Power transmission line icing prediction method, system and device based on multi-layer perceptron and medium

Abstract

The invention relates to the technical field of icing prediction, and discloses a power transmission line icing prediction method, system and device based on a multilayer perceptron, and a medium, and the method comprises the steps: collecting the multi-source data of the icing of a power transmission line in a research region, carrying out the first processing, and constructing a three-dimensional sample matrix according to the data after the first processing; extracting local spatio-temporal features of the meteorological parameters by using a first attention mechanism, performing cross-modal interaction on meteorological, geographic and line features by using a second attention mechanism, and calculating attention weights among the features; fusing the local space-time features and the attention weight to generate a high-order feature vector; and splicing the high-order feature vectors and the original key features, inputting the spliced high-order feature vectors and the original key features into a multi-layer perceptron model for prediction, and outputting an icing thickness prediction result. According to the method, the limitation of a traditional model in the aspects of feature extraction and multi-source data fusion is broken through, and the accuracy and reliability of icing prediction are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of transmission line icing prediction technology, and in particular to a method, system, device and medium for transmission line icing prediction based on a multilayer sensor. Background Technology

[0002] Icing on transmission lines is one of the common natural disasters faced by power systems. The continuous accumulation of ice significantly increases the mechanical load on conductors and towers, easily leading to conductor galloping, hardware damage, insulator flashover, and even catastrophic tower collapses and line breaks. Therefore, accurate prediction of icing is a crucial prerequisite for proactively deploying de-icing measures and ensuring the safe and stable operation of the power grid.

[0003] Currently, mainstream icing prediction methods mainly include numerical model prediction methods based on physical mechanisms, statistical model prediction methods relying on weather forecasts, and data-driven model prediction methods using traditional machine learning algorithms. However, these existing methods all have certain shortcomings: physical models are computationally complex and extremely sensitive to the accuracy of input parameters, making it difficult to meet real-time prediction requirements; statistical models rely on the accuracy of weather forecasts, and forecast errors can be amplified under complex terrain or localized sudden weather conditions, leading to a sharp drop in reliability; traditional machine learning models, when dealing with the highly nonlinear, multi-temporal-scale coupled meteorological-geographical-line status features involved in icing formation, often suffer from insufficient feature extraction capabilities and are prone to getting trapped in local optima, resulting in limited prediction accuracy and weak generalization ability, especially in capturing extreme icing events.

[0004] Therefore, in order to address the problems of insufficient feature extraction quality and insufficient attention to effective information in the sequence in existing icing prediction methods, this invention provides a transmission line icing prediction model based on multilayer perceptron (MLP). By coupling sample convolution and interactive neural networks, it can better capture the short-term and long-term information relationships in the time features and improve prediction accuracy. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for predicting icing of transmission lines based on a multilayer perceptron, which solves the problems of insufficient feature extraction quality and insufficient attention to effective sequence information in existing icing prediction methods.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting icing on transmission lines based on a multilayer sensor, comprising: Multi-source data on icing of transmission lines within the study area were collected and processed in the first stage. A three-dimensional sample matrix was then constructed based on the processed data. Based on the three-dimensional sample matrix, the local spatiotemporal features of meteorological parameters are extracted using the first attention mechanism, and the cross-modal interaction of meteorological, geographical and route features is performed using the second attention mechanism to calculate the attention weights between features. The local spatiotemporal features and the attention weights are fused to generate a higher-order feature vector; The higher-order feature vector is concatenated with the original key features and then input into the multilayer perceptron model for prediction, outputting the ice thickness prediction result.

[0008] As a preferred embodiment of the transmission line icing prediction method based on a multilayer perceptron described in this invention, the step of extracting local spatiotemporal features of meteorological parameters using a first attention mechanism includes: The three-dimensional sample matrix is ​​input into a one-dimensional convolutional layer, and sliding convolution calculation is performed along the time dimension to capture local abrupt change patterns of meteorological parameters. Multi-scale convolutional kernels are used to extract local features at different time scales in parallel, with the kernel size set to three specifications: short-term mode, medium-term mode, and long-term mode. The extracted multi-scale local features are adaptively weighted and fused using the first weight vector; The first enhancement operation is performed on the fused features based on the gated convolution mechanism. The enhanced features are then used to detect meteorological abrupt changes. The feature information of key time steps is highlighted through the attention weighting mechanism, and local spatiotemporal features are output.

