Optimizing Parameter Extraction in Grid Information Models Based on Improved Convolutional Neural Networks

Abstract

With the rapid advancement of digital technology, three-dimensional designs of Grid Information Models (GIMs) are increasingly applied in the power industry. Addressing the challenges of extracting key parameters during the GIM’s process for power grid equipment, this paper explores an innovative approach that integrates artificial intelligence with image recognition technologies into power design engineering. The traditional methods of “template matching, feature extraction and classification, and symbol recognition” have enabled the automated processing of electrical grid equipment engineering drawings, allowing for the extraction of key information related to grid equipment. However, these methods still rely on manually designed and selected feature regions, which limits their potential for achieving full automation. This study introduces an optimized algorithm that combines enhanced Convolutional Neural Networks (CNNs) with Depth-First Search (DFS) strategies, and is specifically designed for the automated extraction of crucial GIM parameters from power grid equipment. Implemented on the design schematics of power engineering projects, this algorithm utilizes an improved CNN to precisely identify component symbols on schematics, and continues to extract text data associated with these symbols. Utilizing a scene text detector, the text data are matched with corresponding component symbols. Finally, the DFS strategy is applied to identify connections between these component symbols in the diagram, thus facilitating the automatic extraction of key GIM parameters. Experimental results demonstrate that this optimized algorithm can accurately identify basic GIM parameters, providing technical support for the automated extraction of parameters using the GIM. This study’s recognition accuracy is 91.31%, while a traditional CNN achieves 71.23% and a Faster R-CNN achieves 89.59%. Compared to existing research, the main innovation of this paper lies in the application of the combined enhanced CNN and DFS strategies for the extraction of GIM parameters in the power industry. This method not only improves the accuracy of parameter extraction but also significantly enhances processing speed, enabling the rapid and effective identification and extraction of critical information in complex power design environments. Moreover, the automated process reduces manual intervention, offering a novel solution in the field of power design. These features make this research broadly applicable and of significant practical value in the construction and maintenance of smart grids.

1. Introduction

Advancements in neural networks and other cutting-edge technologies have enabled the widespread application of image recognition in various real-world scenarios [1]. Text recognition, facial recognition, smart healthcare, non-destructive testing, and agricultural product classification are just a few examples [2,3]. In this context, the power industry has also emerged as a crucial domain for the application of image recognition technology, offering innovative solutions in various aspects. By leveraging image recognition techniques, we can monitor the real-time status of power equipment and swiftly identify any faults [4]. Additionally, image recognition technology exhibits tremendous potential in the recognition and extraction of design drawings and information within electric power systems.

Information technology’s rapid development has greatly propelled the level of intelligence in the power grid while also fueling the growing demand for digitization in grid engineering. Harnessing advanced digital technologies such as computers and communications is crucial for making traditional power systems smarter and more efficient [5]. The integration of new technologies can also enhance the competitiveness of industry products [6]. In terms of operations and maintenance, countries like the United States and Canada have already embraced smart meter systems to analyze users’ power consumption patterns and detect energy theft incidents [7]. Some substations in China have started to combine GIS and digital twin technology to simulate and model equipment in order to monitor the operational status of equipment in real time [8]. Additionally, the application of 3D modeling technology in power systems, as opposed to traditional 2D drawings, has effectively eliminated inefficiencies and the high occurrence of human errors [9]. Since the release of the Digitalized Design Application Guidelines for Substations and Power Transformation Projects by the State Grid Corporation of China in 2018, the Grid Information Model (GIM) has gained widespread attention as a digital model that integrates all aspects of power grid engineering, including grid elements and geographic information. The GIM, with its visual, standardized, and information-based characteristics, not only facilitates effective information sharing in grid projects but also enhances the synergy of information throughout the entire project lifecycle, thus improving engineering management capabilities. The GIM can describe the three-dimensional information and engineering attributes of substations, converter stations, transmission lines, and cable projects through one or more files. Moreover, it plays a crucial role in deepening the integration of “information flow and business flow” within grid projects [10,11,12].

The GIM relies primarily on key technologies such as 3D modeling, model coding, and databases, with 3D modeling being a crucial step in mapping grid information to the digital model. Currently, the 3D modeling of grid engineering projects is mainly based on 2D drawings. Designers use software like Computer-Aided Design (CAD) to create 2D design drawings, which are then manually inputted into 3D modeling software during the digital delivery phase, including parameters such as equipment information and topological structure. This process is time-consuming and inefficient, requiring significant manpower. While there are some plugins available on the market that automate the conversion from 2D to 3D, they have certain limitations, often require substantial investment, and are often unable to accurately identify components in non-standard drawings [13].

