Power grid equipment flood flooding prediction method and equipment based on deep learning

Abstract

The invention provides a power grid equipment flood inundation prediction method and equipment based on deep learning, relates to the technical field of power system safety monitoring, and solves the problems of effectively predicting a power grid equipment flood inundation range and reducing disaster loss, and the adopted scheme is as follows: S1: making a flood inundation model sample library, S2: constructing and training a model, and S3: constructing a flood inundation model sample library. According to the technical scheme, the accuracy is high, the real-time performance is high, the adaptability is good, the adaptability of the model to different regions, different terrains and different flood scenes is enhanced, the flood inundation range can be extracted, the affected power grid equipment can be accurately determined in combination with the power grid equipment information, and the method is suitable for popularization and application. And a comprehensive decision basis is provided for emergency response of an electric power department.

Description

Technical Field

[0001] The present invention relates to the technical field of power system safety monitoring, and in particular to a method and device for predicting flood inundation of power grid equipment based on deep learning. Background Art

[0002] This section aims to provide background or context for the embodiments of the present invention recited in the claims. The descriptions herein may include concepts that can be explored, but not necessarily concepts that have been previously thought of or explored. Therefore, unless otherwise indicated herein, the content described in this section is not prior art to the specification and claims of this application, and is not admitted to be prior art merely by virtue of being included in this section.

[0003] Against the backdrop of global climate change, the frequency and intensity of extreme natural disasters such as floods have shown a significant upward trend. As a key infrastructure for modern society, the safe and stable operation of the power grid is directly related to the normal operation of society. Once a flood hits, power grid equipment is extremely vulnerable to severe damage. For example, transmission towers may be knocked down by fast-flowing water, and substations may be mercilessly flooded. These situations will not only cause large-scale power outages, bringing great inconvenience to residents' lives and seriously affecting industrial production, but may also trigger a series of secondary disasters, such as fires and electric shock accidents, posing a huge threat to social economy and people's lives and property safety. Therefore, quickly and accurately determining the flood inundation range of power grid equipment is of crucial significance for the power department to take effective emergency measures in a timely manner, such as quickly organizing repairs and reasonably transferring important equipment, so as to minimize disaster losses.

[0004] Traditional methods for extracting flood inundation ranges mainly include manual visual interpretation, rule-based algorithms, and simple machine learning models. Although manual visual interpretation performs relatively well in terms of accuracy, its efficiency is extremely low, and it is greatly affected by human subjective factors, making it difficult to meet the urgent need for real-time monitoring. Rule-based algorithms usually construct water body indices based on the spectral characteristics of water bodies to identify water bodies, but in complex and changing geographical environments and various flood scenarios, it is extremely easy to make misjudgments and omissions. Simple machine learning models, such as support vector machines, although improving the degree of automation to a certain extent, have limited ability to model complex non-linear relationships and poor adaptability to different terrains, different types of water bodies, and complex remote sensing image data. In addition, these traditional methods often cannot fully explore and utilize massive remote sensing data and geographical information data, resulting in difficulty in accurately extracting the flood inundation range of power grid equipment.

[0005] Therefore, how to effectively predict the flood inundation range of power grid equipment and reduce disaster losses is a technical problem that needs to be solved. Summary of the Invention

[0006] To solve the problem of how to effectively predict the flood inundation range of power grid equipment in the prior art and reduce disaster losses, the purpose of the present invention is to provide a flood inundation prediction method for power grid equipment based on deep learning, an electronic device, and a computer-readable storage medium.

[0007] To solve the above technical problems, in a first aspect, according to some embodiments, the present invention provides a flood inundation prediction method for power grid equipment based on deep learning, including:

[0008] S1: Making a flood inundation model sample library, specifically including: collecting multi-source remote sensing image data, including optical remote sensing images and radar remote sensing images, obtaining the geographical location information of power grid equipment, power line distribution data, and topographic and geomorphic data; performing preprocessing on the multi-source remote sensing image data, including radiometric correction, geometric correction, and atmospheric correction, and performing format conversion and coordinate unification processing on the geographical information data of the power grid equipment; adopting a combination of manual annotation and semi-automatic annotation, and using an image annotation tool to annotate the flood water bodies in the optical remote sensing image data and the geographical information data of the power grid equipment to obtain an initial model sample library; training the initial model sample library according to the building samples in the optical remote sensing images to obtain a flood inundation model sample library, and converting the annotation results in the final model sample library into a format suitable for input to a deep learning model;

