Multi-modal data power grid line planning method
Through deep learning algorithms, the problem of insufficient data processing limitations and flexibility in power grid line planning is solved, and more efficient and accurate line planning is achieved.
Patent Information
- Application Number
- CN202411053377.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-02
AI Technical Summary
The existing grid line planning methods have shortcomings in data processing limitations and flexibility, making it difficult to effectively integrate and analyze information from different data sources, limiting the accuracy and efficiency of the planning.
The power grid line planning method adopts multimodal data, integrates and preprocesses data from satellites, drones and ground sensors through deep learning algorithms, and combines geographic information systems to design data processing, feature selection, attention mechanism and semantic segmentation. The model of the prediction module is trained using innovative loss functions to generate an optimized line planning scheme.
It improves the accuracy and efficiency of power grid line planning, can make more comprehensive use of the relationship between complex data, reduce labor costs, and provide more flexible and efficient planning solutions.
Smart Images

Figure CN119026287A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power grid line planning, and in particular to a power grid line planning method based on multi-modal data. Background Art
[0002] Grid line planning is a complex problem in power system engineering, which needs to consider geographical, environmental and technical factors. With the development of remote sensing technology and automated tools, such as satellite images and drone aerial photography, a large amount of high-resolution data is provided for grid line planning. These data potentially contain key geographical and environmental information required for planning. Grid planning methods can extract effective information from massive and complex data and provide more scientific and economical line planning solutions.
[0003] Deep learning technology, especially convolutional neural network (CNN) and semantic segmentation algorithms, has shown extremely high efficiency and accuracy in the field of image recognition and processing. By applying these technologies, key signs such as terrain features in images can be automatically identified, providing accurate data support for line planning. The technical basis behind the method of power grid route planning mainly includes advanced architectures and algorithms for semantic segmentation in deep learning.
[0004] The Chinese patent with announcement number CN110532327B discloses "an artificial intelligence power grid line planning method", which uses historical power grid data to train artificial intelligence models, integrates multi-source data through artificial intelligence models to generate optimized design solutions, and emphasizes improving planning speed through automated processing. However, there are deficiencies in deep feature extraction and advanced analysis capabilities for complex data, which limits the accuracy and efficiency of planning.
[0005] Therefore, the existing power grid planning methods have deficiencies in data processing limitations and flexibility, which greatly limits efficiency and adaptability. Traditional methods are usually difficult to effectively integrate and analyze information from different data sources. This difficulty in data integration leads to limitations in analytical capabilities and makes it difficult to fully utilize the relationships between complex data, which in turn affects the accuracy and comprehensiveness of planning. Summary of the invention
[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a power grid line planning method based on multimodal data to solve the problems raised in the above-mentioned background technology. The present invention integrates and preprocesses information from multiple data sources through a deep learning algorithm, and combines it with a geographic information system to perform path planning for power grid lines.
[0007] In order to achieve the above object, the present invention is implemented by the following technical solution: a multi-modal data power grid line planning method, comprising the following steps:
[0008] Step 1: Integrate and filter data sources of various modalities, including satellite shaking images, high-resolution views taken by drones, and meteorological data from ground sensors;
[0009] Step 2: Process the collected multimodal data to obtain data information that can be applied to the model;
[0010] Step 3: Send the processed data to the route planning model for training, and design a loss function to train the model;
[0011] Step 4: Use the trained model to plan routes. The route planning displays the planned routes on a map.
[0012] Furthermore, the model in step 2 includes four modules, a data processing module, a feature selection module, an attention mechanism module and a semantic segmentation prediction module, and step 2 includes the following sub-steps:
[0013] (1.1) Perform preliminary integration of the collected multimodal data and standardize the input data through the data processing module;
[0014] (1.2) Processing and optimizing the features of the image based on the collected data through the feature extraction module;
[0015] (1.3) The extracted information is passed through the attention mechanism to enhance feature selection and capture key information;
[0016] (1.4) Design and optimize the loss function;
[0017] (1.5) The processed data is sent to the semantic segmentation module to predict and analyze the optimal path of the power grid line.
