A Substation Simulation Modeling Method and System

Abstract

The present invention belongs to the field of simulation modeling, and discloses a substation simulation modeling method, which includes: using voxel-constrained hierarchical clustering to segment the cloud scene of the substation site to obtain clusters containing one or more devices; using the Dense-SSD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster; using ICP to register each type of model to the target device point cloud and determine the device type by matching the distances of points; obtaining the pose of the device model in the substation scene from the calculation results of ICP to place the model into the actual substation scene. The present invention not only identifies the type of the device, but also obtains the pose of the model in the substation scene.

Description

Technical Field

[0001] The present invention belongs to the field of simulation modeling, and particularly relates to a substation simulation modeling method and system. Background Art

[0002] Using point cloud to reconstruct the 3D model of a substation is crucial for the operation of the smart grid. Its main goal is to quickly capture the device point cloud data and align the models of each device in the large-scale noisy point cloud scene of the substation. However, due to the low efficiency of traditional anti-noise clustering methods and the challenges in accurately classifying electrical devices with similar appearances, substation reconstruction needs to be improved. Automatic laser scanning modeling is a common method for substation simulation modeling, but it mainly has two limitations. One is that the massive point cloud data of the substation contains complex noises, which seriously increases the computational consumption when segmenting the target device PCD, resulting in mis-segmentation. The other is the lack of sufficient training samples to train a powerful classification network to successfully distinguish many devices with similar appearances.

[0003] With the rapid development of the smart grid, the demand for quickly performing 3D scene modeling of substations is increasing. Compared with traditional computer-aided design (CAD) modeling and oblique photography methods, the 3D laser scanning modeling method stands out by placing the existing model into the simulation scene and using the device point cloud. This method provides high accuracy and excellent scene integrity, effectively meeting various high-precision modeling requirements.

[0004] Currently, the modeling process mainly relies on manual processing. First, the device point cloud in the point cloud scene is segmented; next, the specific type of device is manually judged according to practical experience; then, the corresponding device model is found in the model library and manually dragged into the scene.

[0005] The whole process revolves around two key issues: (1) extracting the target device point cloud and (2) identifying the device model represented by the point cloud.

[0006] (1) Clustering algorithm for extracting the target device point cloud

[0007] Many useful point cloud clustering algorithms for extracting a single target from a point cloud scene have been published. These methods are mainly divided into two categories: distance information-based clustering algorithms and density information-based clustering algorithms. However, the actual effects of both types of algorithms are affected by noises.

[0008] (1.1) Distance information-based clustering algorithm

[0009] For distance - based clustering algorithms, the clustering results mainly depend on the pre - set distance threshold. The premise for obtaining better clustering results is that the intervals of each target cluster have fixed rules. When the point - cloud scene contains a large amount of noise, the noise will connect the target clusters, making it difficult to determine the distance threshold.

[0010] (1.2) Density - based clustering algorithms

[0011] Density - based clustering algorithms can, to a certain extent, solve the problems caused by noise. Its main principle is to judge noise by relying on the difference between the density of target point clouds and the density of noise point clouds. The radius filter is the most commonly used denoising method. It takes a point to be determined as the center, controls the size of the space by setting the radius, and analyzes the average distance from each point in the space to the center point of the space to measure the density of the space. A threshold determines whether a point is noise. This method can only determine noise through a fixed global threshold and is not suitable for the change of noise density.

[0012] On this basis, the statistical filter adaptively distinguishes noises with different density changes by determining the average distance information of the global neighborhood and obtaining the standard - deviation distribution. Statistical outlier removal is a typical statistical filter. However, the neighborhood space is still controlled by a fixed radius. When evaluating noise point clouds, some point clouds in the target may be judged as noise, thus possibly wrongly removing valid point clouds. Density - based spatial clustering of applications with noise (DBSCAN) is an algorithm that can cluster target point clouds of any shape in a noise point - cloud scene and can be well applied to the point - cloud clustering task of complex - shaped devices. However, the computational efficiency of DBSCAN is not high, and it is not suitable for substation scenarios where the point - cloud data volume reaches hundreds of megabytes.

[0013] (2) Identifying the device models represented by point clouds

[0014] Object - recognition algorithms based on neural networks have developed rapidly and can be used to recognize two - dimensional or three - dimensional targets (two recognition schemes), but they largely rely on establishing training samples. There are many types and models of power equipment in substations. Establishing sufficient training samples for each device is a very large project and is difficult to complete in a short time. Especially for power equipment of the same category but different types, their differences are usually difficult to intuitively discover. Generating training samples and establishing corresponding neural networks for minor differences is a huge challenge that requires the cooperation of professionals with knowledge of power and artificial intelligence. The huge workload of generating training samples directly limits the mature application of 3D object - recognition algorithms (recognizing three - dimensional targets) in substation equipment recognition.