[0009] S2: Model construction and training, specifically including: constructing a hybrid deep learning model based on DeepLabv3+ and U-Net, using a convolutional neural network CNN as the backbone network, and using a pyramid pooling module PPM or an atrous spatial pyramid pooling module ASPP for feature fusion; for the DeepLabv3+ model, an encoder-decoder structure is adopted. The encoding part uses the DeepLab3 model as the encoder, and the decoding part uses a cascaded structure as the decoder. The main body of the encoder is a deep convolutional neural network DCNN with atrous convolution and a spatial pyramid pooling module with atrous convolution, and multi-scale information is introduced to make the extracted feature map have rich semantic information and rich details. For the decoder, first select features from the low level, compress the channels of the low-level features with 1x1 convolution, and then upsample the output of the encoder to make its resolution consistent with the low-level features, realizing the further fusion of the low-level features and high-level features; the U-Net model adopts a symmetric encoder-decoder structure, and splices and fuses the low-level features and high-level features through deconvolution operations; in the model, there is a multi-modal data fusion layer, which encodes the preprocessed power grid equipment coordinate information and terrain height information as feature vectors, and splices them in the channel dimension with the feature map of the remote sensing image at a specific layer; use the flood inundation model sample library to train the hybrid deep learning model, and optimize the loss function to improve the flood water body boundary segmentation accuracy;

[0010] S3: Model deployment and prediction, specifically including: inputting the real-time obtained remote sensing image into the trained hybrid model, and calculating and outputting the probability distribution map of the flood water body through forward propagation; performing threshold segmentation and morphological operations on the probability distribution map to generate a binary map of the flood inundation range; performing a spatial overlay analysis on the extracted binary map of the flood inundation range and the geographical location information of the power grid equipment to determine the power grid equipment affected by the inundation, and calculating the inundation degree and inundation range parameters;

[0011] S4: Result post-processing, specifically including: performing raster vectorization processing on the analyzed binary map of the flood inundation range to generate flood contour vector data in the shapefile format, and optimizing the water body edge regularity through manual correction to obtain the flood inundation prediction result of the power grid equipment.

[0012] Optionally, in some embodiments, the optical remote sensing image includes high-resolution series satellite images and Sentinel series satellite images.

[0013] Optionally, in some embodiments, the format suitable for input to the deep learning model is a semantic segmentation label map.

[0014] Optionally, in some embodiments, the training of the initial model sample library further includes:

[0015] Perform rotation and blurring enhancement operations on the original optical remote sensing image data to generate an enhanced training sample library.

[0016] Optionally, in some embodiments, training the initial model sample library according to the building samples in the optical remote sensing image to obtain a flood inundation model sample library specifically includes:

[0017] Use ArcGIS software to open the optical remote sensing image data, manually draw water body vectors, and set the attribute value of the drawn water body vectors to 1;

[0018] Save and output the vector data in PNG format, where the water body pixel value is 1 and the non-water body pixel value is 0;

[0019] Crop the PNG format into water body label tile data of 256×256 size;

[0020] Preprocess the original optical remote sensing image to obtain original tile data corresponding to the size of the water body label tile data, and save the file names in one-to-one correspondence with the file names of the water body label tile data as the flood inundation model sample library.

[0021] Optionally, in some embodiments, preprocessing the original optical remote sensing image to obtain original tile data corresponding to the size of the water body label tile data specifically includes:

[0022] Output the original optical remote sensing image and save it in JPG format, where the JPG format is 3-channel 24-bit;

[0023] Crop the JPG format image to obtain original tile data of 256×256 size.

[0024] Optionally, in some embodiments, for the pyramid pooling module ASPP, first perform convolution operations on features using dilated convolutions with different dilation factors, and then fuse the information of dilated convolutions with different dilation factors to retain object features of different scales.

[0025] Optionally, in some embodiments, the number of different dilation factors is 4.

[0026] In a second aspect, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described in any one of the first aspects above.

[0027] In a third aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method according to any one of the above first aspects are implemented.

