An Improved FSLSTM Icing Prediction Method Based on Multi-Source Data and Dynamic Periodic Attention

Abstract

This invention relates to the field of disaster prevention and mitigation technology for power transmission lines. A multi-source data and dynamic periodic attention-enhanced FSLSTM icing prediction method is presented, characterized by the following steps: 1) multi-source data acquisition and preprocessing; 2) dynamic periodic segmentation and attention-enhanced modeling; 3) scenario-based data augmentation and model training optimization; 4) system prediction and dynamic adjustment: based on the trained model, preprocessed multi-source data is input, and after dynamic periodic segmentation, attention feature extraction and fusion, the predicted icing thickness is output; the prediction error and changes in meteorological and line status parameters are monitored in real time, and the model's periodic calculation window and attention weight allocation logic are dynamically adjusted to adapt to the icing variation patterns of different scenarios. This method can achieve high-precision and high-generalization prediction of icing thickness under complex climate scenarios, providing reliable technical support for disaster prevention and mitigation of power transmission lines.

Description

Technical Field

[0001] This invention relates to the field of disaster prevention and mitigation technology for power transmission lines, specifically to an improved FSLSTM overhead transmission line icing prediction method that integrates multi-source data fusion, dynamic periodic adjustment, and attention mechanisms, applicable to high-precision icing thickness prediction scenarios under complex meteorological conditions. Background Technology

[0002] Overhead transmission lines are the core infrastructure of the power system, and their safe and stable operation is directly related to the reliability of energy supply and the order of social production and life. In recent years, extreme weather events have occurred frequently around the world, and accidents such as line tripping, tower collapse, and power outages caused by icing disasters on transmission lines are commonplace. For example, the icing disaster caused by the strong cold wave in North China in 2023 resulted in the failure and shutdown of four 500kV high-voltage transmission lines, causing more than 100,000 households to experience power outages for more than 12 hours, with direct economic losses exceeding ten million yuan. Therefore, achieving high-precision and early prediction of the thickness of icing on transmission lines is a core requirement for disaster prevention and mitigation work in the power industry, and also a key technical support for ensuring the safety of the power grid.

[0003] Currently, transmission line icing prediction technologies are mainly divided into three categories: physical mechanism models, traditional machine learning models, and conventional neural network models. However, all of them have significant technical limitations and cannot meet the high-precision prediction requirements under complex weather scenarios. Regarding physical mechanism models, existing technologies such as the dry icing growth model proposed by GOODWIN and the numerical icing growth model established by MAKKONEN are based on thermodynamics and fluid dynamics theories, constructing prediction formulas by analyzing the physical relationship between meteorological factors such as temperature, humidity, and wind speed and icing growth. However, these models require coupling with a large number of meteorological and geographical parameters, making it difficult to establish accurate mapping functions in practical applications. Furthermore, they have poor fitting ability for abrupt icing processes under extreme weather conditions, with prediction errors generally exceeding 20%, thus limiting their practicality.

[0004] Traditional machine learning models are commonly used in the industry, with typical methods including grey prediction models, bat algorithm-optimized extreme learning machines, hidden Markov models, and least squares support vector machines (LSSVM). For example, the small-sample grey prediction model proposed by Chen Jie et al. achieves prediction by extrapolating the trend of icing data, but this model is only suitable for icing scenarios with stable changes. When faced with complex situations such as sudden temperature drops and abrupt changes in wind speed, the prediction accuracy drops significantly. Sun et al. used wavelet transform denoising combined with bat algorithm-optimized extreme learning machines, which can improve noise suppression capabilities to some extent, but it is still limited by the bottleneck of machine learning algorithms in modeling high-dimensional time series. It cannot capture the nonlinear coupling relationship between icing and meteorological factors, and its adaptability to multiple climate regions is poor, with prediction error fluctuations exceeding 15% in different regions.

[0005] With the development of deep learning technology, conventional neural network models such as backpropagation (BP) neural networks, recurrent neural networks (RNNs), basic long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) are gradually being applied to the field of icing prediction. The BP neural network icing dance prediction model proposed by Liao Zheng et al. relies on a single meteorological feature input and suffers from the problem of local optima. Liu Zhaolun et al. used a basic LSTM to predict road icing, which alleviated the gradient vanishing problem of RNNs, but still could not simultaneously capture the local fluctuation characteristics and global trends of icing data, resulting in a lag in response to sudden changes in icing thickness. Li Bo et al. combined principal component analysis with CNNs to construct an icing prediction model, which could extract local features, but lacked the ability to mine the icing cycle patterns over time, leading to insufficient long-term prediction accuracy.

[0006] In summary, existing icing prediction technologies have significant shortcomings in data dimensional completeness, dynamic periodic adaptability, utilization of core features, and enhanced sample correlation, making it difficult to meet the high-precision icing prediction requirements of different climatic regions and complex meteorological conditions. Therefore, developing an icing prediction method that can integrate multi-source data, dynamically capture periodic features, strengthen the weight of core factors, and improve sample correlation has become an urgent technical problem to be solved in the power industry. Summary of the Invention

[0007] The purpose of this invention is to address at least one of the aforementioned technical deficiencies. This invention provides a multi-source data and dynamic periodic attention-based improved FSLSTM icing prediction method. This method aims to overcome the limitations of existing prediction technologies, such as insufficient data dimensionality, inadequate periodic adaptation, and inefficient utilization of core features. It achieves high-precision and high-generalization prediction of icing thickness under complex climate scenarios, providing reliable technical support for disaster prevention and mitigation of power transmission lines.

[0008] To achieve the above objectives, the technical solution adopted by this invention is a multi-source data and dynamic periodic attention-based improved FSLSTM icing prediction method, characterized by the following steps: 1) Multi-source data acquisition and preprocessing: Collect basic data such as icing thickness, temperature, and humidity, and add two new types of line status data: real-time conductor tension and tower tilt angle, to construct a multi-dimensional dataset of "meteorology-icing-line status"; after removing outliers, use weighted average interpolation of 5 adjacent time steps to fill in missing values, and use Z-Score normalization to eliminate dimensional differences, thereby improving data quality and consistency; 2) Dynamic Periodic Segmentation and Attention Enhancement Modeling: Taking the current prediction time point as the endpoint, 20 time steps are taken as the sliding calculation window. The core ice-covering period is extracted in real time through Fourier transform, and the multi-source data is segmented according to the dynamic period. A factor attention layer is embedded between the convolutional layer of FSLSTM (Feature Selection Long Short-Term Memory Network) and the iterative layer of LSTM (Basic Long Short-Term Memory Network) to calculate the feature contribution of each input factor and assign dynamic weights. The LSTM cell state update formula is improved to enhance the influence of core factors on the prediction results. 3) Contextualized data augmentation and model training optimization: The samples are divided into two scenarios, cold and dry and mild and humid, according to meteorological characteristics. Within the same scenario, a linear interpolation strategy with scenario weights is used to generate augmented samples to improve the correlation between samples. The AdamW (adaptive moment estimation with decoupled weight decay) optimizer and MSE (mean squared error) loss function are used to train the model. An appropriate learning rate and number of training iterations are set to ensure stable convergence of the model and suppress overfitting. 4) System prediction and dynamic adjustment: Based on the trained model, the preprocessed multi-source data is input, and after dynamic periodic segmentation, attention feature extraction and fusion, the predicted value of icing thickness is output; the prediction error and changes in meteorological and line status parameters are monitored in real time, and the periodic calculation window and attention weight allocation logic of the model are dynamically adjusted to adapt to the icing change patterns of different scenarios.

[0009] The beneficial effects of this invention are as follows: 1. Improve prediction accuracy: Multi-source data fusion supplements the coupling relationship between icing and line status, dynamic periodic adjustment adapts to time-varying characteristics, and attention mechanism strengthens the utilization of core factors, significantly reducing prediction error. Compared with traditional methods, MSE can be reduced by 15%-20%.

[0010] 2. Enhanced generalization ability: The scenario-based data augmentation strategy avoids cross-scenario sample interference, and the AdamW optimizer suppresses the risk of overfitting, reducing the fluctuation of the prediction error of the model in different climate regions by 10%-15%, making it more adaptable.

[0011] 3. Practical for engineering applications: New line status data can be obtained through existing online monitoring systems without the need for additional hardware deployment; the model parameter adjustment is highly automated and requires no manual intervention, making it easy to promote and apply in actual transmission line operation and maintenance, and providing accurate basis for disaster prevention and mitigation decision-making. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a specific implementation of the FSLSTM icing prediction method based on multi-source data and dynamic periodic attention according to the present invention.

[0014] Figure 2 This is a performance comparison chart for different numbers of convolutional layers.

[0015] Figure 3 This is a performance comparison chart for different numbers of hidden layers.