[0009] As a preferred embodiment of the transmission line icing prediction method based on multilayer perceptron described in this invention, the method of utilizing a second attention mechanism to perform cross-modal interaction of meteorological, geographical, and line features, and calculating the attention weights between features, includes: The first transformation operation converts the input features into corresponding query vectors, key vectors, and value vectors. The dot product similarity between features is calculated based on the query vector and the key vector, and the initial attention weights are obtained through dimensional scaling. The initial attention weights are converted into a probability distribution; the value vectors are weighted and summed based on the converted attention weights to obtain the feature representation after cross-modal interaction. The dynamic weight relationship between meteorological, geographical, and route features is extracted from the feature representation after the cross-modal interaction and used as the attention weight.

[0010] The beneficial effects of this preferred technical solution are: to realize cross-modal interaction of meteorological, geographical and route features by using a multi-head attention mechanism, thereby deeply exploring the intrinsic correlation and key coupling relationship between each feature.

[0011] As a preferred embodiment of the transmission line icing prediction method based on a multilayer perceptron described in this invention, the output icing thickness prediction result includes: Based on the prediction error triggering model weight recalibration mechanism, when the prediction error exceeds a preset threshold, the convolutional layer weights and attention mechanism parameters are adjusted. The calibrated multilayer perceptron model outputs the predicted ice thickness; The corresponding risk level is determined based on the predicted ice thickness, and the prediction result containing both ice thickness and risk level is finally output.

[0012] The beneficial effects of this preferred technical solution are: by continuously optimizing model parameters through a dynamic calibration mechanism, accurate prediction of icing thickness and risk level can be achieved.

[0013] As a preferred embodiment of the transmission line icing prediction method based on multilayer perceptron described in this invention, the step of collecting multi-source data on transmission line icing within the study area and performing a first processing includes: Multi-source data on icing cases of transmission lines within the study area were collected, including geographic information, line information, and meteorological parameters, and daily meteorological reanalysis data were obtained for the same period. Using the location of the transmission tower as a spatial reference point, meteorological station data within a 1km radius are aggregated, and the inverse distance weighting method is used for data fusion. The timestamps of heterogeneous data are unified to the same sampling frequency, and the missing values ​​are filled by spatiotemporal KNN interpolation based on the synchronous data of neighboring towers.

[0014] As a preferred embodiment of the transmission line icing prediction method based on multilayer perceptron described in this invention, the construction of the three-dimensional sample matrix includes: The data after the first processing is standardized, and discrete geographic features are embedded into a 32-dimensional dense vector based on a sliding time window; a three-dimensional sample matrix is ​​constructed based on the 32-dimensional dense vector, and the dimensions of the three-dimensional sample matrix are (number of samples, time step, number of features).

[0015] As a preferred embodiment of the transmission line icing prediction method based on a multilayer perceptron described in this invention, the multilayer perceptron model includes: Set up a network structure that includes four hidden layers; The ReLU function is used as the activation function in the hidden layer. An adaptive moment estimation algorithm is used as the network optimizer, and the network is divided into training and testing sets according to a preset ratio for model training and validation.

[0016] Secondly, the present invention provides a transmission line icing prediction system based on a multilayer sensor, comprising: The data acquisition and processing module is used to collect multi-source data on icing of transmission lines in the study area and perform first processing, and construct a three-dimensional sample matrix based on the first processed data. The feature extraction module is used to extract local spatiotemporal features of meteorological parameters based on the three-dimensional sample matrix using a first attention mechanism, and to perform cross-modal interaction of meteorological, geographical and route features using a second attention mechanism, and to calculate the attention weights between features. The feature recombination module is used to fuse the local spatiotemporal features and the attention weights to generate a high-order feature vector; The prediction output module is used to concatenate the high-order feature vector with the original key features and input them into the multilayer perceptron model for prediction, and output the ice thickness prediction result.