[0018] Furthermore, the process of standardizing input data in the data processing module includes the following sub-steps:
[0019] (1.2.1) Identify missing values in the data and decide whether to fill or delete them. The missing original values are replaced with the mean value. The Z-score method is used to identify outliers. The formula is:
[0020]
[0021] Where Z represents the standardized score of the observation, x represents the observation, and x and μ represent the mean and standard deviation of all observations in the data, respectively;
[0022] (1.2.2) The data is converted into a distribution with zero mean and unit variance, and the range is adjusted to 0 to 1. The normalization formula is:
[0023]
[0024] Among them, x norm represents the normalized value of the observation value, x represents the observation value, max(x) and min(x) represent the maximum and minimum values of the observation value of u respectively;
[0025] (1.2.3) Merge the standardized and normalized datasets into a single data frame, ensuring that all data are aligned in terms of timestamps and spatial locations.
[0026] Furthermore, the feature selection module captures effective features of the data through an encoder and a decoder, including the following sub-steps:
[0027] (1.3.1) Input the processed data tensor and use multiple 3×3 convolution kernels for feature extraction. Each convolution kernel slides on the input feature map with a stride of 1 to perform the convolution operation.
[0028] (1.3.2) Apply the ReLU function to the convolution output feature map. The ReLU function sets the activation value of negative numbers to 0, and the positive numbers remain unchanged, introducing nonlinearity;
[0029] (1.3.3) Use a 2×2 pooling window to downsample the feature map. The window moves on the feature map with a stride of 2, retaining only the maximum value of each window and reducing the spatial dimension of each feature value.
[0030] Furthermore, the steps of the attention mechanism module operating the feature map are as follows:
[0031] (1.4.1) Input the encoder’s feature map into multiple neural network layers and perform a softmax operation to ensure that the sum of all weights is 1;
[0032] (1.4.2) Weight the feature representation of the encoder module and its corresponding attention weight feature, and merge all weighted features. The formula is as follows:
[0033] z=∑α i h i
[0034] Where z is the weighted feature combination, α i is the characteristic h i The attention weight of
[0035] (1.4.3) All weighted features are merged and fused to obtain the final fused features.
[0036] Furthermore, the segmentation prediction module is used to accurately identify the categories of different regions and obtain effective path planning, which specifically includes the following sub-steps:
[0037] (1.5.1) Perform final preprocessing on the input fusion features and ensure that the data format is suitable for the model through standardization;
[0038] (1.5.2) Input the processed image data and annotated data into the model for training, and train the model to learn and identify the features of different regions;
[0039] (1.5.3) Use the trained model to perform semantic segmentation on the image of the entire planning area and generate a detailed feature map.
[0040] Furthermore, the loss function used in the training process in step 2 consists of three parts, including the loss function L formula of the image feature extraction module, the loss function L of the attention mechanism module attetion Formula, is the loss function L of the semantic segmentation prediction module segmentaiton formula.
[0041] Furthermore, the loss function L of the image feature extraction module is as follows:
[0042] L=αL contnt +βL structure
[0043]
[0044] Where L content Represents content loss, which measures the difference between the reconstructed image and the original image at the pixel level; L structure represents the structural loss, which measures the preservation of features at a higher level structure; x i and represents the i-th pixel value in the original image and the reconstructed image; N represents the pixel overview in the image; F l (x) and Represent the feature maps of the original image and the reconstructed image at layer l respectively; M l Represents the total number of feature map elements in layer l.
[0045] Furthermore, the loss function L of the attention mechanism module is attetion The formula is as follows:
[0046] L attention =L target +λL entropy
[0047]
[0048] Where L target represents the target loss, which measures the difference between the attention-weighted features and the specific task labels; L entropyrepresents the attention entropy loss, which distracts the model’s attention; λ is a hyperparameter; y o,c represents the indicator value of the true label of the oth sample in category c; p o,c represents the probability that the model predicts the oth sample as category c; a i represents the attention weight of the i-th feature; N is the number of all attention weights.
[0049] Furthermore, the loss function L of the semantic segmentation prediction module is segmentation The formula is as follows:
[0050] L segmentation =L entropy +γL dice
[0051]
[0052] Where L entropy represents the cross entropy loss, indicating the accuracy of pixel-level classification; L dice represents the Dice loss, which calculates the predicted and true coefficients; λ represents the trade-off convention, which is used to balance the cross entropy loss and the Dice loss; H, W, C represent the height, width and number of categories of the image respectively, y ijc represents the indicator value of the true label of the pixel located in the i-th row and j-th column in category c; p ijc represents the probability that the model predicts that the pixel located in the i-th row and j-th column is of category c; y ij represents the true label of the pixel located at the i-th row and j-th column; Represents the corresponding binary label predicted by the model.
