module based on gated convolution, which is used to encode the input images of terrain and geographical information and the power communication network image with a mask to obtain high-order features, and decode the rough feature map to extract multi-scale visual features; this module uses gated convolutional layers to understand the effective distribution within the channels, which is crucial for leveraging the context details in rugged or sparsely populated areas.

[0026] An inflated gated convolutional module for combining the obtained high - order features with the power communication network feature map to further update and obtain rough features .

[0027] A graph convolutional neural network branch for extracting spatial topological features from the input road network graph closely related to the power communication network, and capturing the spatial correlation between the road network and the power communication network; the goal of this graph convolutional neural network branch is to use the rich spatial topological structure of the road network to enhance the feature representation of the power communication network.

[0028] In the embodiment of the present invention, the graph convolutional neural network branch includes: A graph construction module for constructing a local graph of the road network within the image sample area for each image sample according to the input road network graph closely related to the power communication network; the input is the global graph G of the urban road network encoded by the node set V global and the edge set E global . Meanwhile, guided by the image channel R with the same resolution as the power communication network in the area to be studied, for each image sample, find its corresponding longitude and latitude range in R road . Find the existing road nodes within this range, add them to the node set V index of the local graph. At the same time, find all the roads of each node in the area and add them to the edge set E index to construct the local graph of the road network within the image sample area. For more convenient subsequent processing, the present invention converts the data into the data format of the graph defined in the geometri library in python, that is, defines the graph using the node and edge sets, and uses the Euclidean distance between the two end points of each edge as the weight of the edge. global global global global

[0029] An ROI construction and initialization module for expanding each road node in the constructed local graph into a region of interest pixel area (ROI) and initializing the features of this pixel area to obtain the initialized local graph; for the initialization of the local graph, the present invention uses the rough features obtained by the above - mentioned pattern learning module for assignment. Specifically, through the ROI align algorithm, extract a feature block with a pixel size from the corresponding rough features at the position of each road node in the local graph as the initial feature of the node. This algorithm can ensure the precise spatial alignment between the road node and the feature map, retain the correct spatial relationship, and this precise alignment is crucial for maintaining the integrity of the topological structure in the graph.

[0030] A spatial feature extraction and update module for updating and extracting features in the obtained initialized local graph to obtain spatial topological features.

[0031] In the embodiment of the present invention, the spatial feature extraction and update module includes: A graph convolutional neural network module for updating and extracting features in the obtained initialized local graph according to the graph convolutional neural network to obtain updated node information. Given that the road network exhibits a more complex and diverse structure than the power communication network, and the spatial influences of different road nodes are different, the present invention uses a graph attention network to dynamically identify the influence of adjacent nodes on the current node and optimize the feature aggregation process. This mechanism can improve key road structure information during node feature update while suppressing noise and irrelevant information. Specifically, GAT calculates the attention coefficient to measure the influence between nodes v and neighbors u . The calculation formula is as follows: ; where W is a learnable weight matrix, is a weight vector with the same dimension as the node feature, || represents concatenation, Leaky Re Lu is an activation function, is the relevant feature of node v , is the relevant feature of node u .

[0032] Next, use the attention coefficient to calculate the weighted average of the adjacent node features to update the feature of node v . The calculation formula is as follows: ; where is the attention coefficient, N ( v ) represents the set of neighbor nodes of v , soft max normalizes the attention coefficient to attention weights, W is a learnable weight matrix, is the relevant feature of node u , is the new feature of node v updated according to the neighbor node features.

[0033] A feature mapping module for mapping the updated node information back to the rough feature map according to the node positions to obtain spatial topological features.

[0034] A feature fusion and enhancement module for fusing and enhancing the extracted visual features and spatial topological features to generate a physical path of the power communication network in the area to be planned; the architecture diagram of the feature fusion and enhancement module in the embodiment of the present invention is as shown in Figure 2 shown, specifically including: A Unet-style learning module for learning visual features through an encoder-decoder to obtain feature maps at different scales, and in the decoding stage, using skip connections by the decoder to splice feature maps of corresponding sizes to enhance feature information and obtain a visual feature map; and obtaining the final physical path of the power communication network based on this visual feature map; since the extracted visual features and spatial topological features contain both high-dimensional and low-dimensional information, directly processing on the original image will lead to suboptimal results. Therefore, this encoder-decoder structure enhances the feature reconstruction ability through skip connections, not only retaining more original information and high-resolution details, but also significantly improving the quality of feature prediction and generation, ensuring the adaptability and accuracy of the model in complex scenarios.