[0015] Common two-dimensional image recognition algorithms can identify multiple targets simultaneously and have been used to identify equipment in substations. The Single Shot MultiBox Detector (SSD) is an object detection method that combines speed and accuracy and is widely used in various fields of image recognition. However, its drawback is that it is not robust enough for small targets because the shallow feature maps are not powerful enough. The base network was changed from VGG to ResNet to improve the representational ability of deeper networks. This direction is important for improving the prediction ability of SSD. As a better CNN model than ResNet, DenseNet provides greater improvement in SSD prediction (identifying two-dimensional targets).

[0016] Through the above analysis, the problems and deficiencies of the existing technology are as follows: The two key problems of extracting the point cloud of target equipment and identifying the equipment model represented by the point cloud have not been solved in the simulation modeling of substations. Summary of the Invention

[0017] In view of the problems existing in the prior art, the present invention provides a substation simulation modeling method and system.

[0018] The present invention provides a substation simulation modeling method, which specifically includes:

[0019] S1: Use Voxel Constrained Hierarchical Clustering (VCHC) to segment the cloud scene at the substation site to obtain clusters containing one or more devices;

[0020] S2: Use the Dense-SSD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster;

[0021] S3: Retrieve the corresponding different types of models from the model library according to the device category, use ICP to register each type of model to the target device point cloud, and determine the device type by matching the distances of the points;

[0022] S4: Obtain the pose of the device model in the substation scene from the calculation results of ICP to place the model into the actual substation scene.

[0023] Further, the S1 specifically includes:[[]]

[0024] According to the characteristics of the point cloud noise and the vertical continuity of the device point cloud, use the clustering results at different plane levels to constrain each level to identify and remove the noise point cloud;

[0025] Project the point clouds at the same elevation interval onto the same plane and use the voxelization clustering algorithm for clustering to obtain the clustering results of the plane point clouds in different elevation intervals.

[0026] Further, the S2 specifically includes: adding a network for extracting deep features on the basis of SSD, which predicts the device type based on the point cloud image of the power equipment; in this detection framework, DenseNet is used to extract deep features, and then the object frame proposal strategy and frame regression algorithm in the Dense-SSD algorithm are combined to reconstruct an end-to-end object recognition and detection network;

[0027] Further, the Dense-SSD algorithm has a high detection accuracy for small objects, can identify the segmented power equipment and the parts around the equipment that do not belong to the equipment, including pedestrians and other equipment, and reduces the adverse impact of over-segmentation on equipment recognition.

[0028] Further, the Dense-SSD algorithm is divided into two parts: a feature extraction part and a prediction output part.

[0029] Further, the feature extraction part has a DenseNet structure, including a stem block, four dense blocks, two transition layers and two transition layers with pooling layers; in the stem block, in order to reduce the information loss of the original input image, several small convolutional kernels are used to replace the large convolutional kernels, and the original design in DenseNet (7×7 convolutional layer, stride = 2; 3×3 max pooling layer, stride = 2) is changed to a combination of four 3×3 convolutional layers and 2×2 max pooling layers; the transition layer between two adjacent dense blocks consists of a 1×1 convolutional layer and a 2×2 max pooling layer, realizing model compression and reducing the number of feature maps output by each dense block; the transition w / o pooling layer ensures an increase in the number of dense blocks without reducing the resolution of the final feature map, and consists of 1×1 convolutional layers.

[0030] Further, the prediction output part retains the structure of the prediction layer of the SSD network, predicts the prediction confidence of all object categories and the position offset value of the prediction frame, combines the side output feature maps of the stem block and the dense blocks as shallow information, and combines the final output feature map of the network as deep information. The shallow features and deep features jointly determine the prediction value to achieve depth supervision;

[0031] To reduce the computational load, downsampling and merging are sequentially performed on each feature map;

[0032] The downsampling includes a max pooling layer and a convolutional layer; in the whole structure, the dense block is the core part. The connection mode introduces direct connections from any layer to all subsequent layers. The outputs of the upper layers are x0, x1, x2, x3, and x4. The connection structure between the outputs consists of a batch normalization (BN) layer, a rectified linear unit (ReLU), and a 3×3 inclusive (conv). The dense block realizes the reuse of shallow feature maps by establishing shortcut connections between all previous and subsequent layers, which can solve the problem that the features of small targets become less obvious as the depth increases and improve the accuracy of small target detection.