[0017] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of a transmission line icing prediction method based on a multilayer perceptron.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a transmission line icing prediction method based on a multilayer perceptron.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention employs multi-scale convolution combined with an attention mechanism to extract local spatiotemporal features of meteorological parameters, and utilizes a multi-head attention mechanism to achieve cross-modal interaction of meteorological, geographical, and route features, thereby deeply exploring the intrinsic correlations and key coupling relationships between various features. This invention fuses the extracted high-order features with the original key features and inputs them into a multilayer perceptron model. By introducing a more complex deep network structure, it significantly enhances the model's ability to automatically learn deep and abstract features from multi-source data. This invention combines advanced optimization algorithms and activation function design to effectively overcome gradient problems and improve the stability and convergence of model training; combined with a dynamic calibration mechanism, it continuously optimizes model parameters to achieve accurate prediction of icing thickness and risk level. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall process logic of a transmission line icing prediction method based on a multilayer sensor provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for predicting icing on transmission lines based on a multilayer perceptron is provided, such as... Figure 1 The specific steps shown are as follows: S100: Collect multi-source data on icing of transmission lines within the study area and perform the first processing; construct a three-dimensional sample matrix based on the data after the first processing. S200: Based on a three-dimensional sample matrix, it uses a first attention mechanism to extract local spatiotemporal features of meteorological parameters, and a second attention mechanism to perform cross-modal interaction on meteorological, geographical and route features, and calculates the attention weights between features; S300: It fuses local spatiotemporal features and attention weights to generate a high-order feature vector; S400: The high-order feature vector is concatenated with the original key features and then input into the multilayer perceptron model for prediction, outputting the ice thickness prediction result.

[0024] It should be noted that, to address the issues of insufficient feature extraction quality and inadequate attention to effective sequence information in existing icing prediction methods, steps S100-S400 employ multi-scale convolution combined with an attention mechanism to extract local spatiotemporal features of meteorological parameters. A multi-head attention mechanism is used to achieve cross-modal interaction between meteorological, geographical, and route features, thereby deeply exploring the intrinsic correlations and key coupling relationships between various features. The extracted high-order features are fused with the original key features and input into a multilayer perceptron model. By introducing a more complex deep network structure, the model's ability to automatically learn deep and abstract features from multi-source data is significantly enhanced. Combined with advanced optimization algorithms and activation function design, gradient problems are effectively overcome, improving the stability and convergence of model training. A dynamic calibration mechanism continuously optimizes model parameters, achieving accurate prediction of icing thickness and risk level.

[0025] Example 2, based on the previous example, provides a specific implementation of a transmission line icing prediction method based on a multilayer sensor, to illustrate the technical means used in this method.

[0026] In this embodiment of the invention, step S100 involves collecting multi-source data on icing of transmission lines within the study area and performing a first processing step, then constructing a three-dimensional sample matrix based on the processed data. Specifically, the study area was set as Yunnan Province. Ice accretion cases of power transmission lines in Yunnan over the past 10 years, along with corresponding meteorological elements during these periods, were collected through tensile sensor observations or manual ice observation. The collected multi-source data included geographic information, line information, and meteorological parameters. Meteorological elements included variables such as 2-meter air temperature, 2-meter specific humidity, 10-meter wind speed, precipitation, and surface air pressure. Simultaneously, daily meteorological reanalysis data for the same period were acquired. Specifically, the first processing of the collected multi-source data includes: Using the location of the transmission tower as a spatial reference point, meteorological station data within a 1km radius are aggregated, expressed by the following formula: Among them, V i It is the value of the i-th weather station, d i It is the distance to the transmission tower, and p is a power parameter (usually taken as 2); at the same time, the timestamps of heterogeneous data are unified to the same sampling frequency (e.g., 10 minutes / time), and the missing values ​​are filled by spatiotemporal KNN interpolation based on the synchronous data of neighboring towers.

[0027] The numerical features are standardized using the Z-Score, and the calculation formula is as follows: in, The characteristic mean, The standard deviation is denoted as .

[0028] In an optional embodiment, the first processing step may also be to establish a spatial interpolation model based on a geographic information system, combine digital elevation data to perform terrain correction and spatial downscaling processing on meteorological elements to generate a rasterized meteorological field with higher spatial resolution, and then extract the accurate meteorological sequence of the transmission tower location through bilinear interpolation.

[0029] In another optional embodiment, the first processing step may also employ a deep representation learning method based on variational autoencoders to jointly encode and decouple features from multi-source heterogeneous data, complete data alignment and missing value reconstruction in the latent space, and preserve the semantic correlation between various features, ultimately outputting a feature representation with unified dimensions and enhanced semantics.