[0016] Figure 4 This is a box plot of the prediction error of the present invention.

[0017] Figure 5 This is a comparison chart of different training iterations of the present invention.

[0018] Figure 6 This is a comparison chart of the prediction results of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Part One: The complete steps of the method of the present invention are as follows: like Figure 1 As shown, an improved FSLSTM icing prediction method using multi-source data and dynamic periodic attention is proposed. The method includes the following specific steps: 1. Step S101: Multi-source data acquisition stage This step focuses on "full-dimensional feature coverage, high-precision synchronous acquisition, and multi-source data collaborative adaptation" as its core objectives. It constructs a three-tiered acquisition architecture of "distributed sensing - edge aggregation - cloud storage." Through multi-device collaborative linkage, dynamic frequency adjustment, and secure transmission mechanisms, it achieves accurate acquisition of meteorological, icing, and line status data, providing high-quality raw data for subsequent preprocessing and modeling. The specific details are as follows: The three-tier data acquisition architecture and communication mechanism consist of: a bottom layer deploying distributed sensing nodes, each covering 1-2 levels of power lines, integrating multiple types of sensors to achieve regional data acquisition; a middle layer edge gateway responsible for data preprocessing, temporary storage, and protocol conversion, employing dual-mode communication between dedicated power OPGW (optical fiber composite overhead ground wire) fiber and 5G industrial private network to ensure real-time and stable data transmission, with end-to-end latency controlled within 500ms; and a top layer cloud server receiving data through an encrypted tunnel, using a distributed database for storage and indexing, supporting high-concurrency data writing and fast querying, while reserving interfaces for connection with power transmission line operation and maintenance platforms to achieve data sharing.

[0021] Multiple data acquisition schemes are employed. Basic data acquisition: Ice thickness adopts a combination of "indirect calculation + direct calibration": 0.1-level tension sensors (0-50kN) and dual-axis tilt sensors (±5°, 0.001° resolution) are installed at the conductor suspension points to calculate the equivalent ice thickness through a mechanical model; ultrasonic sensors (0-100mm, error ≤±0.5mm) are deployed at the midpoint of the conductor and calibrated every 30 minutes to ensure that the absolute error does not exceed 1mm. In terms of meteorological data, a PT100 platinum resistance temperature sensor (-40℃~60℃, ±0.2℃) is installed inside a radiation shield; a capacitive humidity sensor (0-100% RH, ±2% RH) is integrated with it; a three-dimensional ultrasonic anemometer (0-60m / s, ±0.1m / s) at the top of the tower calculates the average wind speed every 10 minutes; a silicon photovoltaic cell light sensor (0-200000Lux, ±5%) and an absolute pressure sensor (50-110kPa, ±0.1kPa) synchronously collect environmental parameters.

[0022] Line status data acquisition: Conductor tension is monitored using fiber optic grating sensors (0-30kN, 0.1 grade), installed at the tension clamps. Wavelength drift characteristics are used to mitigate electromagnetic interference, and a temperature compensation algorithm eliminates zero-point drift. Tower tilt angle is measured using a three-axis gyroscope sensor (±10°, 0.001° resolution, 10Hz sampling rate), installed in the upper middle part of the tower. A data fusion algorithm is used to eliminate wind-induced vibration interference and extract the steady-state tilt component, accurately reflecting the impact of icing loads.

[0023] A dynamic acquisition frequency and synchronization mechanism are implemented, with the acquisition frequency adaptively adjusted according to meteorological conditions: 30 minutes / time for normal operating conditions (temperature > 0℃, wind speed < 3m / s); 20 minutes / time for conditions prone to icing (temperature ≤ 0℃ and wind speed ≥ 3m / s); and 10 minutes / time for icing thickness exceeding 10mm. Synchronization employs a combination of GPS (Global Positioning System) timing and local clock calibration. All sensors receive GPS timing once per hour, and the edge gateway is equipped with a high-precision clock (drift rate ≤ 1μs / h) to ensure a time difference of ≤ 10ms between different devices, guaranteeing consistency of time-series data.

[0024] Data preprocessing and transmission optimization: The edge gateway performs preliminary processing on the raw data: data exceeding the specified range is removed based on validity criteria; high-frequency fluctuation data is smoothed using moving average filtering; and the data is compressed using the LZ77 (LZ77 lossless data compression algorithm) algorithm (compression ratio ≥ 3:1) before transmission. Data is encrypted using SSL / TLS 1.3 protocol to achieve end-to-end secure transmission; a breakpoint resumption and retransmission mechanism is established, caching data during network interruptions and automatically resuming transmission upon recovery; lost data is supplemented by interpolation between adjacent time steps to ensure data integrity.

[0025] The equipment is designed to be environmentally adaptable. All outdoor equipment uses an IP67 (IP67 protection rating) protective shell, which is rainproof, snowproof, dustproof and corrosion-resistant. The sensor bracket is made of hot-dip galvanized steel to enhance its anti-rust performance. Precision equipment such as fiber optic gratings are equipped with thermal insulation protective sleeves to ensure stable operation in extreme environments of -40℃ to 60℃, adapting to the operation and maintenance needs of power transmission lines in different climate regions and ensuring the continuity and reliability of data collection.

[0026] 2. Step S102: Preprocessing stage.

[0027] This step adopts a two-tier architecture of "local preprocessing at the edge gateway + deep preprocessing on the backend server," focusing on quality optimization and feature adaptation of multi-source data. The specific details are as follows: The two-level preprocessing architecture collaborative mechanism prioritizes real-time preprocessing of the collected data: high-frequency noise in data such as wind speed and tension is initially suppressed through moving average filtering, outliers exceeding the threshold are coarsely screened based on the sensor's rated range, and the LZ77 algorithm is used to compress the data, reducing the transmission bandwidth usage by more than 60%; after receiving the data, the backend server starts a deep preprocessing process, achieving precise optimization of data quality through multi-algorithm fusion. The two-level architecture is linked in real time through a dedicated power communication network to ensure that the preprocessing delay is controlled within 1 second, balancing real-time performance and processing accuracy.

[0028] Multi-dimensional outlier identification and hierarchical processing: Constructing a triple anomaly verification system of "physical constraints - statistical distribution - temporal correlation": At the physical constraint level, data exceeding the equipment range, such as temperatures exceeding -40℃ to 60℃ and conductor tension exceeding 0 to 30kN, are removed; at the statistical distribution level, kernel density estimation combined with the 3σ (3 times standard deviation) criterion is used to mark suspicious outliers that deviate from the core data distribution range and have a kernel density below 0.01; at the temporal correlation level, by calculating the first and second differences between the data and the five adjacent time steps, temporal outliers that logically contradict historical trends and related data are identified. Differentiated processing is implemented for different anomaly types: Deterministic outliers are directly removed and equipment fault information is recorded; suspicious outliers are corrected using the Lowess method of local weighted regression, selecting data from 10 time steps before and after the outlier point for weighted fitting; sudden anomaly data caused by extreme weather are retained and labeled after logical verification, providing key feature support for subsequent modeling.

[0029] Hierarchical intelligent missing value completion strategy: Classified by missing duration and type: Random missing values ​​(≤3 time steps) are handled using weighted average interpolation of adjacent time steps, with weights decreasing exponentially with distance to ensure the completed value closely matches the time series trend; Continuous missing values ​​(3-10 time steps) are completed using a multivariate LSTM (MV-LSTM) model, with related data such as temperature and humidity as input; Periodic missing values ​​(>10 time steps) are completed using historical contemporaneous data migration, aligning time series features through dynamic time warping (DTW) and correcting for current data trend deviations. After completion, the algorithm is iteratively optimized by verifying the deviation with subsequent valid data to ensure that the completion accuracy deviation does not exceed 5%.

[0030] Customized data standardization and dimensional unification: Dedicated standardization schemes are designed for different types of data: Numerical data (temperature, icing thickness, etc.) are normalized using Z-Score, converting the data into a normal distribution with a mean of 0 and a variance of 1, eliminating dimensional differences; Angular data (tower tilt angle) is first converted into [-π, π] radians, and then transformed into two-dimensional features through sine and cosine transformations, retaining periodic information; After standardization of each batch of data, the mean and variance deviations of each feature are checked. If they exceed ±0.05, the global statistics are recalculated to ensure the consistency of data distribution across batches.

[0031] Timing consistency calibration and synchronization optimization: Based on the GPS time stamp standard, the time stamp deviation of different sensors is corrected. When the deviation is large, the data value is adjusted by linear interpolation. The acquisition frequency of 10 minutes / time, 20 minutes / time, and 30 minutes / time is unified to 20 minutes / time. High-frequency data is downsampled by sliding window averaging, and low-frequency data is upsampled by spline interpolation. The timing synchronization of related data is evaluated by cross-correlation coefficient to ensure that the cross-correlation coefficient of core data such as temperature and icing thickness is ≥0.8, laying the foundation for dynamic periodic extraction.

