Real-time transmission line regulation method and system based on cooperation of CNN and transformer

Abstract

The application discloses a power transmission line real-time regulation method and system based on cooperation of CNN and Transformer, relates to the technical field of power transmission line real-time regulation, and comprises the following steps: constructing a multi-modal input stream, extracting feature maps and time sequence features based on a preset double-channel deep interaction network according to the multi-modal input stream, and fitting and constructing the double-channel deep interaction network based on CNN and Transformer; establishing an association matrix based on the feature maps and the time sequence features, wherein the association matrix comprises feature weights in the association matrix adjusted based on a preset cross-modal gating unit; performing feature complementary enhancement on the feature maps and the time sequence features based on the association matrix, generating fusion features; constructing a prediction model according to the fusion features, and outputting a power transmission line regulation strategy. Through weighted adjustment on different modes, the application effectively suppresses mode conflicts and noise interference, improves the robustness of the model, and ensures stability and accuracy in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of real-time control technology for power transmission lines, and more specifically, to a real-time control method and system for power transmission lines based on the collaboration of CNN and Transformer. Background Technology

[0002] With the ongoing advancement of smart grid construction, the safe and stable operation of transmission lines is a core element in ensuring energy supply, making the innovation of their control technology imperative. Traditional transmission line control methods mainly rely on manual inspections and experience-based judgments. When faced with complex and ever-changing operating environments, these methods reveal problems such as response delays and insufficient control precision, making it difficult to meet the real-time and accuracy requirements of automated control systems.

[0003] In recent years, deep learning technology has been gradually applied in the field of power system automation control, but existing single technologies have limitations. Although convolutional neural networks perform well in extracting local features from images and sensor data, they are insufficient in processing long-term time-series data and global information; while Transformer models can effectively mine long-range dependencies in time-series data, they are lacking in the ability to capture local features.

[0004] In actual operation, the status of transmission lines is affected by the interaction of multiple sources of information, such as equipment failures, load fluctuations, and meteorological conditions. Disasters such as winter icing in southern mountainous areas, autumn and winter wildfires, and coastal typhoons place higher demands on the multimodal data processing and real-time decision-making capabilities of automated control systems. The synergistic integration of CNN and Transformer can achieve comprehensive perception and intelligent control of the operating status of transmission lines from multiple dimensions. However, there are still technical bottlenecks in areas such as multi-source data fusion architecture design and real-time analysis algorithm optimization, which urgently need to be overcome to promote the development of power system automation control technology.

[0005] For example, the invention patent announcement CN115657600B discloses a CNN-BLRNET thermal error prediction model, which includes a first input layer, a CNN layer, a BLRNET model, a dropout layer, a flatten layer, a dense layer, and a first output layer arranged sequentially. The BLRNET model includes a second input layer, a ReLU layer, and a second output layer. At least two RESNET units are connected in series between the ReLU layer and the output layer. The RESNET unit includes a direct mapping part and a jump connection part. The direct mapping part includes a convolutional layer I, a max pooling layer I, and a BILSTM layer arranged sequentially. The jump connection part includes a convolutional layer II and a max pooling layer II arranged sequentially. This model can extract the spatial and temporal features of thermal error data, thereby improving prediction accuracy and robustness. This invention also discloses a transfer learning method for the CNN-BLRNET thermal error prediction model, which can improve the model's generalization ability to adapt to different working conditions. This invention also discloses an intelligent integration framework that can provide data services including thermal error control, state monitoring, and fault diagnosis.

[0006] The above-disclosed technical solutions have at least the following technical problems: CNN extracts image icing features, and Transformer processes time-series signals such as meteorology and power. The two are usually modeled independently and fused late, resulting in weak information connection and poor robustness.

[0007] To address the above problems, this invention proposes a solution. Summary of the Invention

[0008] To overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a real-time control method and system for transmission lines based on the collaboration of CNN and Transformer. By using feature fusion based on a gating mechanism, the weighting ratio of image features and temporal features can be dynamically adjusted to solve the problems of the two features being typically modeled independently, fused late, having weak information connection, and poor robustness.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A real-time transmission line control method based on CNN and Transformer collaboration includes the following steps: constructing a multimodal input stream; extracting feature maps and temporal features based on a preset dual-channel deep interactive network according to the multimodal input stream, wherein the dual-channel deep interactive network is constructed based on CNN and Transformer fitting; establishing an association matrix based on the feature maps and temporal features, wherein the association matrix includes adjusting the feature weights in the association matrix based on a preset cross-modal gating unit; performing feature complementarity enhancement on the feature maps and temporal features based on the association matrix to generate fused features; constructing a prediction model based on the fused features and outputting a transmission line control strategy.