[0028] The above technical solutions of the present invention have at least the following beneficial technical effects: 1) High accuracy: Through the powerful feature learning ability of the deep learning model, it can automatically learn the complex features of flood water bodies and other ground objects, effectively improving the accuracy of flood inundation area extraction and reducing misjudgment and missed judgment. 2) Strong real-time performance: The optimized algorithm process and efficient model structure enable the model to quickly process a large amount of remote sensing image data, meeting the real-time monitoring requirements of the power department for flood disasters. 3) Good adaptability: The multi-source data fusion and data enhancement technologies enhance the adaptability of the model to different regions, terrains, and flood scenarios, improving the generalization ability of the model. 4) Comprehensiveness: It can not only extract the flood inundation area, but also combine the power grid equipment information to accurately determine the affected power grid equipment, providing a comprehensive decision-making basis for the emergency response of the power department. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0030] Figure 1 It is a remote sensing monitoring inversion data extraction process for the water body area of the flood inundation area of power grid equipment provided by an embodiment of the present invention.

[0031] Figure 2 It is a technical route for water body extraction provided by an embodiment of the present invention.

[0032] Figure 3 It is a schematic diagram of the production process of a water body sample library provided by an embodiment of the present invention.

[0033] Figure 4 It is a sample example diagram provided by an embodiment of the present invention.

[0034] Figure 5 It is a DeepLabv3+ model architecture diagram provided by an embodiment of the present invention.

[0035] Figure 6 It is a schematic diagram of dilated convolution and conventional convolution (a) conventional convolution (b) dilated convolution provided by an embodiment of the present invention.

[0036] Figure 7It is a schematic diagram of the backbone network of the Xception network provided by an embodiment of the present invention.

[0037] Figure 8 It is a schematic diagram of an ASPP module provided by an embodiment of the present invention.

[0038] Figure 9 It is an architecture diagram of the U-Net network provided by an embodiment of the present invention.

[0039] Figure 10 It is a distribution map of the water body inundation range in Jilin Province in a certain year provided by an embodiment of the present invention.

[0040] Figure 11 It is the predicted monitoring range of the water body inundation of substations in Jilin Province in a certain year provided by an embodiment of the present invention.

[0041] Figure 12 It is the predicted monitoring range of the water body inundation of transmission poles in Jilin Province in a certain year provided by an embodiment of the present invention. Detailed implementation manners

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessarily confusing the concepts of the present invention.

[0044] In the accompanying drawings, a schematic diagram of the layer structure according to an embodiment of the present invention is shown. These figures are not drawn to scale, where for the purpose of clarity, some details are enlarged and some details may be omitted. The various regions, shapes of the layers shown in the figures, and their relative sizes and positional relationships are only exemplary. In practice, there may be deviations due to manufacturing tolerances or technical limitations, and those skilled in the art can design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0045] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0046] In recent years, deep learning technology has made breakthrough progress in the field of computer vision. Its powerful feature learning and pattern recognition capabilities provide a new way of thinking for solving the problem of flood inundation area extraction. However, current research on applying deep learning technology to the extraction of flood inundation areas of power grid equipment is relatively scarce, and there are many problems such as excessively high model complexity, insufficient training data, and poor generalization ability, which cannot meet the stringent requirements of the power industry for accuracy and real-time performance.

[0047] To solve the above problems, embodiments of the present invention provide a method, an electronic device, and a computer-readable storage medium for predicting flood inundation of power grid equipment based on deep learning. This method aims to significantly improve the accuracy and efficiency of extracting the flood inundation area of power grid equipment by leveraging advanced deep learning technology, providing strong support for the emergency response and protection of the power system during flood disasters, and ensuring the stability and reliability of power supply.

[0048] The core objective of the present invention is to overcome the many deficiencies of the prior art and provide a method for extracting the flood inundation area of power grid equipment based on deep learning. This method can fully utilize the powerful feature extraction and classification capabilities of deep learning, organically combine multi-source remote sensing data and geographical information data of power grid equipment, and achieve rapid and accurate extraction of the flood inundation area of power grid equipment. Specifically, by carefully constructing an efficient deep learning model, it automatically learns the complex features of flood water bodies and other ground objects in remote sensing images, significantly improving the extraction accuracy; at the same time, optimizing the method process, and adopting technical means such as parallel computing and distributed storage to improve the computing efficiency to meet the urgent need for real-time monitoring; in addition, through methods such as data augmentation and transfer learning, the generalization ability of the model is enhanced, enabling it to adapt to the task of extracting the flood inundation area of power grid equipment in different regions, different terrains, and different flood scenarios.