Traditional image-based recognition techniques for power grid equipment design drawings primarily employ methods such as template matching, feature extraction, and classification. These techniques aim to extract device parameters and identify interconnection relationships among devices [14,15,16,17].

Template matching is a fundamental technique in the field of image recognition, which involves comparing patterns in engineering design drawings with pre-defined pattern templates [14,15]. By identifying the maximum similarity between the two, this method enables the recognition of power grid equipment in engineering drawings. This approach is favored for its intuitiveness and simplicity. However, it relies heavily on the image quality of the drawings and the accuracy of the pre-defined template library.

Feature extraction and classification techniques are considered to be more complex pathways in the recognition of drawings. The process begins by utilizing image processing techniques to preprocess engineering drawings, accurately extracting key features of power grid equipment. Subsequently, machine learning algorithms such as clustering and decision trees are employed for classification to recognize different categories of power grid devices. Despite its remarkable accuracy and ability to handle complex scenarios, this technique relies on a large amount of training data and computational resources. These requirements may pose limitations in practical applications [16,17].

Symbol recognition involves using detection algorithms to identify symbols such as transformers, casings, and lightning arresters. Additionally, by combining Optical Character Recognition (OCR) technology, text information can be extracted. This process converts symbols and textual content in drawings into structured information for further processing and analysis [17]. Overall, although traditional image-based methods in power grid equipment engineering design drawings enable automated processing and extraction of key information, they still rely on manual design and the manual selection of feature areas, which limits the possibility of achieving complete automation. Furthermore, these traditional methods for recognizing power grid engineering design drawings also need improvements in accuracy and computational efficiency.

Therefore, the accurate and efficient identification and extraction of various types of electrical equipment, along with their corresponding parameter information and interconnection relationships from electrical grid engineering drawings, holds significant importance in the GIM 3D modeling process.

With the widespread application of computer technology and artificial intelligence, advanced image recognition techniques have become an important means for identifying and processing power grid engineering design drawings. This technology allows for the extraction of key parameters, such as equipment information, connection relationships, and operational parameters of the power GIM with higher accuracy and efficiency. This significantly improves the design efficiency and operational performance of power systems [18,19,20]. Currently, commonly used solutions for image recognition include Convolutional Neural Networks (CNNs), Support Vector Machines (SVMs), K-Nearest Neighbor (KNN) classifiers, and Artificial Neural Networks (ANNs) [21]. Among them, CNNs are the most widely used solution in the field of image processing [22]. The Depth-First Search (DFS) algorithm can effectively identify the connection methods of components in engineering drawings. This paper proposes an optimized algorithm for the automatic extraction of key parameters in the power GIM based on an improved Convolutional Neural Network + Depth-First Search (CNN + DFS) approach. The algorithm employs a symbol detection algorithm based on CNN models to recognize electrical component symbols, followed by the identification of textual information within the design diagram and its association with respective component symbols. Finally, it utilizes DFS to identify the connection relationships among component symbols. This algorithm effectively enhances the efficiency and accuracy of grid modeling and is particularly suitable for application in wiring diagrams found in electrical schematics.

This paper analyzes the parameters of drawings in Section 2 and explains the importance of parameter extraction in GIM use. In Section 3, an optimization algorithm for the automatic extraction of key parameters in GIM, based on an improved CNN + DFS method, is introduced. The components of the algorithm are also detailed. Section 4 presents a comparative analysis using an example to demonstrate the performance of the improved algorithm against traditional algorithms. Finally, Section 5 summarizes the research findings and discusses the limitations of the study.

2. Analysis of Key Parameters of GIM

2.1. Key Parameters of GIM

The GIM is an information model for power systems and their equipment, aiming to comprehensively describe the information content and relationships of the grid, equipment, and operations. This model is primarily applied in the field of power transmission and transformation projects, encompassing a wide range of digital design elements, including equipment, structures, and facilities within the substations, as well as towers, conductors, and infrastructure along the transmission lines.