[0033] Another object of the present invention is to provide a substation simulation modeling system for implementing the substation simulation modeling method. The system includes:

[0034] A segmentation module, configured to segment the cloud scene of the substation site using voxel-constrained hierarchical clustering (VCHC) to obtain a cluster containing one or more devices;

[0035] An identification module, connected to the segmentation module, uses the Dense-SSD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster;

[0036] A retrieval module, connected to the identification module, is configured to retrieve corresponding different types of models from the model library according to the device categories, register each type of model to the target device point cloud using ICP, and determine the device type by matching the distances of the points;

[0037] An application module, connected to the retrieval module, obtains the pose of the device model in the substation scene from the calculation results of ICP to place the model into the actual substation scene.

[0038] Another object of the present invention is to provide a computer device. The computer device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the substation simulation modeling method.

[0039] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the substation simulation modeling method.

[0040] Combined with the above technical solutions and the solved technical problems, the advantages and positive effects of the technical solution to be protected by the present invention are:

[0041] First, the present invention proposes a model-driven three-dimensional laser scanning scene reconstruction method for a substation and verifies the effectiveness of this method by measuring the point cloud data of the substation. The VCHC clustering method is used to segment the cloud scene of the entire substation. While removing the high-density noise points, the sparse device point clouds are completely retained, effectively avoiding over-segmentation. The segmentation accuracy rate reaches over 90%. The classification method combining Dense-SSD and ICP is used to quickly identify the types of power equipment, and the accuracy rate is 92.86%. Dense-SSD can accurately identify multiple device categories in the segmented clusters, effectively avoiding inaccurate identification caused by insufficient segmentation. ICP can not only distinguish the obvious differences in the appearance of different types of devices but also accurately place the model into the substation scene.

[0042] (1) Use the cross-constrained hierarchical clustering algorithm (VCHC) to solve the "clustering segmentation" technical problem

[0043] The present invention proposes a cross-constrained hierarchical clustering algorithm (VCHC) based on traditional voxel clustering. This algorithm reduces the dimension of 3D point cloud information to 2D, which not only improves the calculation efficiency but also effectively removes noise. The algorithm uses a constraint criterion to accurately determine noise and can process the point cloud of a large substation scene containing hundreds of megabytes of data at one time. In this cross-constrained hierarchical clustering algorithm based on traditional voxel clustering, the 3D point cloud data is reduced to multiple 2D point cloud projection data, and then each 2D point cloud data is clustered. If there are point cloud clusters in all slices at the same horizontal position, these clusters are retained. Otherwise, the clusters at this horizontal position are regarded as noise and removed.

[0044] The cross-constrained hierarchical clustering algorithm (VCHC) includes the following two strategies:

[0045] (a) The hierarchical clustering strategy improves the calculation efficiency by reducing the calculation dimension, enabling the proposed method to process the point cloud of a large substation scene containing hundreds of megabytes of data at one time.

[0046] (b) The cross-constrained strategy improves the calculation accuracy of segmentation by enhancing the anti-noise ability, ensuring that in the substation point cloud scene with high-density noise, the point clouds of power equipment are completely retained, effectively avoiding over-segmentation.

[0047] (2) Solve the "accurate identification of device models" technical problem through Dense-SSD + ICP

[0048] The present invention applies the improved SSD structure (Dense-SSD) to the image recognition of power equipment and quickly identifies the main equipment categories in the cluster. For devices of the same category but different models, the Iterative Closest Point (ICP) algorithm is used for fine registration, and the device models are finely distinguished by matching point distances to obtain the world coordinates of the models in the scene, and the models are placed in the scene.

[0049] Second, the present invention not only identifies the type of equipment, but also obtains the pose of the model in the substation scene. The processing results of the on-site measurement data of the substation show that the automatic modeling method of the present invention can accurately model 92.86% of the equipment in the substation.

[0050] The present invention is a fast model construction framework for the point cloud scene of large substations, which includes using the cross-constrained hierarchical clustering algorithm (VCHC) to realize the clustering segmentation of the point cloud of power equipment in the substation scene with large noise; and using the Dense-SSD+ICP multi-level recognition scheme to accurately identify the models of power equipment.