[0053] Beneficial effects of the present invention:
[0054] 1. This multimodal data power grid line planning method can extract richer and more discriminative features from multimodal data through deep learning and attention mechanism, and obtain more data information.
[0055] 2. This multimodal data grid line planning method uses an innovative loss function and an optimized neural network architecture to provide higher accuracy in identification and planning.
[0056] 3. This multimodal data power grid line planning method can reduce labor costs through data processing and feature learning, and provide a more efficient and flexible solution for line planning. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 A technical solution flow chart of a multi-modal data power grid line planning method of the present invention;
[0058] Figure 2 A schematic diagram of a module for multimodal data power grid line planning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.
[0060] See also Figure 1 to Figure 2 The present invention provides the following technical solution: a multi-modal data power grid line planning method. Specifically comprising the following steps:
[0061] (1) Process the collected multimodal data to obtain data information that can be applied to the model.
[0062] The data is fed into the route planning model for training, and the loss function is designed to train the model.
[0063] (1.1) Perform preliminary integration of the collected multimodal data and standardize the input data through the data processing module.
[0064] (1.2) The collected data is processed and optimized based on the feature extraction module.
[0065] (1.3) The extracted information is passed through the attention mechanism to enhance feature selection and capture key information.
[0066] (1.4) Design and optimize loss function
[0067] (1.5) The processed data is sent to the semantic segmentation module to predict and analyze the optimal path of the power grid line.
[0068] The data processing module standardizes the input data, including the following sub-steps:
[0069] (1.2.1) Identify missing values in the data and decide whether to fill or delete them. The missing original values are replaced with the mean value. The Z-score method is used to identify outliers. The formula is:
[0070]
[0071] Where Z represents the standardized score of the observation, x represents the observation, and x and μ represent the mean and standard deviation of all observations in the data, respectively.
[0072] (1.2.2) The data is converted into a distribution with zero mean and unit variance, and the range is adjusted to 0 to 1. The normalization formula is:
[0073]
[0074] Among them, x norm represents the normalized value of the observation, x represents the observation, max(x) and min(x) represent the maximum and minimum values of the observation, respectively.
[0075] (1.2.3) Merge the standardized and normalized datasets into a single data frame, ensuring that all data are aligned in terms of timestamps and spatial locations.
[0076] The feature selection module captures the effective features of the data through the encoder and decoder. It includes the following sub-steps:
[0077] (1.3.1) Input the processed data tensor and use multiple 3×3 convolution kernels for feature extraction. Each convolution kernel slides on the input feature map with a stride of 1 to perform the convolution operation.
[0078] (1.3.2) Apply the ReLU function to the convolution output feature map. The ReLU function sets the activation value of negative numbers to 0, and the positive numbers remain unchanged, introducing nonlinearity;
[0079] (1.3.3) Use a 2×2 pooling window to downsample the feature map. The window moves on the feature map with a stride of 2, retaining only the maximum value of each window and reducing the spatial dimension of each feature value;
[0080] The attention module directly operates on the feature map. It includes the following sub-steps:
[0081] (1.4.1) The encoder’s feature map is input into multiple neural network layers, and a softmax operation is performed to ensure that the sum of all weights is 1.
[0082] (1.4.2) Weight the feature representation of the encoder module and its corresponding attention weight feature, and merge all weighted features. The formula is as follows:
[0083] z=Σα i h i
[0084] Where z is the weighted feature combination, α i is the characteristic h i The attention weight.
[0085] (1.4.3) All weighted features are merged and fused to obtain the final fused features.
[0086] Among them, the semantic segmentation prediction module accurately identifies the categories of different regions and obtains effective path planning.
[0087] (1.5.1) The input fusion features are finally preprocessed and standardized to ensure that the range of data formats is suitable for the model.
[0088] (1.5.2) The processed image data and annotated data are input into the model for training, and the training model learns and recognizes the features of different regions.
[0089] (1.5.3) Use the trained model to perform semantic segmentation on the image of the entire planning area and generate a detailed feature map
[0090] (2) Training the designed power grid line planning model;
[0091] Specifically, the loss function used in the training process consists of three parts:
[0092] (2.1) The first part is the image feature extraction module loss function L formula is as follows:
[0093] L=αL contnt +βL structure
[0094]
[0095]
[0096] Where L content Represents the content loss, which measures the difference between the reconstructed image and the original image at the pixel level. structure Represents the structural loss, which measures the preservation of features at higher levels of structure. i and represents the i-th pixel value in the original image and the reconstructed image. N represents the pixel overview in the image. l (x) and Represent the feature maps of the original image and the reconstructed image at layer l respectively. l Represents the total number of feature map elements in layer l.