[0032] Closed-loop quality assessment and optimization mechanism: Establish a multi-dimensional assessment index system, including data integrity (≥99%), anomaly handling accuracy (≥98%), completion accuracy (deviation ≤5%), and time sequence synchronization (timestamp deviation ≤5ms). If the index does not meet the standards, the problematic link is automatically located, and the corresponding preprocessing steps are re-executed to form a closed-loop process of "processing-assessment-optimization" to ensure that the output data meets the modeling requirements.

[0033] 3. Step S103: Dynamic periodic segmentation stage.

[0034] This step focuses on "accurate capture of temporal features, real-time adaptation of periodic information, and collaborative segmentation of multi-source data." Through a comprehensive mechanism of "intelligent periodic extraction - dynamic window adjustment - multi-dimensional segmentation optimization," it overcomes the limitations of traditional fixed-period segmentation, providing segmented data that closely aligns with the icing change patterns for attention-enhanced FSLSTM (Feature Selection Long Short-Term Memory) modeling. The specific details are as follows: A dynamic periodicity extraction system was constructed, with multi-scale periodicity analysis based on preprocessed multi-source fusion data (meteorological data, icing data, and line status data). A combined algorithm of Fast Fourier Transform (FFT) and Wavelet Packet Transform (WPT) was employed to deeply mine periodic characteristics. First, FFT was used to convert the time-domain data to the frequency-domain data, calculating the magnitude of each frequency component to initially screen for potential periods. Then, WPT was used to finely decompose the high-frequency components, capturing subtle fluctuations within short periods and compensating for the limitations of FFT in identifying the periods of non-stationary signals.

[0035] The sliding calculation window dynamic adjustment strategy sets a sliding calculation window with the current forecast time point as the endpoint. The window length adopts an adaptive adjustment mechanism: it is dynamically adjusted according to the current meteorological complexity. When parameters such as temperature and wind speed fluctuate drastically (variance exceeding 1.5 times the historical average), the window length is set to 20 time steps (corresponding to 400 minutes) to ensure the capture of short-term rapid changes in cycles. When meteorological conditions are stable, the window length is expanded to 30 time steps (corresponding to 600 minutes) to fully explore long-term cycle patterns. Every time the window advances by one forecast time step, the cycle extraction process is automatically re-executed to achieve real-time updates of the cycle.

[0036] The core cycle selection and verification were based on frequency domain modulus sorting, and the cycles corresponding to the top 3 largest modulus values ​​were selected as candidate core cycles. Through time series correlation verification, the mutual information value between the candidate cycles and the changes in icing thickness was calculated, and invalid cycles with mutual information values ​​below 0.6 were eliminated. Cross-validation was performed by combining the change cycles of line status data (conductor tension, tower tilt angle) to ensure that the selected core cycles simultaneously adapt to the coupled change patterns of meteorological and line status. Finally, 3 effective core cycles were determined for segmentation.

[0037] Multi-source data collaborative segmentation is implemented, and adaptive data length padding addresses the length differences across different core periods by employing a "smart completion + truncation optimization" strategy to unify data length: for data whose length cannot be divided by the period, a generative completion method based on Generative Adversarial Networks (GANs) is used to supplement virtual data that conforms to the temporal trend; for data exceeding a set threshold after completion, a slope-based truncation algorithm is used to retain the core trend portion and avoid segmentation deviations caused by data redundancy. During the padding process, the proportion of virtual data is strictly controlled to not exceed 5% to ensure the authenticity of the segmented data.

[0038] The hierarchical and segmented execution logic implements hierarchical segmentation based on the scale differences of the core cycle: long-cycle (>60 minutes) data is used to capture long-term trend characteristics of icing, employing a non-overlapping segmentation method to ensure that each data segment fully covers one cycle; medium-cycle (30-60 minutes) data is used to mine medium-term fluctuation patterns, employing a semi-overlapping segmentation (50% overlap rate) to strengthen the correlation between adjacent segments; short-cycle (<30 minutes) data is used to capture instantaneous mutation characteristics, employing a fully overlapping segmentation (80% overlap rate) to avoid missing key mutation information. Each level of segmented data is labeled with cycle attributes to facilitate targeted feature extraction in subsequent modeling.

[0039] The segmented quality verification mechanism constructs a dual-indicator verification system of "cycle matching degree + trend consistency": cycle matching degree is evaluated by calculating the spectral similarity between the segmented data and the corresponding core cycle; if the similarity is lower than 0.85, the data is re-segmented; trend consistency is verified by comparing the changing trends of data such as icing thickness and conductor tension within the segment using the Pearson correlation coefficient; if the correlation coefficient is lower than 0.7, the segmentation window position is adjusted. When verification fails, the system automatically backtracks to the cycle extraction stage to re-optimize the core cycle, forming a closed-loop process of "extraction-segmentation-verification-optimization".

[0040] Segmented data enhancement and adaptation processing, and secondary noise suppression within segments: To address the high-frequency noise remaining in the segmented data, an adaptive wavelet threshold denoising algorithm is adopted. The threshold is dynamically adjusted according to the noise intensity of each segment, suppressing noise interference while preserving detailed features, thus ensuring the purity of the segmented data.

[0041] Cross-segment feature association enhancement constructs a segment association matrix by calculating the feature similarity of adjacent segment data. For adjacent segments with low similarity, a feature transfer algorithm is used to supplement association information, enhancing temporal continuity. Simultaneously, timestamps and period labels are added to each segment data, facilitating accurate identification of segment attributes during model training and prediction, and improving feature utilization efficiency.

[0042] The modeling input adaptation optimization, based on the input requirements of the attention-enhanced FSLSTM, transforms the segmented data at each level into a feature matrix of uniform dimension. Through dimension normalization, it ensures that segmented data from different periods and scales can be collaboratively input into the model, laying the foundation for subsequent fusion and extraction of local and global features. The entire dynamic segmentation process is executed automatically, with the processing latency for each batch of data controlled within 200ms, balancing processing accuracy and real-time performance.

[0043] 4. Step S104: Attention Enhancement Modeling Stage.

[0044] This step focuses on "precisely focusing on core features, enhancing model generalization ability, and deeply mining temporal correlations" as its core objectives. It constructs an integrated process of "attention-enhanced modeling - scenario-based data augmentation - dynamic optimization iteration." Through structural innovation of the FSLSTM model and customized design of data augmentation strategies, it breaks through the limitations of traditional modeling methods. Specific details are as follows: The Attention-Enhanced FSLSTM (Att-FSLSTM) modeling system employs a four-level structure: a 1D convolutional layer, a factor attention layer, an LSTM iterative layer, and a feature fusion layer. The 1D convolutional layer (3 layers, kernel size 3) extracts local spatiotemporal features from segmented data across different periods, introducing nonlinear transformations through activation functions to enhance feature representation. The factor attention layer, a core innovative module, calculates contribution weights based on the feature variance of each input factor (temperature, humidity, wire tension, etc.), normalizes them using a softmax function, and assigns dynamic weights to achieve focused attention on core factors. The LSTM iterative layer (3 hidden layers) incorporates a weight update mechanism, optimizing the cell state update formula as follows: In the formula, This represents the cell state at the current time step t; Output the forget gate at the current time step t; This represents the cell state at the previous time step t-1; The input gate output is the current time step t; The candidate cell states at the current time step t; This represents the average attention weight at the current time step t.

[0045] The influence of weights on temporal feature capture is enhanced; the feature fusion layer adopts a weighted summation strategy to fuse local and global features and output a high-dimensional comprehensive feature vector.

[0046] The model parameter dynamic optimization mechanism employs a combined strategy of "grid search + Bayesian optimization" to determine the optimal parameters: First, grid search is used to initially screen candidate parameter ranges (learning rate 0.0005-0.002, batch size 16-64, LSTM hidden unit number 64-256); then, based on the Bayesian optimization algorithm, the parameter combination is iteratively optimized with the mean squared error (MSE) of the validation set as the objective function, ultimately determining the optimal configuration of learning rate 0.001, batch size 32, and hidden unit number 128. An early stopping mechanism is introduced during training, automatically stopping training when the validation set error fails to decrease for 10 consecutive rounds to avoid overfitting; simultaneously, a gradient clipping strategy is used to limit the gradient norm to within 5, mitigating the gradient explosion problem.