[0051] Third, as the creative auxiliary evidence of the claims of the present invention, it is also reflected in the following important aspects:

[0052] (1) The expected benefits and commercial values after the transformation of the technical solution of the present invention are:

[0053] Currently, the modeling of the substation point cloud scene is still manually modeled, and the cost is about 200,000 yuan per substation. Through the fast modeling of the present invention, the modeling cost per substation can be reduced to 20,000 yuan. The cost saved for each substation modeling is 180,000 yuan. More than 2,000 modeled substations can be completed every year, saving a total of 360 million yuan in modeling costs.

[0054] The traditional modeling time of the substation point cloud scene is 4 - 6 months. Through fast modeling, the modeling of each substation is shortened to 3 - 5 days. It provides an efficient platform for the establishment and implementation of other substation visualization solutions.

[0055] (2) The technical solution of the present invention fills the technical gaps in the domestic and international industries:

[0056] Fast modeling of the cross-voltage-level substation point cloud scene, and the proposed VCHC is applicable to the point cloud scenes of substations with voltage levels such as 500kV, 220kV, 35kV, and 10kV.

[0057] (3) Whether the technical solution of the present invention solves the technical problems that people have always been eager to solve but have never succeeded in:

[0058] In the substation point cloud scenario, the point cloud data volume is large and the noise interference is serious. The density of some noise point clouds is greater than that of the equipment point clouds, and the efficiency of the manual segmentation scheme is low. It is difficult for traditional automatic segmentation schemes to achieve efficient and accurate segmentation. The proposed VCHC scheme solves this problem.

[0059] There are many types of substation equipment, and the similarity of equipment of the same model is high. Identification requires professional technical personnel with high professional ability to search a large number of models from the model library and compare them repeatedly, resulting in low efficiency. The proposed Dense-SSD+ICP multi-level identification scheme solves this problem.

[0060] (4) Does the technical solution of the present invention overcome technical prejudice?

[0061] The applicability of each point cloud segmentation and clustering algorithm is limited, and the noise characteristics of point clouds of different substation equipment are different. Therefore, it is difficult to use a single segmentation and clustering algorithm to extract the point clouds of all power equipment in all voltage level areas. The present invention overcomes this technical prejudice through VCHC.

[0062] Traditionally, it is considered that the method of using artificial intelligence to automatically identify power equipment is only applicable to equipment types. For similar-looking equipment with the same type but different models, when there is a lack of a large number of training samples, the recognition accuracy of the neural network for equipment models is low. Therefore, there are many artificial intelligence automatic recognition methods based on neural networks, which are difficult to play a role in the field of power equipment model recognition. The present invention overcomes this technical prejudice through the Dense SSD+ICP multi-level recognition scheme. Description of the Drawings

[0063] Figure 1 is the automatic modeling workflow diagram of the substation point cloud scenario provided by the embodiment of the present invention.

[0064] Figure 2 is the schematic diagram of longitudinal constraint provided by the embodiment of the present invention.

[0065] Figure 3 is the intercepted key point cloud data and elevation interval division diagram within the 220kV area provided by the embodiment of the present invention.

[0066] Figure 4 is the intercepted key point cloud data and elevation interval division diagram within the 220kV area provided by the embodiment of the present invention.

[0067] Figure 5 is the constraint process and clustering result diagram of the measured point cloud scenario in the 220kV area provided by the embodiment of the present invention; where a is the bounding box of the cluster in two elevation intervals; b is the new bounding box generated by merging the clusters in the two intervals after applying cross-constraint; c is the top view of the measured point cloud data.

[0068] Figure 6 It is a clustering result graph obtained by applying DBSCAN to the 35kV area using different neighborhood radii ε and minimum neighborhood point numbers MinPts values provided by the embodiments of the present invention; among them, a, b, and c respectively show the DBSCAN clustering results with the minimum neighborhood point number of 30 and ε being 0.1m, 0.5m, and 1.0m respectively; d, e, and f respectively show the DBSCAN clustering results with ε = 0.5m and the minimum neighborhood point numbers being 10, 50, and 80 respectively.

[0069] Figure 7 It is a VCHC segmentation clustering result graph of the 35kV area provided by the embodiments of the present invention.

[0070] Figure 8 It is a device category classification result graph provided by the embodiments of the present invention; a is the recognition result graph of the insulating post, b is the recognition result graph of the disconnector, c is the recognition result graph of the circuit breaker; the value above the candidate box is the confidence level.

