A machine learning driven power line icing prediction method considering the influence of high and low air meteorological factors and terrain conditions, medium and program product

Abstract

The application discloses a machine learning driven electric wire icing prediction method considering the influence of high and low air meteorological factors and terrain conditions, a medium and a program product, and relates to the cross technical field of power meteorological disaster prediction and intelligent calculation. The method faces the power transmission line corridor, obtains icing event records and ground observation, multi-pressure layer reanalysis or numerical prediction and digital elevation model data; a time window is set at the event time and spatial neighborhood rules are used to complete space-time registration; a multi-dimensional feature vector containing low-altitude meteorological factors, high-altitude meteorological factors and pressure layer difference / gradient derived features and micro-terrain parameters is constructed; the icing thickness is divided into light and heavy icing, and the no-icing sample is completed under the same season or meteorological neighborhood constraint, and the prediction model is trained and adjusted in combination with the uneven processing; the icing occurrence probability and no / light / heavy grade results are output, which are used for line icing early warning, dispatching and operation and maintenance decision-making, and improve the early warning reliability and interpretability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power meteorological disaster prediction and intelligent computing. It involves the fusion of multi-source meteorological data, spatiotemporal registration, multi-dimensional feature construction and machine learning modeling prediction of icing risk in transmission line corridors. Specifically, it is a machine learning-driven method, medium and program product for predicting power line icing that takes into account the influence of high and low atmospheric meteorological factors and terrain conditions. Background Technology

[0002] Transmission lines are prone to icing under low temperature, high humidity, and precipitation-induced icing conditions in winter. This leads to a series of power grid safety risks, including increased conductor tension, changes in sag, insulator flashover, hardware damage, and tower collapse, and may trigger large-scale power outages. Especially in coastal and mountainous areas, the interaction of land and sea air masses, near-surface wind fields, and orographic uplift jointly influence water vapor transport and supercooled water droplet formation. The icing process exhibits sudden, localized, and highly nonlinear characteristics, making early identification and precise prediction of icing risks a crucial technical foundation for power grid disaster prevention and mitigation.

[0003] Existing icing forecasting technologies mainly include physical models, statistical empirical methods, and post-processing of numerical weather prediction products. Physical models can characterize collision efficiency, freezing fraction, and convective heat transfer, but they are highly dependent on the accuracy of input factors and often face problems such as difficulty in obtaining parameters and long computational chains. Statistical empirical methods are simple to implement but struggle to express multi-factor coupling and nonlinear effects, and are insufficiently adaptable to regional differences and interannual variations. Diagnostic methods based on numerical weather prediction (NWP) outputs are limited by model resolution and physical process parameterization; errors still exist in characterizing near-surface supercooled liquid water and boundary layer stability, and they are insufficient in representing the local topographic effects of transmission line corridors.

[0004] In recent years, research on icing prediction based on statistical learning and machine learning has gradually increased. However, existing methods still have several problems: First, icing samples are characterized by spatiotemporal scarcity and class imbalance, making it easy for models to under-identify strong icing events, affecting operational reliability. Second, different data sources (such as meteorological station observations and power fault records) differ in spatiotemporal representativeness, observation scope, and error structure. Simple fusion may introduce bias and reduce generalization stability. Third, the icing process is constrained by both high- and low-altitude meteorological conditions and topography. Existing methods still lack sufficient granularity in characterizing topography and near-surface local effects, making it difficult to meet the needs of refined early warning at the line scale. Fourth, the reliability, threshold controllability, and cross-year and cross-regional applicability of existing models under multi-level output still need further improvement.

[0005] In existing technologies, such as Chinese patent CN117010196B, a method for forecasting power line icing based on data assimilation and artificial intelligence is disclosed. This method improves the accuracy of meteorological fields through mesoscale models and three-dimensional variational assimilation, but it lacks sufficient exploration of upper-level factors related to atmospheric vertical stratification and its micro-topographic representation is relatively simplistic. CN120688704A discloses an icing growth prediction method considering the influence of micro-topography, which uses tower structure perturbations to correct local wind fields, but it heavily relies on real-time sensor monitoring and lacks deep integration of macro-meteorological dynamics and micro-topographic parameters. These methods all have limitations in the synergistic representation of multi-scale factors and struggle to capture the synergistic driving mechanism of upper-level circulation and local microclimate on icing.

[0006] In summary, existing icing prediction technologies still suffer from problems such as sample scarcity and imbalance, generalization instability caused by differences in data sources, insufficient representation of local topographic effects, and difficulty in balancing operational constraints of multi-level early warning systems under the background of multi-source observation and reanalysis. Therefore, achieving reliable prediction and representation of power line icing risk under complex terrain and multi-scale meteorological backgrounds is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] (a) Purpose of the invention To address the aforementioned deficiencies and shortcomings of existing technologies, this invention aims to provide a machine learning-driven method, medium, and program product for predicting power line icing that considers the influence of upper and lower atmospheric meteorological factors and topographic conditions. By performing spatiotemporal registration and consistency processing on multi-source meteorological observations, reanalysis of upper-air elements, and micro-topographic parameters, a feature set capable of characterizing boundary layer thermal and humidity states, water vapor transport conditions, and topographically induced local differences is constructed. Even under conditions of scarce icing samples and uneven category distribution, stable modeling and prediction of icing levels are achieved. This enables icing risk classification output and early warning support for transmission line corridors, thereby improving prediction accuracy, critical level recall capability, and cross-regional / cross-time applicability, and reducing the impact of missed and false alarms on power grid operation safety.