[0097] (2.2) The second part is the loss function L of the attention mechanism module attetion The formula is as follows:
[0098] L attention =L target +λL entropy
[0099]
[0100] Where L target represents the target loss, which measures the difference between the attention-weighted features and the task-specific labels. entropyrepresents the attention entropy loss, which disperses the model's attention. λ is a hyperparameter. o,c Represents the indicator value of the true label of the oth sample in category c. o,c Indicates the probability that the model predicts that the oth sample is of category c. i represents the attention weight of the ith feature. N is the number of all attention weights.
[0101] (2.3) The third part is the loss function L of the semantic segmentation prediction module segmentation The formula is as follows:
[0102] L segmentation =L entropy +γL dice
[0103]
[0104] Where L entropy represents the cross entropy loss, which indicates the accuracy of pixel-level classification. dice represents the Dice loss, which calculates the coefficients of prediction and truth. λ represents the trade-off convention, which is used to balance the cross entropy loss and Dice loss. H, W, C represent the height, width and number of categories of the image respectively, y ijc represents the indicator value of the true label of the pixel located in the i-th row and j-th column in category c. ijc Represents the probability that the model predicts that the pixel located in the i-th row and j-th column is of category c. ij Denotes the true label of the pixel located at the i-th row and j-th column. Represents the corresponding binary label predicted by the model.
[0105] (3) Use the trained multimodal data to generate a predicted line planning path.
[0106] Specifically, by combining deep learning technology with data processing technology, data from multiple sources such as satellites, drones, and ground sensors can be effectively integrated. Through the line optimization model, data preprocessing and reasoning are performed on the planned area to generate a predicted line planning path. This method can not only improve the accuracy and response speed of power grid line planning, but also improve work efficiency and decision-making quality.
[0107] The basic principles and main features of the present invention and the advantages of the present invention are shown and described above. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention.
[0108] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A method for planning power grid lines based on multimodal data, characterized in that: The following steps are involved: Step 1: Integrate and filter data sources of various modalities, including satellite shaking images, high-resolution views taken by drones, and meteorological data from ground sensors; Step 2: Process the collected multimodal data to obtain data information that can be applied to the model; Step 3: Send the processed data to the route planning model for training, and design a loss function to train the model; Step 4: Use the trained model to plan routes. The route planning displays the planned routes on a map.
2. A method for planning a power grid line based on multimodal data according to claim 1, characterized in that: The model in step 2 includes four modules: data processing module, feature selection module, attention mechanism module and semantic segmentation prediction module, and step 2 includes the following sub-steps: (1.1) Perform preliminary integration of the collected multimodal data and standardize the input data through the data processing module; (1.2) Processing and optimizing the features of the image based on the collected data through the feature extraction module; (1.3) The extracted information is passed through the attention mechanism to enhance feature selection and capture key information; (1.4) Design and optimize the loss function; (1.5) The processed data is sent to the semantic segmentation module to predict and analyze the optimal path of the power grid line.
3. A method for planning a power grid line based on multimodal data according to claim 2, characterized in that: The process of standardizing input data in the data processing module includes the following sub-steps: (1.2.1) Identify missing values in the data and decide whether to fill or delete them. The missing original values are replaced with the mean value. The Z-score method is used to identify outliers. The formula is: Where Z represents the standardized score of the observation, x represents the observation, and x and μ represent the mean and standard deviation of all observations in the data, respectively; (1.2.2) The data is converted into a distribution with zero mean and unit variance, and the range is adjusted to 0 to 1. The normalization formula is: Among them, x norm represents the normalized value of the observation value, x represents the observation value, max(x) and min(x) represent the maximum and minimum values of the observation value of u respectively; (1.2.3) Merge the standardized and normalized datasets into a single data frame, ensuring that all data are aligned in terms of timestamps and spatial locations.