[0047] A multi-task joint training strategy is employed to construct a joint training framework consisting of a "main task (ice thickness prediction) + auxiliary task (line condition anomaly identification)". The main task uses MSE as the loss function to optimize the accuracy of ice thickness prediction; the auxiliary task uses a binary classification cross-entropy loss function to identify abnormal states of conductor tension and tower tilt angle, and utilizes the gradient information of the auxiliary task to improve the sensitivity of the main task to key features. A dynamic weight allocation mechanism is adopted, with the main task weight set to 0.7 and the auxiliary task weight to 0.3 in the initial training phase. As training iterates, the weight is gradually adjusted to 0.9 for the main task and 0.1 for the auxiliary task to ensure that the model prioritizes the core prediction target.

[0048] The scenario-based data augmentation solution employs a stratified sample classification mechanism based on the K-means clustering algorithm to segment the original samples into three typical scenarios: "cold and dry," "mild and humid," and "low temperature and strong wind," using temperature, humidity, and wind speed as clustering features. This ensures that the meteorological feature similarity of samples within the same scenario is higher than 0.8. The distribution characteristics of statistical data are then analyzed separately for each scenario to provide a basis for subsequent targeted augmentation.

[0049] The weighted mixup enhancement strategy performs a mixup operation within similar scenarios, introducing scene weights γ and dynamic interpolation coefficients. Scene weights are determined based on the proportion of samples from each scene (cold and dry 0.6, mild and humid 0.3, low temperature and strong wind 0.1) to ensure that the enhanced samples closely match the actual scene distribution; interpolation coefficients Following a Beta (5,5) distribution, new samples are generated through linear interpolation, using the following formula: In the formula, For the generated enhanced feature samples; For the generated enhanced label samples; As scene weight; For dynamic interpolation coefficients ( Specifically, it follows a Beta(5,5) distribution. , Two types of original feature samples; , To and , The corresponding original label sample.

[0050] During the augmentation process, the number of new samples is controlled to be 1.5 times that of the original samples to avoid data redundancy.

[0051] To enhance sample quality verification, a two-dimensional verification system of "distribution consistency + logical rationality" is constructed: distribution consistency is determined by the KS test to assess the distribution difference between the enhanced sample and the original sample; a test statistic greater than 0.05 indicates compliance. Logical rationality is verified through rule-based checks to eliminate contradictory samples. Samples that fail verification are automatically discarded and regenerated to ensure the effectiveness of the enhanced samples.

[0052] For model training and adaptation optimization, the AdamW optimizer was used instead of the traditional Adam optimizer in the training optimizer and loss function design. A weight decay coefficient of 0.001 was added to suppress overfitting of model parameters. The loss function adopted was a "MSE+MAPE weighted combined loss", with weights set to 0.6 and 0.4 respectively, balancing overall control of prediction error and precise constraint of relative error. During training, a cosine annealing strategy was used, with an initial learning rate of 0.001, which decayed to 0.8 of the current value every 50 iterations to ensure stable model convergence.

[0053] The cross-scene transfer adaptation mechanism addresses the scene differences across different climate regions by introducing a transfer learning strategy: A model trained in a cold, dry scene is used as the pre-trained model. Fine-tuning is then performed on a small number of samples in mild, humid, low-temperature, and windy scenes. The parameters of the bottom convolutional layers are frozen, and only the parameters of the attention layer and the top fully connected layer are updated, reducing the training data requirements for new scenes. During fine-tuning, a domain-adaptive loss function is used to reduce the distributional differences between the source and target scenes, thereby improving the model's cross-scene adaptability.

[0054] Real-time model performance monitoring and iterative training process involve real-time monitoring of MSE, MAE, and MAPE metrics on the training and validation sets, plotting learning curves and error distribution histograms to identify underfitting (high errors on both training and validation sets) or overfitting (low error on the training set, high error on the validation set). For underfitting, the number of hidden units in the network is increased or the training iterations are extended; for overfitting, a Dropout layer (dropout rate of 0.2) is introduced or the number of augmented samples is increased, forming a closed-loop iterative process of "training-monitoring-optimization" to ensure optimal model performance.

[0055] 5. Step S105: Scenario-based data augmentation and model training optimization, system prediction and dynamic adjustment stage.

[0056] This step focuses on "efficient model convergence, controllable prediction accuracy, and dynamic adaptation to different scenarios," constructing a complete system encompassing "refined training strategies, real-time prediction engines, and closed-loop optimization mechanisms." Through multi-dimensional optimization during training and intelligent adaptation in the prediction phase, it ensures stable and efficient model operation under various climatic scenarios. Specific details are as follows: A refined model training system was constructed, employing a "stratified sampling + time-series preservation" strategy for training data partitioning and adaptation. Multi-source data was divided into training, validation, and test sets in a 7:2:1 ratio. The training set covers all scenarios to ensure comprehensive learning, the validation set is co-distributed with the training set for monitoring training effectiveness, and the test set retains complete time-series segments to simulate actual predictions. To address the issue of imbalanced scenario samples, weighted sampling was used, assigning a 1.5x weight to scarce scenario samples such as those from low temperatures and strong winds to prevent the model from favoring scenarios it already has.

[0057] The multi-stage training strategy is implemented in three phases: The warm-up phase (first 50 epochs) uses a small learning rate of 0.0005 and large batch sizes of 64 to quickly adapt to the data distribution; the core training phase (epochs 50-250) uses a learning rate of 0.001 and standard batch sizes of 32, combining the AdamW optimizer with the MSE+MAPE combined loss function to deeply optimize parameters; the fine-tuning phase (epochs 250-300) reduces the learning rate to 0.0001, freezes the underlying convolutional layers, and only updates the parameters of the attention and fully connected layers. A gradient accumulation mechanism is also introduced, accumulating gradients every 4 batches before updating, improving training stability.

[0058] A multi-dimensional monitoring system is constructed to monitor the training process and implement an early stopping mechanism. This system tracks changes in MSE, MAE, MAPE, and the loss function in real time, and plots learning curves and error heatmaps. A dual early stopping condition is set: if the validation set error does not decrease for 10 consecutive epochs, or if the difference between the validation set error and the training set error exceeds 0.05 (indicating overfitting), the system immediately stops and backtracks to the optimal parameters. A model snapshot is saved every 50 epochs to record parameters and performance for easy backtracking and management.

[0059] The intelligent prediction engine is designed and implemented, including data preprocessing and adaptation before prediction. After the prediction data is input, real-time preprocessing is performed: Z-score normalization is performed using standardized parameters during the training phase to ensure consistent distribution; the core period is updated through a dynamic period extraction algorithm to generate adapted segmented data; and a feature alignment mechanism is used to ensure that the dimension and order of the input features match the training data to avoid bias.

[0060] The multi-modal prediction inference mechanism employs a dual-modal inference engine: a fast prediction mode uses a pruned, lightweight model while retaining the core feature extraction module, with latency controlled within 100ms; the accurate prediction mode targets critical periods such as rapid icing growth, calling the full model and combining it with a sliding window strategy to continuously update prediction results. Simultaneously, a weighted average is used to fuse predictions from adjacent time steps to reduce random errors.

[0061] The post-processing and correction of prediction results employs multiple mechanisms: unreasonable values ​​are eliminated based on physical constraints; deviations are adjusted using a linear formula incorporating historical data from the same period; and cross-validation is performed using line status data, triggering a secondary prediction when trends contradict each other. The corrected results include confidence intervals based on the statistical distribution of errors, providing a reference for operational and maintenance decisions.

[0062] A dynamic optimization and iteration mechanism is implemented, and error feedback optimization is established to create an error feedback closed loop. The sources of deviation are analyzed by scenario and time period. If the MAPE of a certain scenario exceeds 10%, incremental training is automatically triggered, adding new data to the training set, retraining the attention layer and top-level parameters, and updating the dynamic periodic extraction threshold to optimize the segmentation strategy.

[0063] Cross-scene transfer adaptation targets new routes or climate regions. Using the existing model as a pre-trained foundation, it is fine-tuned with at least 500 new scene data points, freezing the underlying feature extraction module and updating only the upper-layer parameters. A domain-adaptive loss function is employed to reduce scene distribution differences and decrease data collection requirements.

[0064] Model performance evaluation and updates are conducted monthly using both test set data and operational data. An update is initiated if core metrics decline by more than 15%. Updated models are released in a phased, gradual manner, initially on a small scale for pilot testing before full deployment once the criteria are met. The entire process is fully automated, with manual intervention triggered only in cases of anomalies, adapting to the operational needs of transmission lines.

[0065] Through the synergistic effect of five steps (S101-S105), data acquisition lays the foundation for accuracy, preprocessing refines data quality, dynamic segmentation uncovers temporal patterns, modeling enhances feature utilization, and training prediction enables efficient inference. Ultimately, this achieves high-precision icing prediction, strong scenario adaptability, and low maintenance costs, providing reliable technical support for power transmission line disaster prevention.