[0008] (II) Technical Solution To achieve the objective of this invention and solve its technical problems, the present invention adopts the following technical solution: The first objective of this invention is to provide a machine learning-driven method for predicting power line icing that considers the influence of upper and lower atmospheric meteorological factors and terrain conditions, comprising at least the following steps: SS1. Data Acquisition and Sample Event Construction: Acquire records of icing events along the target transmission line corridor and acquire multi-source data corresponding to each icing event, including surface meteorological observation data, multi-bar reanalysis data, numerical forecast data, and digital elevation model (DEM) data; SS2. Spatiotemporal registration of transmission line corridors: Based on the occurrence time of each icing event, a time matching window Δt is set, and multi-source data within Δt is extracted for time matching; based on the spatial location of each icing event, a neighborhood matching rule is established, and multi-source data is spatially registered with the transmission line corridor to form an input sample corresponding to each icing event and to construct a training sample set. SS3. Construction of Multidimensional Feature Vectors: Construct multidimensional feature vectors and establish a correlation with icing events. The stratigraphic state features are low-altitude meteorological factors extracted from surface meteorological observation data. The upper and lower stratification structure features include upper-altitude meteorological factors extracted from multi-pressure layer reanalysis or numerical forecast data, as well as derived features that characterize the vertical structure based on differential or gradient operations between different pressure layers. The topographic condition features are micro-topographic parameters extracted from DEM data. SS4. Icing Level Labeling and Sample Completion: Icing events are classified into light icing and heavy icing according to icing thickness. Samples of the no-icing category are constructed based on the time window constraint of the same season as the icing event or the neighborhood constraint of the same meteorological conditions, forming a multi-level training label set including no icing, light icing and heavy icing. SS5. Predictive Model Training and Parameter Tuning: Perform class imbalance processing on the multi-level training label set to ensure that the distribution of sample numbers in each class meets the preset training balance conditions; establish a machine learning prediction model with multi-dimensional feature vectors as input and icing level as output, and use a combination of parameter search and validation evaluation to train and tune the machine learning prediction model. SS6. Icing Prediction Output: Obtain corresponding multi-source data for the target transmission line corridor at the time to be predicted and use it as the sample to be predicted. Repeat steps SS2~SS3 to obtain the multi-dimensional feature vector to be predicted, input it into the trained model, and output the probability distribution and corresponding level results of the sample to be predicted belonging to no icing, light icing and heavy icing, forming power line icing early warning information.

[0009] The second objective of this invention is to provide a computer program product, including computer instructions, for executing the steps of the above-described machine learning-driven power line icing prediction method that considers the influence of high and low atmospheric factors and terrain conditions.

[0010] The third objective of this invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described machine learning-driven power line icing prediction method that considers the influence of high and low atmospheric meteorological factors and terrain conditions.

[0011] (III) Technical Effects Compared with the prior art, the machine learning-driven method, medium, and program product for predicting power line icing that considers the influence of high and low atmospheric meteorological factors and terrain conditions of the present invention have the following beneficial technical effects: (1) This invention is aimed at the transmission line corridor scenario. It integrates ground low-altitude meteorological observation, multi-pressure layer reanalysis or numerical forecast upper-altitude meteorological factors and micro-topographic parameters derived from digital elevation model. Through spatiotemporal registration, it constructs training samples that correspond one-to-one with icing events. It also uses differential or gradient calculations of different pressure layers to form stratified structure derived features, thereby more completely depicting the common constraints of temperature and humidity wind field and vertical thermal structure on icing occurrence, improving the ability to characterize sudden and local icing processes, and enhancing the accuracy and stability of icing level prediction.

[0012] (2) This invention constructs non-icing samples by constraining the same seasonal time window or the same meteorological conditions neighborhood, and combines class imbalance processing and parameter tuning to alleviate the bias learning problem caused by the scarcity of icing samples and the lack of samples of severe icing; and outputs the probability of icing occurrence and the classification results of no icing, light icing and severe icing, so that the warning threshold has controllability and business interpretability, reduces the risk of missed reporting of severe icing and improves the generalization ability of cross-regional and cross-time period applications. Attached Figure Description

[0013] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. Hereinafter, embodiments of the invention will be described in detail with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of a machine learning-driven method for predicting power line icing. Figure 2 This is a schematic diagram illustrating the implementation process of the graded training and progressive optimization strategy. Detailed Implementation

[0014] This invention aims to provide a machine learning-driven method, medium, and program product for predicting power line icing risk in transmission line corridors, integrating low-altitude meteorological factors, upper-altitude meteorological stratification factors, and terrain conditions. This method predicts the icing risk level of transmission line corridors and outputs results for no icing, light icing, and heavy icing. To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this invention, and are exemplary, intended to explain the invention, and should not be construed as limiting the invention. Based on the embodiments 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.

[0015] Example 1: Machine Learning-Driven Method for Predicting Power Line Icing like Figure 1 As shown in the embodiments of the present invention, the machine learning-driven power line icing prediction method considering the influence of high and low atmospheric meteorological factors and terrain conditions mainly includes the following steps in its implementation: SS1. Data Acquisition and Sample Event Construction: Acquire icing event records for the target transmission line corridor and obtain multi-source data corresponding to each icing event, including surface meteorological observation data, multi-barometric reanalysis data, numerical weather prediction data, and DEM data. Icing event records must include at least the time of icing occurrence, the latitude and longitude of the line, the icing thickness, and the icing type. The icing occurrence time is used as the time index for multi-source data extraction, and the latitude and longitude information is used as the positioning reference for spatial registration of the transmission line corridor. Furthermore, consecutive icing observations of the same line segment in adjacent time periods are merged into a single icing event according to a pre-defined event merging rule. This reduces the impact of duplicate annotations on the independence of training samples and improves the representativeness of event-level samples. The event merging rule includes at least time interval thresholds, spatial distance thresholds, and icing thickness continuity constraints to minimize the impact of duplicate annotations on sample independence and improve the representativeness of event-level samples.