Therefore, crucial parameters in building the GIM for the grid include engineering information of substations, parameter information of equipment, the parameters of substation structures and facilities, and relevant parameters of towers, conductors, and foundations along the transmission lines. Taking substations as an example, their three-dimensional modeling involves not only geometric models but also physical models. Thus, the parameter encompasses not only the geometric information describing the shape, structure, and dimensions of physical objects but also the parameters of engineering equipment, materials, structures, and other facilities.

2.2. Representation of Key Parameters in GIM in Engineering Design Drawings

The acquisition of key parameters is crucial in the construction of the GIM for the power grid, and these parameters are primarily obtained from engineering design drawings. Typically, engineering design drawings provide detailed information about the electrical characteristics, connection methods, and installation positions of grid equipment. These pieces of information have a close correspondence with the parameters in the GIM. For example, information such as equipment models and rated voltages from the design drawings can directly correspond to the equipment types and voltage parameters in the GIM. Additionally, parameters regarding equipment installation positions from the design drawings can be directly applied to the equipment location parameters in the GIM, enabling an accurate representation of the layout of the grid equipment.

Therefore, the crucial parameters of the GIM can be found in the engineering design drawings. Applying artificial intelligence technology to automatically identify and extract the key parameters for GIM use from the design drawings, instead of relying on tedious manual data entry, is beneficial in reducing labor and time costs, improving the accuracy and consistency of parameter input, and subsequently enhancing the speed and quality of GIM establishment.

2.3. Characteristics of Representation in Power Grid Equipment Engineering Design Drawings

Taking the electrical main single-line diagram of a substation as an example, as shown in Figure 1, this diagram primarily includes equipment layout, electrical connections, equipment parameter tables, and other additional notes. Among them, the equipment layout is the main part of the drawing, describing the overall structural layout of the substation. This includes not only the power supply locations of electrical equipment, such as transformers and reactive power compensation devices, but also the routing of power transmission lines and the distribution and usage of customer equipment. The electrical connections between equipment characterize the interconnection relationships, providing detailed information on the starting and ending points of various power transmission lines, as well as the specifications and ratings of the lines. The equipment parameter table records the electrical parameters of each device, including rated voltage, rated power, protection level, insulation level, and other relevant information. Meanwhile, other additional notes are usually found in blank areas of the drawing and are provided in the form of annotations using text or symbols, offering supplementary explanations for special design requirements or equipment specifications. Figure 1 shows the electrical circuit of a converter station in a certain ultra-high-voltage direct current (UHVDC) transmission project. This circuit, starting from the 66 kV busbar, sequentially passes through a single grounding disconnector, a current transformer, a station service transformer, and a zinc oxide arrester, before being routed via a 10 kV cable to the 10 kV distribution room of the converter station.

3. Optimization Algorithm for Automatic Extraction of Key Parameters in GIM Based on Improved CNN + DFS

Given the limitations of traditional methods for extracting information from drawings, this paper aims to develop a new method that can accurately and efficiently extract key parameter information for the GIM in a fully automated manner. Taking into consideration the unique characteristics of design drawings, this paper combines a CNN with a Depth-First Search algorithm to propose an optimized algorithm for automatically extracting key parameters for the GIM. This method aims to improve the accuracy and efficiency of information extraction and overcome the shortcomings of traditional methods.

In this paper, we optimize an algorithm for automatically extracting key parameters for the GIM, based on a CNN, which consists of three main modules: component symbol detection, text recognition and association, and connection detection. Focusing on the processing of the electrical main single-line diagram of a substation in the grid engineering drawings, this technique first crops the substation’s electrical main single-line diagram to generate images with consistent dimensions. Next, an improved CNN model is used to recognize and extract information from the component symbols. Subsequently, the text in the drawing is detected using a neural network algorithm, allowing for the establishment of a scene text detector to ensure the correct association between the text and the corresponding component symbols. Finally, a Depth-First Search method is employed to identify the connections between the various component symbols in the engineering drawings.

3.1. Component Symbol Detection

The component symbol detection module employs a CNN model to automatically identify the component symbols in electrical elementary drawings of substation main wiring diagrams. These drawings consist of the symbol elements of various electrical equipment, such as transformers, reactors, capacitors, and distribution devices, along with their corresponding device name labels, which are combined with electrical wiring to create the substation’s main wiring diagram. Therefore, this paper focuses on achieving precise detection of these component symbols.

