Power transmission line real-time regulation and control method and system based on CNN (Convolutional Neural Network) and Transform cooperation

Abstract

The invention discloses a power transmission line real-time regulation and control method and system based on CNN and Transform cooperation, and relates to the technical field of power transmission line real-time regulation and control, and the method comprises the following steps: constructing a multi-modal input stream, respectively extracting a feature map and a time sequence feature according to the multi-modal input stream based on a preset dual-channel deep interaction network, and carrying out the calculation of the feature map and the time sequence feature; the dual-channel deep interactive network is constructed based on CNN and Transform fitting; establishing an incidence matrix based on the feature map and the time sequence features, wherein the incidence matrix comprises feature weights in the incidence matrix adjusted based on a preset cross-modal gating unit; performing feature complementary enhancement on the feature map and the time sequence features based on the incidence matrix to generate fusion features; and constructing a prediction model according to the fusion features, and outputting a power transmission line regulation and control strategy. According to the method, through weighted adjustment of different modes, mode conflicts and noise interference are effectively suppressed, the robustness of the model is improved, and the stability and accuracy in a complex environment are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time control of transmission lines, and more specifically, to a real-time control method and system for transmission lines based on collaboration between CNN and Transformer. Background Art

[0002] As smart grid construction continues to advance, the safe and stable operation of transmission lines, a core component of ensuring energy supply, necessitates urgent innovation in control technology. Traditional transmission line control methods rely primarily on manual inspections and empirical judgment. These methods, faced with complex and changing operating environments, suffer from response delays and insufficient control accuracy, making them unable to meet the real-time and accuracy requirements of automated control.

[0003] In recent years, deep learning technology has been gradually applied in the field of power system automation and control, but existing single technologies have limitations. While convolutional neural networks excel at extracting local features from images and sensor data, they struggle with long time series data and global information. Transformer models, while effective at mining long-range dependencies in time series data, lack the ability to capture local features.

[0004] In actual operation, the status of transmission lines is influenced by the interaction of multiple sources of information, including 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 Transforme enables comprehensive perception and intelligent control of transmission line operating conditions from multiple dimensions. However, technical bottlenecks remain in the design of multi-source data fusion architectures and the optimization of real-time analysis algorithms, and breakthroughs are urgently needed to advance the development of power system automation and control technology.

[0005] For example, the invention patent publication number 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. The BLRNET model includes a second input layer, a ReLU layer, and a second output layer, with at least two RESNET units connected in series between the ReLU layer and the output layer. The RESNET unit includes a direct mapping portion and a skip connection portion. The direct mapping portion includes a convolutional layer I, a maximum pooling layer I, and a BILSTM layer, and the skip connection portion includes a convolutional layer II and a maximum pooling layer II. This model can extract spatial and temporal features of thermal error data, thereby improving prediction accuracy and robustness. The 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 operating conditions. The invention also discloses an intelligent integration framework that can provide data services including thermal error control, condition monitoring, and fault diagnosis.

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

[0007] In view of the above problems, the present invention proposes a solution. Summary of the Invention

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

[0009] To achieve the above object, the present invention provides the following technical solutions: A real-time control method for transmission lines based on the collaboration of CNN and Transformer includes the following steps: constructing a multimodal input stream, extracting feature maps and time series features based on the multimodal input stream based on a preset dual-channel deep interaction network, wherein the dual-channel deep interaction network is constructed based on CNN and Transformer fitting; establishing an association matrix based on the feature maps and time series features, wherein the association matrix includes feature weights in the association matrix adjusted based on preset cross-modal gating units; performing feature complementarity enhancement on the feature maps and time series 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, and the specific steps are: preprocessing the visible light image and infrared image in the multimodal input stream; extracting local texture features through channel-by-channel convolution and point-by-point convolution based on a deep separable convolutional neural network, and performing multi-channel feature fusion to obtain image features; during the convolution process, dynamically setting the receptive field size of the convolution kernel according to the local texture complexity of different areas in the image, assigning weights to the image features, and obtaining an enhanced spatial feature map; at multiple scale levels of the network, respectively extracting spatial feature maps at different resolutions, and fusing feature maps of different scales through feature splicing to form a multi-scale feature map set.

