Power transmission line icing thickness prediction method, system and device based on CNN-BiLSTM

Abstract

The application discloses a CNN-BiLSTM-based power transmission line icing thickness prediction method, system and equipment, belongs to the technical field of power system disaster prevention, and comprises the following steps: S1, sequence data is formed according to historical multi-dimensional meteorological data collected at different time and historical icing data corresponding to the historical multi-dimensional meteorological data; S2, training samples are obtained by adaptively segmenting and intercepting the sequence data through a sliding window method; S3, an initial icing thickness prediction model is constructed based on CNN and BiLSTM, and the initial icing thickness prediction model is trained through the training samples to obtain a final icing thickness prediction model; and S4, real-time multi-dimensional meteorological data is input into the final icing thickness prediction model to obtain target icing thickness. The technical problem that the prior art is difficult to improve the icing thickness prediction precision is solved.

Description

Technical Field

[0001] This invention relates to the field of power system disaster prevention technology, specifically to a method, system, and equipment for predicting the icing thickness of transmission lines based on CNN-BiLSTM. Background Technology

[0002] Winter icing on transmission lines can easily cause major accidents such as power grid outages and tower collapses, seriously threatening the safe operation of the power grid. Uneven icing increases the mechanical load on conductors, making them more prone to breakage, and the weight of the ice layer can cause towers to become unbalanced and collapse. Icing also reduces the electrical performance of insulators, increasing the risk of short circuits. These consequences not only endanger power grid safety but may also cause prolonged power outages, affecting the normal operation of society. Therefore, it is urgent to develop methods for predicting line icing thickness, to assess icing thickness and risks in advance, to provide a scientific basis for power grid operation and maintenance, to reduce the frequency of power grid accidents in winter, and to improve the safety and stability of power grid operation. Currently, icing thickness prediction methods are mainly divided into two categories: one is based on physical models of thermodynamic equilibrium, such as the Makkonen model, which establishes an icing rate equation and combines parameters such as line radius, temperature, wind speed, and precipitation to model the mechanism. While these methods possess clear physical significance, they are limited by factors such as insufficient meteorological monitoring network coverage and complex mountainous terrain. Furthermore, when wind speed measurement errors are large, the prediction of icing thickness shows significant deviations, resulting in poor practical engineering applicability. Secondly, data-driven models, such as single convolutional neural networks or bidirectional recurrent networks, can only achieve spatial feature extraction or temporal evolution modeling independently. They do not comprehensively depict the icing evolution patterns under the coupled effects of multidimensional meteorological factors, thus leading to limited accuracy in icing thickness prediction. Therefore, improving the accuracy of icing thickness prediction is a pressing technical challenge that needs to be addressed by existing technologies. Summary of the Invention

[0003] To address the technical problem of imbalanced icing time-series samples and insufficient characterization of icing evolution patterns in existing technologies, which makes it difficult to improve the accuracy of icing thickness prediction, this invention provides a CNN-BiLSTM-based method, system, and device for predicting icing thickness of transmission lines. By using a sliding window method to adaptively segment and extract training samples from the sequence data, the imbalance of icing time-series samples is improved. Furthermore, by combining the advantages of CNN in multi-dimensional feature extraction and BiLSTM in time-series modeling, the complex dynamic characteristics of the icing process are accurately fitted from two levels: data optimization and model structure design. This solves the technical problem of existing technologies' inability to improve the accuracy of icing thickness prediction.

[0004] To address the aforementioned technical problems, this invention provides a method for predicting icing thickness on transmission lines based on CNN-BiLSTM, comprising the following steps: S1: The sequence data is composed of historical multidimensional meteorological data collected at different times and the corresponding historical icing data. S2: Training samples are obtained by adaptively segmenting and truncating the sequence data using the sliding window method; S3: Construct an initial ice thickness prediction model based on CNN and BiLSTM, and train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model; S4: Input the collected real-time multi-dimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

[0005] Preferably, the sequence data is composed of historical multidimensional meteorological data collected at different times and historical icing data corresponding to the historical multidimensional meteorological data, including: S11: Perform anomaly processing on historical multidimensional meteorological data and historical icing data, and perform time alignment processing on historical multidimensional meteorological data and historical icing data. S12: The historical multidimensional meteorological data after time alignment processing is spliced ​​and integrated with the historical icing data in chronological order to form sequence data; The historical multidimensional meteorological data includes historical temperature, historical relative humidity, historical wind speed, and historical atmospheric pressure.