Firstly, the dataset undergoes symbol cropping. This dataset comprises two parts: the component symbol dataset and the composite symbol label dataset. We extract these from the collection of electrical main wiring drawings at the substation, marking all component symbols and their composite symbols in each electrical main wiring diagram. Given the significant differences between component symbols and composite symbols in the drawings, we choose an appropriate pixel window for individual marking and cropping of all electrical main wiring diagrams. Additionally, addressing situations where some drawings lack corresponding equipment text labels for component symbols or where their semantics cannot be clearly identified, we establish an “unsigned” cropping dataset to store cropped images in such special cases, as shown in Figure 2.

Building upon the compiled symbol-cropped dataset, this paper employs a CNN model for image classification to determine whether input images contain component symbols, composite symbols, or neither. Recognizing the complexity and diversity of graphical symbols, we propose an enhanced CNN architecture for symbol detection, as illustrated in Figure 3. In the figure, dashed boxes represent a part as a whole module, arrows represent the direction of tensor transmission, and asterisk represents the number of executions.

The equivalent calculation formula for the convolution operation is shown in Equation (1) [19]:

$$C^{\left(k\right)}=B^{\left(k\right)}\ast A^{(k-1)}+H^{\left(k\right)}$$

(1)

In the equation, $A^{(k-1)}$ represents the input image of the k-th layer, $B^{\left(k\right)}$ is the weight matrix of the convolutional kernel, $H^{\left(k\right)}$ is the bias vector of the k-th layer, and $C^{\left(k\right)}$ is the output image after the convolution operation of the k-th layer [19].

In the enhanced CNN architecture proposed in this paper for symbol detection, the overall framework consists of four parts: the input end, Backbone network, Neck, and prediction sections. The input terminal receives images as input for the algorithm. The Backbone is typically a pre-trained Convolutional Neural Network with strong capabilities in extracting image features. The Neck further processes or merges the features extracted by the Backbone, with common structures such as FPN, PAN, ASPP, etc. The prediction module is used for object detection, usually comprising classification and regression branches to predict the class and position of the target, and box objects. These four components work together to form the basic architecture of image object detection algorithms, enabling the algorithm to automatically extract target location and class information from images. Specifically, in the input end of the network, considering the limited number of training samples, the Mosaic data augmentation method is employed. This technique randomly selects several images from the training set and recombines them using methods such as scaling, cropping, and rearrangement, effectively enhancing the network’s robustness. Additionally, during the network training process, the auto-anchor mechanism is utilized to automatically determine the optimal anchor boxes for the current training set, accelerating the convergence speed of the network.

In the Backbone section, the architecture is constructed using slice convolution and Cross Stage Partial (CSP) modules. Slice convolution processes images through channel shuffling operations, integrating odd rows/odd columns, odd rows/even columns, even rows/even columns, and even rows/odd columns before conducting convolution operations, aiming to preserve the original data characteristics of graphical symbols to the maximum extent. The CSP module first divides the underlying feature maps into two parts and then effectively merges them using skip connections, thereby significantly reducing the computational complexity of the network while maintaining the accuracy of symbol detection.

In the Neck section, the network structure integrates the advantages of Feature Pyramid Networks (FPNs) and Path Aggregation Network (PANs). The FPN structure propagates high-level feature information in a top-down manner, transmitting strong semantic features through upsampling, and then integrates with the PAN structure, which can transmit strong localization features in a bottom-up manner. Through the combination of these two structures, feature fusion is performed on different feature layers of the Backbone and different detection layers, enhancing the network’s ability to extract features related to graphical symbols.

In the prediction section, the bounding box loss function adopts the Generalized Intersection over Union Loss (GIoU_Loss) function, which effectively enhances the model’s performance without increasing costs, as shown in Equation (2) [23]:

$$L_{GIoU}=1-IoU+\frac{\mid C\backslash (A\cup B)\mid }{\mid C\mid }$$

(2)

In this equation, A represents the area of the ground truth box; B represents the area of the predicted box; C represents the area of the minimum rectangle containing both A and B; and IoU is the Intersection over U nion of A and B [23].

The improved model structure and its parameter settings are shown in the

Table 1. Network structure and parameter setting of model.