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

[0012] In a preferred embodiment, the feature weights in the association matrix are adjusted based on a preset cross-modal gating unit, specifically: the similarity between the multi-scale feature map and the meteorological-electricity time series features is calculated to construct a correlation matrix; a gating value is generated through an activation function based on the weighted sum of the association matrix and the pre-acquired weight matrix; and the time series features and image features after gated adjustment are weighted and fused according to the gating value to obtain weighted features.

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

[0014] In a preferred embodiment, the feature purification of the preliminary fusion features through the adversarial network is specifically as follows: taking the preliminary fusion features as input, extracting high-order feature representations, and generating purified features; evaluating the purified features through the discriminator, and outputting a discriminant score; defining the generator loss and the discriminator loss, and alternately training the generator and the discriminator based on the discriminant score, continuously improving the feature purification effect, and finally obtaining the fusion features after the modal conflict is eliminated.

[0015] In a preferred embodiment, the purified features are guided and enhanced based on the de-icing field knowledge graph, specifically: determining entity nodes and their association relationships; embedding encoding the knowledge graph, converting entity nodes and relationships into low-dimensional vector representations; embedding the fused features and the converted knowledge graph into a vector, and calculating the correlation scores between the features and each knowledge entity through the attention mechanism; and weighting and adjusting the features according to the correlation scores to generate enhanced fused features.

[0016] In a preferred embodiment, the output transmission line control strategy specifically comprises: classifying the predicted line safety risk levels and determining the current line status; Based on the current status, multiple transmission line control strategies are generated, and the transmission line control strategies are optimized to generate control instructions. The transmission line control strategies include generating corresponding de-icing operation instructions based on the predicted ice thickness value and the current operating parameters of the line, and dynamically adjusting the maximum allowable load, current limit and operation warning threshold of the line based on the predicted ice thickness value and meteorological-load development trends.

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

[0018] 2. By combining the collaborative work of CNN and Transformer, this method can deeply extract multi-dimensional feature information across both temporal and spatial 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 characteristics of image and time series data, enhancing the accuracy of ice thickness prediction and power safety risk assessment, and improving the real-time and accuracy of transmission line control strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the real-time control method for power transmission lines based on the collaboration of CNN and Transformer in the present invention; Figure 2 This is a structural diagram of the real-time control system for power transmission lines based on the collaboration of CNN and Transformer in the present invention. DETAILED DESCRIPTION

[0020] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] Example 1, Figure 1 The present invention provides a real-time control method for transmission lines based on the collaboration of CNN and Transformer, which includes the following steps: S1: construct a multimodal input stream, and extract feature maps and temporal features based on the multimodal input stream based on a preset dual-channel deep interaction network, wherein the dual-channel deep interaction network is constructed based on CNN and Transformer fitting; The multimodal input stream is constructed by synchronously collecting first data of the power transmission line.

[0022] The first data includes a visible light image sequence, an infrared image sequence, meteorological time series data, and power load time series data; The multimodal input stream is constructed by synchronously collecting first data of the power transmission line, specifically: Combining satellite remote sensing, drone inspections, and visible light cameras, micro-meteorological sensors, and power load monitoring terminals deployed along the 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 the transmission lines; The sampling period of each collection device is uniformly set, and time synchronization is performed based on the global positioning system time synchronization module or network time protocol mechanism to ensure that the timestamps of various data are consistent; In each sampling period, the visible light image, infrared image, meteorological data vector and power load time series data vector at the corresponding timestamp are packaged to construct a multimodal data unit with a unified timestamp; Perform quality checks on multimodal data units. If data is missing or abnormal, forward filling, interpolation compensation, or missing marking are used to handle it. The multimodal data units of consecutive time slices are stacked in time order to form a spatiotemporally aligned multimodal input stream.