[0006] Preferably, the anomaly processing of historical multidimensional meteorological data and historical icing data, and the time alignment processing of historical multidimensional meteorological data and historical icing data, include: S111: Based on historical multidimensional meteorological data collected from different locations, anomaly detection is performed on the historical multidimensional meteorological data of the target location, and the historical multidimensional meteorological data of the target location is corrected using the spatial interpolation method with historical multidimensional meteorological data collected from different locations. S112: Take the collection time period of the historical icing data of the target location as the target time set. If the historical multidimensional meteorological data corresponding to the target time set is obtained in the corrected historical multidimensional meteorological data, then execute S113. Otherwise, based on the auxiliary multidimensional meteorological data before and after the target time set for which no corresponding historical multidimensional meteorological data has been obtained, use linear interpolation to obtain the historical multidimensional meteorological data corresponding to the target time set, and execute S113. S113: Associate the historical multidimensional meteorological data corresponding to the target time set with the historical icing data of the target time set at the target location.

[0007] Preferably, the step of adaptively segmenting and extracting training samples from sequence data using the sliding window method includes: S21: Extract the first and second rates of change of ice thickness from historical icing data, extract the short-term fluctuation variance of meteorological characteristics from historical multidimensional meteorological data, construct an icing sensitivity index based on the first rate of change, the second rate of change, and the short-term fluctuation variance, and classify the sequence data based on the icing sensitivity index. S22: Obtain the sliding window length and sliding step size based on the characteristics of different types of sequence data, and segment the sequence data according to the sliding window length and sliding step size, using the sequence data in each segment as a group of training samples.

[0008] Preferably, the initial icing thickness prediction model based on CNN and BiLSTM includes: The initial ice thickness prediction model consists of a CNN containing two one-dimensional convolutional layers and corresponding ReLU activation functions, and a BiLSTM with two stacked bidirectional LSTM layers. The first one-dimensional convolutional layer in the one-dimensional convolutional layer includes 64 convolutional kernels of size 3, and the second one-dimensional convolutional layer in the one-dimensional convolutional layer includes 32 convolutional kernels of size 3; each bidirectional LSTM structure includes 128 hidden neurons.

[0009] Preferably, the step of training the initial icing thickness prediction model using training samples to obtain the final icing thickness prediction model includes: The training samples are normalized, and the normalized training samples are reconstructed into an adaptation tensor. The adaptation tensor is input into the CNN in the initial ice thickness prediction model to obtain local correlation features related to ice thickness. The local correlation features are folded, compressed and flattened to obtain the adaptation vector. The adaptation vector is input into the BiLSTM in the initial ice thickness prediction model to obtain comprehensive temporal features. The comprehensive temporal features are input into the fully connected layer to obtain the predicted icing thickness. If the predicted icing thickness does not match the actual icing thickness, the parameters of the initial icing thickness prediction model are corrected based on the thickness difference between the predicted and actual icing thicknesses and the icing sensitivity index to obtain the final icing thickness prediction model. Otherwise, the initial icing thickness prediction model is used as the final icing thickness prediction model.

[0010] Preferably, the step of correcting the parameters of the initial icing thickness prediction model based on the thickness difference between the predicted and actual icing thickness and the icing sensitivity index to obtain the final icing thickness prediction model includes: The thickness difference weights are obtained based on the icing sensitivity index corresponding to the training samples, and the thickness difference values ​​are weighted and summed based on the thickness difference weights to obtain the total loss value. The parameters of the initial icing thickness prediction model are corrected based on the total loss value to obtain the final icing thickness prediction model.