[0030] In this embodiment of the invention, constructing a three-dimensional sample matrix based on the first processed data includes: The data after the first processing is standardized, and discrete geographic features are embedded into a 32-dimensional dense vector based on a sliding time window (e.g., 6 hours, i.e., 36 time steps, assuming one data point every 10 minutes). A three-dimensional sample matrix is ​​constructed based on the 32-dimensional dense vector, and the dimensions of the three-dimensional sample matrix are (number of samples, time step, number of features).

[0031] It should be noted that step S100 above systematically collects and integrates heterogeneous data from multiple sources, such as meteorological, geographical, and line status data, to construct a structured three-dimensional sample matrix. This not only unifies the spatiotemporal reference of the data but also enhances the integrity and consistency of the data through spatial aggregation and missing value imputation, laying a solid foundation for accurately capturing the complex environmental factors that contribute to icing formation.

[0032] In this embodiment of the invention, step S200, based on a three-dimensional sample matrix, uses a first attention mechanism to extract local spatiotemporal features of meteorological parameters, and uses a second attention mechanism to perform cross-modal interaction on meteorological, geographical, and route features, calculating the attention weights between features, including the following sub-steps B1 and B2: In B1: The local spatiotemporal features of meteorological parameters are extracted using the first attention mechanism; the detailed steps include: The three-dimensional sample matrix is ​​input into a one-dimensional convolutional layer, and sliding convolution calculation is performed along the time dimension to capture the local abrupt change pattern of meteorological parameters. Multi-scale convolutional kernels are used to extract local features at different time scales in parallel, with the kernel size set to three specifications: short-term mode, medium-term mode, and long-term mode. The extracted multi-scale local features are adaptively weighted and fused using the first weight vector; The first enhancement operation is performed on the fused features based on the gated convolution mechanism. The enhanced features are then used to detect meteorological abrupt changes. The feature information of key time steps is highlighted through the attention weighting mechanism, and local spatiotemporal features are output.

[0033] Specifically, a 1D convolutional layer is used to slide along the time dimension to capture local abrupt change patterns in meteorological parameters, outputting meteorological spatiotemporal features from cross-modal feature interactions and feature recombination. Let the input tensor be X∈RN×T×F, where N is the number of samples, T is the time step, and F is the number of features. The number of channels is 64, and the stride is 1. The output meteorological features are calculated as follows: in, This represents the convolution operation, where W is the convolution kernel weight matrix, b is the bias, and σ is the activation function.

[0034] Specifically, feature extraction capabilities are enhanced by introducing multi-scale, gating mechanisms, and attention enhancement while maintaining the core structure. The number of cores is set to 3 (short-term mode), 5 (medium-term mode), and 7 (long-term mode), and multi-scale convolution calculations are performed. Simultaneously, the three sets of meteorological features are adaptively fused using the following formula: Where α, β, and γ represent the learnable channel weight vectors, which are modified based on the calculation results.

[0035] In an optional embodiment, the first weight vector may also employ a dynamic weight generation method based on a channel attention mechanism, capturing channel-level statistical features through global average pooling, and then generating an adaptive weight vector related to the input content via a fully connected layer and a non-linear activation function.

[0036] In another alternative embodiment, the first weight vector may also introduce a temperature-regulated softmax transformation to probabilistically remap the initial weights. By controlling the concentration of the weight distribution through a learnable temperature parameter, the model can dynamically adjust the fusion ratio of multi-scale features according to the importance of the input features.

[0037] Specifically, the first enhancement operation is performed on the fused features based on the gated convolution mechanism. The formula for generating the gated signal is as follows: Where Sigmoid is a non-linear activation function. The gated feature transformation formula is: Furthermore, the two features are connected: In an optional embodiment, the first enhancement operation may also introduce a residual gating mechanism, which adds the original multi-scale fused features to the gated transformation features element-wise through skip connections, thereby enhancing the feature representation ability while preserving the original information, and ensuring training stability through layer normalization.

[0038] In another alternative embodiment, the first enhancement operation may also employ a multi-head gated convolution structure to perform multiple independent gated convolution transformations in parallel, concatenating and linearly projecting the output features of each path, thereby increasing the number of parallel paths to enhance the model's feature enhancement capabilities for complex weather patterns.