[0071] In the figure: 1, segmentation module; 2, recognition module; 3, retrieval module; 4, application module. Specific Embodiments

[0072] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0073] As Figure 1 shown, the embodiments of the present invention provide a substation simulation modeling method, which specifically includes:

[0074] S1: Use voxel-constrained hierarchical clustering (VCHC) to segment the cloud scene of the substation site to obtain clusters containing one or more devices;

[0075] S2: Use the Dense-SSD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster;

[0076] S3: Retrieve corresponding different types of models from the model library according to the device categories, use ICP to register each type of model to the target device point cloud, and determine the device type by matching the distances of the points;

[0077] S4: Obtain the pose of the device model in the substation scene from the calculation results of ICP to place the model into the actual substation scene.

[0078] The specific content of S1 includes:

[0079] According to the characteristics of point cloud noise and the vertical continuity of the device point cloud, the clustering results at different plane levels are used to constrain each level to identify and remove the noisy point cloud;

[0080] Project the point clouds with the same elevation interval onto the same plane and use the voxel clustering algorithm for clustering, and the clustering results of the plane point clouds in different elevation intervals can be obtained.

[0081] As Figure 2 shown, (a) shows that the entire area can be divided into three elevation intervals: h0, h1, and h2; (b) indicates that the point cloud data intervals at different elevations are projected onto three planes. The gray area represents the noisy point cloud clusters, and the purple area represents the non-noisy point cloud clusters. The vertical dotted lines connect the unified horizontal position units

[0082] As Figure 3 shown, (a) focuses on the interception of point cloud data; (b) after dividing the elevation intervals of the key point cloud data and using cross-constraints to remove the noisy point cloud clusters, the remaining point cloud clusters need to be merged.

[0083] The S2 specifically includes: adding a network for extracting deep features on the basis of SSD, and constructing a new neural network, similar to Dense SSD (DSOD), which predicts the device type based on the point cloud image of the power equipment; in this detection framework, deep features are extracted through DenseNet, and then the object frame proposal strategy and frame regression algorithm in the Dense-SSD algorithm are combined to reconstruct an end-to-end object recognition and detection network;

[0084] The Dense-SSD algorithm has a high detection accuracy for small objects, can identify the segmented power equipment and the parts around the equipment that do not belong to the equipment, including pedestrians and other equipment, and reduces the adverse effects of over-segmentation on equipment recognition.

[0085] The Dense-SSD algorithm is divided into two parts: the feature extraction part and the prediction output part.

[0086] The feature extraction part has a DenseNet structure, including a stem block, four dense blocks, two transition layers, and two transition layers with pooling layers; in the stem block, in order to reduce the information loss of the original input image, several small convolutional kernels are used to replace the large convolutional kernels, and the original design in DenseNet (7×7 convolutional layer, stride = 2; 3×3 max pooling layer, stride = 2) is changed to a combination of four 3×3 convolutional layers and a 2×2 max pooling layer; the transition layer between two adjacent dense blocks consists of a 1×1 convolutional layer and a 2×2 max pooling layer, which compresses the model and reduces the number of feature maps output by each dense block; the transition without pooling layer ensures an increase in the number of dense blocks without reducing the resolution of the final feature map, and consists of a 1×1 convolutional layer.

[0087] The prediction output part retains the structure of the prediction layer of the SSD network, predicts the prediction confidence of all object categories and the position offset values of the prediction frames, combines the side output feature maps of the stem block and the dense blocks into shallow information, and combines the final output feature map of the network into deep information. The shallow features and deep features jointly determine the prediction values to achieve depth supervision;

[0088] To reduce the computational amount, downsampling and merging are performed on each feature map in turn;

[0089] The downsampling includes a max pooling layer and a convolutional layer; in the whole structure, the dense block is the core part, and the connection mode introduces direct connections from any layer to all subsequent layers. The outputs of the upper layers are x0, x1, x2, x3, and x4, and the connection structure between the outputs consists of a batch normalization (BN) layer, a rectified linear unit (ReLU), and a 3×3 convolution (conv). The dense block realizes the reuse of shallow feature maps by establishing shortcut connections between all previous layers and subsequent layers, which can solve the problem that the features of small targets become less obvious as the depth increases and improve the accuracy of small target detection.

[0090] As Figure 4 shown, an embodiment of the present invention provides a substation simulation modeling system for implementing the substation simulation modeling method. The system includes:

[0091] A segmentation module 1, which is used to segment the cloud scene at the substation site using voxel-constrained hierarchical clustering (VCHC) to obtain clusters containing one or more devices;

[0092] An identification module 2, connected to the segmentation module 1, uses the Dense-SSD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster;

[0093] The retrieval module 3, connected to the recognition module 2, is used to retrieve corresponding different types of models from the model library according to the device category, register each type of model into the target device point cloud using ICP, and determine the device type by matching the distances of points.