4. The method for planning a power grid line based on multimodal data according to claim 2, characterized in that: The feature selection module captures effective features of data through an encoder and a decoder, including the following sub-steps: (1.3.1) Input the processed data tensor and use multiple 3×3 convolution kernels for feature extraction. Each convolution kernel slides on the input feature map with a stride of 1 to perform the convolution operation. (1.3.2) Apply the ReLU function to the convolution output feature map. The ReLU function sets the activation value of negative numbers to 0, and the positive numbers remain unchanged, introducing nonlinearity; (1.3.3) Use a 2×2 pooling window to downsample the feature map. The window moves on the feature map with a stride of 2, retaining only the maximum value of each window and reducing the spatial dimension of each feature value.
5. The method for planning a power grid line based on multimodal data according to claim 2, characterized in that: The steps of the attention mechanism module operating the feature map are as follows: (1.4.1) Input the encoder’s feature map into multiple neural network layers and perform a softmax operation to ensure that the sum of all weights is 1; (1.4.2) Weight the feature representation of the encoder module and its corresponding attention weight feature, and merge all weighted features. The formula is as follows: z=∑α i h i Where z is the weighted feature combination, α i is the characteristic h i The attention weight of (1.4.3) All weighted features are merged and fused to obtain the final fused features.
6. A method for planning a power grid line based on multimodal data according to claim 2, characterized in that: The segmentation prediction module is used to accurately identify the categories of different regions and obtain effective path planning, which specifically includes the following sub-steps: (1.5.1) Perform final preprocessing on the input fusion features and ensure that the data format is suitable for the model through standardization; (1.5.2) Input the processed image data and annotated data into the model for training, and train the model to learn and identify the features of different regions; (1.5.3) Use the trained model to perform semantic segmentation on the image of the entire planning area and generate a detailed feature map.
7. The method for planning a power grid line based on multimodal data according to claim 1, characterized in that: The loss function used in the training process in step 2 consists of three parts, including the loss function L formula of the image feature extraction module and the loss function L of the attention mechanism module. attetion Formula, is the loss function L of the semantic segmentation prediction module segmentation formula.
8. A method for planning a power grid line based on multimodal data according to claim 7, characterized in that: The formula of the loss function L of the image feature extraction module is as follows: L=αL contnt +βL structure Where L content Represents content loss, which measures the difference between the reconstructed image and the original image at the pixel level; L structure represents the structural loss, which measures the preservation of features at a higher level structure; x i and represents the i-th pixel value in the original image and the reconstructed image; N represents the pixel overview in the image; F l (x) and Represent the feature maps of the original image and the reconstructed image at layer l respectively; M l Represents the total number of feature map elements in layer l.
9. A method for planning a power grid line based on multimodal data according to claim 8, characterized in that: The loss function L of the attention mechanism module attetion The formula is as follows: THE attention =L target +λL entropy Where L target represents the target loss, which measures the difference between the attention-weighted features and the specific task labels; L entropy Represents the attention entropy loss, distracting the model’s attention; λ is a hyperparameter; y o,c represents the indicator value of the true label of the oth sample in category c; p o,c represents the probability that the model predicts the oth sample as category c; a i represents the attention weight of the i-th feature; N is the number of all attention weights.
10. A method for planning a power grid line based on multimodal data according to claim 9, characterized in that: The loss function L of the semantic segmentation prediction module is segmentation The formula is as follows: THE segmentation =L entropy +γL dice Where L entropy represents the cross entropy loss, indicating the accuracy of pixel-level classification; L dice represents the Dice loss, which calculates the predicted and true coefficients; λ represents the trade-off convention, which is used to balance the cross entropy loss and the Dice loss; H, W, C represent the height, width and number of categories of the image respectively, y ijc represents the indicator value of the true label of the pixel located in the i-th row and j-th column in category c; p ijc represents the probability that the model predicts that the pixel located in the i-th row and j-th column is of category c; y ij represents the true label of the pixel located at the i-th row and j-th column; Represents the corresponding binary label predicted by the model.
Citation Information
Patent Citations
An Artificial Intelligence-Based Power Grid Line Planning Method
CN110532327B
Power line intelligent planning method fusing multi-source multi-dimensional heterogeneous big data
CN113361786A
Remote sensing image road segmentation method combining super-resolution and attention mechanism
CN113888550A
Intelligent power transmission line path planning method fusing pixel-level image segmentation and ant colony algorithm
CN116740091A
Pulmonary nodule intelligent grading method and system based on multi-modal feature fusion
CN116883768A
Cited By
Automatic adjustment method and system for geological division of power transmission and transformation project
CN120561624A