[0066] In one embodiment, the number of hidden layers and internal convolutional layers in an FSLSTM model jointly determine its complexity. Too few layers result in insufficient model parameters, making it difficult to capture correlations between data points; while too many layers make the model overly complex, increasing training difficulty. To select the optimal number of hidden and convolutional layers, this study uses mean squared error (MSE) as the evaluation metric. Detailed training results can be found in [link to relevant documentation]. Figure 2 and Figure 3 .

[0067] from Figure 5 It can be seen that as the number of training iterations increases, the mean squared error first decreases, but when the number of training iterations exceeds 300, the mean squared error begins to rise again.

[0068] Figure 4Analysis shows that, compared to other models, the proposed prediction model performs better in fitting real icing thickness data, especially in cases of abrupt changes in icing thickness. Traditional prediction models often fail to predict accurately, while the proposed model can adapt to such abrupt changes, maintaining a low prediction error. This advantage is mainly attributed to two factors: First, the decomposition module effectively reduces the complexity of input features, enabling the model to extract features more accurately; second, the proposed FSLSTM can deeply explore the correlation between icing thickness and meteorological factors, and the application of the Mixup algorithm enriches the dataset. These factors collectively improve the accuracy of icing thickness prediction.

[0069] Furthermore, Figure 6 The results show that the prediction error of the model in this paper corresponds to a relatively flat box, and the error midline is closer to 0. This indicates that the fluctuation of its prediction error is minimal, reflecting the stability and reliability of the model.

[0070] Part Two The inventive point of this invention: 1. Multi-source data fusion and intelligent acquisition mechanism Cross-dimensional data acquisition system: Innovatively introduces line status data such as real-time conductor tension and tower tilt angle, and constructs a three-dimensional input feature of "meteorology-icing-line status" with traditional meteorological and icing data, filling the gap in existing technologies that do not link the coupling relationship between line response and icing, and improving the completeness of feature dimensions.

[0071] Dynamic adaptive acquisition strategy: The acquisition frequency is intelligently adjusted based on meteorological conditions (temperature, wind speed). The acquisition density is automatically increased during the stages when icing is likely to form or grow rapidly. Combined with a dual synchronization mechanism of GPS time synchronization and local high-precision clock calibration, the timing consistency and timeliness of multi-source data are ensured.

[0072] 2. Refined data preprocessing techniques Triple anomaly verification and hierarchical processing: Construct a multi-dimensional anomaly identification system based on "physical constraints, statistical distribution, and temporal correlation". Differentiated strategies such as elimination, local weighted regression correction, and burst marker retention are adopted for different types of outliers, taking into account both data purification and key feature retention.

[0073] Hierarchical missing value intelligent completion: Based on the duration and type of missing values, a hierarchical scheme is adopted, which combines weighted average interpolation, multivariate LSTM prediction, and historical data migration combined with dynamic time warping (DTW) correction. With error feedback and iterative optimization, the accuracy and temporal rationality of the completed data are ensured.

[0074] Customized standardization solution: Design exclusive standardization strategies for numerical and angular data. Angular data retains periodicity through trigonometric function transformation. At the same time, establish a cross-batch data distribution verification mechanism to avoid the impact of dimensional differences and distribution offsets on model training.

[0075] 3. Dynamic Periodic Segmentation and Attention Enhancement Modeling Adaptive period extraction mechanism: Fast Fourier Transform (FFT) combined with wavelet packet transform (WPT) is used to mine multi-scale periodic features. The analysis range is dynamically adjusted by sliding calculation window and the core period is updated in real time, which solves the problem that traditional fixed period cannot adapt to time-varying features in different scenarios.

[0076] Hierarchical collaborative segmentation strategy: Implement non-overlapping, semi-overlapping, and fully overlapping hierarchical segmentation based on the difference in cycle scale, and combine the closed loop of dual indicators of cycle matching degree and trend consistency to ensure that the segmented data fits the ice cover change pattern and lays the foundation for feature extraction.

[0077] Attention-enhanced FSLSTM (Att-FSLSTM) structure: A factor attention layer is embedded between the convolutional layers and iterative layers of FSLSTM. The contribution weight of each input factor is dynamically calculated based on the feature variance, the cell state update formula of LSTM is optimized, the feature utilization efficiency of core factors such as temperature and tension is enhanced, and the parameters of multiple convolutional layers and hidden layers are optimized to take into account both local and global correlation capture.

[0078] 4. Contextualized data augmentation and efficient training strategies Targeted data augmentation for similar scenarios: Innovatively, K-means clustering is used to divide meteorological scenarios, and Mixup interpolation with scenario weights is performed within the same scenario to avoid cross-scenario sample interference, improve the correlation between augmented samples and actual icing patterns, and alleviate the problem of data scarcity.

[0079] Multi-stage joint training mechanism: Design a three-stage training process of warm-up-core-fine-tuning, combining AdamW optimizer, combined loss function (MSE+MAPE) and gradient accumulation strategy, and introducing early stopping mechanism and model snapshot management to balance training efficiency, stability and generalization ability.

[0080] Joint optimization of main and auxiliary tasks: Construct a joint training framework of "main task of icing thickness prediction + auxiliary task of line status anomaly identification", and use the gradient information of auxiliary tasks to improve the sensitivity of the main task to key features, thereby further optimizing the model performance.

[0081] 5. Intelligent Prediction and Dynamic Iteration System Dual-mode predictive inference engine: Provides two modes for different needs: fast prediction (lightweight model pruning, low latency) and accurate prediction (full model + sliding window rolling update), combined with multi-stage result correction (physical constraints + historical comparison + line cross-validation) to improve prediction reliability and provide confidence intervals.

[0082] Error feedback and incremental training closed loop: Real-time statistical analysis of prediction errors for different scenarios, automatic triggering of incremental training for high-error scenarios, updating only the attention layer and top-level parameters, combined with iterative optimization of parameters extracted dynamically to achieve continuous improvement in model performance.

[0083] Cross-scene transfer adaptation technology: Based on the transfer learning strategy, the pre-trained model is fine-tuned with a small amount of data in the new scene, the underlying feature extraction module is frozen, and the domain adaptive loss function is used to reduce the scene distribution difference, thereby reducing the data collection cost and adaptation difficulty in the new environment.

[0084] The embodiments in this specification are described in conjunction with the accompanying drawings. Since the effects are shown in the drawings, the descriptions are relatively simple. For specific methods, please refer to the descriptions in the detailed implementation section.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A multi-source data and dynamic cycle attention improved FSLSTM ice accretion prediction method, characterized by Includes the following steps: 1) Multi-source data acquisition and preprocessing: Collect basic data such as icing thickness, temperature, and humidity, and add two new types of line status data: real-time conductor tension and tower tilt angle, to construct a multi-dimensional dataset of "meteorology-icing-line status"; after removing outliers, use weighted average interpolation of 5 adjacent time steps to fill in missing values, and use Z-Score normalization to eliminate dimensional differences, thereby improving data quality and consistency; 2) Dynamic Periodic Segmentation and Attention Enhancement Modeling: Taking the current prediction time point as the endpoint, 20 time steps are taken as the sliding calculation window. The core ice-covering period is extracted in real time through Fourier transform, and the multi-source data is segmented according to the dynamic period. A factor attention layer is embedded between the convolutional layer and the iterative layer of FSLSTM to calculate the feature contribution of each input factor and assign dynamic weights. The LSTM cell state update formula is improved to enhance the influence of core factors on the prediction results. 3) Contextualized data augmentation and model training optimization: The samples are divided into two scenarios, cold and dry and mild and humid, according to meteorological characteristics. Within the same scenario, a linear interpolation strategy with scenario weights is used to generate augmented samples to improve the correlation between samples. The AdamW optimizer and MSE loss function are used to train the model. The appropriate learning rate and number of training iterations are set to ensure stable convergence of the model and suppress overfitting. 4) System prediction and dynamic adjustment: Based on the trained model, the preprocessed multi-source data is input, and after dynamic periodic segmentation, attention feature extraction and fusion, the predicted value of ice thickness is output. Real-time monitoring of forecast errors and changes in meteorological and line status parameters; dynamic adjustment of the model's periodic calculation window and attention weight allocation logic to adapt to the icing change patterns in different scenarios.

