Intelligent optimization method of data-driven extra-high voltage line deicing and melting method

Abstract

The invention provides an intelligent optimization method of a data-driven extra-high voltage line melting and deicing method, and the method comprises the steps: carrying out the space-time alignment and preprocessing of multi-source data after the collection is completed because the data from different sources have significant differences in the aspects of sampling frequency, time scale and space coordinate expression; in the time dimension, resampling and interpolation processing are carried out on data of different sampling periods by constructing a unified time axis, so that the data are mapped to the same time scale; in the spatial dimension, a spatial coordinate projection method is adopted, meteorological network data, line towers and section positions are subjected to correlation mapping, the high-precision effect prediction performance is shown as the fact that the system has excellent overall prediction capacity after fusing a multi-source data set, and compared with a single data source decision, the high-precision effect prediction performance is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of ultra-high voltage transmission lines, and more particularly to a data-driven intelligent optimization method for de-icing of ultra-high voltage transmission lines. Background Technology

[0002] Currently, ultra-high voltage (UHV) transmission lines face severe threats to safe operation during winter icing disasters. Existing de-icing methods include current-based de-icing, laser de-icing, drone de-icing, manual de-icing, robotic de-icing, electromagnetic pulse de-icing, and ground wire anti-icing reinforcement. These methods vary in terms of technicality, economy, safety, effectiveness, and timeliness, lacking a unified and scientific evaluation and optimization mechanism. Traditional evaluation methods rely heavily on expert experience, resulting in strong subjectivity, slow response, and difficulty in dynamically adapting to complex environmental changes. With the development of big data and artificial intelligence technologies, there is an urgent need to construct a data-driven intelligent optimization method to achieve scientific, rapid, and dynamic decision-making for de-icing methods. Existing single-category de-icing methods still face many challenges: Traditional methods have significant limitations and lack universality. For example, electric de-icing requires power outages or significant adjustments to the line's operation, affecting power supply reliability, and is ineffective for excessively thick or asymmetrical icing. Drone and robot de-icing are limited by payload capacity, flight time, and complex geographical and meteorological conditions (such as strong winds and dense fog), making them unsuitable for large-scale, long-distance line de-icing tasks. Manual de-icing is inefficient, carries extremely high safety risks, and cannot operate promptly in extreme weather conditions. No single method can independently address all icing scenarios. Currently, minute-level weather forecasts (ice thickness, temperature, humidity, wind speed, maximum wind deflection angle, maximum wind tilt angle), simulation data, real-time line loads, and geographic information system terrain data have not been coupled with the effectiveness models of different de-icing methods, resulting in decision-making lagging behind environmental changes and making it impossible to achieve precise interventions that are "preventive" or "adaptive". The selection and decision-making of existing de-icing methods largely rely on human experience, lacking an optimal selection mechanism. Traditional practices are usually based on expert experience or fixed rules, which have limitations such as strong subjectivity, slow response speed, and difficulty in adapting to complex and changing field environments.

[0003] Therefore, there is an urgent need for an intelligent optimization method that can integrate multi-source historical data, rely on machine learning algorithms, and has high interpretability to guide the decision-making of de-icing on UHV lines, thereby improving the timeliness, safety, and economy of power grid de-icing work. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by providing a data-driven intelligent optimization method for de-icing of ultra-high voltage power lines, and to establish a unified, dynamic, and quantifiable effect evaluation and comparison model for de-icing methods.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an intelligent optimization method for data-driven de-icing of ultra-high voltage power lines, comprising: multi-source heterogeneous data fusion processing, specifically: Data from different sources differ significantly in sampling frequency, time scale, and spatial coordinate representation. After collection, spatiotemporal alignment and preprocessing of multi-source data are required. In the time dimension, by constructing a unified time axis, data from different sampling periods are resampled and interpolated to map to the same time scale. In the spatial dimension, the spatial coordinate projection method is used to associate and map meteorological network data with the location of power line towers and sections; To reduce the impact of measurement noise on subsequent feature extraction, a Kalman filter is introduced to denoise and estimate the state of the continuous time series. The state update process is represented as follows: ; in, For the first The state vector at each moment; This is the state transition matrix; This is process noise; For observation vectors; The observation matrix; To observe noise; Provided that the data quality meets the requirements, the multi-source features are normalized and dimensionality reduced to eliminate the influence of differences in dimensions and scales on the fusion results. The normalization process is as follows: ; in, These are the normalized eigenvalues; These are the original eigenvalues; This represents the mean of the original feature samples; is the standard deviation of the original feature samples.