[0023] The dual-channel deep interaction network is constructed based on CNN and Transformer fitting, specifically: 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. It embeds variable convolution kernel components, feature distillation units, and spatiotemporal memory modules to extract local texture and multi-scale spatial features.

[0024] Time series pipeline (Transformer): A hierarchical Transformer structure is built, including multiple Transformer encoder modules enhanced with causal convolution. Each layer uses a local attention mechanism to capture the temporal dependency characteristics of meteorological and power load time series data, and multiple layers are stacked to expand the receptive field.

[0025] Joint training goal setting: Set the regression target (such as ice thickness prediction) and the classification target (such as risk level), and feed the feature vectors output by the two channels into a unified prediction head; Construct a joint loss function, which usually includes regression loss (such as MSE), classification loss (such as cross entropy), and feature consistency regularization between channels; Adding auxiliary supervision signals, such as guiding the image channel to focus on ice-covered areas and guiding the time series channel to focus on key change intervals, can improve learning effects.

[0026] Network fitting and optimization: Initialization parameters: Use pre-trained models (such as CNN on ImageNet, Transformer on time series models) or random initialization to set weights; Joint training: Based on historical labeled multimodal data, CNN and Transformer parameters are optimized simultaneously through backpropagation to enable collaborative learning; Alternating adjustment: To avoid training conflicts between the two channels, an alternating optimization strategy is adopted, first freezing one channel and training the other channel, and then switching; Introduction of gating mechanism: Gating weights are introduced in the feature fusion stage to automatically learn the contribution ratio of image and time series features based on training samples to achieve dynamic weighting.

[0027] Fitting effect evaluation and structure iteration: Evaluate the expressiveness of the output features of the two channels on the validation set, and adjust hyperparameters such as convolution kernel size, number of Transformer layers, and attention window size; Use visualization methods (such as CAM heat maps and attention weight distribution) to analyze the model's focus areas and further optimize the network structure and training process.

[0028] The dual-channel deep interaction network includes an image channel and a time sequence channel; The feature map is extracted through the image channel, specifically: Preprocessing of visible light and infrared images in multimodal input streams; Based on the depthwise separable convolutional neural network, local texture features are extracted through channel-by-channel convolution and point-by-point convolution, and multi-channel feature fusion is performed. Batch normalization layers and Swish activation functions are inserted between every two layers of convolution units to accelerate training and enhance nonlinear modeling capabilities to obtain image features. During the convolution process, the receptive field size of the convolution kernel is dynamically set according to the local texture complexity of different areas in the image, and the image features are assigned weights to obtain the enhanced spatial feature map; At multiple scale levels of the network, spatial feature maps at different resolutions are extracted respectively, and feature maps of different scales are fused through feature splicing to form a multi-scale feature map set.

[0029] The preprocessing of the visible light image and the infrared image in the multimodal input stream is specifically as follows: Normalize the input image to ensure that the input image size is consistent; Histogram equalization technology is used to enhance the contrast of the image and highlight the ice-covered area and background details.

[0030] In this embodiment, the receptive field size of the convolution kernel is dynamically set to adaptively capture small-scale initial ice cover and large-scale heavy ice cover; The depth-wise separable convolutional neural network includes: Variable convolution kernel component: Dynamically adjusts the convolution kernel size from 3×3 to 7×7 according to current meteorological conditions; Feature Distillation Unit: This unit uses channel attention to filter key ice cover features and suppress interfering noise such as clouds and fog. Spatiotemporal memory module: stores the feature change patterns of the past 24 hours, providing a comparison benchmark for current feature extraction.

[0031] The time series features are extracted through the time series channel, specifically: Resample and time-align the multimodal input stream to obtain an aligned multidimensional time series signal; Based on time causal convolution, local features of multi-dimensional time series signals are extracted to obtain local features; The local features are input into the Transformer encoder. At each encoder layer, the time series data is divided based on the sliding window technology, and the 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. Stack several Transformer encoder layers in layers to form a hierarchical Transformer structure; The multimodal input stream is processed based on the hierarchical Transformer structure to obtain meteorological-power time series characteristics.