2. The multi-source data and dynamic cycle attention improved FSLSTM icing prediction method according to claim 1, characterized in that The method includes the following specific steps: Step S101: Multi-source data acquisition stage This step focuses on "full-dimensional feature coverage, high-precision synchronous acquisition, and multi-source data collaborative adaptation" as its core objectives. It constructs a three-tiered acquisition architecture of "distributed sensing - edge aggregation - cloud storage." Through multi-device collaborative linkage, dynamic frequency adjustment, and secure transmission mechanisms, it achieves accurate acquisition of meteorological, icing, and line status data, providing high-quality raw data for subsequent preprocessing and modeling. Specific details are as follows: The three-tier data acquisition architecture and communication mechanism consist of: a bottom layer deploying distributed sensing nodes, each covering 1-2 levels of power lines, integrating multiple types of sensors to achieve regional data acquisition; a middle layer edge gateway responsible for data preprocessing, temporary storage, and protocol conversion, employing dual-mode communication of dedicated power OPGW fiber optic cables and 5G industrial private networks to ensure real-time and stable data transmission, with end-to-end latency controlled within 500ms; and a top layer cloud server receiving data through an encrypted tunnel, using a distributed database for storage and indexing, supporting high-concurrency data writing and fast querying, while reserving interfaces for connection with power transmission line operation and maintenance platforms to achieve data sharing. Multiple data acquisition schemes are employed. Basic data acquisition, including ice thickness, utilizes a combination of indirect calculation and direct calibration: a 0.1-grade tension sensor and a dual-axis tilt sensor are installed at the conductor suspension point to calculate the equivalent ice thickness using a mechanical model; an ultrasonic sensor is deployed at the midpoint of the conductor and calibrated every 30 minutes to ensure an absolute error of no more than 1 mm; for meteorological data, a PT100 platinum resistance temperature sensor is installed inside a radiation shield; a capacitive humidity sensor is integrated with it; a three-dimensional ultrasonic anemometer at the top of the tower collects average wind speed every 10 minutes; and a silicon photovoltaic cell light sensor and an absolute pressure sensor simultaneously collect environmental parameters. Line status data acquisition: Conductor tension is monitored by fiber optic grating sensors installed at tension clamps. Wavelength drift characteristics are used to resist electromagnetic interference, and temperature compensation algorithms are combined to eliminate zero-point drift. Tower tilt angle is monitored by three-axis gyroscope sensors installed in the upper middle part of the tower. Wind vibration interference is eliminated by data fusion algorithms, and steady-state tilt angle components are extracted to accurately reflect the impact of icing load. The system features a dynamic acquisition frequency and synchronization mechanism. The acquisition frequency is adaptively adjusted according to meteorological conditions: 30 minutes / time under normal operating conditions; 20 minutes / time when icing is likely to form; and 10 minutes / time when the icing thickness exceeds 10mm. Synchronization uses a combination of GPS timing and local clock calibration. All sensors receive GPS timing once per hour, and the edge gateway is equipped with a high-precision clock to ensure that the acquisition time difference between different devices is ≤10ms, thus guaranteeing the consistency of time-series data. Data preprocessing and transmission optimization: The edge gateway performs preliminary processing on the raw data: data exceeding the range is removed by validity judgment, high-frequency fluctuating data is smoothed by moving average filtering, and data is compressed using the LZ77 algorithm before transmission; data is encrypted using SSL / TLS 1.3 protocol to achieve end-to-end secure transmission; a breakpoint resumption and retransmission mechanism is established, data is cached when the network is interrupted, and automatic resumption is performed after recovery; lost data is supplemented by interpolation of adjacent time steps to ensure data integrity; The equipment is designed to be environmentally adaptable. All outdoor equipment uses an IP67-rated housing, which is rainproof, snowproof, dustproof, and corrosion-resistant. The sensor bracket is hot-dip galvanized to enhance its rust resistance. Precision equipment such as fiber optic gratings are equipped with thermal insulation protective sleeves to ensure stable operation in extreme environments of -40℃ to 60℃, adapting to the maintenance needs of power transmission lines in different climate regions and ensuring the continuity and reliability of data collection. Step S102: Preprocessing stage; This step adopts a two-tier architecture of "local preprocessing at the edge gateway + deep preprocessing on the backend server," focusing on quality optimization and feature adaptation of multi-source data. The specific details are as follows: Two-level preprocessing architecture collaborative mechanism: The edge gateway prioritizes real-time preprocessing of the collected data: high-frequency noise in data such as wind speed and tension is initially suppressed by moving average filtering, abnormal values ​​exceeding the threshold are coarsely screened based on the sensor's rated range, and the data is compressed using the LZ77 algorithm, reducing the transmission bandwidth usage by more than 60%. After receiving the data, the backend server initiates a deep preprocessing process. Through the fusion of multiple algorithms, it achieves precise optimization of data quality. The two-level architecture is linked in real time through a dedicated power communication network to ensure that the preprocessing delay is controlled within 1 second, balancing real-time performance and processing accuracy. Multi-dimensional outlier identification and hierarchical processing: Constructing a triple anomaly verification system of "physical constraints - statistical distribution - temporal correlation": At the physical constraint level, data exceeding the equipment range, such as temperatures exceeding -40℃~60℃ and conductor tension exceeding 0~30kN, are removed; At the statistical distribution level, kernel density estimation combined with the 3σ criterion is used to mark suspicious outliers that deviate from the core interval of the data distribution and have a kernel density of less than 0.