Index	Input Layer	Number of Modules	Number of Parameters	Module Name	Parameter Settings
0	−1	1	3520	Conv	[3, 32, 6, 2, 2]
1	−1	1	20,736	Rep_Block	[32, 64, 3, 2]
2	−1	1	18,816	C3	[64, 64, 1]
3	−1	1	82,432	Rep_Block	[64, 128, 3, 2]
4	−1	1	74,560	MT_Block	[128, 128, 1]
5	−1	1	328,704	Rep_Block	[128, 256, 3, 2]
6	−1	1	296,704	MT_Block	[256, 256, 1]
7	−1	1	1,312,768	Rep_Block	[256, 512, 3, 2]
8	−1	1	1,182,976	MT_Block	[512, 512, 1]
9	−1	1	656,896	SPPF	[512, 512, 5]
10	−1	1	131,584	Conv	[512, 256, 1, 1]
11	−1	1	0	Upsampling	[None, 2, ‘nearest’]
12	[1, 6]	1	0	Concat	[1]
13	−1	1	391,984	C3	[512, 256, 1, False]
14	−1	1	33,024	Conv	[256, 128, 1, 1]
15	−1	1	0	Upsampling	[None, 2, ‘nearest’]
16	[1, 4]	1	0	Concat	[1]
17	−1	1	90,880	C3	[256, 128, 1, False]
18	−1	1	6448	CA_Block	[128, 128, 8]
19	−1	1	147,712	Conv	[128, 128, 3, 2]
20	[1, 14]	1	0	Concat	[1]
21	−1	1	296,448	C3	[256, 256, 1, False]
22	−1	1	12,848	CA_Block	[256, 256, 16]
23	−1	1	590,336	Conv	[256, 256, 3, 2]
24	[1, 10]	1	0	Concat	[1]
25	−1	1	1,182,720	C3	[512, 512, 1, False]
26	−1	1	25,648	CA_Block	[512, 512, 32]
27	[18, 22, 26]	1	24,273	Detect	[4,[[…], […], […]], […]]

.

The symbol cropping dataset was randomly divided into a training set (60%), a validation set (20%), and a test set (20%) for training the network model. The training results are shown in Figure 4. It can be observed that, in the initial stages of training, the model loss exhibits significant variation, indicating that the model is unable to effectively recognize unseen new data. This is because the model is just starting to learn and does not yet have a thorough understanding of the data. This phenomenon is common in most machine learning models, especially when the training dataset is complex or has many features. However, as the number of training epochs increases, the loss variation decreases significantly, and the tendency of the model to overfit is reduced. This indicates that the model continuously learns and adapts to the training data, gradually reducing the prediction error. The continuous decline in training loss is expected, as the model’s performance on the training set is continuously optimized. Therefore, by comparing the loss changes between the training set and the validation set, it can be observed that using the enhanced CNN model proposed in this paper for symbol detection makes it easier for the model to learn effective features, thereby improving the model’s performance. The decline in training loss indicates that the model is better at fitting the training data, while the decline in validation loss shows that the model possesses good generalization ability. The combination of both indicates that the improved Convolutional Neural Network model can more effectively extract features for symbol detection, enhancing the overall performance of the model.

It is worth noting that, to effectively use the trained CNN model for component symbol detection in new drawings, three prediction feature layers of different scales are designed in this study to adapt to various sizes of electrical components. Considering the high sparsity of most drawings, the module automatically filters out over 90% of empty windows and symbols lacking clear semantic representation, classifying them into the “unsigned” category. Finally, a non-maximum suppression algorithm is employed to merge the detection results from multiple prediction feature layers, ensuring that the output layer presents only unique and accurate detected symbols. This effectively avoids the issue of repetitive recognition of the same symbol in different feature layers.

In the end, the component symbol detection module creates thresholds for the predicted probabilities based on preset thresholds and categorizes them into discrete symbols. Image windows with prediction probabilities below all thresholds are defined as non-symbols. Figure 5 and

Table 2. Symbol detection results after inputting the drawing.

Symbol Number	Symbol Name
(1)	Wind farm output circuit
(34)	Reactor circuit
(5)	transformer
(24)	Lightning arrester
(31)	CT
(42)	CT
(59)	CT

shows the symbol detection results after inputting a specific drawing segment. It can be observed that all preset symbols are accurately classified, and composite symbols are also accurately recognized.