[0032] Each layer of encoder models the dependency between the current time step and the adjacent time step through the self-attention mechanism, increases the receptive field through dilated convolution, and improves the dependency pattern of long sequences.

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

[0034] S2, establishing a correlation matrix based on the feature graph and the temporal features, wherein the correlation matrix includes adjusting feature weights in the correlation matrix based on a preset cross-modal gating unit; The feature weights in the correlation matrix are adjusted based on the preset cross-modal gating unit, specifically: Calculate the similarity between the multi-scale feature map and the meteorological-power time series features, and construct the correlation matrix; Generate a gating value through an activation function based on the weighted sum of the association matrix and the pre-acquired weight matrix; The gated temporal features and image features are weighted and fused according to the gate value to obtain weighted features.

[0035] The correlation matrix is ​​specifically:

[0036] in, and are the indexes of temporal features and image features respectively, and are vector representations of temporal features and image features, respectively. is the similarity between temporal features and image features.

[0037] The gate value is specifically:

[0038] The weighted features are specifically:

[0039] in, is the gate value, is the sigmoid activation function, is the incidence matrix, is the pre-obtained weight matrix, is the weighted feature, is the time series feature, is the image feature.

[0040] S3, based on the correlation matrix, performs feature complementation enhancement on the feature map and time series features to generate fusion features; The feature map and time series features are complementary enhanced based on the correlation matrix to generate fusion features, specifically: In the Transformer decoding layer, the temporal features are weightedly fused with the enhanced information of the 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 ability of image features to respond to temporal changes is enhanced, and new enhanced image features are obtained: Perform weighted fusion of the enhanced temporal features and image features to obtain preliminary fusion features; Based on the features initially fused by the correlation matrix, an adversarial network is used to perform feature purification. The adversarial network includes a feature generator and a feature discriminator. Through adversarial training optimization, the noise interference between the modalities is eliminated to obtain the purified fusion features. The knowledge graph of the power transmission line de-icing domain is embedded into the feature space. Through entity relationship modeling and attention mechanism, the purified fusion features are guided to enhance the dimensions that are highly relevant to the domain knowledge. A gating mechanism is adopted to dynamically adjust the contribution ratio of adversarial purification features and knowledge graph enhancement features during the fusion process, ensuring that the fused features fully retain the effective feature information related to actual physical laws while removing noise, thereby improving the accuracy and robustness of the overall modeling.

[0041] Furthermore, the preliminary fusion features are purified by the adversarial network, specifically: Taking the preliminary fused features as input, it goes through several convolutional layers, normalization layers, and nonlinear activation layers to extract high-order feature representations and generate purified features. Based on the purified features, determine whether they contain modal conflict or noise interference signals and output the discrimination score; Define the generator loss and discriminator loss. The generator goal is to generate high-quality, non-modal conflict features, and the discriminator goal is to distinguish the original fused features from the purified features. The generator and discriminator are trained alternately based on the discriminant scores to continuously improve the feature purification effect, and finally the fusion features are obtained after the modal conflict is eliminated.

[0042] Furthermore, based on the de-icing domain knowledge graph, the purified features are guided and enhanced, specifically: Determine entity nodes (such as temperature, humidity, wind speed, load, current, ice thickness, etc.) and their relationships (for example, low temperature and high humidity are prone to ice accumulation, and increased ice accumulation leads to increased load); Embed the knowledge graph to convert entity nodes and relationships into low-dimensional vector representations to facilitate alignment with fusion features; Combine the fusion features with the knowledge graph embedding vector through the attention mechanism, calculate the correlation score between the features and each knowledge entity, and strengthen the feature components that are highly correlated with the actual physical relationship; The features are weighted and adjusted according to the correlation scores to enhance the feature signals while suppressing irrelevant or noise components to generate enhanced fusion features.

[0043] The weighted fusion of the purification features and the enhancement features through the gating mechanism has the following benefits: Dynamically control information flow to adapt to different sample characteristics: Traditional concatenation or weighted averaging is a static fusion that applies a one-size-fits-all approach to all samples. Adding a gating mechanism dynamically adjusts the ratio of cleansed to enhanced features based on the feature distribution of the current input sample. For samples with high noise, gating will tend to retain cleansed features; for samples with weak features but clear patterns, gating will favor enhanced features.