01. At the temporal correlation level, by calculating the first and second differences between the data and the five adjacent time steps, temporal outliers that logically contradict historical trends and related data are identified. Differentiated processing is implemented for different anomaly types: deterministic anomalies are directly removed and equipment failure information is recorded; suspicious anomalies are corrected by local weighted regression scatter smoothing method, selecting data from 10 time steps before and after the anomaly point as the center and weighting the fitting; sudden anomaly data caused by extreme weather are retained and labeled after logical verification, providing key feature support for subsequent modeling. Hierarchical missing value intelligent completion strategy: Classification and processing according to missing duration and type: Random missing values ​​are interpolated by weighted average of adjacent time steps, with the weight decreasing exponentially with distance to ensure that the completed value conforms to the time series trend; Continuous missing data is predicted and filled using a multivariate LSTM model with related data such as temperature and humidity as input. Periodic missing data is addressed using a historical data migration method. This method aligns temporal characteristics through dynamic time warping and corrects for deviations in the current data trend to generate complete values. After completion, the algorithm is iteratively optimized by verifying the deviation with subsequent valid data to ensure that the completion accuracy deviation does not exceed 5%. Customized data standardization and dimensional unification: Customized standardization schemes are designed for different data types: Numerical data is normalized using Z-Score, converting it into a normal distribution with a mean of 0 and a variance of 1, eliminating dimensional differences; Angular data is first converted to [-π, π] radians, then transformed into two-dimensional features through sine and cosine transformations, preserving periodic information; After standardization of each batch of data, the mean and variance deviations of each feature are checked. If the deviation exceeds ±0.05, the global statistics are recalculated to ensure consistency in data distribution across batches. Timing consistency calibration and synchronization optimization: Based on the GPS time stamp standard, the time stamp deviation of different sensors is corrected. When the deviation is large, the data value is adjusted by linear interpolation. The acquisition frequency of 10 minutes / time, 20 minutes / time, and 30 minutes / time is unified to 20 minutes / time. High-frequency data is downsampled by sliding window averaging, and low-frequency data is upsampled by spline interpolation. The timing synchronization of related data is evaluated by cross-correlation coefficient to ensure that the cross-correlation coefficient of core data such as temperature and icing thickness is ≥0.8, laying the foundation for dynamic periodic extraction. Closed-loop quality assessment and optimization mechanism: Establish a multi-dimensional assessment index system, including data integrity, anomaly handling accuracy, completion accuracy, and time sequence synchronization; if the index does not meet the standards, automatically locate the problematic link, re-execute the corresponding preprocessing steps, and form a closed-loop process of "processing-assessment-optimization" to ensure that the output data meets the modeling requirements; Step S103: Dynamic periodic segmentation stage; This step focuses on "accurate capture of temporal features, real-time adaptation of periodic information, and collaborative segmentation of multi-source data." Through a comprehensive mechanism of "intelligent periodic extraction - dynamic window adjustment - multi-dimensional segmentation optimization," it overcomes the limitations of traditional fixed-period segmentation, providing segmented data that closely aligns with the icing change patterns for attention-enhanced FSLSTM modeling. The specific details are as follows: The dynamic periodicity extraction system is constructed based on multi-scale periodicity analysis. Using preprocessed multi-source fused data as the analysis object, a combined algorithm of Fast Fourier Transform (FFT) and Wavelet Packet Transform (WPT) is employed to achieve in-depth mining of periodic features. First, the time-domain data is converted to frequency-domain data using FFT, and the magnitude of each frequency component is calculated to initially screen potential periods. Then, WPT is used to finely decompose high-frequency components, capturing subtle fluctuations within short periods and compensating for the shortcomings of FFT in identifying the periods of non-stationary signals. The sliding calculation window dynamic adjustment strategy sets the sliding calculation window with the current forecast time point as the endpoint. The window length adopts an adaptive adjustment mechanism: it is dynamically adjusted according to the current meteorological complexity. When temperature and wind speed parameters fluctuate drastically, the window length is set to 20 time steps to ensure the capture of short-term rapid changes in the cycle; when the meteorological conditions are stable, the window length is expanded to 30 time steps to fully explore long-term cycle patterns; every time the window advances by one forecast time step, the cycle extraction process is automatically re-executed to achieve real-time updates of the cycle. The core cycle selection and verification were based on frequency domain modulus sorting, and the cycles corresponding to the top 3 largest modulus values ​​were selected as candidate core cycles. Through time series correlation verification, the mutual information value between the candidate cycles and the changes in icing thickness was calculated, and invalid cycles with mutual information values ​​below 0.6 were eliminated. Cross-validation was performed by combining the change cycle of line status data to ensure that the selected core cycles simultaneously adapt to the coupled change law of meteorology and line status. Finally, 3 effective core cycles were determined for segmentation. Multi-source data collaborative segmentation is implemented, and the data length is adaptively filled. To address the length differences of different core periods, a strategy of "intelligent completion + truncation optimization" is adopted to unify the data length: for data whose length cannot be divided by the period, virtual data that conforms to the time series trend is supplemented by a generative completion method based on generative adversarial networks. For data that exceeds the set threshold after completion, a slope-based truncation algorithm is used to retain the core trend part and avoid segmentation deviation caused by data redundancy; during the filling process, the proportion of virtual data is strictly controlled to not exceed 5% to ensure the authenticity of segmented data; The hierarchical and segmented execution logic implements hierarchical segmentation based on the scale differences of the core cycle: long-cycle data is used to capture the long-term trend characteristics of icing, and a non-overlapping segmentation method is adopted to ensure that each segment of data fully covers one cycle; medium-cycle data is used to mine medium-term fluctuation patterns, and a semi-overlapping segmentation method is adopted to strengthen the correlation between adjacent segments; short-cycle data is used to capture instantaneous mutation characteristics, and a fully overlapping segmentation method is adopted to avoid missing key mutation information; each level of segmented data is labeled with cycle attributes to facilitate targeted feature extraction in subsequent modeling. The segmented quality verification mechanism constructs a dual-indicator verification system of "cycle matching degree + trend consistency": the cycle matching degree is evaluated by calculating the spectral similarity between the segmented data and the corresponding core cycle, and if the similarity is less than 0.85, the data is re-segmented; Trend consistency is verified by comparing the changing trends of data such as icing thickness and conductor tension within the segment and using the Pearson correlation coefficient. If the correlation coefficient is lower than 0.7, the segment window position is adjusted. If the verification fails, it automatically backtracks to the cycle extraction stage to re-optimize the core cycle, forming a closed-loop process of "extraction-segmentation-verification-optimization". Segmented data enhancement and adaptation processing, secondary noise suppression within segments: To address the high-frequency noise remaining in the segmented data, an adaptive wavelet threshold denoising algorithm is adopted. The threshold is dynamically adjusted according to the noise intensity of each segment, suppressing noise interference while preserving detailed features, thus ensuring the purity of the segmented data. Cross-segment feature association enhancement constructs a segment association matrix by calculating the feature similarity of adjacent segment data. For adjacent segments with low similarity, a feature transfer algorithm is used to supplement association information and enhance temporal continuity. At the same time, timestamps and period labels are added to each segment data to facilitate the model to accurately identify segment attributes during training and prediction, thereby improving feature utilization efficiency. The modeling input adaptation optimization, based on the input requirements of attention-enhanced FSLSTM, converts the segmented data at each level into a feature matrix of a unified dimension. Through dimension normalization, it ensures that segmented data of different periods and scales can be collaboratively input into the model, laying the foundation for the subsequent fusion and extraction of local and global features. The entire dynamic segmentation process is executed automatically, with the processing latency of each batch of data controlled within 200ms, balancing processing accuracy and real-time performance. Step S104: Attention Enhancement Modeling Stage; This step focuses on "precisely focusing on core features, enhancing model generalization ability, and deeply mining temporal correlations" as its core objectives. It constructs an integrated process of "attention-enhanced modeling - scenario-based data augmentation - dynamic optimization iteration." Through structural innovation of the FSLSTM model and customized design of data augmentation strategies, it breaks through the limitations of traditional modeling methods. Specific details are as follows: The attention-enhanced FSLSTM modeling system adopts a four-level structure of "1D convolutional layer - factor attention layer - LSTM iterative layer - feature fusion layer" for the hierarchical feature extraction structure design model: the 1D convolutional layer extracts local spatiotemporal features for segmented data of different periods, and introduces nonlinear transformation through activation function to enhance feature expression ability; The factor attention layer, as a core innovative module, calculates the contribution weights based on the feature variance of each input factor, and assigns dynamic weights after normalization using the softmax function, thus achieving focused attention on core factors. The LSTM iterative layer incorporates a weight update mechanism, optimizing the cell state update formula as follows: wherein, is the cell state for the current time step t; is the forget gate output for the current time step t; is the cell state for the previous time step t-1; is the input gate output for the current time step t; is the candidate cell state for the current time step t; is the attention weight mean for the current time step t; Strengthening the impact of weights on temporal feature capture; The feature fusion layer employs a weighted summation strategy to fuse local and global features, outputting a high-dimensional comprehensive feature vector. The model parameter dynamic optimization mechanism adopts a combination strategy of "grid search + Bayesian optimization" to determine the optimal parameters: first, grid search is used to initially screen out the candidate range of parameters; then, based on the Bayesian optimization algorithm, the parameter combination is iteratively optimized with the mean square error of the validation set as the objective function, and finally the optimal configuration of learning rate 0.001, batch size 32, and number of hidden units 128 is determined. An early stopping mechanism is introduced during training. Training is automatically stopped when the validation set error does not decrease for 10 consecutive rounds to avoid overfitting. At the same time, a gradient clipping strategy is adopted to limit the gradient norm to within 5 to alleviate the gradient explosion problem. A multi-task joint training strategy is used to construct a joint training framework of "main task + auxiliary task": the main task uses MSE as the loss function to optimize the prediction accuracy of icing thickness; the auxiliary task uses a binary classification cross-entropy loss function to identify abnormal states of conductor tension and tower tilt angle, and uses the gradient information of the auxiliary task to improve the sensitivity of the main task to key features; a dynamic weight allocation mechanism is adopted, with the main task weight set to 0.7 and the auxiliary task weight to 0.3 in the early stage of training, and gradually adjusted to 0.9 for the main task and 0.1 for the auxiliary task as training iterations, to ensure that the model focuses on the core prediction target first; The scenario-based data augmentation solution employs a sample hierarchical classification mechanism based on the K-means clustering algorithm to divide the original samples into scenarios: using temperature, humidity, and wind speed as clustering features, the samples are divided into three typical scenarios: "cold and dry," "mild and humid," and "low temperature and strong wind," ensuring that the meteorological feature similarity of samples in the same scenario is higher than 0.8; and statistical distribution characteristics are collected separately for each scenario to provide a basis for subsequent targeted augmentation. The weighted Mixup enhancement strategy performs Mixup operation within the same scene, introducing scene weight γ and dynamic interpolation coefficient The scene weight is determined according to the proportion of the number of samples of each scene, ensuring that the enhanced samples conform to the actual scene distribution; the interpolation coefficient obeys Beta (5,5) distribution, and new samples are generated by linear interpolation, and the formula is: In the formula, For the generated enhanced feature samples; For the generated enhanced label samples; As scene weight; These are dynamic interpolation coefficients. Specifically, it follows a Beta(5,5) distribution; , Two types of original feature samples; , To and , The corresponding original label sample; During the augmentation process, the number of new samples is controlled to be 1.5 times that of the original samples to avoid data redundancy; To enhance sample quality verification, a dual-dimensional verification system of "distribution consistency + logical rationality" is constructed: distribution consistency is determined by the KS test to judge the distribution difference between the enhanced sample and the original sample, and the test statistic is greater than 0.05 to be considered qualified; logical rationality is verified by rule verification to eliminate contradictory samples; samples that fail the verification are automatically discarded and regenerated to ensure the effectiveness of the enhanced samples; Model training and adaptation optimization: The AdamW optimizer was used instead of the traditional Adam optimizer for training optimizer and loss function design. A weight decay coefficient of 0.001 was added to suppress overfitting of model parameters. The loss function adopted "MSE+MAPE weighted combination loss" with weights set to 0.6 and 0.4 respectively, which takes into account both the overall control of prediction error and the precise constraint of relative error. During training, a learning rate cosine annealing strategy is used, with an initial learning rate of 0.001, which is decayed to 0.8 of the current value every 50 iterations to ensure stable convergence of the model; The cross-scene transfer adaptation mechanism addresses the scene differences in different climate regions by introducing a transfer learning strategy: using a model trained in a cold and dry scene as a pre-trained model, fine-tuning it with a small number of samples in mild, humid, low-temperature, and windy scenes, freezing the parameters of the bottom convolutional layers, and only updating the parameters of the attention layer and the top fully connected layer to reduce the training data requirements in new scenes; during the fine-tuning process, a domain-adaptive loss function is used to reduce the distribution differences between the source scene and the target scene and improve the model's cross-scene adaptability. Real-time monitoring of model performance and iterative training process involves real-time monitoring of MSE, MAE, and MAPE metrics on the training and validation sets, plotting learning curves and error distribution histograms to identify underfitting or overfitting states. For underfitting, the number of hidden units in the network is increased or the number of training iterations is extended. For overfitting, Dropout layers are introduced or the number of augmented samples is increased, forming a closed-loop iterative process of "training-monitoring-optimization" to ensure that the model performance reaches its optimal level. Step S105: Scenario-based data augmentation and model training optimization, system prediction and dynamic adjustment stage; This step focuses on "efficient model convergence, controllable prediction accuracy, and dynamic scene adaptation" as its core objectives. It constructs a comprehensive system encompassing "refined training strategies, real-time prediction engines, and closed-loop optimization mechanisms." Through multi-dimensional optimization during training and intelligent adaptation in the prediction phase, it ensures stable and efficient model operation under different climatic scenarios. The specific details are as follows: A refined model training system is constructed, and the training data is hierarchically divided and adapted using a "stratified sampling + time-series preservation" strategy. Multi-source data is divided into training, validation, and test sets in a 7:2:1 ratio. The training set covers the entire scenario to ensure comprehensive learning, the validation set is co-distributed with the training set for monitoring training effectiveness, and the test set retains complete time-series segments to simulate actual predictions. To address the problem of imbalanced scenario samples, weighted sampling is used, assigning 1.5 times the weight to scarce scenario samples such as low temperature and strong winds to avoid the model being biased towards advantageous scenarios. The multi-stage training strategy is implemented in three phases: the warm-up phase uses a small learning rate of 0.0005 and a large batch size of 64 to quickly adapt to the data distribution; the core training phase uses a learning rate of 0.001 and a standard batch size of 32, combining the AdamW optimizer with the MSE+MAPE combined loss function to deeply optimize the parameters; the fine-tuning phase reduces the learning rate to 0.0001, freezes the bottom convolutional layers, and only updates the parameters of the attention layer and fully connected layer; at the same time, a gradient accumulation mechanism is introduced, accumulating gradients every 4 batches before updating, improving training stability. A multi-dimensional monitoring system is constructed for training process monitoring and early stopping mechanism to track changes in MSE, MAE, MAPE and loss function in real time, and to draw learning curves and error heatmaps. Set dual early stopping conditions: if the validation set error does not decrease for 10 consecutive rounds, or the difference between the validation set error and the training set error exceeds 0.05, immediately stop and backtrack to the optimal parameters; save a model snapshot every 50 rounds to record parameters and performance for easy backtracking management; The intelligent prediction engine is designed and implemented, including data preprocessing and adaptation before prediction. After the prediction data is input, real-time preprocessing is performed: Z-score normalization is performed using standardized parameters during the training phase to ensure consistent distribution; and the core period is updated through a dynamic period extraction algorithm to generate adapted segmented data. By leveraging feature alignment mechanisms, we ensure that the dimensions and order of the input features match those of the training data, thus avoiding bias. The multi-mode prediction inference mechanism is designed with a dual-mode inference engine: the fast prediction mode uses a pruned lightweight model, retains the core feature extraction module, and controls the latency within 100ms; the accurate prediction mode targets key periods such as the rapid growth period of icing, calls the complete model and combines a sliding window strategy to update the prediction results in a rolling manner; at the same time, a weighted average is used to fuse the prediction values ​​of adjacent time steps to reduce random errors. Post-processing and correction of prediction results employs multiple mechanisms to correct the prediction results: unreasonable values ​​are eliminated based on physical constraints; deviations are adjusted using linear formulas in conjunction with historical data from the same period; cross-validation is performed using line status data, and secondary prediction is initiated when trends contradict each other. The corrected results include confidence intervals based on the statistical distribution of errors, providing a reference for operation and maintenance decisions; The dynamic optimization and iteration mechanism predicts error feedback and establishes an error feedback closed loop, analyzing the source of deviation by scenario and time period. If the MAPE of a certain scenario exceeds 10%, incremental training is automatically triggered, adding new data to the training set, retraining the attention layer and top-level parameters, and updating the dynamic periodic extraction threshold to optimize the segmentation strategy. Cross-scene transfer adaptation targets new routes or climate regions. Based on the existing model as the pre-training foundation, it fine-tunes by inputting no less than 500 new scene data, freezes the underlying feature extraction module, and only updates the upper-layer parameters; it adopts a domain-adaptive loss function to reduce scene distribution differences and reduce data collection requirements. Model performance evaluation and updates are conducted monthly using both test sets and operational data. If core metrics drop by more than 15%, an update is initiated. Updated models are released in a phased manner, first on a small scale for pilot testing, and then fully deployed after meeting the standards. The entire process is fully automated, with manual intervention only triggered in case of anomalies, adapting to the operational needs of power transmission lines.