[0010] In a preferred embodiment, the feature map is extracted through image channels, specifically by: preprocessing the visible light image and infrared image in the multimodal input stream; extracting local texture features through channel-wise convolution and point-wise convolution based on a deep separable convolutional neural network, and performing multi-channel feature fusion to obtain image features; dynamically setting the receptive field size of the convolution kernel according to the local texture complexity of different regions in the image during the convolution process, assigning weights to the image features to obtain an enhanced spatial feature map; extracting spatial feature maps at different resolutions at multiple scale levels of the network, and fusing feature maps of different scales through feature concatenation to form a multi-scale feature map set.

[0011] In a preferred embodiment, the time-series features are extracted through a time-series channel. Specifically, the steps are as follows: preprocessing the multimodal input stream to obtain a multidimensional time-series signal; extracting local features from the multidimensional time-series signal based on temporal causal convolution to obtain local features; inputting the local features into a Transformer encoder; dividing the time-series data at each encoder layer using a sliding window technique and calculating local attention within each window; obtaining the attention output for the current time step by weighted summation of local time steps; stacking several Transformer encoder layers hierarchically based on the attention output to form a hierarchical Transformer structure; and processing the multimodal input stream based on the hierarchical Transformer structure to obtain meteorological-power time-series features.

[0012] In a preferred embodiment, adjusting the feature weights in the correlation matrix based on a preset cross-modal gating unit specifically involves: calculating the similarity between multi-scale feature maps and meteorological-electricity time-series features to construct a correlation matrix; generating a gating value through an activation function based on the weighted sum of the correlation matrix and a pre-acquired weight matrix; and weighting and fusing the gating-adjusted time-series features and image features according to the gating value to obtain weighted features.

[0013] In a preferred embodiment, the step of performing complementary feature enhancement on feature maps and temporal features based on the correlation matrix to generate fused features specifically involves: in the Transformer decoding layer, the temporal features are weighted and fused with the enhanced information of image features through a gating mechanism during the decoding process to obtain new enhanced temporal features; by adding temporal feature information to the upsampling layer of the CNN network, the responsiveness of image features to temporal changes is enhanced to obtain new enhanced image features; the enhanced temporal features and image features are weighted and fused to obtain preliminary fused features; the preliminary fused features are cleaned through an adversarial network to suppress intermodal interference noise; the cleaned features are guided and enhanced based on the de-icing domain knowledge graph; and the cleaned features and enhanced features are weighted and fused through a gating mechanism to generate the final fused features.

[0014] In a preferred embodiment, the feature purification of the preliminary fused features through an adversarial network specifically involves: taking the preliminary fused features as input, extracting high-order feature representations to generate purified features; evaluating the purified features through a discriminator and outputting a discrimination score; defining generator loss and discriminator loss, and training the generator and discriminator alternately based on the discrimination score to continuously improve the feature purification effect, ultimately obtaining the fused features after modality conflict elimination.

[0015] In a preferred embodiment, the step of guiding and enhancing the purified features based on the knowledge graph of the de-icing domain specifically involves: determining entity nodes and their relationships; embedding and encoding the knowledge graph to convert entity nodes and relationships into low-dimensional vector representations; embedding the fused features and the converted knowledge graph into vectors, and calculating the relevance scores between the features and each knowledge entity through an attention mechanism; and weighting the features according to the relevance scores to generate enhanced fused features.

[0016] In a preferred embodiment, the output transmission line control strategy specifically involves: classifying the predicted line safety risk level and determining the current line status; Based on the current status, multiple transmission line control strategies are generated, and the optimal transmission line control strategy is selected to generate control instructions. The transmission line control strategy includes generating corresponding de-icing operation instructions based on the predicted icing thickness value and the current line operating parameters, and dynamically adjusting the maximum allowable load, current limit and operation early warning threshold of the line based on the predicted icing thickness value and the weather-load development trend.