[0049] The water body extraction algorithm based on deep learning mainly uses the semantic segmentation algorithm of ground objects to extract water bodies in high-resolution remote sensing images. By enhancing the attention to comprehensive features such as image texture and spatial features, the influence of buildings, shadows, etc. on the water body extraction result is effectively reduced, and the result accuracy is improved. The model input is optical multi-spectral data, and the output is the extraction result of the flood inundation area of power grid facilities, supporting various output formats such as raster and vector. The present invention combines artificial intelligence and the DeepLabv3+ model of deep learning, and proposes a method for quickly identifying water bodies based on satellite remote sensing, realizing high-frequency and full-coverage intelligent extraction of risk information such as the flood inundation area of power grid equipment, the area of water accumulation areas, and the spatial distribution of water depth.

[0050] The technical solution of the present invention mainly includes four key parts: 1. Sample library production; 2. Model construction and training; 3. Model deployment and prediction; 4. Result post-processing. The technical route is as Figure 1 shown.

[0051] The following is illustrated by examples.

[0052] S1: Create a flood inundation model sample library, specifically including: collecting multi-source remote sensing image data, including optical remote sensing images and radar remote sensing images, obtaining the geographical location information of power grid equipment, power line distribution data, and topographic and geomorphic data; performing preprocessing on the multi-source remote sensing image data, including radiometric correction, geometric correction, and atmospheric correction, and performing format conversion and coordinate unification processing on the geographical information data of the power grid equipment; adopting a combination of manual annotation and semi-automatic annotation, and using an image annotation tool to annotate the flood water bodies in the optical remote sensing image data and the geographical information data of the power grid equipment to obtain an initial model sample library; training the initial model sample library according to the building samples in the optical remote sensing images to obtain a flood inundation model sample library, and converting the annotation results in the final model sample library into a format suitable for input to a deep learning model.

[0053] S2: Model construction and training, specifically including: constructing a hybrid deep learning model based on DeepLabv3+ and U-Net, using a convolutional neural network CNN as the backbone network, and using a pyramid pooling module PPM or an atrous spatial pyramid pooling module ASPP for feature fusion; for the DeepLabv3+ model, adopting an encoder-decoder structure, using the DeepLab3 model as the encoder in the encoding part, and using a cascaded structure as the decoder in the decoding part. The main body of the encoder is a deep convolutional neural network DCNN with atrous convolution and a spatial pyramid pooling module with atrous convolution, and multi-scale information is introduced to make the extracted feature maps have rich semantic information and rich details. The decoder selects features from the low levels, compresses the channels of the low-level features with 1x1 convolution, and then upsamples the output of the encoder to make its resolution consistent with the low-level features, realizing the further fusion of the low-level features and the high-level features; the U-Net model adopts a symmetric encoder-decoder structure, and fuses the low-level features and the high-level features by concatenating through deconvolution operations; there is a multi-modal data fusion layer in the model, encoding the preprocessed coordinate information and terrain height information of the power grid equipment as feature vectors, and performing channel dimension concatenation with the feature maps of the remote sensing images at a specific layer; training the hybrid deep learning model with the flood inundation model sample library, and optimizing the loss function to improve the flood water body boundary segmentation accuracy.

[0054] S3: Model deployment and prediction, specifically including: inputting the real-time acquired remote sensing images into the trained hybrid model, and calculating and outputting the probability distribution map of flood water bodies through forward propagation; performing threshold segmentation and morphological operations on the probability distribution map to generate a binary map of the flood inundation area; performing spatial overlay analysis on the extracted binary map of the flood inundation area and the geographical location information of power grid equipment to determine the power grid equipment affected by inundation, and calculating the inundation degree and inundation range parameters;

[0055] S4: Result post-processing, specifically including: performing raster vectorization processing on the analyzed binary map of the flood inundation area to generate flood contour vector data in the shapefile format, and optimizing the water body edge regularity through manual correction to obtain the flood inundation prediction result of the power grid equipment.

[0056] Specifically, in the training stage, on the one hand, the pascalvoc2012 water body semantic segmentation dataset is adopted; on the other hand, to further meet the training requirements of the model, water body samples are collected according to the annotations to complete the construction of the sample library. At the same time, data augmentation strategies such as random rotation, scaling, cropping, adding noise, etc. are adopted to better train and optimize the model. Finally, through continuous training and optimization, a water body extraction model with strong generalization performance is obtained. In the prediction stage, the deployed model is used to predict the relevant images. Through the image input operation, the images are sent into the model for prediction processing. Finally, through the regularization and raster-to-vector algorithms for the prediction impact, the vector patch data of the water surface is obtained, specifically referring to Figure 2 as shown.