3. The multi-source data and dynamic cycle attention improved FSLSTM ice accretion prediction method according to claim 2, characterized in that In step S101, multiple data acquisition schemes are implemented, including basic data acquisition: 0.1 grade tensile sensor: 0-50kN; dual-axis tilt sensor: ±5°, 0.001° resolution; ultrasonic sensor: 0-100mm, error ≤±0.5mm; PT100 platinum resistance temperature sensor: -40℃~60℃, ±0.2℃; capacitive humidity sensor: 0-100% RH, ±2% RH; three-dimensional ultrasonic anemometer: 0-60m / s, ±0.1m / s; silicon photovoltaic cell illumination sensor: 0-200000Lux, ±5%; absolute pressure sensor: 50-110kPa, ±0.1kPa. Line status data acquisition: Fiber Bragg grating sensor: 0-30kN, 0.1 grade; Three-axis gyroscope sensor: ±10°, 0.001° resolution, 10Hz sampling rate; Dynamic acquisition frequency and synchronization mechanism, the acquisition frequency is adaptively adjusted according to meteorological conditions: normal operating conditions: temperature > 0℃, wind speed < 3m / s; icing-prone operating conditions: temperature ≤ 0℃ and wind speed ≥ 3m / s; high-precision clock: drift rate ≤ 1μs / h; Data preprocessing and transmission optimization: The edge gateway performs preliminary processing on the raw data: compression using the LZ77 algorithm: compression ratio ≥3:

1.

4. The multi-source data and dynamic cycle attention improved FSLSTM ice accretion prediction method according to claim 2, characterized in that In step S102, the following criteria apply: random missing data: ≤3 time steps; continuous missing data: 3-10 time steps; periodic missing data: >10 time steps; numerical data: temperature, icing thickness; angular data: tower tilt angle; data integrity: ≥99%; anomaly handling accuracy: ≥98%; completion accuracy: deviation ≤5%.

5. The multi-source data and dynamic periodic attention improved FSLSTM icing prediction method according to claim 2, characterized in that... In step S103, temperature and wind speed parameters fluctuate drastically: the variance exceeds 1.5 times the historical average; the window length is set to 20 time steps: corresponding to 400 minutes; the window length is extended to 30 time steps: corresponding to 600 minutes; combined with line status data: conductor tension, tower tilt angle; long period: >60 minutes; medium period: 30-60 minutes; semi-overlapping segmentation: overlap rate 50%; Short cycle: <30 minutes, full overlap segmentation: overlap rate 80%.

6. The multi-source data and dynamic periodic attention improved FSLSTM icing prediction method according to claim 2, characterized in that... In step S104, there are 3 1D convolutional layers with a kernel size of 3, based on various input factors: temperature, humidity, and wire tension; 3 hidden layers in the LSTM iterative layer; and parameter candidate ranges: learning rate 0.0005-0.002, batch size 16-64, and number of LSTM hidden units 64-256. Main task: Predicting icing thickness; Auxiliary task: Identifying abnormal line conditions; Scene weights are determined based on the proportion of samples in each scene: cold and dry 0.6, mild and humid 0.3, low temperature and strong wind 0.1; Identification model underfitting: both training and validation set errors are too high; Overfitting: low training set error and high validation set error; Dropout layer: dropout rate 0.

2.

7. The multi-source data and dynamic periodic attention improved FSLSTM icing prediction method according to claim 2, characterized in that... In step S105, the warm-up phase consists of the first 50 rounds; the core training phase consists of rounds 50-250. Fine-tuning phase: 250-300 rounds.

Images