3.2. Text Recognition and Association

In the electrical primary wiring diagrams of substations, textual elements play a crucial role. They identify and describe various electrical equipment types, voltage levels, and design annotations within the drawings. These texts typically include the names of devices and the parameters of the devices in parameter tables placed next to the electrical equipment. Together, this information clearly defines the engineering and parameter details of the electrical equipment. Therefore, the accurate identification and parsing of these texts, along with the correct association with recognized component symbols on the drawings, are essential for the automatic extraction of key parameters in the GIM. The lack of this step would significantly reduce the extraction effectiveness.

Therefore, to ensure the effective identification of text in power grid engineering drawings and its accurate association with corresponding component symbols, this paper introduces a specially designed scene text detector based on a neural network algorithm in the optimization of key parameter automatic extraction for the GIM. The detector utilizes a CNN to generate bounding boxes indicating the positions of text in the image. It then identifies relevant text based on the proximity threshold between component symbols and text bounding boxes. Finally, Tesseract OCR is employed for text recognition. For text boxes in different orientations, a detection enhancement method is applied to extract text from various directions. The accuracy of the text content is ensured by comparing multiple extraction results of the same text box. Subsequently, these texts are associated with the corresponding symbols, achieving a structured representation of symbols and text information, as illustrated in Figure 6 and

Table 3. Symbol recognition results after inputting the drawing.

Text Number	Recognized Text Content	Text Number	Recognized Text Content	Text Number	Recognized Text Content
(T-169)	[Jin Guan ] Metal oxide arrester	(T-170)	WA.ST1	(T-171)	YH10W-204/532
(T-172)	F2	(T-173)	2× (NAHLGJQ-1440/120)	(T-174)	[Heng Bian] Three-phase integrated on-site assembly transformer
(T-239)	1000/1000/334 MVA	(T-240)	525/230 ± 2 × 2.5%/66 kV	(T-241)	YN, a0, d11, ODAF
(T-242)	U1-3 = 67%, U1-2 = 20%	(T-243)	F1	(T-244)	U2-3 = 40%
(T-245)	High voltage bushing CT:5P20/0.2/0.5	(T-246)	5P20	(T-247)	0.2
(T-248)	0.5	(T-249)	2000/1 A 15 VA	(T-250)	Medium voltage bushing CT: 5P20/0.2
(T-251)	4000/1 A 15 VA	(T-252)	Low voltage bushing CT: TPY/TPY/5P20/0.2	(T-253)	5P20
(T-254)	0.2	(T-255)	0.2	(T-256)	5P20
(T-257)	TPY	(T-258)	TPY	(T-259)	0.2S
(T-260)	5P20	(T-261)	TPY	(T-262)	TPY
(T-263)	F1	(T-264)	Q22	(T-265)	Q11
(T-266)	Q21	(T-267)	T2	(T-268)	Q1
(T-269)	T1	(T-270)	WC1.W01	(T-375)	4000/1A 15VA
(T-376)	72.5 kV, 5000 A	(T-377)	2000/4000/1 A 15 VA	(T-378)	No.1 main transformer
(T-379)	[Jin Guan] Metal oxide arrester	(T-380)	YH20W-420/1046 kV	(T-381)	2× (NAHLGJQ-1440/120)
(T-382)	2× (NAHLGJQ-1440/120)	(T-383)	[Jin Guan] Zinc oxide lightning arrester YH5W-96/250	(T-384)	[Ping Gao] Isolation switch (double grounding) GW4-72.5DDW
(T-385)	72.5 kV, 5000 A, 50 kA (3 s), 125 kA (Peak)	(T-386)	Bushing current transformer	(T-387)	5P30/0.2S/0.2
(T-388)	2000-4000/1A/1A/1A	(T-389)	15VA/5VA/15VA	(T-390)	Tank-type circuit breaker LW24-72.5
(T-391)	72.5 kV, 4000 A,40 kA (3 s), 100 kA (Peak)	(T-392)	Bushing current transformer	(T-393)	TPY/TPY/5P30
(T-394)	2000-4000/1 A/1 A/1 A	(T-395)	15 VA/15 VA/15 VA	(T-396)	Heat-resistant aluminum alloy conductor 2× (NAHLGJQ-1440/120)

and

Table 4. Symbol recognition and text association results after inputting drawings.