[0011] Preferably, the step of correcting the parameters of the initial icing thickness prediction model based on the total loss value to obtain the final icing thickness prediction model includes: The parameter gradient of the initial icing thickness prediction model is calculated through backpropagation based on the total loss value. The parameters are then iteratively updated according to the gradient descent direction to obtain the final icing thickness prediction model.

[0012] By adopting the above technical solution, the present invention has the following advantages: By using the sliding window method to adaptively segment and extract training samples from sequence data, the imbalance problem of ice-covering time-series samples is improved. Combining the advantages of CNN in multi-dimensional feature extraction and BiLSTM in time-series modeling, the complex dynamic characteristics of the ice-covering process are accurately fitted from two aspects: data optimization and model structure design. This solves the technical problem that existing technologies cannot improve the accuracy of ice thickness prediction. Specifically, during the initial training of the ice thickness prediction model, the total loss is obtained by weighted summation of the prediction errors of various training samples. By assigning higher weights to high-sensitivity training samples, their prediction errors account for the largest proportion of the total loss, thus generating a larger parameter gradient during backpropagation. Using this gradient to guide model parameter updates, the model can be guided to prioritize reducing the prediction errors of high-sensitivity samples during optimization. This effectively alleviates the problem of insufficient model learning for abrupt changes in conditions due to an excessively high proportion of stable samples, improves the model's ability to fit the patterns of drastic ice thickness changes, and ultimately enhances the overall prediction accuracy of ice thickness.

[0013] This invention also provides a CNN-BiLSTM-based transmission line icing thickness prediction system, applicable to the aforementioned CNN-BiLSTM-based transmission line icing thickness prediction method, comprising: The sequence data acquisition module is used to construct sequence data based on historical multidimensional meteorological data collected at different times and the corresponding historical icing data. The training sample acquisition module is used to adaptively segment and truncate the sequence data using the sliding window method to obtain training samples. The model acquisition module is used to build an initial ice thickness prediction model based on CNN and BiLSTM, and to train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model. The target data acquisition module is used to input the collected real-time multidimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

[0014] By adopting the above technical solution, the present invention has the following advantages: By using the sliding window method to adaptively segment and extract training samples from sequence data, the imbalance problem of ice-covering time-series samples is improved. Combining the advantages of CNN in multi-dimensional feature extraction and BiLSTM in time-series modeling, the complex dynamic characteristics of the ice-covering process are accurately fitted from two aspects: data optimization and model structure design. This solves the technical problem that existing technologies cannot improve the accuracy of ice thickness prediction.

[0015] The present invention also provides a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the CNN-BiLSTM-based transmission line icing thickness prediction method. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0017] Figure 1 This is a flowchart illustrating the method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to the present invention. Figure 2 This is a schematic diagram of the structure of the initial icing thickness prediction model of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0020] Example 1: like Figure 1 As shown, the method for predicting icing thickness of transmission lines based on CNN-BiLSTM includes the following steps: S1: The sequence data is composed of historical multidimensional meteorological data collected at different times and the historical icing data corresponding to the historical multidimensional meteorological data.

[0021] As an optional embodiment, the step of constructing sequence data based on historical multidimensional meteorological data collected at different times and historical icing data corresponding to the historical multidimensional meteorological data includes: S11: Perform anomaly processing on historical multidimensional meteorological data and historical icing data, and perform time alignment processing on historical multidimensional meteorological data and historical icing data. S12: The historical multidimensional meteorological data after time alignment processing is spliced ​​and integrated with the historical icing data in chronological order to form sequence data; The historical multidimensional meteorological data includes historical temperature, historical relative humidity, historical wind speed, and historical atmospheric pressure.