[0094] The application module 4, connected to the retrieval module 3, obtains the pose of the device model in the substation scenario from the calculation result of ICP to place the model into the actual substation scenario.

[0095] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the substation simulation modeling method.

[0096] An embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the substation simulation modeling method.

[0097] It should be noted that the embodiments of the present invention can be implemented through hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips and transistors, or field-programmable gate arrays and programmable logic devices, can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0098] Some positive effects have been achieved during the research and development or use of the embodiments of the present invention, and it indeed has great advantages compared with the prior art. The following content is described in combination with the data and charts of the test process.

[0099] The segmentation and clustering results of VCHC:

[0100] The present invention compares the clustering method with the DBSCAN and voxelization clustering methods in terms of segmentation time and accuracy on point cloud data of different voltage levels. The segmentation accuracy is the ratio of the number of correctly segmented devices to the total number of segmented devices. The comparison results are shown in Table 1.

[0101] Table 1 Comparison of the Efficiency and Accuracy of Different Segmentation and Clustering Methods for Substation Point Cloud Scenes at Different Voltage Levels

[0102]

[0103] According to the calculation time results in Table 1, voxel clustering is the least efficient scene segmentation method. The calculation efficiency of DBSCAN is higher than that of voxel clustering, but as the voltage level increases, the point cloud scene becomes larger and the number of point clouds increases, and the advantage of DBSCAN becomes less and less obvious. The calculation efficiency of VCHC is significantly higher than that of the other two methods, and the calculation time does not change significantly with the change of voltage level, indicating that this method is less affected by the number of point clouds. The high segmentation calculation efficiency of VCHC may come from its special dimensionality reduction processing.

[0104] In the point cloud of the substation scene, the sparse and scattered noise point clouds in the substation can be easily identified and removed by traditional denoising methods. However, there are two types of noise that are difficult to identify: super-density observation noise and construction equipment point clouds. When these two types of noise are not identified, the clustering and segmentation algorithm is likely to cluster them into the point cloud clusters, thus hindering equipment identification. The present invention uses the point cloud clustering of a 220 kV substation to illustrate the ability of the denoising method provided by the present invention to suppress this type of noise. As Figure 5 shown, the present invention selects point cloud data at two elevation intervals, one is the interval of 1 - 2 m, and the other is the interval of 2 - 3 m, for voxel clustering on the plane. The two-dimensional clustering results are as Figure 5 (a) shown. The outer bounding boxes of these two elevation intervals cross multiple times in the horizontal direction. The results of the cross-constraint calculation are as Figure 5 (b) shown. However, there are also non-intersecting areas, and six example areas in these places are marked in Figure 5 . From the top view of the measured point cloud, that is Figure 5 (c), obvious noise point clouds can be seen in the marked areas 3 - 6. These noise point clouds are dense and have a large expansion range, and it is difficult to distinguish them from the equipment point clouds by density and expansion range. In the marked areas 1 and 2, the construction of most equipment has not been completed, and there are only concrete pillars supporting the equipment on the substation site. This observation also explains why there are more clusters in the 1 - 2 m interval than in the 2 - 3 m interval.

[0105] Comparison with DBSCAN with Different Parameters:

[0106] The comparison method is DBSCAN because this density-based clustering method can effectively remove noise and avoid the unfair clustering segmentation tasks that may occur in other clustering methods that require prior noise removal. The present invention records their segmentation accuracies and compares them with the segmentation accuracy of the proposed VCHC method. Comparisons were made for the 35 kV and 220 kV regions (Table 2).