[0016] In this embodiment of the invention, the ground meteorological observation data and the low-altitude meteorological factors subsequently extracted from it include air temperature, relative humidity, dew point temperature, and wind speed; the multi-pressure layer reanalysis or numerical forecast data are derived from reanalysis data or numerical weather prediction products, including hourly atmospheric element field data of multiple pressure layers near the ground and at least 200-1000 hPa, and the upper-altitude meteorological factors extracted from it include temperature, specific humidity or relative humidity, and wind field elements, covering at least two different pressure layers. The derived features include one or more of the temperature difference, humidity difference, or wind speed difference between different pressure layers to characterize the constraint effect of water vapor transport, vertical movement, and stratification stability on icing; the DEM data is 30m or higher resolution topographic data, covering the target transmission line corridor and its surrounding influence area, and the micro-topographic parameters extracted from it include at least one of altitude, slope, topographic relief, or windward index, and can further extract topographic gradient features along the line direction to characterize the local differences caused by abrupt changes in windward and leeward winds.

[0017] It should be noted that this step establishes an icing event merging rule to address the pseudo-independence problem caused by repeated sampling of the same icing process in traditional methods. For example, when icing is observed in multiple consecutive time periods along the same railway line and meteorological conditions evolve continuously, these are treated as different observation times of the same event. Only the sample at the peak of the icing thickness is retained, or a single sample is constructed using statistical features within the time window (such as maximum, mean, and cumulative values). This significantly improves the event representativeness of the training samples and the generalization ability of the model, effectively avoiding the risk of overfitting.

[0018] It should also be noted that this invention focuses on multi-source data acquisition and organization to meet modeling requirements: Low-altitude meteorological factors emphasize acquiring near-surface temperature, humidity, and wind field information close to the railway line's altitude, prioritizing gridded near-surface elements observed by stations near the railway corridor or spatially interpolated to characterize supercooled water droplet formation and near-surface stability changes; Upper-altitude meteorological factors emphasize acquiring multi-layered element fields from 200 to 1000 hPa, ensuring key layers simultaneously cover temperature, humidity, and wind fields to support the construction of vertical structural features such as interlayer temperature and humidity differences and wind speed differences, characterizing water vapor transport and stratification constraints; Topographic conditions focus on extracting micro-topographic parameters such as altitude, slope, undulation, and windward index from the DEM, and clarifying their statistical scope at the icing event location point and its neighborhood. Through the coordinated acquisition and consistent organization of these three types of data—low-altitude, high-altitude, and topographic—a stable input foundation is provided for subsequent spatiotemporal registration and feature construction.

[0019] SS2. Spatiotemporal registration of transmission line corridors: Based on the time of occurrence of each icing event, a time matching window Δt is set, and multi-source data within Δt is extracted for time matching. Neighborhood matching rules are established according to the spatial location of each icing event, and multi-source data is spatially registered with the transmission line corridor to form input samples that correspond one-to-one with each icing event and to construct a training sample set.

[0020] In this embodiment of the invention, the time matching window Δt is adaptively set according to the rate of change of meteorological elements before and after the icing event. This rate of change is calculated by the difference between low-altitude meteorological factors or high-altitude meteorological factors in adjacent time periods. When the rate of change exceeds a preset threshold, the coverage of Δt is expanded to ensure that the multi-source data used for modeling can include key meteorological evolution information of the icing triggering and enhancement stages. The neighborhood matching rule includes the definition of a spatial buffer centered on the spatial location of the icing event. Within the spatial buffer, reanalysis or numerical forecast grid elements and topographic elements are aggregated and extracted. The aggregation and extraction includes at least one of the mean, maximum, minimum, or quantile values ​​to construct a spatial registration result that can characterize the local environment of the transmission line corridor and suppress the error caused by the insufficient representativeness of a single grid point. For elements with missing measurements, neighborhood interpolation at the same time period or constrained interpolation of similar elements is used to ensure the integrity of the input.

[0021] Preferably, spatial registration discretizes the transmission line corridor into multiple sampling points distributed along the line direction, and performs neighborhood matching rules on each sampling point to obtain corresponding multi-source data. Furthermore, spatial registration adopts a resolution consistency strategy, resampling high-space point data and digital elevation model data on a unified spatial benchmark, and performing consistency compensation on the errors introduced by resampling. The consistency compensation establishes a deviation correction term by statistically analyzing the registration residuals of historical samples in the same area, and establishes temperature deviation term, humidity deviation term, and wind speed deviation term according to the element type. The deviation correction term is applied to the corresponding element to reduce the systematic deviation caused by cross-data source resolution differences.

[0022] It should be noted that spatiotemporal registration is a fundamental step in this invention to achieve collaborative modeling of high and low altitudes and terrain. Through adaptive Δt expansion, key evolutions can be covered when elements change rapidly due to cold air intrusion, topographic uplift, etc.; through aggregation such as buffers and quantiles, the representativeness bias of the reanalysis / forecast grid on mountain corridors can be reduced. Using a unified spatial benchmark and bias correction term allows different data sources to form comparable features at the same corridor scale, reducing systematic errors introduced by resolution differences.

[0023] SS3. Construction of Multidimensional Feature Vectors: Multidimensional feature vectors are constructed and correlated with icing events. The stratigraphic features are low-altitude meteorological factors extracted from surface meteorological observation data. The upper and lower strata structure features include upper-altitude meteorological factors extracted from multi-pressure layer reanalysis or numerical forecast data, as well as derived features representing vertical structures constructed based on differential or gradient operations between different pressure layers. The topographic features are micro-topographic parameters extracted from DEM data.