[0057] The following is a detailed description.

[0058] 1. Sample library production

[0059] Data acquisition: Multisource data of the study area is acquired from satellite remote sensing platforms, aerial remote sensing equipment, and the databases of power departments. Professional remote sensing image processing software (such as ENVI, ERDAS, etc.) and geographic information system software (such as ArcGIS, etc.) can be used for data preprocessing and annotation work.

[0060] (1) Multisource data collection

[0061] The collected multisource remote sensing image data includes optical remote sensing images (such as high-resolution series satellite images, Sentinel series satellite images), radar remote sensing images, etc. to obtain rich flood information. At the same time, the geographical location information of power grid equipment, power line distribution data, and topographic and geomorphic data (such as digital elevation model DEM) are collected to provide comprehensive data support for subsequent analysis.

[0062] (2) Data preprocessing

[0063] Perform preprocessing operations such as radiometric correction, geometric correction, and atmospheric correction on remote sensing images to eliminate noise, geometric deformation, and atmospheric interference in the images and improve image quality. For power grid equipment and topographic data, perform format conversion, coordinate unification, etc. to enable effective integration with remote sensing image data.

[0064] (3) Data annotation and sample library production

[0065] Adopt a combination of manual annotation and semi-automatic annotation to annotate flood water bodies and power grid equipment in remote sensing images. Use professional image annotation tools such as LabelImg, CVAT, etc. to mark the boundaries of flood inundation areas and the locations of power grid equipment, and convert the annotation results into a format suitable for input to deep learning models, such as semantic segmentation label maps.

[0066] To ensure the effectiveness of the model, a large number of remote sensing image building samples are required to train the model. Based on this, this solution designs a method for producing a high-resolution remote sensing image water body sample library based on the DeepLabv3+ model. At the same time, image enhancement methods such as image rotation and blurring are designed to increase the generalization ability of the model on different remote sensing images.

[0067] Figure 3 For the water body sample library production method.

[0068] First, use ArcGIS software to open the remote sensing image, manually draw the water body vector, set the vector attribute value to 1, and then save and output the vector data as a PNG format. This raster image is saved as a single-channel 8-bit PNG image, where the water body pixel value is 1 and the non-water body (background) pixel value is 0. Then, crop it into 256×256-sized sample label tile data.

[0069] At the same time, save and output the original remote sensing image as a JPG format. This image is a 3-channel 24-bit JPG image, and then crop it into 256×256-sized original image tile data. The file names of the original image tile data and the label data tile data correspond one by one.

[0070] Specifically, refer to Figure 4 , which is an example diagram of the produced samples (white represents water bodies, and black represents other surfaces such as buildings and shadows).

[0071] 2. Model construction and training

[0072] A deep learning environment can be built on a GPU cluster, and a deep learning model can be implemented using the Python language and a deep learning framework. According to the characteristics of the dataset and the performance of the model, adjust the hyperparameters of the model, and perform multiple trainings and validations until the model converges and reaches the expected performance indicators.

[0073] Backbone network selection: Select the Convolutional Neural Network (CNN) that performs excellently in the field of computer vision as the backbone network, which can automatically extract high-level semantic features of images. At the same time, combined with attention mechanisms such as SE-Net and CBAM, the model can pay more attention to the features related to floods and power grid equipment and suppress irrelevant information.

[0074] Feature fusion module design: Design a dedicated feature fusion module to fuse feature maps of different scales. For example, use the Pyramid Pooling Module (PPM) or the Atrous Spatial Pyramid Pooling Module (ASPP) to fuse features with different receptive fields, so as to obtain more comprehensive context information and improve the recognition ability for flood areas and power grid equipment of different sizes.

[0075] Multi-modal data fusion layer: To make full use of the information of multi-source data, a multi-modal data fusion layer is constructed. The preprocessed remote sensing image data is fused with the geographical information data of power grid equipment and the topographic and geomorphic data. For example, the coordinate information and terrain height information of power grid equipment are encoded as feature vectors and spliced and fused with the feature map of the remote sensing image at a specific layer, so that the model can comprehensively consider various factors to extract the flood inundation range.

[0076] (1) DeepLabv3+ model

[0077] Figure 5 It is the architecture diagram of the DeepLabv3+ model.