[0035] A convolutional attention module for fusing features from different sources, specifically for fusing and enhancing the visual feature map, the power communication network feature map and the spatial topological features to generate a physical path of the power communication network in the area to be planned. Since the sources of these features and their focuses of attention are different, directly connecting them may lead to semantic and spatial misalignment, making it difficult for the model to learn effective feature expressions. To solve this problem, the present invention adopts a convolutional attention module (CBAM module). The CBAM module dynamically highlights relevant features while using its channel and spatial attention mechanisms to suppress less important features. This method can alleviate the misalignment and redundancy problems, ensure effective feature fusion and reduce noise.

[0036] In the embodiment of the present invention, the convolutional attention module includes: A channel attention module for adaptively assigning weights to different channels, enhancing relevant features and suppressing irrelevant or redundant features according to the importance of each channel in the visual feature map, the power communication network feature map and the spatial topological features, thereby enhancing the feature expression ability between channels; A spatial attention module for adaptively capturing significant regions in the spatial dimension of the visual feature map, the power communication network feature map and the spatial topological features, enhancing the perception ability of key positions by calculating the weight distribution in the spatial dimension, optimizing feature expressions, and fusing each feature to generate a physical path of the power communication network in the area to be planned.

[0037] In the embodiment of the present invention, the feature map output by the convolutional attention module is calculated as follows: ; Among them, M c represents the channel attention weight, M s represents the spatial attention weight, and ⊙ represents element-wise multiplication. F concat is the aggregated feature obtained by directly splicing feature maps from different sources. F output is the feature map obtained after being enhanced by the channel attention module.

[0038] By using the CBAM module, features can be effectively refined, noise can be reduced, and feature representation can be improved.

[0039] The discriminator is used to evaluate the generated physical path of the power communication network, generate an evaluation result, which conforms to the general process of the generative adversarial network, and dynamically adjusts and updates the parameters of the convolutional neural network branch and the graph convolutional neural network branch according to the image loss value calculated by the loss calculation unit until the image loss value converges, so as to obtain the optimal physical path of the power communication network; continuously improve the accuracy and rationality of path generation.

[0040] The loss calculation unit is used to compare and calculate the generated evaluation result with the real path data (i.e., data that does not need to be repaired and has a complete physical path) to obtain the image loss value.

[0041] The overall loss function used in the loss calculation unit of the present invention is a weighted combination of three partial losses, specifically including reconstruction loss, adversarial loss, and structural similarity index loss. The overall loss function is expressed as: ; Among them, , and are hyperparameters that control the importance of each component. Through multiple experiments and adjustments, these parameters are set to , and , is the overall loss, L recon is the reconstruction loss; L adv is the adversarial loss, L ssim is the structural similarity index loss; The reconstruction loss ( L recon ) is mainly used to measure the generated image ( I result ) and the ground truth image ( Itruth ) The difference between them is calculated using the mean absolute error (MAE), and the specific calculation formula is as follows: ; where N is the total number of pixels in the image, I result,i and I truth,i are the pixel values of the generated result image and the actual real image respectively.

[0042] The adversarial loss ( L adv ) evaluates the authenticity of the generated image through a generative adversarial network (GAN). Its calculation formula is: ; where I truth represents the real image, I gen represents the generated image, D (·) is the discriminator.

[0043] The structural similarity index loss ( L ssim ) is used to evaluate the structural similarity between the generated image ( I result ) and the actual real image ( I truth ). This loss function helps to ensure the structural integrity of the generated image. The definition of the structural similarity index loss is: ; where I result and I truth are the generated image and the real image respectively, SSIM is the calculation function of the structural similarity index.