[0024] In this embodiment of the invention, after constructing the multidimensional feature vector, the statistical correlation between the training sample set and each candidate feature is quantitatively evaluated and a feature importance ranking is formed. This includes at least correlation analysis and significance testing. The correlation analysis is used to calculate the correlation coefficient between the icing thickness and each candidate feature. The significance test is used to determine whether the correlation coefficient meets the preset significance level. Candidate features that do not meet the preset significance level or whose absolute value of the correlation coefficient is lower than the preset threshold are eliminated. At the same time, redundancy suppression is performed on highly collinear feature pairs to reduce the impact of multicollinearity on training stability. This reduces feature redundancy and input noise while ensuring the icing level prediction performance.

[0025] It should be noted that the multidimensional feature vector constructed in this invention integrates three heterogeneous information sources: the ground layer, upper and lower strata, and micro-topography, overcoming the limitation of traditional icing prediction relying solely on a single meteorological observation. By introducing differential features of upper and lower strata (such as vertical temperature gradient and vertical humidity transport intensity), this invention can characterize the physical constraints of atmospheric stratification stability on the formation and transport of supercooled water droplets. Combined with micro-topography parameters, it achieves a quantitative characterization of the effects of topographic forced lifting and local wind convergence on icing enhancement, improving the model's physical interpretation ability and prediction accuracy for icing events under complex terrain conditions.

[0026] SS4. Icing Level Labeling and Sample Completion: Based on the thickness of the ice cover, icing events are classified into light icing and heavy icing. Based on the constraints of the same seasonal time window or the same meteorological conditions neighborhood constraints as the icing events, samples of the no-icing category are constructed to form a multi-level training label set including no icing, light icing and heavy icing.

[0027] Preferably, the icing level is divided into three categories based on the icing thickness: no icing, light icing, and heavy icing. Light icing corresponds to an icing thickness greater than 0 and not greater than 5 mm, while heavy icing corresponds to an icing thickness greater than 5 mm. The construction of the no-icing category samples must satisfy at least one of the following constraints: the same seasonal time window constraint or the meteorological condition neighborhood constraint. The same seasonal time window constraint selects the period in the same month or adjacent month as the icing event that did not occur as the no-icing sample. The same meteorological condition neighborhood constraint selects the period in which the temperature is in the range of 0-5 ℃ and the relative humidity is higher than 70% but no icing occurs as the no-icing sample. Furthermore, the high- and low-altitude stratification structure characteristics corresponding to the selected samples are required to be within a preset compatibility range to improve the comparability between the no-icing category samples and the icing category samples.

[0028] It should be noted that the core of the non-icing sample completion is to construct control samples with "similar meteorological backgrounds but no icing", rather than arbitrarily selecting non-icing periods. Therefore, it is necessary to apply seasonal constraints or meteorological neighborhood constraints, and introduce high and low air stratification structure compatibility conditions to exclude pseudo-negative samples with significantly different physical backgrounds. At the same time, it is necessary to combine spatial neighborhood exclusion, line maintenance or observation defect elimination rules to reduce label noise, so that the multi-level training labels are closer to the operational discrimination boundary.

[0029] SS5. Predictive Model Training and Parameter Tuning: Class imbalance processing is performed on the multi-level training label set to ensure that the distribution of sample numbers in each class meets the preset training balance condition; a machine learning prediction model is established with multi-dimensional feature vectors as input and icing level as output, and the machine learning prediction model is trained and its parameters are tuned by combining parameter search and validation evaluation. The sample division for training, validation and testing follows the principle of time independence or winter independence.

[0030] In this embodiment of the invention, the class imbalance handling is a combined balance constraint handling during the training phase, including: dividing the training sample set into a training subset and a validation subset according to a stratified sampling method, so that each icing level has sample coverage in both the training subset and the validation subset; performing synthetic minority class oversampling only on the lightly iced and / or heavily iced samples in the training subset to generate synthetic samples, the oversampling is based on the neighborhood interpolation of the minority class samples in the feature space to generate new samples, and increasing the number of minority class samples to a level that meets the preset balance condition, so that the ratio of the number of minority class samples to the number of majority class samples is not less than a preset ratio threshold; setting class loss weights for each icing level in the loss function of the model training, the weights being inversely proportional to the sample frequency of the corresponding class or determined according to a preset cost coefficient; validating and evaluating the model obtained by sampling and loss weighting training on the validation subset, and selecting the final training configuration with the macro average F1 value and / or the recall rate of the heavily iced class as constraint indicators to balance the overall stability and the target of controlling the underreporting of heavily iced samples.

[0031] As a preferred approach, the machine learning prediction model employs at least one of the extreme gradient boosting algorithm, random forest algorithm, or backpropagation (BP) neural network algorithm. When two or more models are established simultaneously, the classification prediction effects of no icing, light icing, and heavy icing are compared and evaluated based on the confusion matrix. The final model is determined based on at least one of the following evaluation indicators: Threat Score (TS) score, false negative rate, and false alarm rate, in order to meet the power disaster prevention business's requirements for controlling false negatives of heavy icing. The indicators related to heavy icing are set as priority constraints to improve availability in risk scenarios.

[0032] Furthermore, in step SS5, the parameter search and validation evaluation uses grid search or random search to traverse the model hyperparameters and evaluates candidate parameter combinations using cross-validation. Under the premise that the candidate parameter combinations meet the preset training balance conditions, the set of parameters that maximizes the macro-average F1 value or weighted F1 value on the validation set is selected as the tuning result to improve the overall stability and class balance of multi-level icing prediction. Moreover, the parameter search and validation evaluation adopts a training and validation partitioning method that is divided by time period or by winter, so that the validation subset is independent of the training subset in time, and the cross-winter validation results are used as constraints for model parameter tuning to improve the generalization stability of the model under cross-year climate background changes, thereby improving the availability of icing level early warning for transmission line corridors.