[0078] The DeepLabv3+ model is the latest version of the DeepLab series and is developed based on DeepLabv3. It solves problems such as the loss of edge information in segmented objects and unclear edges in segmentation results. At the same time, it introduces a special encoder-decoder structure. The encoding part uses the DeepLab3 model as the encoder, and the decoding part introduces a cascaded structure as the decoder, enabling DeepLabv3+ to successfully integrate the advantages of two major types of models commonly used in semantic segmentation. The main body of its Encoder is a deep convolutional neural network DCNN with dilated convolution and an atrous spatial pyramid pooling module (ASPP), which introduces multi-scale information, making the extracted feature maps have rich semantic information and more details at the same time. In the Decoder module, features are first selected from the low levels, and the features at the low levels are compressed in channels using 1x1 convolutions (originally 256 channels or 512 channels), thereby reducing the proportion of the low levels. This is because the high-level features obtained by the encoder have richer information, so the features of the encoder should have a higher proportion, which is also beneficial for calculation. Finally, the output of the encoder is upsampled to make its resolution consistent with the low-level features, realizing the further fusion of the low-level features and the high-level features, thereby improving the accuracy of the segmentation boundary and enhancing the expression ability of the model.

[0079] ①Dilated convolution

[0080] Figure 6 It is a comparison diagram of dilated convolution and conventional convolution.

[0081] The main purpose of dilated convolution is to reduce the downsampling rate during feature extraction in the encoding stage while ensuring the receptive field of features. This is beneficial for extracting multi-scale information of images, making the extracted feature maps have rich semantic information and more details at the same time, ensuring that the model can learn more rich details and increasing the quality of feature extraction of the model.

[0082] ②DCNN module

[0083] Figure 7 It is a schematic diagram of the backbone network of the Xception network.

[0084] The DCNN is an Xception backbone network with an input stream, an intermediate stream, and an output stream. In the input stream, separable convolutions are mainly used to extract features from the input image. The separable convolutions process features in the spatial dimension and the channel dimension, enabling the full decoupling of spatial information and channel information, reducing the number of parameters, and improving the model training efficiency. At the same time, Xception draws on the idea of the Residual Network (ResNet) and introduces the design of residual connections to ensure the feature extraction effect, so that the extraction effect of the next layer of the network is not worse than that of the previous layer, and the transmission of effective features is well preserved. In the intermediate stream stage, it is mainly processed by repeatedly stacking separable convolutions based on the residual idea. Finally, in the output stream, the results generated by the intermediate stream are used for further feature extraction.

[0085] In the Xception network, when using dilated convolutions to extract image features, the downsampling rate of the image is ensured, and the receptive field of the features is expanded, so that the model can learn more detailed features of the image. Finally, multi-scale information of the image is obtained, making the extracted feature map have rich semantic information while also being more detailed.

[0086] ③ ASPP module

[0087] Since objects of the same class may have different proportions in the image, the invariant convolution operation of conventional convolutions may cause information loss during the convolution process for objects of different proportions. The ASPP module cleverly uses dilated convolutions with different dilation factors to perform convolution operations on the features, and then by fusing the information of dilated convolutions with different dilation factors, the features of objects of different proportion sizes are retained, enabling the network to learn more multi-scale information. By considering the retention of information of different object proportions during the convolution process, ASPP can improve the accuracy to a certain extent.

[0088] Figure 8 Describes the process of the ASPP module extracting and fusing features through different dilation factors. It mainly uses dilated convolutions with 4 different dilation factors to perform feature extraction operations on the feature image respectively.

[0089] In the decoding stage, the low-level features extracted by 1x1 convolutions are stacked with the high-level features of the encoded layer after 4-fold upsampling to obtain the abstract information of the image, and finally a segmentation map of the same size as the input image is obtained through the convolutional layer and the upsampling layer.

[0090] (2) U-Net network model

[0091] The network structure diagram is as Figure 9 shown.

[0092] Net is a network structure extended based on FCN and is a U-shaped network structure for multi-object tasks. Compared with CNN and FCN, using the U-Net network can more accurately evaluate the location information of features and speed up the network training speed. U-Net can well solve the loss of details lost due to downsampling operations, such as boundary information, etc., which is crucial for this fully connected prediction task like semantic segmentation. This network presents a symmetric structure, with the upsampling operation and the downsampling operation each performed 4 times, including 23 convolutional layers (18 of which are 3×3 convolutional layers, 4 are 2×2 convolutional layers, and 1 is 1×1 convolutional layer). Each time the image is downsampled, the number of image channels is doubled through convolutional operations. When using deconvolution operations for upsampling, the number of image channels is halved, and the result of deconvolution is concatenated with the feature map obtained from the left path.