[0107] Table 2 Comparison of segmentation accuracies of different methods in substation point cloud scenarios

[0108]

[0109] The number of segmentation clusters obtained by the proposed VCHC method is closest to the actual number of modeled devices. Secondly, the number of segmentation clusters obtained by DBSCAN with the neighborhood radius ε set to 1.0 m is the closest. However, when the neighborhood radius ε is set to 0.1 m and 0.5 m, the number of obtained segmentation clusters is significantly higher than the actual number of devices. This is because there is a large amount of noise in the clustering results, causing DBSCAN to wrongly cluster the noise as valid device point clouds. Even with the VCHC method that is closest to the actual number of devices, there are still some clusters formed by noise point clouds and some over-segmented clusters. Finally, VCHC achieved a segmentation accuracy of 83% in the 35 kV substation point cloud scenario. For the 220 kV substation point cloud scenario, the accuracy increased to 96%. This increase is because compared with the 35 kV substation, the device sizes in the 220 kV substation are generally larger, and the noise is more easily and effectively identified. On the other hand, for DBSCAN, although the number of segmentation clusters obtained with the neighborhood radius ε of 1.0 m is relatively close to the number of modeled devices, the number of well-formed device point cloud clusters is still quite low. In the 35 kV substation scenario, they only account for about 15%, while in the 220 kV substation scenario, they only account for about 27%.

[0110] Comparative analysis of VCHC:

[0111] In DBSCAN, when setting the neighborhood radius ε, it must be made larger than the minimum distance between points so that the points are within the neighborhood. To ensure effective noise reduction, the neighborhood radius should not be too large either, because an overly large radius may retain too much noise, thus hindering the clustering process. The present invention conducted two groups of comparisons: one group compared the clustering results under different neighborhood radii ε, and the other group compared the clustering results under different minimum neighborhood point numbers (MinPts), as Figure 6 shown.

[0112] Figure 6The clustering results obtained by applying DBSCAN to the 35 kV area using different neighborhood radii ε and minimum neighborhood point numbers MinPts values; among them, (a), (b), and (c) respectively show the DBSCAN clustering results with a minimum neighborhood point number of 30 and ε of 0.1 m, 0.5 m, and 1.0 m respectively; while (d), (e), and (f) respectively show the DBSCAN clustering results with ε = 0.5 m and minimum neighborhood point numbers of 10, 50, and 80 respectively. When comparing the clustering results of different minimum neighborhood point numbers MinPts, using a neighborhood radius ε of 0.5 m and MinPts of 10, 50, and 80 respectively, the following was observed: From the clustering results, for the same neighborhood radius, different MinPts values led to relatively consistent clustering numbers and overall segmentation. However, the choice of MinPts determines whether the clustering represents device points or noise. It can be seen that the segmentation results of different MinPts values show slight differences( Figure 6 ). However, as MinPts increases, more device points are wrongly deleted, and setting MinPts too low will lead to an increase in the retained noise points. The segmentation results obtained by the VCHC method exhibit extremely small residual noise, and the clustering results are clear and clean( Figure 7 ).

[0113] Classification results:

[0114] Use Dense-SSD to classify and identify power equipment categories, and use ICP registration to classify and identify equipment models. The classification results are shown in Table 3. The classification accuracy is the ratio of the number of correctly classified devices to the total number of devices. In the substation point cloud scenario, the number of devices to be modeled and the number of correctly modeled devices are shown in Table 3.

[0115] Table 3 Comparison of recognition accuracies of different methods in the substation point cloud scenario

[0116]

[0117] The present invention uses the point cloud clusters in the VCHC segmentation results as the basis for recognition, and compares the recognition results of the "category recognition and type matching" method with the results of traditional VGG, AlexNet, ResNet, and DenseNet models (Table 3). In the category recognition stage, the recognition results of all methods change very little, remaining at about 96% - 97%. This indicates that neural networks based on traditional CNNs can effectively identify different device categories according to the appearance of the devices.

[0118] When it comes to cases where the categories are the same but the types are different, that is, the overall appearance is similar but the details are different, the fine - matching method provided by the present invention significantly widens the gap with other traditional methods in terms of recognition accuracy. The method provided by the present invention achieves a final recognition accuracy of over 95%. For the 220 kV substation scenario, the recognition accuracy reaches 97%. In contrast, the recognition accuracy of other neural networks is around 30%.

[0119] In the actual modeling process, it is often very difficult to distinguish current transformers, voltage transformers, and capacitor voltage transformers based solely on point - cloud images. Distinguishing these transformers requires the assistance of on - site photos and CAD construction drawings. Therefore, the present invention places these transformers in the same category and further differentiates them during the matching process. The category classification results of the three main devices are shown in Figure 8 as Figure 8 shown, where the candidate boxes precisely outline the spatial positions of the devices and give relatively high confidence values, that is, the confidence values can be directly used to distinguish the categories of the devices.

[0120] In the entire substation scenario, six types of insulating supports, eight types of circuit breakers, and 26 types of disconnectors can be observed. The more categories there are, the more difficult the neural network training process becomes, and the lower the predicted confidence value. Therefore, the confidence of the insulating supports is the highest, while that of the disconnectors is the lowest.