[0033] More preferably, the training and optimization of the machine learning prediction model adopts a hierarchical training and incremental optimization strategy, such as... Figure 2 As shown, its implementation includes at least the following sub-steps: S501. Hierarchical Label Reconstruction: The multi-level training label set is reconstructed into first-level labels and second-level labels according to the discrimination level. The first-level labels are binary classification labels for no icing and icing, and the second-level labels are binary classification labels for light icing and heavy icing. Training sample sets and validation sample sets corresponding to the first-level labels and the second-level labels are constructed respectively. S502. First-level model training and parameter tuning: Establish a first-level prediction model with multi-dimensional feature vectors as input and first-level labels as output. Tune the model parameters by combining parameter search and validation evaluation. Use the optimal preset evaluation index on the validation sample set as the model parameter selection criterion to obtain the trained first-level prediction model. S503. First-level threshold setting and error constraint: Based on the probability of icing occurrence output by the validation sample set, the first-level judgment threshold θ1 is set, and the false alarm rate of samples without icing does not exceed the preset upper limit is used as a constraint. The threshold that makes the recall capability of the icing category reach the optimal value is selected from the candidate threshold set that meets the constraint. S504. Construction of the ice-covered sub-sample set: Construct the ice-covered sub-sample set based on the sample subset determined by the first-level prediction model as ice-covered or the sample subset labeled as ice-covered by the training labels; S505. Second-level model training and parameter tuning: Based on the icing subsample set, a second-level prediction model is established with multi-dimensional feature vectors as input and second-level labels as output. The model parameters are tuned by combining parameter search and validation evaluation. The optimal preset evaluation index is used as the model parameter selection criterion to obtain the trained second-level prediction model. S506. Second-level threshold setting and error cost constraint: Based on the probability of severe icing output by the verification sample set, the second-level judgment threshold θ2 is set, and the false negative rate of severe icing does not exceed the preset upper limit is taken as the priority constraint. Among the candidate thresholds that meet the priority constraint, the threshold that makes the comprehensive discrimination performance of light icing and severe icing reach the optimal value is selected as θ2. S507. Hierarchical Serial Reasoning and Output: The first and second level prediction models are connected in series to form a hierarchical icing level prediction model. The sample to be predicted first passes through the first level prediction model and is judged as icy or not icy based on the threshold θ1. Only when it is judged as icy does it pass through the second level prediction model and is judged as lightly icy or heavily icy based on the threshold θ2. Finally, the prediction results and their corresponding prediction probability values ​​are output, which include three levels: no icing, lightly icing, and heavily icing. This achieves icing early warning output with controllable hierarchical thresholds and constrained error costs.

[0034] As a preferred option, step SS5 further includes an online model update and adaptive learning mechanism: setting model performance monitoring metrics, including prediction accuracy, recall, and critical success index (CSI) within a rolling time window; when the monitoring metrics continuously decline beyond a preset degradation threshold, it is determined that the model has experienced performance degradation or data distribution drift, triggering an incremental model update process, merging newly acquired icing event data with the original training sample set, and updating the model parameters using an incremental learning algorithm or model retraining method; through the online learning mechanism, the model continuously adapts to the long-term evolution trend of climate change and observation data distribution, and performs independent verification after the update to prevent degradation updates from being introduced into the business system.

[0035] It should be noted that the combined class imbalance handling and hierarchical training strategy proposed in this invention solves the dual challenges of class imbalance and heterogeneous error costs in multi-level icing prediction. Through a combination of constraints including oversampling, loss weighting, and hierarchical validation, sufficient learning of the minority classes (especially heavy icing) is ensured. Hierarchical training decomposes the three-class classification problem into two sub-tasks: icing identification and level differentiation. Each level independently optimizes the threshold and loss function, achieving refined control over the false alarm rate for no-icing and the false negative rate for heavy icing, significantly improving the operational availability and disaster prevention effectiveness of the early warning system.

[0036] SS6. Icing Prediction Output: To obtain the corresponding multi-source data of the target transmission line corridor at the time of prediction and use it as the sample to be predicted, repeat steps SS2~SS3 to obtain the multi-dimensional feature vector to be predicted, input it into the trained machine learning prediction model, and output the probability distribution and corresponding level results of the sample to be predicted belonging to the categories of no icing, light icing and heavy icing, thus forming power line icing early warning information.

[0037] In this embodiment of the invention, the probability of icing occurrence is obtained from the posterior probability of the categories output by the machine learning prediction model. The posterior probability of the categories is the probability distribution of the sample to be predicted belonging to the categories of no icing, light icing, and heavy icing. Before output, the posterior probability of the categories is subjected to probability calibration processing to ensure that it meets the preset consistency requirements with the actual frequency of icing occurrence. This includes at least: obtaining the original posterior probability of the categories of the model and the corresponding real labels based on the validation sample set, and establishing a mapping relationship from the original probability to the calibrated probability; applying the mapping relationship to the original posterior probability of the samples to be predicted to obtain the calibrated posterior probability of the categories; and setting the classification judgment thresholds for no icing, light icing, and heavy icing based on the calibrated icing occurrence probability to achieve controllable output of thresholds for multi-level early warning results.