[0044] Example 2: The embodiment of the present invention provides a method for automatically planning the physical path of a power communication network, including the following steps: S1. Extract multi-scale visual features from the input images of terrain and geographical information to provide environmental perception information for the physical path planning of the power communication network. Specifically, it includes the following steps: S11. Convert the input images of terrain and geographical information into feature maps with resolutions of 256x256, 64x64, and 16x16 through a pooling layer respectively. At the same time, gradually add the features on the generated feature maps with resolutions of 64x64 and 16x16 to the generated feature map with a resolution of 256x256, and fuse all the features to obtain a multi-scale feature map rich in multi-scale information; S12. Multiply the obtained multi-scale feature map with the image points of the physical path of the power communication network to obtain a power communication network feature map ; S13. Encode the input images of terrain and geographical information and the physical path image of the power communication network with a mask to obtain high-order features, and combine the high-order features with the power communication network feature map to obtain a rough feature map , and then decode the rough feature map to extract multi-scale visual features; S2. Extract spatial topological features from the input road network diagram closely related to the power communication network, and capture the spatial correlation between the road network and the power communication network, which specifically includes the following steps: S21. For each image sample, construct a local graph of the road network within the image sample area according to the input road network diagram closely related to the power communication network; S22. Expand each road node in the constructed local graph into a pixel area of interest, and initialize the features of the pixel area to obtain an initialized local graph; S23. Update and extract the features in the obtained initialized local graph to obtain spatial topological features.

[0045] S3. Fuse and enhance the extracted visual features and spatial topological features to generate the physical path of the power communication network in the area to be planned; S4. Evaluate the generated physical path of the power communication network to generate an evaluation result; S5. Compare and calculate the generated evaluation result with the real path data to obtain an image loss value; S6. According to the calculated image loss value, continuously update the parameters of the convolutional neural network branch and the graph convolutional neural network branch, and dynamically adjust and update the parameters of the convolutional neural network branch and the graph convolutional neural network branch until the image loss value converges to obtain the optimal physical path of the power communication network.

[0046] Next, study the performance of the power communication network physical path automatic planning system (AutoEPCNRoute) provided in Embodiment 1 of the present invention 1. Experimental settings: (1)Dataset construction: The experimental dataset of the present invention was created in cooperation with the State Grid Corporation of China (SGCC) to meet the specific requirements of the Xi'an area in Shaanxi Province, China, and the self-made dataset was used to evaluate the proposed AutoEPCNRoute method of the present invention. The structure of the dataset is as Figure 3 shown.

[0047] From Figure 3 it can be seen that the dataset mainly includes the following contents: Topographic information: It includes a digital elevation model (DEM) image with a resolution of 12.5 meters, a slope image, a point, and four shadow images at different angles.

[0048] Geographic information: It covers an image of the urban land distribution composed of building facilities, lakes, and other urban infrastructure.

[0049] Power communication network information: It includes an image of the physical path of the urban power communication network and a guiding point image marking the intersection points.

[0050] These information are comprehensively generated into a high-dimensional multi-channel image, which together with the graphic data of the road network structure in the corresponding area constitutes the dataset. All images and graphics have been processed by geospatial alignment. During the sampling process, 594 samples were randomly cropped from the high-dimensional multi-channel image, and the size of each sample is 256×256 pixels. Through the graph construction method, a graph representing the road network was created for each sample. At the same time, data augmentation techniques such as rotation and flipping were used to expand the samples to 2376 for model training and validation to fully evaluate the performance of the model.

[0051] (2)Experimental condition setting: The present invention randomly divides the above-mentioned 2376 samples into a training set and a validation set, among which 1900 samples are used for training and 476 samples are used for validation. During the training process, the Adam optimizer is used for parameter optimization. The training is carried out for a total of 50 epochs, and the initial learning rate is , and the weight decay is also set to , and the batch size is 12. After the 10th epoch, the learning rate decays by a factor of 10. The entire training process is completed on a GTX 4090 GPU with 24GB video memory, ensuring high computing performance and model training efficiency.

[0052] Currently, there is no existing model developed for the research task of the present invention. Therefore, the AutoEPCNRoute model proposed in Embodiment 1 of the present invention is compared with some of the most advanced street network generation and image inpainting solutions, which are closely related to the research of the present invention, including DeepStreet, Topo-aware, DeepFill, and AOT-GAN. At the same time, since the CNN branch of the Topo-aware model is the same as that of the AutoEPCNRoute model of the present invention, the Topo-aware model is regarded as the baseline model (Base) of the present invention, and three ablation experiments are carried out. Specifically, the effectiveness of the method provided by the present invention is evaluated by adding or removing the GCN branch, the Unet-style learning module, and the CBAM module. The three models corresponding to the three ablation experiments are respectively: Base + GCN branch (Base+GCN-Branch): The GCN branch is added to the baseline model, the features in the graph structure are fused by simple concatenation, and processed through a basic encoder-decoder architecture; Base + GCN branch + Unet-style learning module (Base+GCN-Branch+Unet-style): This variant modifies the encoder-decoder architecture into a Unet-style structure, and still maintains simple concatenation for feature fusion; Base + GCN branch + CBAM module (Base+ GCN-Branch+CBAM): In this variant, the basic encoder-decoder architecture is retained, but the CBAM module is added for feature fusion.