[0022] Specifically, the anomaly processing of historical multidimensional meteorological data and historical icing data, and the time alignment processing of historical multidimensional meteorological data and historical icing data, include: S111: Based on historical multidimensional meteorological data collected from different locations, anomaly detection is performed on the historical multidimensional meteorological data of the target location, and the historical multidimensional meteorological data of the target location is corrected using the spatial interpolation method with historical multidimensional meteorological data collected from different locations. S112: Take the collection time period of the historical icing data of the target location as the target time set. If the historical multidimensional meteorological data corresponding to the target time set is obtained in the corrected historical multidimensional meteorological data, then execute S113. Otherwise, based on the auxiliary multidimensional meteorological data before and after the target time set for which no corresponding historical multidimensional meteorological data has been obtained, use linear interpolation to obtain the historical multidimensional meteorological data corresponding to the target time set, and execute S113. S113: Associate the historical multidimensional meteorological data corresponding to the target time set with the historical icing data of the target time set at the target location.

[0023] In this embodiment, historical multidimensional meteorological data is acquired through unmanned meteorological stations deployed along the transmission line corridor. Considering that transmission lines are mostly located in mountainous areas with dispersed monitoring points and complex terrain, individual monitoring points are prone to data anomalies, missing time sequences, and time synchronization issues. Furthermore, meteorological data and icing data are collected from separate devices, which can lead to timestamp mismatches. Therefore, anomaly detection and correction are first achieved through spatial comparison of multiple stations. Then, the missing meteorological data is supplemented and time alignment is completed based on the collection time of the icing data, ensuring reliable, complete, and strictly synchronized data, laying the data foundation for subsequent high-precision forecasts. It is important to emphasize that time alignment refers to temporally matching and associating historical multidimensional meteorological data within 24 hours with historical icing data at the next moment, establishing a temporal correspondence where meteorological data precedes icing data. For example, the historical multidimensional meteorological data from 0:00 to 24:00 on day T is time-aligned with the historical icing data at 1:00 on day T+1.

[0024] In this embodiment, taking the icing monitoring scenario of a power transmission line in a mountainous area as an example, the target location is point A in the middle section of the line, with two auxiliary meteorological monitoring stations, B and C, deployed around it. The data collection period is 24 consecutive hours with a time interval of 1 hour. During anomaly detection and spatial interpolation correction, temperature, relative humidity, wind speed, and atmospheric pressure data for point A and auxiliary stations B and C at the same time period are acquired. By comparing the numerical distribution of the three data points, abnormal data are identified at point A: abnormally high wind speed in the 5th hour and abnormally low temperature in the 12th hour. Inverse distance weighted spatial interpolation is used, and normal meteorological data from stations B and C are used to correct the abnormal data at point A, resulting in complete and reliable multidimensional meteorological time-series data for point A. During linear interpolation to complete missing data, 24 time points from the historical icing data of point A are used as the target time set. Upon verification, it is found that the corrected meteorological data is missing two sets of data for the 8th and 18th hours. Using adjacent meteorological data from the 7th and 9th hours and the 17th and 19th hours respectively as references, linear interpolation was employed to calculate the temperature, humidity, wind speed, and atmospheric pressure at the missing time points, thus completing the time-series meteorological data. Collaborative anomaly detection and spatial interpolation correction using multi-site meteorological data effectively eliminated anomalous noise introduced by single-point monitoring, significantly improving the accuracy and reliability of meteorological data at the target location. For periods with missing meteorological data, linear interpolation was performed based on adjacent valid meteorological data before and after the missing time points to fill the temporal gaps, ensuring the integrity and continuity of the sequence data and providing a high-quality data foundation for subsequent model training.

[0025] S2: Adaptive segmentation and truncation of sequence data using the sliding window method to obtain training samples.

[0026] The step of adaptively segmenting and extracting training samples from sequence data using the sliding window method includes: S21: Extract the first and second rates of change of ice thickness from historical icing data, extract the short-term fluctuation variance of meteorological characteristics from historical multidimensional meteorological data, construct an icing sensitivity index based on the first rate of change, the second rate of change, and the short-term fluctuation variance, and classify the sequence data based on the icing sensitivity index. S22: Obtain the sliding window length and sliding step size based on the characteristics of different types of sequence data, and segment the sequence data according to the sliding window length and sliding step size, using the sequence data in each segment as a group of training samples.