[0038] As a preferred approach, the transmission line corridor is divided into multiple continuous spatial units, and the prediction results of icing probability and icing level for each spatial unit are subjected to spatial consistency constraints, including at least one of neighborhood consistency test or majority voting correction of the prediction results of adjacent spatial units, in order to reduce isolated false alarms and form continuous and interpretable icing early warning information along the line corridor.

[0039] It should be noted that step SS6 not only outputs category results but also improves the reliability and operational applicability of the prediction results through probability calibration and threshold control. Probability calibration establishes a mapping relationship between the output probability of the original model and the actual frequency of occurrence, eliminating probability distortion caused by the model's inherent overconfidence or calibration bias, and ensuring that the output probability accurately reflects the true risk level. Spatial consistency constraints, through neighborhood consistency checks and majority voting mechanisms, utilize the spatial continuity constraints of transmission line corridors to effectively suppress isolated anomaly prediction points, generating physically reasonable and spatially coherent early warning information, significantly improving the interpretability of early warning results and the credibility of power grid dispatching decisions.

[0040] Example 2: Application Case Based on the above embodiment 1, this embodiment 2 takes the icing service of a power transmission line in a certain province as the application scenario, constructs a three-classification forecast model for power line icing that considers high and low atmospheric meteorological factors and terrain conditions, and verifies the forecast performance through historical samples.

[0041] Following step SS1, icing data from stations with icing observation projects within the target area from 2003 to 2024, along with concurrent line icing fault records, were acquired. Simultaneously, hourly ERA5 reanalysis data (0.25°×0.25°, covering near-surface to 200–1000 hPa, variables including geopotential height, wind speed and direction, vertical velocity, specific humidity, relative humidity, and temperature) were obtained. 30 m resolution DEM topographic data was also acquired for topographic factor extraction. A sample set was established by integrating the station icing samples (417 records) and the line icing samples (107 cases). To avoid class imbalance, a corresponding number of non-icing events were selected as negative samples. The training set was constructed with 80% of the data, and the validation set with 20%. Output labels were categorized into three classes: no icing, light icing (0–5 mm), and heavy icing (>5 mm).

[0042] Based on steps SS2 and SS3, a time matching window Δt = ±3h was set using the time of the icing event as the benchmark, and multi-source data within the window were extracted for time matching. For spatial registration, a spatial buffer with a radius of 10km was established centered on the location of the icing event, and mean aggregation was performed on the reanalysis grid data and topographic features. The constructed multidimensional feature vector includes eight key factors: air temperature, relative humidity, dew point temperature, wind speed (surface layer state characteristics), 850-900hPa temperature difference, 850-925hPa temperature difference (high and low altitude stratification characteristics), altitude, and slope (topographic condition characteristics).

[0043] Correlation analysis (Table 1 below) shows that wind speed, temperature, dew point temperature, relative humidity, the two types of upper-level temperature differences, altitude, and slope are statistically correlated with icing thickness (significance threshold 0.01). Further analysis reveals that: when the line is icing, the air temperature is mainly distributed between -5 and 5 ℃, the relative humidity is mostly above 60% and 70%, and the wind speed mostly falls within the range of 0.1 to 10; the two types of upper-level temperature differences mostly fall between -4 and 4 ℃ and -6 and 6 ℃, respectively; the altitude of the line icing is mostly between 0 and 800 m, and the slope is mostly between 0 and 30.

[0044] Table 1. Correlation between ice thickness and various influencing factors Note: This indicates that the value passed the significance test at 0.01.

[0045] A random forest was used to build a three-class classification model for icing, and SMOTE was used to handle class imbalance. Hyperparameters were determined through grid search. The out-of-bag error rate was used as the criterion for hyperparameter tuning. By comprehensively comparing the error trends of different numbers of decision trees and feature splits, the optimal parameters were determined to be 4 feature splits, 100 decision trees, and 1 minimum number of leaves. As a control, a BP neural network three-class classification model was constructed: an 8-node input layer, two hidden layers with 128 and 64 nodes respectively, ReLU activation, and Softmax output; a learning rate of 0.001, a batch size of 256, a maximum iteration count of 300, and early stopping on the validation set was introduced to suppress overfitting.

[0046] Table 2. Accuracy, Root Mean Square Error, and Mean Absolute Error of Random Forest and Neural Network Models As shown in Table 2, for line samples, the random forest model has an accuracy of 0.75, a MAE of 0.35, and an RMSE of 0.742, which is better than the BP model (accuracy 0.429, MAE 0.81, RMSE 1.134); for station samples, the random forest model has an accuracy of 0.846, a MAE of 0.154, and an RMSE of 0.392, which also shows good performance.

[0047] Table 3. TS score, false negative rate, and empty report rate of random forest and neural network models As shown in Table 3, the Random Forest model achieved a TS score of 0.791, a false negative rate of 0.068, and a false alarm rate of 0.161 for light icing events (0-5 mm); and a TS score of 0.419, a false negative rate of 0.133, and a false alarm rate of 0.552 for heavy icing events (>5 mm). In comparison, the BP model had a higher false negative rate (0.545) for heavy icing events, while the Random Forest model maintained a false negative rate of approximately 0.1, which better meets the requirements for reducing false negatives in power disaster prevention scenarios. Therefore, in this embodiment 2, the Random Forest model was ultimately adopted as the operational forecasting model. In actual operation, the false alarm characteristics were combined to classify and manually verify the heavy icing threshold to achieve a balance between early warning availability and resource investment. According to step SS6, the trained Random Forest model was deployed to a provincial transmission line icing early warning platform to make real-time predictions for the winter of 2024 (December to February of the following year). Practical application results show that the system successfully provided early warning for 86% of severe icing events, with a false alarm rate of less than 15%, meeting the business needs of the power grid dispatching department and providing effective technical support for disaster prevention and mitigation of transmission lines.