Symbol Number	Symbol Name	Associated Text	Symbol Number	Symbol Name	Associated Text
(23)	NoSymbol	None	(5)	Transformer	[T-174]
(16)	GRD	None	(31)	Switch	[T-268, T-390, T-391]
(11)	Disconnector	[T-264, T-265, T-266, T-384, T-385]	(31)	CT	[T-250, T-251, T-253, T-254]
(14)	Disconnector	None	(42)	CT	[T-245, T-246,T-247, T-248, T-249]
(17)	Arrester	[T-169, T-172]	(43)	CT	[T-252,T-255, T-256, T-257, T-258, T-375]
(24)	Arrester	[T-243, T-397]	(44)	CT	[T-267, T-386, T-387, T-388, T-389]
(25)	Arrester	[T-383, T-263]	(45)	CT	[T-269, T-392, T-393, T-394, T-395]
(39)	Arrester	[T-379]	(59)	CT	[T-259, T-260, T-261, T-262, T-376, T-377]

.

3.3. Connection Detection

In the field of power grid engineering, the interconnection of various grid devices is primarily achieved through transmission lines such as cables and conductors. In engineering drawings, the physical connections between devices are represented by solid line segments. The connection detection module, building upon the symbol detection step, further identifies which symbols are interconnected by these lines. This allows for the determination of the spatial and connection relationships between power grid devices. This process is crucial for reconstructing the mutual relationships between devices and establishing the hierarchical structure of power grid assets.

The optimization extraction method proposed in this paper achieves symbol connection detection through a graph search algorithm. The thresholded images of device component symbols are treated as graphical representations, with each pixel in the image corresponding to a node in the graph. Each node contains information about its color and whether it belongs to a device component symbol based on its thresholded pixel intensity. The connections between adjacent pixels form the edges of the graph, with each node having at most eight edges. Within this framework, device component symbols are represented in the graph by a group of nodes corresponding to the pixels forming the symbol.

Utilizing the graphical representation described above, the connection detection module employs a DFS strategy to identify connections between device component symbols. For each detected symbol, a DFS process is initialized from a node within the symbol. This process traverses along the black nodes in the graph while simultaneously recording and tracking encountered connections between symbols along its path. The search automatically terminates once all valid paths have been explored.

Figure 7 illustrates the results of connection detection on image segments, where the source symbols are highlighted in red, and the connected symbols are highlighted in green. From the figure, it can be observed that, apart from the labeled “unsigned” clipped set, all detected symbols are determined to be connected to the source symbol.

4. Example Analysis

In order to verify the effectiveness of the proposed method, this study used 100 electrical main wiring diagrams of a substation with voltage levels of 220 kV and above. The proposed optimized algorithm based on the improved CNN + DFS method for key parameter automatic extraction of a power grid GIM was used for recognition and key element extraction of the diagrams, and the effectiveness of the method was evaluated. The diagrams were divided into training and test sets in a ratio of 4:1. After the experiment, the confusion matrix, F1-score curve, and Mean Average Precision (mAP) curve were drawn to analyze the prediction ability and results of the model from multiple aspects.

The confusion matrix in Figure 8 presents the confusion matrix of the recognition system, reflecting the recognition performance and confusion degree of different symbol categories. The recognition of unsigned and component symbols is less affected by interference and is not easily confused with the background or other categories. This indicates that the features of these two types of symbols are more distinct and less affected by background interference. Therefore, we can conclude that the feature extraction and classification algorithms for unsigned and component symbols are highly reliable in practical applications. On the other hand, composite symbols easily interfere with the background, with a probability of 22% of being misclassified as background. This higher misclassification rate suggests that the features of composite symbols are more easily confused with background features, leading to lower recognition accuracy. This may be due to the complex structure of composite symbols and the diverse background environments.

To effectively quantify the validity of the recognition results, this paper establishes the following evaluation criteria: a detected symbol is considered correctly identified if its category matches the ground truth label and if the bounding box of the detected symbol has at least a 0.5 Intersection over Union (IOU) with the bounding box of the ground truth symbol. The mAP@0.5 metric measures the average precision of the model at an Intersection over Union (IoU) threshold of 0.5, that is, the accuracy when the bounding boxes detected by the model overlap by 50% with the true bounding boxes. Generally, mAP@0.5 is one of the commonly used evaluation metrics in object detection tasks, utilized to gauge the precision and accuracy of the model. The results indicate that the validation set, comprising 20 sets of diagrams with 400 different categories of symbols, achieves an mAP@0.5 of 0.832 when using a classification probability threshold of 0.5 for symbol recognition, as illustrated in Figure 9. Although the improved CNN performs well on the test set, the performance of the modified model on other types of drawings has not yet been tested, and further experiments are needed to determine this.