[0027] Understandably, the first-order rate of change of ice thickness characterizes the rate of change of ice thickness, while the second-order rate of change characterizes the acceleration of the change in ice thickness. Both are used together to determine whether ice is in a state of rapid thickening or rapid melting. Short-time variance characterizes the intensity of fluctuations in meteorological data over a short period; greater meteorological fluctuations indicate more drastic environmental changes and a greater likelihood of abrupt changes in ice thickness, while less drastic changes indicate a relatively stable environment and a relatively unchanged ice condition. Understandably, the faster the change in ice thickness and the more intense the meteorological fluctuations, the higher the ice sensitivity index. The sequence data are categorized as high-sensitivity, medium-sensitivity, and low-sensitivity. High-sensitivity sequence data is characterized by a sudden increase or decrease in ice thickness and drastic fluctuations in meteorological parameters; medium-sensitivity sequence data is characterized by slow ice growth and minor meteorological fluctuations; and low-sensitivity sequence data is characterized by stable ice thickness and stable meteorological parameters. The sliding window length and sliding step size decrease sequentially from low-sensitivity to medium-sensitivity to high-sensitivity sequence data. In this embodiment, considering the significant differences in icing thickness and meteorological parameters across different time periods, the redundancy of data during stable periods, and the sparse features during periods of abrupt changes, using a fixed window for sample extraction would lead to insufficient learning of the patterns of sudden icing changes in the model. Therefore, an icing sensitivity index is first constructed using the first and second rates of change of icing thickness and the variance of short-term meteorological fluctuations to quantitatively distinguish the degree of sequence change. Then, differentiated window parameters are assigned according to categories, allowing the sampling strategy to adaptively match the severity of data changes. This optimizes the sample structure from the source, thereby improving the prediction accuracy of the final icing thickness prediction model.

[0028] S3: Construct an initial ice thickness prediction model based on CNN and BiLSTM, and train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model.

[0029] Specifically, the initial icing thickness prediction model based on CNN and BiLSTM includes: The initial ice thickness prediction model consists of a CNN containing two one-dimensional convolutional layers and corresponding ReLU activation functions, and a BiLSTM with two stacked bidirectional LSTM layers. The first one-dimensional convolutional layer in the one-dimensional convolutional layer includes 64 convolutional kernels of size 3, and the second one-dimensional convolutional layer in the one-dimensional convolutional layer includes 32 convolutional kernels of size 3; each bidirectional LSTM structure includes 128 hidden neurons.

[0030] In this embodiment, the structure of the initial icing thickness prediction model is as follows: Figure 2 As shown, the CNN consists of two consecutive one-dimensional convolutional layers. The first convolutional layer uses 64 kernels of size 3 with a stride of 1, employs Same padding to keep the sequence length unchanged, and is followed by the ReLU activation function. The second convolutional layer uses 32 kernels of size 3, also with ReLU activation.

[0031] In some embodiments, training the initial icing thickness prediction model with training samples to obtain the final icing thickness prediction model includes: The training samples are normalized, and the normalized training samples are reconstructed into an adaptation tensor. The adaptation tensor is input into the CNN in the initial ice thickness prediction model to obtain local correlation features related to ice thickness. The local correlation features are folded, compressed and flattened to obtain the adaptation vector. The adaptation vector is input into the BiLSTM in the initial ice thickness prediction model to obtain comprehensive temporal features. The comprehensive temporal features are input into the fully connected layer to obtain the predicted icing thickness. If the predicted icing thickness does not match the actual icing thickness, the parameters of the initial icing thickness prediction model are corrected based on the thickness difference between the predicted and actual icing thicknesses and the icing sensitivity index to obtain the final icing thickness prediction model. Otherwise, the initial icing thickness prediction model is used as the final icing thickness prediction model.