[0017] The technical effects and advantages of the real-time control method and system for power transmission lines based on the collaboration of CNN and Transformer in this invention are as follows: 1. This invention utilizes a gating mechanism-based feature fusion method to dynamically adjust the weighting ratio of image features and temporal features, thereby adaptively optimizing model performance in the face of different environments and sample characteristics. By adjusting the weighting of different modalities (such as image, weather, and load), modal conflicts and noise interference are effectively suppressed, improving the robustness of the model and ensuring stability and accuracy in complex environments.

[0018] 2. This invention, by combining the collaborative work of CNN and Transformer, can deeply extract multi-dimensional feature information across spatiotemporal scales. CNN is responsible for extracting spatial features from images and performing multi-scale processing, while Transformer excels at modeling long-term dependencies in time-series data. Through a dual-channel deep interactive network, this method effectively combines the features of images and time-series data, enhancing the accuracy of icing thickness prediction and power safety risk assessment, and improving the real-time performance and accuracy of transmission line control strategies. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the real-time control method for power transmission lines based on the collaboration of CNN and Transformer according to the present invention. Figure 2 This is a schematic diagram of the structure of the real-time control system for power transmission lines based on the collaboration of CNN and Transformer according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, Figure 1 This invention presents a real-time control method for power transmission lines based on the collaboration of CNN and Transformer, comprising the following steps: S1, construct a multimodal input stream, and extract feature maps and temporal features based on a preset dual-channel deep interactive network according to the multimodal input stream. The dual-channel deep interactive network is constructed based on CNN and Transformer fitting. The multimodal input stream is constructed by synchronously acquiring the first data from the transmission line.

[0022] The first data includes visible light image sequences, infrared image sequences, meteorological time-series data, and power load time-series data; The multimodal input stream is constructed by synchronously acquiring the first data from the transmission line, specifically as follows: Combining satellite remote sensing, drone inspections, and visible light cameras, micro-meteorological sensors, and power load monitoring terminals deployed along power transmission lines, it is used to collect visible light image sequences, infrared image sequences, meteorological time-series data, and power load time-series data of power transmission lines, respectively. The sampling period of each data acquisition device is set uniformly, and time synchronization is performed based on the GPS time synchronization module or network time protocol mechanism to ensure that the timestamps of various data are consistent. Within each sampling period, the visible light image, infrared image, meteorological data vector, and power load time series data vector under the corresponding timestamp are packaged to construct a multimodal data unit with a unified timestamp. The quality of multimodal data units is checked. If there are missing or abnormal data, forward filling, interpolation compensation or missing data marking methods are used to handle them. Multimodal data units of consecutive time slices are stacked in chronological order to form a spatiotemporally aligned multimodal input stream.

[0023] The dual-channel deep interactive network is constructed based on CNN and Transformer fitting, specifically as follows: Channel structure design and initialization: Image Channel (CNN Part): Design a CNN backbone network structure suitable for infrared / visible light images, which includes multiple convolutional layers, normalization layers and activation functions, and embeds variable convolutional kernel components, feature distillation units and spatiotemporal memory modules to extract local texture and multi-scale spatial features.

[0024] Temporal Channel (Transformer Part): A hierarchical Transformer structure is built, including multiple Transformer encoder modules based on causal convolution enhancement. Each layer uses a local attention mechanism to capture the temporal dependency features of meteorological and power load time-series data, and multiple layers are stacked to expand the receptive field.

[0025] Joint training objective setting: Set regression objectives (such as predicting ice thickness) and classification objectives (such as risk level), and feed the feature vectors output from the two channels into a unified prediction head; Constructing a joint loss function typically includes: regression loss (such as MSE), classification loss (such as cross-entropy), and a feature consistency regularization term between channels; Adding auxiliary supervision signals, such as guiding the image channel to focus on icy areas and guiding the time series channel to focus on key change intervals, can improve learning performance.