[0093] 3. Model Deployment and Prediction Images

[0094] After the model is built, configure the corresponding environment to complete the model deployment. Input the data to be predicted into the model to complete the prediction of the flood inundation range of power grid equipment.

[0095] Model Inference: Input the preprocessed real-time remote sensing images into the trained deep learning model. The model calculates through forward propagation and outputs the probability that each pixel belongs to flood water bodies or other ground objects.

[0096] Post-processing: Post-process the probability map output by the model, using methods such as threshold segmentation and morphological operations (such as erosion and dilation) to remove noise and small isolated regions to obtain the final binary map of the flood inundation range.

[0097] Fusion with Power Grid Equipment Information: Conduct spatial overlay analysis on the extracted flood inundation range and the geographical location information of power grid equipment to determine which power grid equipment is within the flood inundation range, and calculate parameters such as the degree and range of inundation to provide accurate data support for the emergency decision-making of the power department.

[0098] For example, the trained model can be applied to real-time remote sensing image data for flood inundation range extraction. By comparing and verifying with the actual flood inundation situation, evaluate the accuracy and reliability of the model. At the same time, collect data under different regions and different flood scenarios to test and optimize the generalization ability of the model. Through the above algorithm for extracting the flood inundation range of power grid equipment based on deep learning, it can provide timely and accurate flood disaster monitoring information for the power department, effectively improve the ability of the power grid to cope with flood disasters, and ensure the safe and stable operation of the power system.

[0099] Taking the 1998 major flood in the Songhua River Basin as an example, the scope of power grid equipment threatened by floods in Jilin Province was intelligently extracted. In practical applications, this algorithm can quickly and accurately identify power grid equipment affected by floods, providing strong support for the emergency repair and equipment transfer of the power department, and effectively reducing the losses caused by flood disasters to the power grid.

[0100] Taking the 1998 flood monitoring changes in Jilin Province as an example, the water body inundation area in Jilin Province in 1998 was 7,668.93 square kilometers, as Figure 10 shown.

[0101] It was monitored and predicted that the number of substations in Jilin Province within the water body inundation area in 1998 was 3,168, as Figure 11 shown.

[0102] It was monitored and predicted that the number of transmission poles in Jilin Province within the water body inundation area in 1998 was 17, as Figure 12 shown.

[0103] From the results, the current solution can effectively predict the flood inundation of power grid equipment.

[0104] 4. Post-processing of Results

[0105] The extraction of the water body in the flood inundation area of power grid equipment often uses pixel-level semantic segmentation technology. The model classifies each pixel to complete the prediction of the water body. In the prediction of the water body edge, the classification ability of the model is often weak, and there will inevitably be cases of pixel misclassification and omission in the extracted patches, resulting in rough edges and inaccurate inflection points in the extracted patches compared with the real ground objects, thus presenting an irregular water body edge. Design water body regularization post-processing operations and manual corrections to make the water body contour more regular and thus closer to the actual shape of the water body in the flood inundation area of power grid equipment. At the same time, in order to make the water body extraction results more convenient to use, integrate raster-vector functions to realize the conversion from raster patches to contour vectors, and finally obtain a shapefile format file.

[0106] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method described in any one of the above embodiments are implemented.

[0107] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any one of the above embodiments are implemented.

[0108] An embodiment of the present invention further provides a computer program product, including a computer program, where the computer program is stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, so that the electronic device executes the steps of the method according to any one of the above embodiments.