[0032] The min-max normalization method is used to scale the parameters in the training samples to the [0,1] interval. The calculation formula is as follows: , This represents the normalized parameter value. This represents the parameter values ​​before normalization. and These represent the minimum and maximum values ​​of the parameters in the training samples, respectively. Understandably, the adaptation tensor is specifically a four-dimensional tensor (number of training samples × time step × feature dimension × number of channels) to meet the input requirements of CNN. CNN is specifically used to extract local correlation features of meteorological parameters in the time dimension. Local correlation features refer to the mutual influence relationships between multiple meteorological parameters such as temperature, relative humidity, wind speed, and atmospheric pressure at the same moment or within a very short time window, as well as the local coupling rules between the combination of these meteorological parameters and changes in ice thickness. For the local correlation features extracted by CNN, a global average pooling layer is first used for feature folding and compression, and then a Flatten layer is used to flatten them into a one-dimensional feature vector, obtaining an adaptation vector adapted to the BiLSTM input structure. In BiLSTM, the forward LSTM layer propagates forward from the beginning to the end of the sequence, capturing the cumulative effect of historical weather conditions; the backward LSTM layer propagates backward from the end of the sequence, fusing predictive information about future weather trends. The hidden states output by the two layers are concatenated at each time step to form a comprehensive temporal feature. The fully connected layers nonlinearly transform the comprehensive temporal feature output by BiLSTM through two fully connected layers. The first fully connected layer contains 64 neurons activated by ReLU, and the second fully connected layer contains 1 neuron (corresponding to the predicted ice thickness) and uses the softmax activation function to obtain the target ice thickness. By combining the advantages of CNN in multidimensional feature extraction with BiLSTM in temporal modeling, the accuracy of ice thickness prediction is improved.

[0033] In some embodiments, the step of correcting the parameters of the initial icing thickness prediction model based on the thickness difference between the predicted and actual icing thickness and the icing sensitivity index to obtain the final icing thickness prediction model includes: The thickness difference weights are obtained based on the icing sensitivity index corresponding to the training samples, and the thickness difference values ​​are weighted and summed based on the thickness difference weights to obtain the total loss value. The parameters of the initial icing thickness prediction model are corrected based on the total loss value to obtain the final icing thickness prediction model.

[0034] Specifically, the step of correcting the parameters of the initial icing thickness prediction model based on the total loss value to obtain the final icing thickness prediction model includes: The parameter gradient of the initial icing thickness prediction model is calculated through backpropagation based on the total loss value. The parameters are then iteratively updated according to the gradient descent direction to obtain the final icing thickness prediction model.

[0035] Considering that conventional models treat all sample errors equally, they are easily dominated by the large proportion of stable icing samples, leading to excessively large prediction errors for critical and dangerous conditions such as sudden increases and decreases in icing thickness. Therefore, by assigning higher weights to the thickness differences corresponding to high-sensitivity training samples, their prediction errors account for the largest proportion of the total loss, thus generating a larger parameter gradient during backpropagation. Using this gradient to guide model parameter updates, the model can be guided to prioritize reducing the prediction errors of high-sensitivity samples during optimization, effectively alleviating the problem of insufficient learning of abrupt changes in conditions due to an excessively high proportion of stable samples, improving the model's ability to fit the patterns of drastic icing changes, and ultimately improving the overall prediction accuracy of icing thickness. The expression for the model's loss function is: , This represents the total loss value. Indicates the number of training samples. The weights represent the thickness differences corresponding to the k-th training sample. This represents the predicted icing thickness obtained by training using the k-th training sample. This represents the actual icing thickness corresponding to the kth training sample.