[0026] Network fitting and optimization: Initialization parameters: Set weights using a pre-trained model (such as a CNN on ImageNet or a Transformer on a temporal model) or random initialization; Joint training: Based on historical labeled multimodal data, the parameters of CNN and Transformer are optimized simultaneously through backpropagation, enabling them to learn together; Alternating adjustment: To avoid training conflicts between the two channels, an alternating optimization strategy is adopted, which first freezes one channel to train the other channel, and then switches them; Gating mechanism introduction: Gating weights are introduced in the feature fusion stage to automatically learn the contribution ratio of images and temporal features based on training samples, thereby achieving dynamic weighting.

[0027] Fit evaluation and structural iteration: The expressive power of the output features of the two channels was evaluated on the validation set, and hyperparameters such as kernel size, number of Transformer layers, and attention window size were adjusted. Visualization methods (such as CAM heatmaps and attention weight distribution) can be used to analyze the regions of interest of the model, thereby further optimizing the network structure and training process.

[0028] The dual-channel deep interactive network includes an image channel and a temporal channel; The feature map is extracted through image channels, specifically: Preprocess the visible light and infrared images in the multimodal input stream; Based on a deep separable convolutional neural network, local texture features are extracted through channel-wise convolution and point-wise convolution, and multi-channel feature fusion is performed. Batch normalization layers and Swish activation functions are inserted between every two convolutional units to accelerate training and enhance nonlinear modeling capabilities, thereby obtaining image features. During the convolution process, the size of the receptive field of the convolution kernel is dynamically set according to the local texture complexity of different regions in the image, and weights are assigned to the image features to obtain the enhanced spatial feature map. Spatial feature maps at different resolutions are extracted at multiple scale levels of the network, and feature maps at different scales are fused together to form a multi-scale feature map set through feature stitching.

[0029] The preprocessing of the visible light and infrared images in the multimodal input stream specifically includes: The input images are normalized to ensure that the input images are of consistent size; Histogram equalization is used to enhance image contrast and highlight icy areas and background details.

[0030] In this embodiment, dynamically setting the receptive field size of the convolution kernel is for adaptive capture of small-scale initial icing and large-scale thick icing. The depthwise separable convolutional neural network includes: Variable kernel component: dynamically adjusts the kernel size from 3×3 to 7×7 according to current weather conditions; Feature distillation unit: Filters key icing features through channel attention to suppress interference noise such as clouds and fog; Spatiotemporal memory module: Stores feature change patterns over a 24-hour period of time, providing a benchmark for current feature extraction.

[0031] The temporal features are extracted through temporal channels, specifically as follows: The multimodal input stream is resampled and time-aligned to obtain an aligned multidimensional time-series signal; Local features are extracted from multidimensional time-series signals based on temporal causal convolution. Local features are input into the Transformer encoder. In each encoder layer, the temporal data is divided based on the sliding window technique, and local attention is calculated within each window. The attention output of the current time step is obtained by weighted summation of the local time steps. Several Transformer encoder layers are stacked hierarchically to form a hierarchical Transformer structure; The meteorological-electricity time series features are obtained by processing the multimodal input stream based on a hierarchical Transformer structure.

[0032] Each encoder layer models the dependency between the current time step and adjacent time steps through a self-attention mechanism, and increases the receptive field by expanding the convolution to improve the handling of long sequence dependency patterns.

[0033] Using a one-dimensional convolution kernel ensures that the convolution operation depends only on the data of the current time step and the time steps before it. By stacking multiple causal convolutional layers, the receptive field can be expanded to capture a longer range of local dependencies.

[0034] S2, establish an association matrix based on feature maps and temporal features, wherein the association matrix includes feature weights adjusted based on preset cross-modal gating units; The adjustment of feature weights in the correlation matrix based on a preset cross-modal gating unit specifically involves: Calculate the similarity between multi-scale feature maps and meteorological-electricity time-series features, and construct an association matrix; The gate value is generated by weighting the correlation matrix and the pre-obtained weight matrix using an activation function. The time-series features and image features after gating are weighted and fused based on the gating value to obtain the weighted features.

[0035] The correlation matrix is ​​specifically as follows:

[0036] in, and These are indices for temporal features and image features, respectively. and These are vector representations of temporal features and image features, respectively. This represents the similarity between temporal features and image features.

[0037] The specific gate value is:

[0038] The weighted features are specifically:

[0039] in, This is the gate value. It is the sigmoid activation function. It is an incidence matrix. For the pre-obtained weight matrix, As a weighted feature, As a time series feature, These are image features.