[0109] Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0110] It should be understood that the above specific embodiments of the present invention are only used for exemplary illustration or explanation of the principle of the present invention, and do not constitute a limitation to the present invention. Therefore, any modifications, equivalent replacements, improvements, etc. made without departing from the spirit and scope of the present invention shall be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A flood inundation prediction method for power grid equipment based on deep learning, characterized in that, Including: S1: Create a flood inundation model sample library, specifically including: collecting multi-source remote sensing image data, including optical remote sensing images and radar remote sensing images, obtaining the geographical location information of power grid equipment, power line distribution data, and topographic and geomorphic data; performing preprocessing such as radiometric correction, geometric correction, and atmospheric correction on the multi-source remote sensing image data, and performing format conversion and coordinate unification processing on the geographical information data of the power grid equipment; using a combination of manual annotation and semi-automatic annotation, and through an image annotation tool, annotating the flood water bodies in the optical remote sensing image data and the geographical information data of the power grid equipment to obtain an initial model sample library; training the initial model sample library according to the building samples in the optical remote sensing images to obtain a flood inundation model sample library, and converting the annotation results in the final model sample library into a format suitable for input to a deep learning model; S2: Model construction and training, specifically including: constructing a hybrid deep learning model based on DeepLabv3+ and U-Net, using a convolutional neural network CNN as the backbone network, and using a pyramid pooling module PPM or an atrous spatial pyramid pooling module ASPP for feature fusion; for the DeepLabv3+ model, using an encoder-decoder structure, with the encoding part using the DeepLab3 model as the encoder and the decoding part using a cascaded structure as the decoder. The main body of the encoder is a deep convolutional neural network DCNN with atrous convolution and a spatial pyramid pooling module with atrous convolution, and multi-scale information is introduced to make the extracted feature maps have rich semantic information and detailed details. For the decoder, first select features from the low levels, compress the channels of the low-level features with 1x1 convolution, and then upsample the output of the encoder to make its resolution consistent with the low-level features, realizing the further fusion of the low-level features and the high-level features; the U-Net model uses a symmetric encoder-decoder structure, and splices and fuses the low-level features and the high-level features through deconvolution operations; there is a multi-modal data fusion layer in the model, encoding the preprocessed coordinate information and topographic height information of the power grid equipment as feature vectors, and splicing them in the channel dimension with the feature maps of the remote sensing images at a specific layer; using the flood inundation model sample library to train the hybrid deep learning model, and optimizing the loss function to improve the flood water body boundary segmentation accuracy; S3: Model deployment and prediction, specifically including: inputting the real-time obtained remote sensing image into the trained hybrid model, and calculating and outputting the probability distribution map of the flood water body through forward propagation; performing threshold segmentation and morphological operations on the probability distribution map to generate a binary map of the flood inundation range; performing spatial overlay analysis on the extracted binary map of the flood inundation range and the geographical location information of the power grid equipment to determine the power grid equipment affected by inundation, and calculating the inundation degree and inundation range parameters; S4: Post - processing of results, specifically including: performing raster - vectorization processing on the analyzed binary map of the flood inundation range to generate flood contour vector data in the shapefile format, and optimizing the regularity of the water body edge through manual correction to obtain the flood inundation prediction result of the power grid equipment.

2. The method according to claim 1, wherein The optical remote - sensing images include high - resolution series satellite images and Sentinel series satellite images.

3. The method according to claim 1, wherein The format suitable for input to the deep - learning model is a semantic segmentation label map.

4. The method according to claim 1, characterized in that The training of the initial model sample library further includes: Performing rotation and blurring enhancement operations on the original optical remote - sensing image data to generate an enhanced training sample library.

5. The method according to claim 1, wherein The training of the initial model sample library according to the building samples in the optical remote - sensing image to obtain a flood inundation model sample library specifically includes: Using ArcGIS software to open the optical remote - sensing image data, manually delineating the water body vector, and setting the attribute value of the delineated water body vector to 1; Saving and outputting the vector data in the PNG format, where the water body pixel value is 1 and the non - water body pixel value is 0; Cropping the PNG format into water body label tile data of size 256×256; Pre - processing the original optical remote - sensing image to obtain original tile data corresponding to the size of the water body label tile data, and saving the file names in one - to - one correspondence with the file names of the water body label tile data as the flood inundation model sample library.

6. The method according to claim 5, wherein The pre - processing of the original optical remote - sensing image to obtain original tile data corresponding to the size of the water body label tile data specifically includes: Outputting the original optical remote - sensing image and saving it in the JPG format, where the JPG format is 3 - channel 24 - bit; Cropping the JPG - formatted image to obtain original tile data of size 256×256.

7. The method according to claim 1, characterized in that, The pyramid pooling module ASPP first performs convolution operations on features using dilated convolutions with different dilation factors, and then fuses the information of dilated convolutions with different dilation factors to retain object features of different scales.

8. The method according to claim 7, wherein The number of different dilation factors is 4.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 - 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 - 7.

Images