[0039] Specifically, the enhanced features are used to detect meteorological abrupt changes, and an attention-weighted mechanism is employed to highlight feature information at key time steps. The formula is as follows: Where M represents the detection function. The final output is the local spatiotemporal features: In an optional embodiment, the first attention mechanism can also capture global temporal dependencies by constructing a time-series self-attention module, calculating the autocorrelation matrix between time steps, and using a sliding window mechanism to focus on local key time periods, thereby achieving collaborative perception of long- and short-term patterns of meteorological sequences.

[0040] In another alternative embodiment, the first attention mechanism may also employ a spatiotemporal graph attention network, which constructs meteorological stations as graph-structured nodes, represents spatial correlation through edge connections, and performs attention weighting synchronously in the spatiotemporal dimension, thereby simultaneously capturing the abrupt changes in meteorological parameters in time and space.

[0041] In B2: A second attention mechanism is used to perform cross-modal interaction of meteorological, geographical, and route features, and attention weights between features are calculated; detailed steps include: The first transformation operation converts the input features into corresponding query vectors, key vectors, and value vectors. The dot product similarity between features is calculated based on the query vector and the key vector, and the initial attention weights are obtained by dimensional scaling. The initial attention weights are converted into a probability distribution; the value vectors are then weighted and summed based on the converted attention weights to obtain the feature representation after cross-modal interaction. The dynamic weight relationship between meteorological, geographical, and route features is extracted from the feature representation after cross-modal interaction and used as attention weights.

[0042] Specifically, the input features are transformed into a query vector (Q), a key vector (K), and a value vector (V) through a linear transformation based on the feature projection transformation.

[0043] In an optional embodiment, the first transformation operation may also employ a factorization transformation based on low-rank projection, which reduces computational complexity by decomposing the fully connected weight matrix into the product of two low-rank matrices while maintaining the expressive power of the feature projection.

[0044] In another alternative embodiment, the first transformation operation may also introduce a multi-head grouped linear transformation, which groups the input features by channel dimension and then performs linear projection in parallel, thereby enhancing the model's adaptability to different semantic features by increasing the diversity of transformation paths.

[0045] Specifically, attention weights are calculated based on the similarity between the query and the key. The formula for calculating the attention weight matrix is ​​as follows: in, Calculate the dot product similarity between features; dk is the dimension of the key and query vectors, used for scaling to prevent gradient vanishing; softmax converts the similarity into a probability distribution.

[0046] Specifically, feature relevance is extracted from attention weights, expressed by the following formula: Where h represents the number of heads, A k That is, the attention weight matrix.

[0047] In an optional embodiment, the second attention mechanism can also be a hierarchical cross-modal attention mechanism. First, self-attention calculation is performed within each modality to extract key information within the modality. Then, interactive fusion between modalities is achieved through a cross-attention layer, thereby establishing a hierarchical association mapping between meteorological, geographical and route features.

[0048] In another alternative embodiment, the second attention mechanism may also employ a memory-enhanced attention network, which introduces trainable memory units to store typical cross-modal interaction patterns and queries the memory fragment most relevant to the current input when calculating attention weights, thereby enhancing the model's ability to model complex feature relationships.

[0049] It should be noted that step S200 innovatively adopts a combination of multi-scale convolutional structure and attention mechanism to extract discriminative local spatiotemporal features from meteorological data. At the same time, it explicitly models the dynamic coupling relationship between meteorological, geographical and line features through cross-modal attention mechanism, which can adaptively focus on key feature interactions and significantly improve the model's ability to characterize complex nonlinear correlations in the ice formation process.

[0050] In this embodiment of the invention, step S300, which fuses local spatiotemporal features and attention weights to generate a high-order feature vector, includes: Specifically, local spatiotemporal features and attention weights are concatenated, and the concatenated features are normalized to generate a high-order feature vector.

[0051] It should be noted that step S300 above generates a high-order feature vector representing the ice formation mechanism by deeply fusing local spatiotemporal features with cross-modal attention weights. This not only preserves the detailed information of the original data but also strengthens the synergistic effect between different modal features, making the final feature representation more discriminative.

[0052] In this embodiment of the invention, step S400 concatenates the high-order feature vector with the original key features and inputs the result into the multilayer perceptron model for prediction, outputting the ice thickness prediction result, including: Specifically, the higher-order feature vector is concatenated with the original key features and then input into a multilayer perceptron (MLP). Let the reconstructed feature be F. fused ∈RN×T×D, the original key feature is K∈R N×T×M (M is the number of key features), then the concatenated feature representation is: F enhanced =[F fused ,K] The dimensions become (N, T, D+M).