[0040] S3, based on the correlation matrix, performs complementary feature enhancement on feature maps and temporal features to generate fused features; The step of performing complementary feature enhancement on feature maps and temporal features based on the correlation matrix to generate fused features is as follows: In the Transformer decoding layer, temporal features are weighted and fused with the enhanced information of image features through a gating mechanism during the decoding process to obtain new enhanced temporal features; By incorporating temporal feature information into the upsampling layer of a CNN network, the responsiveness of image features to temporal changes is enhanced, resulting in new enhanced image features: The enhanced temporal features and image features are weighted and fused to obtain preliminary fused features; Based on the features fused from the initial correlation matrix, an adversarial network is used for feature purification. The adversarial network includes a feature generator and a feature discriminator. Through adversarial training and optimization, noise interference between modalities is eliminated to obtain the purified fused features. The knowledge graph of the power transmission line de-icing field is embedded into the feature space. Through entity relationship modeling and attention mechanism, the cleaned fused features are enhanced in dimensions that are highly related to domain knowledge. By employing a gating mechanism, the contribution ratio of adversarial purification features and knowledge graph enhancement features is dynamically adjusted during the fusion process. This ensures that while removing noise, the fused features fully retain effective feature information related to actual physical laws, thereby improving the accuracy and robustness of the overall modeling.

[0041] Furthermore, the feature purification of the preliminary fused features through adversarial networks specifically involves: Using the initial fused features as input, the high-order feature representation is extracted through several convolutional layers, normalization layers, and nonlinear activation layers to generate purified features. Based on the purified features, determine whether it contains modal conflict or noise interference signals, and output a discrimination score; Define generator loss and discriminator loss. The generator aims to generate high-quality, modal conflict-free features, while the discriminator aims to distinguish between the original fused features and the cleaned features. By alternating training of the generator and discriminator based on the discrimination score, the feature purification effect is continuously improved, and finally the fused features after modality conflict elimination are obtained.

[0042] Furthermore, the enhancement of the purified features based on the knowledge graph of the de-icing domain specifically includes: Identify physical nodes (such as temperature, humidity, wind speed, load, current, icing thickness, etc.) and their relationships (e.g., low temperature and high humidity easily lead to icing, and increased icing leads to increased load). The knowledge graph is embedded and encoded to transform entity nodes and relationships into low-dimensional vector representations, so as to facilitate alignment with fused features; By combining fused features with knowledge graph embedding vectors through an attention mechanism, the correlation score between features and each knowledge entity is calculated, thereby strengthening feature components that are highly correlated with actual physical relationships. The features are weighted and adjusted based on the correlation score to enhance the feature signal while suppressing irrelevant or noise components, thus generating enhanced fused features.

[0043] The weighted fusion of purification and enhancement features through a gating mechanism has the following advantages: Dynamically controlling information flow to adapt to different sample characteristics: Traditional concatenation or weighted averaging is a static fusion, applying a one-size-fits-all approach to all samples. By adding a gating mechanism, the ratio of cleaned features to enhanced features can be dynamically adjusted based on the feature distribution of the current input samples. For noisy samples, the gating mechanism tends to retain cleaned features; for samples with weak features but clear patterns, the gating mechanism strengthens enhanced features.

[0044] Suppressing modal conflicts and improving fusion robustness: Cleaned features are typically low in noise but may lack sufficient information; augmented features are rich in information but may introduce modal conflicts or artifacts. Gating mechanisms can adaptively balance information content and purity during feature fusion, dynamically weakening the influence of conflicting features. This effectively avoids the negative interference introduced by augmented features and improves the overall stability of the model.

[0045] Optimize convergence speed and final performance: Because the feature fusion process is gated, optimization will not allow the cleanup branch and the enhancement branch to interfere with each other, reducing gradient conflicts, improving training convergence speed, and improving final metrics (such as accuracy and F1 score).

[0046] S4. Construct a prediction model based on the fusion characteristics and output the transmission line control strategy.