[0053] The present invention evaluates the performance of the generated physical path of the power communication network from multiple perspectives. First, the present invention uses the mean absolute error (MAE) and the mean square error (MSE) to measure the pixel-level differences. At the same time, the structural similarity index (SSIM) is used to evaluate the image quality. In addition, the present invention also uses the Dice coefficient (DICE) and the Jaccard index (JAC) to measure the overlap degree between the generated physical path and the real path, so as to evaluate the generation accuracy.

[0054] 2. Experimental results: The evaluation results of different models are shown in Table 1 below. From Table 1, it can be seen that the AutoEPCNRoute model proposed by the present invention has remarkable performance. Compared with the existing street network generation models, the AutoEPCNRoute model not only achieves the best results in terms of MAE, SSIM, DICE, and JAC, but also demonstrates a significant performance improvement. In particular, compared with the second-best performing model, the AutoEPCNRoute model has improved by 5.4% in MAE, 4.5% in SSIM, 25.7% in DICE, and 38.9% in JAC. Although the performance of the AutoEPCNRoute model in MSE is slightly inferior to the best score, the focus of the present invention is to generate coherent overall results rather than pixel-level reconstruction. Therefore, this result does not weaken the advantages of the method proposed by the present invention. On the other hand, the ablation experiment results show that the complete AutoEPCNRoute model is optimized in generating the physical paths of the power communication network and achieves the best performance.

[0055] Table 1 Evaluation results of different models: ; Figure 4 Figure is the experimental result diagram of the dataset applied to the AutoEPCNRoute model proposed by the present invention. The first column shows the real physical path, and the second column shows the masked area where the physical path needs to be planned to connect the breakpoints. The following columns respectively show the results generated by different models. From (a) to (j) in Figure 4 it can be clearly seen that the physical path generated by the AutoEPCNRoute model provided by the present invention highly coincides with the real path, effectively captures the complexity of the structure, and can generate high-quality physical paths of the power communication network. Figure 4

[0056] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A system for automatically planning physical paths in a power communication network, characterized in that: include: The convolutional neural network branch is used to extract multi-scale visual features from the input images of terrain and geographic information to provide environmental perception information for physical path planning of the power communication network; The graph convolutional neural network branch is used to extract spatial topological features from the input road network graph that is closely related to the power communication network, capturing the spatial correlation between the road network and the power communication network; The feature fusion and enhancement module is used to fuse and enhance the extracted visual features and spatial topological features to generate the physical path of the power communication network in the area to be planned; A discriminator is used to evaluate the generated physical path of the power communication network, generate an evaluation result, and dynamically adjust and update the parameters of the convolutional neural network branch and the parameters of the graph convolutional neural network branch according to the image loss value obtained by the loss calculation unit until the image loss value converges to obtain the optimal physical path of the power communication network; Among them, the loss calculation unit is used to compare and calculate the generated evaluation result with the real path data to obtain the image loss value.

2. According to claim 1, a power communication network physical path automatic planning system is characterized in that: Convolutional neural network branches include: Auxiliary information pre-extraction module, used to convert the input terrain and geographic information images into feature maps with resolutions of 256x256, 64x64 and 16x16 respectively through the pooling layer, and gradually add the features of the feature maps with resolutions of 64x64 and 16x16 to the generated feature map with resolution of 256x256, and fuse all the features to obtain a multi-scale feature map rich in multi-scale information; The geographic merging module is used to multiply the obtained multi-scale feature map with the physical path image of the power communication network to obtain the power communication network feature map ; The pattern learning module is used to encode the input images of terrain and geographic information and the physical path images of the power communication network with masks to obtain high-order features and combine the high-order features with the power communication network feature map. Combined to get a rough feature map , and then the coarse feature map Decode and extract multi-scale visual features.