[0048] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A machine learning-driven method for predicting power line icing that considers the influence of upper and lower atmospheric meteorological factors and terrain conditions, characterized in that, It should include at least the following steps: SS1. Acquire records of icing events along the target transmission line corridor and corresponding multi-source data, including surface meteorological observation data, multi-barosphere reanalysis data, numerical weather prediction data, and DEM data; SS2. Based on the occurrence time of each icing event, a time matching window Δt is set, and multi-source data within Δt is extracted for time matching; based on the spatial location of each icing event, a neighborhood matching rule is established, and the multi-source data is spatially registered with the transmission line corridor; finally, a training sample set is formed. SS3. Construct multidimensional feature vectors and establish a correlation with icing events. The stratigraphic features are low-altitude meteorological factors extracted from surface meteorological observation data. The upper and lower stratification features include upper-altitude meteorological factors extracted from multi-pressure layer reanalysis or numerical forecast data and derived features that characterize the vertical structure based on differential or gradient operations between different pressure layers. The topographic features are micro-topographic parameters extracted from DEM data. SS4. Classify icing events into light icing and heavy icing according to icing level, and construct no-icing category samples based on the same seasonal time window constraint or the same meteorological condition neighborhood constraint as the icing event, forming a multi-level training label set including no icing, light icing and heavy icing. SS5. Perform class imbalance processing on multi-level training label sets, establish a machine learning prediction model with multi-dimensional feature vectors as input and icing level as output, and use a combination of parameter search and validation evaluation to train and tune the parameters of the machine learning prediction model. SS6. Obtain multi-source data corresponding to the target transmission line corridor at the time to be predicted and use it as the sample to be predicted. Repeat steps SS2 to SS3 to obtain the multi-dimensional feature vector to be predicted. Input it into the trained machine learning prediction model and output the probability distribution and corresponding level results of the sample to be predicted belonging to the categories of no icing, light icing and heavy icing.

2. The method according to claim 1, characterized in that, In step SS1, the icing event record includes at least the time information of icing occurrence, the latitude and longitude information of the line, the icing thickness, and the icing type. The icing occurrence time is used as the time index for multi-source data extraction, and the latitude and longitude information is used as the positioning reference for spatial registration of the transmission line corridor. At the same time, continuous icing observations of the same line section in adjacent time periods are merged into the same icing event according to the preset event merging rules.

3. The method according to claim 1 or 2, characterized in that, In steps SS1 and SS3, the surface meteorological observation data and the low-altitude meteorological factors extracted from it include air temperature, relative humidity, dew point temperature, and wind speed; the multi-pressure layer reanalysis or numerical forecast data are derived from reanalysis data or numerical weather prediction products, including hourly atmospheric element field data of multiple pressure layers near the surface and at least 200-1000 hPa, and the upper-altitude meteorological factors extracted from it include temperature, specific humidity or relative humidity, and wind field elements, covering at least two different pressure layers, and the derived features include one or more of the temperature difference, humidity difference, or wind speed difference between different pressure layers; the DEM data covers the target transmission line corridor and its surrounding influence area, and the micro-topographic parameters extracted from it include one or more of the altitude, slope, topographic relief, or windward index.

4. The method according to claim 1, characterized in that, In step SS2, the time matching window Δt is adaptively set according to the rate of change of meteorological elements before and after the icing event. The rate of change is calculated by the difference between low-altitude meteorological factors or high-altitude meteorological factors in adjacent time periods, and the coverage of Δt is expanded when the rate of change exceeds a preset threshold. The neighborhood matching rule includes the definition of a spatial buffer centered on the spatial location of the icing event, and the reanalysis or numerical forecast grid elements and topographic elements are aggregated and extracted within the spatial buffer.

5. The method according to claim 1 or 4, characterized in that, In step SS2, the spatial registration discretizes the transmission line corridor into multiple sampling points distributed along the line direction, and performs neighborhood matching rules on each sampling point to obtain the corresponding multi-source data. Spatial registration employs a resolution consistency strategy, resampling high-space point data and digital elevation model data on a unified spatial benchmark, and performing consistency compensation on the errors introduced by resampling. The consistency compensation establishes a deviation correction term by statistically analyzing the registration residuals of historical samples in the same area, and applies the deviation correction term to the corresponding elements to reduce the systematic deviation caused by cross-data source resolution differences.

6. The method according to claim 1, characterized in that, In step SS3, after the construction of the multidimensional feature vector is completed, the statistical correlation between ice thickness and each candidate feature is quantitatively evaluated based on the training sample set, and the feature importance ranking is formed. This includes at least correlation analysis and significance test. The correlation analysis is used to calculate the correlation coefficient between ice thickness and each candidate feature, and the significance test is used to determine whether the correlation coefficient meets the preset significance level. Candidate features that do not meet the preset significance level or whose absolute value of the correlation coefficient is lower than the preset threshold are eliminated.

7. The method according to claim 1, characterized in that, In step SS4, the icing level is divided into three categories based on the icing thickness: no icing, light icing, and heavy icing. Light icing corresponds to an icing thickness greater than 0 and not greater than 5 mm, while heavy icing corresponds to an icing thickness greater than 5 mm. The construction of the no-icing category samples must satisfy at least one of the following constraints: the same seasonal time window constraint or the meteorological condition neighborhood constraint. The same seasonal time window constraint selects the period in the same month or adjacent month as the icing event that did not occur as the no-icing sample. The same meteorological condition neighborhood constraint selects the period in which the temperature is in the range of 0~5 ℃ and the relative humidity is higher than 70% but no icing occurs as the no-icing sample. The high and low altitude stratification structure characteristics corresponding to the selected samples are required to be within a preset compatibility range.