[0047] The construction of the prediction model based on the fusion features is specifically as follows: Based on the fusion features, a joint learning prediction model framework is built to simultaneously handle the tasks of icing thickness prediction and line safety risk level assessment. The fused features are input into a regression branch, passed through several fully connected layers, and output the predicted ice thickness. The fused features are input into another classification branch, and after feature extraction and classifier layers, the security risk level of the line is output. By combining regression loss (such as mean squared error, MSE) and classification loss (such as cross-entropy), weighting coefficients are set, and the two branches are jointly optimized. The prediction model is trained under supervision based on historical labeled data to obtain a well-trained prediction model.

[0048] The output transmission line control strategy is as follows: Based on the predicted line safety risk level, determine whether the current line status is normal, under warning, or dangerous. When the risk level is warning or danger, the corresponding de-icing operation command is generated based on the predicted icing thickness value and the current line operating parameters, including starting and stopping the de-icing device, selecting the heating / vibration mode and the de-icing duration parameter. When icing does not reach the danger level but there is a slight icing trend, the maximum allowable load, current limit and operation warning threshold of the line are dynamically adjusted or the line is shut down in an emergency, based on the predicted icing thickness and the weather-load development trend. Based on current meteorological forecast information and equipment status data, multiple feasible transmission line control strategies are selected for optimization, with priority given to command schemes that have low energy consumption and high de-icing efficiency. Based on the generated de-icing transmission line control strategy, the weighting ratio of each modal feature is dynamically adjusted through a gating mechanism.

[0049] For areas with thick ice, the weights of image (infrared and visible light) modal features are increased preferentially; In situations with complex meteorological conditions or high loads, priority should be given to increasing the weight of meteorological data (such as temperature, humidity, wind speed, etc.). When the ambient temperature is low and the icing is severe, the gating mechanism will automatically increase the weight of the infrared image in order to better detect the specific distribution and thickness of the ice layer; When the load is high and the weather is relatively stable, the weight of the power load characteristics will be increased, thereby enhancing the responsiveness to grid dispatch.

[0050] The weighting ratio is as follows:

[0051] in, Let be the dynamic gating weight matrix at time step t. For ice thickness, Based on safety risk level, For meteorological data, This is time-series data of power load.

[0052] Example 2, Figure 2 The present invention provides a real-time control system for power transmission lines based on the collaboration of CNN and Transformer, comprising the following modules: Feature extraction module: used to construct a multimodal input stream, and extract feature maps and temporal features based on a preset dual-channel deep interactive network based on the multimodal input stream. The dual-channel deep interactive network is constructed based on CNN and Transformer fitting. Weight adjustment module: used to establish an association matrix based on feature maps and temporal features, wherein the association matrix includes feature weights in the association matrix adjusted based on a preset cross-modal gating unit; Feature fusion module: used to perform complementary enhancement of feature maps and temporal features based on the correlation matrix to generate fused features; Prediction and Control Module: Used to construct a prediction model based on fusion characteristics and output transmission line control strategies.

[0053] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0054] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0055] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0058] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time control method for power transmission lines based on the collaboration of CNN and Transformer, characterized in that, Includes the following steps: A multimodal input stream is constructed, and feature maps and temporal features are extracted based on a preset dual-channel deep interactive network. The dual-channel deep interactive network is constructed based on CNN and Transformer fitting. An association matrix is ​​established based on feature maps and temporal features, wherein the association matrix includes feature weights adjusted based on a preset cross-modal gating unit; Based on the correlation matrix, feature maps and temporal features are complementarized and enhanced to generate fused features. Specifically, in the Transformer decoding layer, the temporal features are weighted and fused with the enhancement information of image features through a gating mechanism during the decoding process to obtain new enhanced temporal features. By incorporating temporal feature information into the upsampling layer of a CNN network, the responsiveness of image features to temporal changes is enhanced, resulting in new enhanced image features. The enhanced temporal features and image features are then weighted and fused to obtain preliminary fused features. An adversarial network is used to cleanse the preliminary fused features, suppressing intermodal interference noise. Based on a de-icing domain knowledge graph, the cleaned features are further enhanced. Finally, a gating mechanism is used to weightedly fuse the cleaned and enhanced features to generate the final fused features. Specifically, the guided enhancement of the purified features involves: identifying entity nodes and their relationships; embedding and encoding the knowledge graph to transform entity nodes and relationships into low-dimensional vector representations; embedding the fused features into the transformed knowledge graph vectors and calculating the relevance scores between the features and each knowledge entity using an attention mechanism; and weighting the features based on the relevance scores to generate enhanced fused features. The entity nodes include temperature, humidity, wind speed, load, current, and icing thickness. The relationships include the relationship between low temperature and high humidity and icing formation, as well as the relationship between increased icing thickness and increased load. A prediction model is constructed based on the final fusion characteristics, and a transmission line control strategy is output.