[0121] Increasing the number of device models of the same category in the training set may make the neural network training more challenging and result in a decrease in confidence, but compared with devices of other categories, the confidence level is still significantly higher. Therefore, the present invention accurately determines its category according to the highest confidence level. Even though the confidence of the disconnector is less than 90%, based on the available confidence level, the present invention can still confidently identify its category.

[0122] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A substation simulation modeling method, characterized in that, The method specifically includes: S1: Use voxel-constrained hierarchical clustering to segment the cloud scene at the substation site to obtain clusters containing one or more devices; S2: Use the DSOD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster; S3: Retrieve corresponding different types of models from the model library according to the device categories, use the iterative closest point algorithm to register each type of model to the target device point cloud, and determine the device type by matching the distances of the points; S4: Obtain the pose of the device model in the substation scene from the calculation results of the iterative closest point algorithm to place the model into the actual substation scene; The S2 specifically includes: adding a network for extracting deep features on the basis of SSD, and constructing a new neural network that predicts the device type based on the point cloud images of power devices; in the detection framework, extract deep features through DenseNet, and then combine the object frame proposal strategy and the frame regression algorithm in the Dense-SSD algorithm to reconstruct an end-to-end object recognition and detection network; The DSOD algorithm is divided into two parts: a feature extraction part and a prediction output part; The feature extraction part has a DenseNet structure, including a stem block, four dense blocks, two transition layers, and two transition layers with pooling layers; in the stem block, in order to reduce the information loss of the original input image, several small convolutional kernels are used to replace the large convolutional kernels, and the original design in DenseNet is changed to a combination of four 3×3 convolutional layers and a 2×2 max pooling layer, where the original design in DenseNet is a 7×7 convolutional layer, stride = 2, 3×3 max pooling layer, stride = 2; the transition layer between two adjacent dense blocks consists of a 1×1 convolutional layer and a 2×2 max pooling layer; the transition without pooling layer ensures an increase in the number of dense blocks without reducing the resolution of the final feature map, and consists of a 1×1 convolutional layer; The prediction output part retains the structure of the prediction layer of the SSD network, predicts the prediction confidence of all object categories and the position offset values of the prediction frames, combines the side output feature maps of the stem block and the dense blocks as shallow information, and combines the final output feature map of the network as deep information; the shallow information and the deep information jointly determine the prediction value to achieve depth supervision; In order to reduce the computational amount, downsample and merge each feature map in turn; The downsampling includes a max pooling layer and a convolutional layer; in the whole structure, the dense block is the core part, and the connection mode introduces direct connections from any layer to all subsequent layers. The outputs of the upper layers are x0, x1, x2, x3, and x4, and the connection structure between the outputs consists of a batch normalization (BN) layer, a rectified linear unit (ReLU), and a 3×3 convolution (conv). The dense block realizes the reuse of shallow feature maps by establishing shortcut connections between all previous and subsequent layers.

2. The substation simulation modeling method according to claim 1, wherein, The S1 specifically includes: According to the characteristics of point cloud noise and the vertical continuity of device point clouds, different levels of clustering results are used to constrain each level to identify and remove noisy point clouds; Project the point clouds with the same elevation interval onto the same plane and use the voxel clustering algorithm for clustering to obtain the clustering results of the plane point clouds in different elevation intervals.

3. A substation simulation modeling system for implementing the substation simulation modeling method according to any one of claims 1-2, characterized in that, The substation simulation modeling system includes: A segmentation module, which is used to segment the cloud scene of the substation site using voxel-constrained hierarchical clustering to obtain clusters containing one or more devices; An identification module, connected to the segmentation module, uses the DSOD algorithm to identify the snapshot images of each point cloud cluster and obtain the device categories and the number of devices included in the cluster; A retrieval module, connected to the identification module, is used to retrieve corresponding different types of models from the model library according to the device categories, register each type of model into the target device point cloud using the iterative closest point algorithm, and determine the device type by matching the distances of the points; An application module, connected to the retrieval module, obtains the pose of the device model in the substation scene from the calculation results of the iterative closest point algorithm to place the model into the actual substation scene.

4. A computer device, characterized in that, The computer device includes a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor is caused to execute the steps of the substation simulation modeling method according to any one of claims 1-2.

5. A computer-readable storage medium storing a computer program, which when executed by a processor, causes the processor to execute the steps of the substation simulation modeling method according to any one of claims 1-2.

Images