3. The automatic physical path planning system for a power communication network according to claim 2 is characterized in that: The pattern learning modules include: The encoding and decoding module based on gated convolution is used to encode the input images of terrain and geographic information and the physical path images of the power communication network with masks to obtain high-order features and to generate coarse feature maps. Decode and extract multi-scale visual features; Dilated gated convolution module to combine the obtained high-order features with the power communication network feature map Combined to get a rough feature map .

4. The automatic physical path planning system for a power communication network according to claim 1 is characterized in that: The graph convolutional neural network branches include: A graph construction module is used to construct a local graph of the road network in the image sample area according to an input road network graph closely related to the power communication network for each image sample; The ROI construction and initialization module is used to expand each road node in the constructed local map into a pixel region of interest and initialize the features of the pixel region to obtain an initialized local map; The spatial feature extraction and updating module is used to update and extract the features in the initialized local graph to obtain spatial topological features.

5. The automatic physical path planning system for a power communication network according to claim 4 is characterized in that: The spatial feature extraction and updating module includes: A graph convolutional network module is used to update and extract features in the initialized local graph according to the graph convolutional neural network to obtain updated node information; Feature mapping module, used to map the updated node information back to the rough feature map according to the position coordinates In the above example, we can obtain the spatial topological characteristics.

6. The automatic physical path planning system for a power communication network according to claim 2, characterized in that: The feature fusion and enhancement module includes: The Unet-style learning module is used to learn visual features through the encoder to obtain feature maps at different scales, and in the decoding stage, the decoder uses jump connections to splice feature maps of corresponding sizes to enhance feature information and obtain visual feature maps; Convolutional attention module, used to obtain visual feature maps and power communication network feature maps The spatial topological features are integrated and enhanced to generate the physical path of the power communication network in the area to be planned.

7. The automatic physical path planning system for a power communication network according to claim 6, characterized in that: The convolutional attention module consists of: Channel attention module, used to adaptively assign weights to different channels, based on visual feature maps, power communication network feature maps and the importance of each channel in the spatial topological features, enhancing relevant features and suppressing irrelevant or redundant features; Spatial attention module to adaptively capture visual feature maps, power communication network feature maps By calculating the weight distribution in the spatial dimension, the perception of key locations is enhanced, the feature expression is optimized, and the features are integrated to generate the physical path of the power communication network in the area to be planned.

8. A method for automatically planning a physical path in a power communication network, characterized in that: The following steps are involved: S1. Extract multi-scale visual features from the input images of terrain and geographic information to provide environmental perception information for physical path planning of power communication networks; S2, extracting spatial topological features from the input road network graph that is closely related to the power communication network, capturing the spatial correlation between the road network and the power communication network; S3, fusing and enhancing the extracted visual features and spatial topological features to generate a physical path of the power communication network in the area to be planned; S4. Evaluate the generated physical path of the power communication network and generate an evaluation result; S5, comparing and calculating the generated evaluation result with the real path data to obtain an image loss value; S6. According to the obtained image loss value, dynamically adjust and update the parameters of the convolutional neural network branch and the parameters of the graph convolutional neural network branch until the image loss value converges to obtain the optimal physical path of the power communication network.

9. A method for automatic planning of physical paths in a power communication network according to claim 8, characterized in that: Step S1 comprises the following steps: S11, converting the input terrain and geographic information images into feature maps with resolutions of 256x256, 64x64 and 16x16 respectively through a pooling layer, and gradually adding the features of the feature maps with resolutions of 64x64 and 16x16 to the generated feature map with a resolution of 256x256 to fuse them, so as to obtain a multi-scale feature map rich in multi-scale information; S12, used to multiply the obtained multi-scale feature map with the physical path image of the power communication network to obtain the power communication network feature map ; S13, encoding the input terrain and geographic information image and the physical path image of the power communication network with the mask to obtain high-order features, and combining the high-order features with the power communication network feature map Combined to get a rough feature map , and then the coarse feature map Decode and extract multi-scale visual features.

10. The method for automatic planning of physical paths in a power communication network according to claim 8, characterized in that: Step S2 comprises the following steps: S21, for each image sample, constructing a local map of the road network in the image sample area according to the input road network map closely related to the power communication network; S22, expanding each road node in the constructed local graph into a pixel region of interest, and initializing the features of the pixel region to obtain an initialized local graph; S23, updating and extracting the features in the initialized local graph to obtain spatial topological features.

Images