[0006] Furthermore, principal component analysis is used to reduce the dimensionality of the high-dimensional feature space. Eigenvalue decomposition is performed using the covariance matrix, and principal components with sufficient cumulative contribution rates are selected. This reduces feature redundancy while preserving key information. The covariance matrix is ​​as follows: ; in, The characteristic covariance matrix; The number of samples; For the first Feature vectors of each sample; It is the mean vector of all sample feature vectors; The fusion process is performed using a weighted mapping method as follows: ; in, This is the final fused feature vector; These are weighting coefficients used to measure the weighting factor of the first... The relative importance of class features in the fusion process; For the first Class feature vectors.

[0007] After feature fusion is completed, the discriminative ability of the fused features is evaluated to verify its ability to distinguish between different icing states and the effectiveness of de-icing schemes. This is measured by the ratio of inter-class divergence to intra-class divergence. The discriminant index is as follows: ; in, As an indicator for feature discrimination capability; The projection direction vector; The inter-class scatter matrix; It is the intra-class scatter matrix.

[0008] Furthermore, it also includes effect evaluation and model optimization, specifically: Using known de-icing performance evaluation results from historical samples as supervision signals, model training is completed by minimizing the error between the predicted result and the actual evaluation result. The overall optimization objective of the model consists of a loss function and a regularization term, and the objective function is as follows: ; in, The objective function is... This represents the actual data value. These are the model's predicted values; This is the loss function for single-sample prediction error, used to measure the deviation between the model output and the actual evaluation results; This is a lateralization function used to constrain the structural complexity of the base learner.

[0009] The loss function guides the model to continuously approximate the actual ice-melting effect evaluation results, enabling the model to have high fitting accuracy. The regularization function, by penalizing the model complexity, suppresses the occurrence of excessive invalid splits or overly deep structures. The regularization function is expressed as: ; in, The regularization coefficient is used. The number of leaf nodes in the base learner; These are the weight parameters corresponding to the leaf nodes.

[0010] Furthermore, the historical sample data is divided into several non-overlapping subsets. In each round of training, one subset is selected as the validation set, and the remaining subsets are used as the training set. This process is repeated multiple times to complete the training and validation. The overall validation error is then: ; in, The overall validation error of cross-validation; The number of folds for cross-validation; For the first The single-round verification error of round verification.

[0011] The beneficial effects of this invention are as follows: A multi-source heterogeneous data fusion processing module is added. This module is used to collect and standardize multi-source data from different systems in real time, constructing a decision feature database. The data includes minute-level weather forecasts (ice thickness, temperature, humidity, wind speed, maximum wind deflection angle, maximum wind tilt angle), simulation data, real-time line loads, and key influencing factors from geographic information system terrain data.

[0012] An ensemble learning algorithm, XGBoost, is employed as the core decision engine. Training samples are constructed using a historical case library, with multi-source data as input features and the comprehensive performance score of different de-icing methods as the training objective. The model's hyperparameters are optimized through cross-validation, resulting in a high-precision de-icing performance evaluation model. This model can learn complex nonlinear relationships, predict the expected performance of each candidate method in a given scenario, and automatically prioritize candidate solutions.

[0013] The high-precision performance of the system demonstrates its excellent overall predictive capability after fusing multi-source datasets, showing a significant improvement compared to decision-making based on a single data source. Thanks to the deep coupling of meteorological forecasts and line monitoring data, the system addresses the pain point of traditional models failing in extreme environments. Its optimal matching capability is greatly enhanced for complex scenarios such as thick icing, asymmetric icing, and strong winds and dense fog, providing accurate and reliable support for UHV de-icing decisions under extreme conditions.

[0014] An interpretability output module has been added to enhance model credibility by integrating SHAP interpretability analysis. It provides confidence intervals for model predictions and offers risk alerts for missing data or boundary scenarios. All results are presented through a visual decision dashboard to assist operations personnel in making final decisions. Attached Figure Description