To measure the robustness of the model, we drew the F1-score curve based on the harmonic mean of precision and recall. As shown in Figure 10, the F1 curves for unsigned, combination symbols, and element symbols all performed well and were not significantly affected by the imbalance of the circuit diagram dataset. This indicates that the model has good robustness and overall performance.

Additionally, from the above figure, it is evident that the performance of unsigned symbol recognition is significantly better than that of component symbol recognition. This discrepancy arises due to the fewer categories and samples for unsigned symbols, as well as the relatively complex composition and representation of component symbols, making them prone to confusion with other symbols. Comparisons between the proposed algorithm and traditional CNN models, as well as the Faster Regions with CNN features (R-CNN) model, are presented in

Table 5. Accuracy of symbol recognition in power grid engineering drawings under different algorithms.

Drawing Component Symbols	Number of Samples	Accuracy of Drawing Symbol Recognition under Different Algorithms
		Traditional CNN	Faster R-CNN	Optimized Algorithm
1	107	66.36	85.98	90.65
2	136	72.06	88.24	91.18
2	143	73.43	90.91	91.61
2	263	69.20	88.97	90.87
2	274	70.80	91.61	91.56
4	304	68.75	86.84	90.79
5	102	74.51	91.18	92.16
3	294	75.17	91.84	91.80
all	1623	71.23	89.59	91.31

. In this study, the traditional Convolutional Neural Network (CNN) model has a parameter count of 229.00 M and a computational cost of 220.0 G. The Faster R-CNN model has a parameter count of 315.00 M and a computational cost of 267.0 G. In contrast, our proposed CNN + DFS model significantly reduces both parameters and computational requirements, with a parameter count of 138.1 M and a computational cost of 89.3 G. The results indicate that, out of a total of 1623 component symbol samples tested, the drawing symbol recognition accuracy using the proposed improved CNN is the highest, achieving an overall accuracy of 91.31%. This surpasses the accuracy of traditional CNN models (71.23%) and the Faster R-CNN model (89.59%), demonstrating the effectiveness of the proposed improved CNN architecture for symbol detection.

In this study, a comparison was conducted among three algorithms. The results indicate that the traditional CNN algorithm yielded lower accuracy in the recognition of all types of component symbols when compared to the other two algorithms, demonstrating inferior performance. In contrast, the proposed optimization algorithm exhibited better performance in the recognition of the majority of component symbols when compared to the Faster R-CNN algorithm. Analysis of the F1-score curve leads to the conclusion that the improved algorithm demonstrates good robustness and overall performance.

5. Conclusions

To address the challenge of automatically extracting key parameters in the GIM, this paper proposes an optimization algorithm based on artificial intelligence image recognition technology, utilizing a CNN model. The proposed approach, which is based on an improved CNN + DFS framework, innovatively enhances traditional CNN methods by incorporating the Mosaic data augmentation technique and the GIoU_Loss loss function. These enhancements aim to improve the effectiveness, robustness, and computational speed of the neural network model. Simulation results demonstrate that compared to traditional CNN and Faster R-CNN algorithms, the proposed optimization algorithm achieves higher accuracy in character recognition and detection.

Compared to traditional manual input and simple image recognition techniques, this algorithm can automatically process design drawings, reducing labor costs and time, while reducing error rates. Moreover, the optimization algorithm introduced in this paper, based on the DFS strategy, enables the detection of connections between device symbols in power grid engineering drawings. This capability facilitates the intelligent recognition of power grid topological architecture in engineering drawings. In summary, the proposed optimization algorithm accurately identifies device information and topological structure information, among other key parameters, from engineering design drawings of power grid devices, thereby providing a reference for the automatic extraction of parameters in GIM modeling.

In addition to its application in GIM use, this key parameter extraction method has various other wide-ranging applications. In scenarios involving blueprint recognition and parameter extraction, as long as the blueprint category is ensured to be wiring diagrams and the extracted parameters are formatted to meet the actual requirements, it can be extended to other scenarios. However, it is important to note that the current algorithm focuses on the automatic extraction of key parameters in the GIM from power grid engineering design drawings and has not yet achieved automatic entry of detected and recognized data into the GIM’s process. This limitation arises because some specialized software does not provide data interfaces to facilitate automated workflows, which introduces significant difficulties. This aspect will be explored in future research endeavors.