8. The method according to claim 1, characterized in that, In step SS5, the class imbalance handling is a combined equalization constraint handling during the training phase, including: The training sample set is divided into a training subset and a validation subset using stratified sampling, so that each icing level has sample coverage in both the training subset and the validation subset. Synthetic minority oversampling is performed only on lightly iced and / or heavily iced samples in the training subset to generate synthetic samples. The oversampling generates new samples based on neighborhood interpolation of minority samples in the feature space and increases the number of minority samples to a level that meets a preset balance condition, such that the ratio of the number of minority samples to the number of majority samples is not lower than a preset ratio threshold. In the loss function of model training, category loss weights are set for each icing level. The category loss weights are inversely proportional to the sample frequency of the corresponding category or determined according to a preset cost coefficient. The model obtained by sampling and loss-weighted training is validated and evaluated on the validation subset, and the final training configuration is selected by using the macro-average F1 score and / or the recall rate of the heavily iced category as constraints.

9. The method according to claim 1 or 8, characterized in that, In step SS5, the machine learning prediction model adopts at least one of the extreme gradient boosting algorithm, random forest algorithm, or BP neural network algorithm. When two or more models are established at the same time, the classification prediction effect of no ice, light ice and heavy ice is compared and evaluated based on the confusion matrix, and the final model is determined according to at least one of the evaluation indicators of TS score, false negative rate and false alarm rate.

10. The method according to claim 1 or 8, characterized in that, In step SS5, the parameter search and validation evaluation uses grid search or random search to traverse the model hyperparameters and evaluates the candidate parameter combinations using cross-validation. Under the premise that the candidate parameter combinations meet the preset training balance conditions, the set of parameters that maximizes the macro-average F1 value or weighted F1 value on the validation set is selected as the tuning result. Furthermore, the parameter search and validation evaluation adopts a training and validation division method that is divided by time period or by winter, and the cross-winter validation results are used as the constraint conditions for model parameter tuning.

11. The method according to claim 1 or 8, characterized in that, In step SS5, the training and optimization of the machine learning prediction model adopts a hierarchical training and incremental optimization strategy, including: S501. Hierarchical Label Reconstruction: The multi-level training label set is reconstructed into first-level labels and second-level labels according to the discrimination level. The first-level labels are binary classification labels for no icing and icing, and the second-level labels are binary classification labels for light icing and heavy icing. Training sample sets and validation sample sets corresponding to the first-level labels and the second-level labels are constructed respectively. S502. First-level model training and parameter tuning: Establish a first-level prediction model with multi-dimensional feature vectors as input and first-level labels as output. Tune the model parameters by combining parameter search and validation evaluation. Use the optimal preset evaluation index on the validation sample set as the model parameter selection criterion to obtain the trained first-level prediction model. S503. First-level threshold setting and error constraint: Based on the probability of icing occurrence output by the validation sample set, the first-level judgment threshold θ1 is set, and the false alarm rate of samples without icing does not exceed the preset upper limit is used as a constraint. The threshold that makes the recall capability of the icing category reach the optimal value is selected from the candidate threshold set that meets the constraint. S504. Construction of the ice-covered sub-sample set: Construct the ice-covered sub-sample set based on the sample subset determined by the first-level prediction model as ice-covered or the sample subset labeled as ice-covered by the training labels; S505. Second-level model training and parameter tuning: Based on the icing subsample set, a second-level prediction model is established with multi-dimensional feature vectors as input and second-level labels as output. The model parameters are tuned by combining parameter search and validation evaluation. The optimal preset evaluation index is used as the model parameter selection criterion to obtain the trained second-level prediction model. S506. Second-level threshold setting and error cost constraint: Based on the probability of severe icing output by the verification sample set, the second-level judgment threshold θ2 is set, and the false negative rate of severe icing does not exceed the preset upper limit is taken as the priority constraint. Among the candidate thresholds that meet the priority constraint, the threshold that makes the comprehensive discrimination performance of light icing and severe icing reach the optimal value is selected as θ2. S507. Hierarchical Serial Reasoning and Output: The sample to be predicted first goes through the first-level prediction model and is judged as icy or not based on the threshold θ1. Only when it is judged as icy does it go through the second-level prediction model and is judged as lightly icy or heavily icy based on the threshold θ2.

12. The method according to claim 1, characterized in that, In step SS6, the probability of icing occurrence is obtained from the posterior probability of the class output by the machine learning prediction model, corresponding to the probability distribution of the sample to be predicted belonging to the categories of no icing, light icing, and heavy icing; and a probability calibration process is performed on the posterior probability of the class before output, including: obtaining the original posterior probability of the class and the corresponding true label of the model based on the validation sample set, and establishing a mapping relationship from the original probability to the calibrated probability; applying the mapping relationship to the original posterior probability of the sample to be predicted to obtain the calibrated posterior probability of the class; and setting the classification threshold for no icing, light icing, and heavy icing based on the calibrated probability of icing occurrence.

13. The method according to claim 1 or 12, characterized in that, In step SS6, the transmission line corridor is divided into multiple continuous spatial units, and the prediction results of icing probability and icing level of each spatial unit are subjected to spatial consistency constraints, including at least one of neighborhood consistency test or majority voting correction of the prediction results of adjacent spatial units.

14. A computer program product, characterized in that, Includes computer instructions for performing the steps of the machine learning-driven power line icing prediction method according to any one of claims 1 to 13, which takes into account the influence of high and low atmospheric factors and terrain conditions.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the machine learning-driven power line icing prediction method according to any one of claims 1 to 13, which takes into account the influence of high and low atmospheric meteorological factors and terrain conditions.

Images