[0044] Suppressing modal conflict and improving fusion robustness: Purified features typically have low noise but may lack information; enhanced features are information-rich but may introduce modal conflict or artifacts. The gating mechanism adaptively balances information content and purity during feature fusion, dynamically reducing the impact of conflicting features. This effectively avoids the negative interference introduced by enhanced features and improves overall model stability.

[0045] Optimizing convergence speed and final performance: Because the feature fusion process is regulated by gating, the purification branch and the enhancement branch will not interfere with each other during optimization, reducing gradient conflicts, improving training convergence speed, and improving final indicators (such as accuracy and F1 score).

[0046] S4, builds a prediction model based on the fusion features and outputs the transmission line control strategy.

[0047] The prediction model is constructed based on the fusion features, specifically: Based on the fusion features, a joint learning prediction model framework is built to simultaneously handle ice thickness prediction tasks and line safety risk level assessment tasks; The fused features are input into a regression branch, which passes through several fully connected layers and outputs the predicted ice thickness value; The fused features are input into another classification branch, and after feature extraction and classifier layer, the safety risk level of the line is output; Combine regression loss (such as mean square error (MSE)) and classification loss (such as cross-entropy), set weight coefficients, and jointly optimize the two branches; The prediction model is supervised and trained based on historical labeled data to obtain a trained prediction model.

[0048] The output transmission line control strategy is specifically as follows: Based on the predicted line safety risk level, determine whether the current line status is normal, warning or dangerous; When the risk level is warning or dangerous, the system generates corresponding de-icing operation instructions based on the predicted ice thickness and the current line operating parameters, including starting and stopping the de-icing device, selecting the heating / vibration mode, and the de-icing duration parameters. When icing does not reach a dangerous level but there is a trend of mild icing, the maximum allowable load, current limit, and operation warning threshold of the line will be dynamically adjusted or the line will be shut down urgently based on the predicted ice thickness value and weather-load development trend; Combining current weather forecast information with equipment status data, multiple feasible transmission line control strategies are optimized, with priority given to command schemes with low energy consumption and high de-icing efficiency. Based on the generated de-icing transmission line control strategy, the weighted proportion of each modal feature is dynamically adjusted through the gating mechanism.

[0049] For areas with thick ice cover, the weight of image (infrared, visible light) modal features is increased first; For situations with complex meteorological conditions or high load, 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 ice coverage is severe, the gating mechanism automatically increases the weight of the infrared image 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 weighted ratio is specifically:

[0051] in, is the dynamic gating weight matrix at time step t, is the ice thickness, is the security risk level, For meteorological data, It is the 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, which includes the following modules: Feature extraction module: used to construct a multimodal input stream, and extract feature maps and temporal features based on the multimodal input stream based on a preset dual-channel deep interaction network. The dual-channel deep interaction network is constructed based on CNN and Transformer fitting; A weight adjustment module is used to establish a correlation matrix based on the feature map and the temporal features, wherein the correlation matrix includes adjusting the feature weights in the correlation matrix based on a preset cross-modal gating unit; Feature fusion module: used to perform feature complementation enhancement on feature maps and time series features based on the correlation matrix to generate fused features; Prediction and control module: used to build a prediction model based on the fusion features and output the transmission line control strategy.

[0053] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

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

[0055] Those skilled in the art will appreciate that the modules and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

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

[0058] Finally: 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 in the scope of protection of the present invention.

Claims

1. A real-time control method for transmission lines based on CNN and Transformer collaboration, characterized by: The following steps are involved: Construct a multimodal input stream, and extract feature maps and temporal features based on the multimodal input stream based on a preset dual-channel deep interaction network, wherein the dual-channel deep interaction network is constructed based on CNN and Transformer fitting; Establishing a correlation matrix based on the feature graph and the temporal features, wherein the correlation matrix includes adjusting feature weights in the correlation matrix based on a preset cross-modal gating unit; Based on the correlation matrix, feature maps and time series features are enhanced complementary to generate fusion features; A prediction model is constructed based on the fusion features, and the transmission line control strategy is output.