[0053] Specifically, the multilayer perceptron model includes: Set up a network structure with four hidden layers, each containing 100 neurons; Use the ReLU function as the activation function in the hidden layer; An adaptive moment estimation algorithm was used as the network optimizer, with a learning rate of 0.1. The network was divided into training and testing sets according to a preset ratio for model training and validation.

[0054] In this embodiment of the invention, the processing of the multilayer perceptron model specifically includes: inputting the concatenated enhanced feature vector into a deep network structure consisting of four fully connected layers, each hidden layer containing 100 neurons and undergoing nonlinear transformation through the ReLU activation function; extracting and combining high-order abstract patterns from the features layer by layer; optimizing the network parameters using an adaptive moment estimation algorithm with a learning rate of 0.1; calculating the prediction results through forward propagation and updating the weights through backpropagation; and finally generating the ice thickness value in the output layer. The entire process achieves continuous calibration of the model parameters through dynamic error monitoring to ensure the stability of the prediction accuracy.

[0055] Specifically, the output icing thickness prediction results include: Based on the prediction error triggering model weight recalibration mechanism, when the prediction error exceeds a preset threshold of 15%, the convolutional layer weights and attention mechanism parameters are adjusted. The calibrated multilayer perceptron model outputs the predicted ice thickness; The corresponding risk level is determined based on the predicted ice thickness, and the final output is a prediction result that includes both the ice thickness and the risk level.

[0056] It should be noted that step S400 above, by concatenating high-order features with the original key features and then inputting the result into a multilayer perceptron, fully utilizes the powerful function approximation capability of deep neural networks. Combined with a dynamic weight calibration mechanism, the model can automatically adjust parameters based on prediction errors, ensuring not only prediction accuracy but also significantly improving the model's adaptability and robustness under different operating conditions, thus achieving accurate prediction of icing thickness and risk level.

[0057] Example 3: This example provides a transmission line icing prediction system based on a multilayer sensor, comprising: The data acquisition and processing module is used to collect multi-source data on icing of transmission lines in the study area and perform the first processing, and construct a three-dimensional sample matrix based on the data after the first processing. The feature extraction module is used to extract local spatiotemporal features of meteorological parameters based on a three-dimensional sample matrix using a first attention mechanism, and to perform cross-modal interaction of meteorological, geographical and route features using a second attention mechanism, and to calculate the attention weights between features. The feature reorganization module is used to fuse local spatiotemporal features and attention weights to generate high-order feature vectors; The prediction output module is used to concatenate the high-order feature vector with the original key features and input them into the multilayer perceptron model for prediction, and output the ice thickness prediction result.

[0058] It should be noted that the technical solution of the transmission line icing prediction system based on multilayer sensor is based on the same concept as the technical solution of the transmission line icing prediction method based on multilayer sensor described above. For details not described in detail in the technical solution of the transmission line icing prediction system based on multilayer sensor in this embodiment, please refer to the description of the technical solution of the transmission line icing prediction method based on multilayer sensor described above.

[0059] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0060] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a transmission line icing prediction method based on a multilayer sensor. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0061] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0062] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0063] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0064] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for power transmission line icing prediction based on a multi-layer perceptron, characterized by, The method comprises the following steps: Collecting multi-source data of power transmission line icing in the research area and performing first processing, and constructing a three-dimensional sample matrix according to the first processed data; Based on the three-dimensional sample matrix, the local spatio-temporal features of meteorological parameters are extracted by using the first attention mechanism, the second attention mechanism is used for cross-modal interaction of meteorological, geographical and line features, and the attention weight between features is calculated; Fuse the local spatio-temporal features and the attention weight to generate a high-order feature vector; The high-order feature vector and the original key feature are spliced and input into a multi-layer perception model for prediction, and the icing thickness prediction result is output.