2. The real-time control method for transmission lines based on the collaboration of CNN and Transformer as described in claim 1, characterized in that, The feature map is extracted through image channels, and the specific steps are as follows: Preprocess the visible light and infrared images in the multimodal input stream; Based on a deep separable convolutional neural network, local texture features are extracted through channel-wise convolution and point-wise convolution, and multi-channel feature fusion is performed to obtain image features; During the convolution process, the size of the receptive field of the convolution kernel is dynamically set according to the local texture complexity of different regions in the image, and weights are assigned to the image features to obtain the enhanced spatial feature map. Spatial feature maps at different resolutions are extracted at multiple scale levels of the network, and feature maps at different scales are fused together to form a multi-scale feature map set through feature stitching.

3. The real-time control method for transmission lines based on the collaboration of CNN and Transformer as described in claim 2, characterized in that, The temporal features are extracted through temporal channels, and the specific steps are as follows: The multimodal input stream is preprocessed to obtain a multidimensional time-series signal; Local features are extracted from multidimensional time-series signals based on temporal causal convolution. Local features are input into the Transformer encoder. In each encoder layer, the temporal data is divided based on the sliding window technique, and local attention is calculated within each window. The attention output of the current time step is obtained by weighted summation of the local time steps. Based on attention output, several Transformer encoder layers are stacked hierarchically to form a hierarchical Transformer structure; The meteorological-electricity time series features are obtained by processing the multimodal input stream based on a hierarchical Transformer structure.

4. The real-time control method for transmission lines based on the collaboration of CNN and Transformer as described in claim 3, characterized in that, The adjustment of feature weights in the correlation matrix based on a preset cross-modal gating unit specifically involves: Calculate the similarity between multi-scale feature maps and meteorological-electricity time-series features, and construct an association matrix; The gate value is generated by weighting the correlation matrix and the pre-obtained weight matrix using an activation function. The time-series features and image features after gating are weighted and fused based on the gating value to obtain the weighted features.

5. The real-time control method for transmission lines based on the collaboration of CNN and Transformer as described in claim 4, characterized in that, The feature purification of the preliminary fused features through adversarial networks specifically involves: Using the initial fused features as input, higher-order feature representations are extracted to generate purified features; The purified features are evaluated by a discriminator, and a discrimination score is output. Define generator loss and discriminator loss, and train the generator and discriminator alternately based on the discrimination score to continuously improve the feature purification effect, and finally obtain the fused features after modality conflict elimination.

6. The real-time control method for transmission lines based on the collaboration of CNN and Transformer as described in claim 5, characterized in that, The output transmission line control strategy is as follows: The predicted line safety risk levels are classified to determine the current line status; Based on the current status, multiple transmission line control strategies are generated, and the optimal transmission line control strategy is selected to generate control instructions. The transmission line control strategy includes generating corresponding de-icing operation instructions based on the predicted icing thickness value and the current line operating parameters, and dynamically adjusting the maximum allowable load, current limit, and operation warning threshold of the line or emergency shutdown of the line based on the predicted icing thickness value and the weather-load development trend.

7. The system of the real-time control method for transmission lines based on the collaboration of CNN and Transformer according to any one of claims 1-6, characterized in that, Includes the following modules: Feature extraction module: used to construct a multimodal input stream, and extract feature maps and temporal features based on a preset dual-channel deep interactive network based on the multimodal input stream. The dual-channel deep interactive network is constructed based on CNN and Transformer fitting. Weight adjustment module: used to establish an association matrix based on feature maps and temporal features, wherein the association matrix includes feature weights in the association matrix adjusted based on a preset cross-modal gating unit; Feature fusion module: used to perform complementary enhancement of feature maps and temporal features based on the correlation matrix to generate fused features; Prediction and Control Module: Used to construct prediction models based on fused features and output transmission line control strategies.

Images