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

3. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 2 is characterized in that: The time series features are extracted through the time series channel, and the specific steps are as follows: Preprocess the multimodal input stream to obtain a multidimensional time series signal; Based on time causal convolution, local features of multi-dimensional time series signals are extracted to obtain local features; The local features are input into the Transformer encoder. At each encoder layer, the time series data is divided based on the sliding window technology, and the 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 the attention output, several Transformer encoder layers are stacked in layers to form a hierarchical Transformer structure; The multimodal input stream is processed based on the hierarchical Transformer structure to obtain meteorological-power time series characteristics.

4. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 3 is characterized in that: The feature weights in the correlation matrix are adjusted based on the preset cross-modal gating unit, specifically: Calculate the similarity between the multi-scale feature map and the meteorological-power time series features, and construct the correlation matrix; Generate a gating value through an activation function based on the weighted sum of the association matrix and the pre-acquired weight matrix; The gated temporal features and image features are weighted and fused according to the gate value to obtain weighted features.

5. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 4 is characterized in that: The feature map and time series features are complementary enhanced based on the correlation matrix to generate fusion features, specifically: In the Transformer decoding layer, the temporal features are weightedly fused with the enhanced information of the 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 ability of image features to respond to temporal changes is enhanced, and new enhanced image features are obtained: Perform weighted fusion of the enhanced temporal features and image features to obtain preliminary fusion features; The initial fusion features are purified through the adversarial network to suppress inter-modal interference noise; Based on the de-icing domain knowledge graph, the purified features are guided and enhanced; The purified features and enhanced features are weightedly fused through a gating mechanism to generate the final fused features.

6. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 5 is characterized in that: The feature purification of the preliminary fusion features through the adversarial network is specifically as follows: Taking the preliminary fused features as input, extract high-order feature representations and generate purified features; Based on the purified features, the discriminator is used to evaluate and output the discrimination score; Define the generator loss and discriminator loss, and alternately train the generator and discriminator based on the discriminant score to continuously improve the feature purification effect, and finally obtain the fusion feature after the modal conflict is eliminated.

7. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 6 is characterized in that: Based on the de-icing domain knowledge graph, the purified features are guided and enhanced, specifically: Determine entity nodes and their relationships; Embed the knowledge graph to convert entity nodes and relationships into low-dimensional vector representations; The fused features and the transformed knowledge graph are embedded into a vector, and the relevance scores between the features and each knowledge entity are calculated through the attention mechanism; The features are weighted and adjusted according to the correlation scores to generate enhanced fusion features.

8. The real-time control method for power transmission lines based on CNN and Transformer collaboration according to claim 7 is characterized in that: The output transmission line control strategy is specifically as follows: Classify the predicted line safety risk levels and determine the current line status; Based on the current status, multiple transmission line control strategies are generated, and the transmission line control strategies are optimized to generate control instructions. The transmission line control strategies include generating corresponding de-icing operation instructions based on the predicted ice thickness value and the current operating parameters of the line, and dynamically adjusting the maximum allowable load, current limit and operation warning threshold or emergency shutdown of the line based on the predicted ice thickness value and meteorological-load development trends.

9. The system for real-time control of power transmission lines based on CNN and Transformer collaboration according to any one of claims 1 to 8, 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 the multimodal input stream based on a preset dual-channel deep interaction network. The dual-channel deep interaction network is constructed based on CNN and Transformer fitting; A weight adjustment module is used to establish a correlation matrix based on the feature map and the temporal features, wherein the correlation matrix includes adjusting the feature weights in the correlation matrix based on a preset cross-modal gating unit; Feature fusion module: used to perform feature complementation enhancement on feature maps and time series features based on the correlation matrix to generate fused features; Prediction and control module: used to build a prediction model based on the fusion features and output the transmission line control strategy.

Images