[0036] To quantitatively analyze the prediction performance of the initial icing thickness prediction model, root mean square error was used. Mean absolute error With the coefficient of determination As an evaluation metric for the performance of the initial icing thickness prediction model: ; ; ; This represents the average actual icing thickness corresponding to all training samples. In this embodiment, the final icing thickness prediction model achieves a determination coefficient of 0.97636 for the training set, with a root mean square error (RMSE) of only 0.068883. The predicted and measured icing thickness values ​​are closely distributed along a 45° diagonal, indicating that the model effectively utilizes the spatiotemporal features of the training samples and possesses good fitting ability. Although the determination coefficient for the test set slightly decreases to 0.87388, the RMSE remains controlled at 0.083528. The data point dispersion band is narrow, with only a slight right skew in the high-icing section, indicating good generalization performance and no significant overfitting. Overall, the final icing thickness prediction model can estimate icing thickness with high accuracy, laying a reliable foundation for subsequent interval prediction and risk warning. In the training set, the predicted and actual icing thickness values ​​show a high degree of consistency in overall trend, demonstrating the model's effective capture of data features during the training phase. The prediction results generally fluctuate around the actual values, indicating that the model has good fitting ability. The test set results also showed high consistency, with predicted values ​​closely following the actual value trends, exhibiting only slight deviations in local areas. This indicates that the model possesses strong generalization ability and can effectively predict ice thickness for new data. Overall, the final ice thickness prediction model demonstrated reliable time-series prediction performance during both the training and testing phases, providing a solid foundation for ice thickness prediction.

[0037] S4: Input the collected real-time multi-dimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

[0038] Understandably, real-time multidimensional meteorological data includes real-time temperature, real-time relative humidity, real-time wind speed, and real-time atmospheric pressure. By using training samples obtained through adaptive segmentation and combining them with a final icing thickness prediction model acquired through CNN and BiLSTM, the accuracy of icing thickness prediction is significantly improved.

[0039] Example 2: This embodiment also provides a CNN-BiLSTM-based transmission line icing thickness prediction system, applicable to the aforementioned CNN-BiLSTM-based transmission line icing thickness prediction method, including: The sequence data acquisition module is used to construct sequence data based on historical multidimensional meteorological data collected at different times and the corresponding historical icing data. The training sample acquisition module is used to adaptively segment and truncate the sequence data using the sliding window method to obtain training samples. The model acquisition module is used to build an initial ice thickness prediction model based on CNN and BiLSTM, and to train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model. The target data acquisition module is used to input the collected real-time multidimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

[0040] Example 3: This embodiment also provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions that the processor can execute. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the CNN-BiLSTM-based transmission line icing thickness prediction method.

[0041] The specific embodiments described above are preferred embodiments of the CNN-BiLSTM-based transmission line icing thickness prediction method, system, and equipment of the present invention, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting icing thickness on transmission lines based on CNN-BiLSTM, characterized in that, Includes the following steps: S1: The sequence data is composed of historical multidimensional meteorological data collected at different times and the corresponding historical icing data. S2: Training samples are obtained by adaptively segmenting and truncating the sequence data using the sliding window method; S3: Construct an initial ice thickness prediction model based on CNN and BiLSTM, and train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model; S4: Input the collected real-time multi-dimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

2. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 1, characterized in that, The sequence data, composed of historical multidimensional meteorological data collected at different times and corresponding historical icing data, includes: S11: Perform anomaly processing on historical multidimensional meteorological data and historical icing data, and perform time alignment processing on historical multidimensional meteorological data and historical icing data. S12: The historical multidimensional meteorological data after time alignment processing is spliced ​​and integrated with the historical icing data in chronological order to form sequence data; The historical multidimensional meteorological data includes historical temperature, historical relative humidity, historical wind speed, and historical atmospheric pressure.

3. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 2, characterized in that, The anomaly processing of historical multidimensional meteorological data and historical icing data, and the time alignment processing of historical multidimensional meteorological data and historical icing data, include: S111: Based on historical multidimensional meteorological data collected from different locations, anomaly detection is performed on the historical multidimensional meteorological data of the target location, and the historical multidimensional meteorological data of the target location is corrected using the spatial interpolation method with historical multidimensional meteorological data collected from different locations. S112: Take the collection time period of the historical icing data of the target location as the target time set. If the historical multidimensional meteorological data corresponding to the target time set is obtained in the corrected historical multidimensional meteorological data, then execute S113. Otherwise, based on the auxiliary multidimensional meteorological data before and after the target time set for which no corresponding historical multidimensional meteorological data has been obtained, use linear interpolation to obtain the historical multidimensional meteorological data corresponding to the target time set, and execute S113. S113: Associate the historical multidimensional meteorological data corresponding to the target time set with the historical icing data of the target time set at the target location.

4. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 1, characterized in that, The step of adaptively segmenting and extracting training samples from sequence data using the sliding window method includes: S21: Extract the first and second rates of change of ice thickness from historical icing data, extract the short-term fluctuation variance of meteorological characteristics from historical multidimensional meteorological data, construct an icing sensitivity index based on the first rate of change, the second rate of change, and the short-term fluctuation variance, and classify the sequence data based on the icing sensitivity index. S22: Obtain the sliding window length and sliding step size based on the characteristics of different types of sequence data, and segment the sequence data according to the sliding window length and sliding step size, using the sequence data in each segment as a group of training samples.

5. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 1, characterized in that, The initial icing thickness prediction model based on CNN and BiLSTM includes: The initial ice thickness prediction model consists of a CNN containing two one-dimensional convolutional layers and corresponding ReLU activation functions, and a BiLSTM with two stacked bidirectional LSTM layers. The first one-dimensional convolutional layer in the one-dimensional convolutional layer includes 64 convolutional kernels of size 3, and the second one-dimensional convolutional layer in the one-dimensional convolutional layer includes 32 convolutional kernels of size 3; each bidirectional LSTM structure includes 128 hidden neurons.

6. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 4, characterized in that, The step of training the initial icing thickness prediction model using training samples to obtain the final icing thickness prediction model includes: The training samples are normalized, and the normalized training samples are reconstructed into an adaptation tensor. The adaptation tensor is input into the CNN in the initial ice thickness prediction model to obtain local correlation features related to ice thickness. The local correlation features are folded, compressed and flattened to obtain the adaptation vector. The adaptation vector is input into the BiLSTM in the initial ice thickness prediction model to obtain comprehensive temporal features. The comprehensive temporal features are input into the fully connected layer to obtain the predicted icing thickness. If the predicted icing thickness does not match the actual icing thickness, the parameters of the initial icing thickness prediction model are corrected based on the thickness difference between the predicted and actual icing thicknesses and the icing sensitivity index to obtain the final icing thickness prediction model. Otherwise, the initial icing thickness prediction model is used as the final icing thickness prediction model.

7. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 6, characterized in that, The step of obtaining the final icing thickness prediction model by correcting the parameters of the initial icing thickness prediction model based on the thickness difference between the predicted and actual icing thickness and the icing sensitivity index includes: The thickness difference weights are obtained based on the icing sensitivity index corresponding to the training samples, and the thickness difference values ​​are weighted and summed based on the thickness difference weights to obtain the total loss value. The parameters of the initial icing thickness prediction model are corrected based on the total loss value to obtain the final icing thickness prediction model.

8. The method for predicting icing thickness of transmission lines based on CNN-BiLSTM according to claim 7, characterized in that, The step of correcting the parameters of the initial icing thickness prediction model based on the total loss value to obtain the final icing thickness prediction model includes: The parameter gradient of the initial icing thickness prediction model is calculated through backpropagation based on the total loss value. The parameters are then iteratively updated according to the gradient descent direction to obtain the final icing thickness prediction model.

9. A CNN-BiLSTM-based transmission line icing thickness prediction system, applicable to the CNN-BiLSTM-based transmission line icing thickness prediction method described in any one of claims 1-8, characterized in that, include: The sequence data acquisition module is used to construct sequence data based on historical multidimensional meteorological data collected at different times and the corresponding historical icing data. The training sample acquisition module is used to adaptively segment and truncate the sequence data using the sliding window method to obtain training samples. The model acquisition module is used to build an initial ice thickness prediction model based on CNN and BiLSTM, and to train the initial ice thickness prediction model with training samples to obtain the final ice thickness prediction model. The target data acquisition module is used to input the collected real-time multidimensional meteorological data into the final icing thickness prediction model to obtain the target icing thickness.

10. A computer device, comprising: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. The processor executes the machine-readable instructions to perform the steps of the CNN-BiLSTM-based transmission line icing thickness prediction method as described in any one of claims 1-8.

Images