2. The method of claim 1, wherein the method is based on a multi-layer perceptron. The method for extracting local spatio-temporal features of meteorological parameters by using the first attention mechanism comprises: Input the three-dimensional sample matrix into a one-dimensional convolution layer, and perform sliding convolution calculation along the time dimension to capture the local mutation mode of meteorological parameters; Parallelly extract local features of different time scales using multi-scale convolution kernels, wherein the convolution kernel size is set to three specifications of short-term mode, medium-term mode and long-term mode; The first weight vector is used for adaptive weighted fusion of the extracted multi-scale local features; Based on the gating convolution mechanism, the first enhancement operation is performed on the fused features, the enhanced features are detected for meteorological mutation, the feature information of the key time step is highlighted through the attention weighting mechanism, and the local spatio-temporal features are output.

3. The method for predicting icing on transmission lines based on a multilayer perceptron as described in claim 2, characterized in that, The method for cross-modal interaction of meteorological, geographical and line features by using the second attention mechanism and calculating the attention weight between features comprises: Convert the input features into corresponding query vectors, key vectors and value vectors through a first transformation operation; Calculate the dot product similarity between features based on the query vectors and key vectors, and obtain the initial attention weight through dimension scaling processing; Convert the initial attention weight into a probability distribution form; based on the converted attention weight, the value vector is weighted and summed to obtain the feature representation after cross-modal interaction; Extract the dynamic weight relationship between meteorological, geographical and line features from the feature representation after cross-modal interaction as the attention weight.

4. The method for predicting icing on transmission lines based on a multilayer perceptron as described in claim 3, characterized in that, The method for outputting the icing thickness prediction result comprises: Based on the prediction error triggering model weight recalibration mechanism, when the prediction error exceeds the preset threshold, the convolution layer weight and the attention mechanism parameter are adjusted; Output the icing thickness prediction value through the calibrated multi-layer perception model; According to the icing thickness prediction value, determine the corresponding risk level, and finally output the prediction result containing the icing thickness and the risk level.

5. The method of claim 1, wherein the method is based on a multi-layer perceptron. The method for collecting multi-source data of power transmission line icing in the research area and performing first processing comprises: Collecting multi-source data of power transmission line icing cases in the research area, including geographical information, line information and meteorological parameters, and obtaining contemporaneous daily meteorological reanalysis data; Taking the power transmission tower position as the spatial reference point, aggregating the data of meteorological stations within a radius of 1km, and using the inverse distance weighting method for data fusion; Uniform the time stamp of heterogeneous data to the same sampling frequency, and use the spatio-temporal KNN interpolation method to fill in the missing values based on the contemporaneous data of adjacent towers.

6. The method of claim 5, wherein the method is based on a multi-layer perceptron. The method for constructing a three-dimensional sample matrix comprises: The first processed data is standardized, and discrete geographical features are embedded into 32-dimensional dense vectors based on a sliding time window; a three-dimensional sample matrix is constructed based on the 32-dimensional dense vectors, and the three-dimensional sample matrix has dimensions of (sample number, time step, feature number).

7. The method of claim 6, wherein the method is based on a multi-layer perceptron. The multi-layer perception model comprises: a network structure comprising four hidden layers is set; a ReLU function is used as an activation function in the hidden layers; an adaptive moment estimation algorithm is used as a network optimizer, and a preset proportion is divided into a training set and a test set for model training and verification.

8. A multi-layer perceptron based transmission line icing prediction system, applying a multi-layer perceptron based transmission line icing prediction method according to any one of claims 1 to 7, characterized in that, comprises: a data acquisition and processing module for acquiring multi-source data of power line icing in a research area and performing first processing, and constructing a three-dimensional sample matrix based on the first processed data; a feature extraction module for extracting local spatio-temporal features of meteorological parameters based on the three-dimensional sample matrix using a first attention mechanism, and performing cross-modal interaction on meteorological, geographical and line features using a second attention mechanism to calculate attention weights between features; a feature reorganization module for fusing the local spatio-temporal features and the attention weights to generate high-order feature vectors; a prediction output module for splicing the high-order feature vectors with original key features and inputting them into a multi-layer perception model for prediction to output an icing thickness prediction result. 9.An electronic device comprising a memory and a processor, the electronic device characterized by: The memory is used to store computer executable instructions, and the processor executes the computer executable instructions to realize the steps of the power line icing prediction method based on the multi-layer perception machine according to any one of claims 1-7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: The computer executable instructions are executed by the processor to realize the steps of the power line icing prediction method based on the multi-layer perception machine according to any one of claims 1-7.

Images