[0015] Figure 1 A flowchart of an intelligent optimization method for a data-driven de-icing method for ultra-high voltage power transmission lines; Figure 2 This is a flowchart of Example 1; Figure 3 shows the simulation data output by the de-icing jump dynamic model of the iron tower conductor system in Example 1; Figure 4 shows the ten-fold cross-validation diagram when K=10; Figure 5 shows the weighting of the influence of each feature on the decision outcome. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Please see Figure 1 A data-driven intelligent optimization method for de-icing of ultra-high voltage power lines includes: multi-source heterogeneous data fusion processing, specifically: Data from different sources differ significantly in sampling frequency, time scale, and spatial coordinate representation. After collection, spatiotemporal alignment and preprocessing of multi-source data are required. In the time dimension, by constructing a unified time axis, data from different sampling periods are resampled and interpolated to map to the same time scale. In the spatial dimension, the spatial coordinate projection method is used to associate and map meteorological network data with the location of power line towers and sections; To reduce the impact of measurement noise on subsequent feature extraction, a Kalman filter is introduced to denoise and estimate the state of the continuous time series. The state update process is represented as follows: ; in, For the first The state vector at each moment; This is the state transition matrix; This is process noise; For observation vectors; The observation matrix; To observe noise; Provided that the data quality meets the requirements, the multi-source features are normalized and dimensionality reduced to eliminate the influence of differences in dimensions and scales on the fusion results. The normalization process is as follows: ; in, These are the normalized eigenvalues; These are the original eigenvalues; This represents the mean of the original feature samples; is the standard deviation of the original feature samples.

[0018] Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional feature space. Eigenvalue decomposition is performed using the covariance matrix, and principal components with satisfactory cumulative contribution rates are selected. This approach preserves key information while reducing feature redundancy. The covariance matrix is ​​as follows: ; in, The characteristic covariance matrix; The number of samples; For the first Feature vectors of each sample; It is the mean vector of all sample feature vectors; The fusion process is performed using a weighted mapping method as follows: ; in, This is the final fused feature vector; These are weighting coefficients used to measure the weighting factor of the first... The relative importance of class features in the fusion process; For the first Class feature vectors.

[0019] After feature fusion is completed, the discriminative ability of the fused features is evaluated to verify its ability to distinguish between different icing states and the effectiveness of de-icing schemes. This is measured by the ratio of inter-class divergence to intra-class divergence. The discriminant index is as follows: ; in, As an indicator for feature discrimination capability; The projection direction vector; The inter-class scatter matrix; It is the intra-class scatter matrix.

[0020] It also includes effect evaluation and model optimization, specifically: Using known de-icing performance evaluation results from historical samples as supervision signals, model training is completed by minimizing the error between the predicted result and the actual evaluation result. The overall optimization objective of the model consists of a loss function and a regularization term, and the objective function is as follows: ; in, The objective function is... This represents the actual data value. These are the model's predicted values; This is the loss function for single-sample prediction error, used to measure the deviation between the model output and the actual evaluation results; This is a lateralization function used to constrain the structural complexity of the base learner.

[0021] The loss function guides the model to continuously approximate the actual ice-melting effect evaluation results, enabling the model to have high fitting accuracy. The regularization function, by penalizing the model complexity, suppresses the occurrence of excessive invalid splits or overly deep structures. The regularization function is expressed as: ; in, The regularization coefficient is used. The number of leaf nodes in the base learner; These are the weight parameters corresponding to the leaf nodes.

[0022] The historical sample data is divided into several non-overlapping subsets. In each training round, one subset is selected as the validation set, and the remaining subsets are used as the training set. This process is repeated multiple times to complete the training and validation rounds. The overall validation error is then: ; in, The overall validation error of cross-validation; The number of folds for cross-validation; For the first The single-round verification error of round verification.

[0023] like Figure 4 As shown, by introducing a cross-validation mechanism, this invention can effectively avoid the model performance bias problem caused by uneven sample distribution or individual extreme working condition samples, so that the constructed de-icing effect evaluation model has good applicability under different icing levels, different meteorological conditions and different combinations of de-icing methods.

[0024] As shown in Figure 2, Example 1: Real-time acquisition of multi-source heterogeneous monitoring data of UHV lines, standardization processing, and construction of a decision feature database; The multi-source heterogeneous monitoring data are as follows: Meteorological monitoring data includes information such as icing thickness, temperature, humidity, wind speed, maximum wind deflection angle, and maximum wind tilt angle, which are used to reflect external environmental conditions.

[0025] As shown in Figure 3, the simulation data comes from the de-icing jump dynamic model of the tower-conductor system built on the ANSYS platform. By simulating the transient shedding, conductor vibration and tension change of the iced conductor during the de-icing process, the jump amplitude, transient stress change and dynamic response characteristics of the conductor are obtained.

[0026] Real-time load data for the line includes changes in conductor tension, sag, and related load response information, which are used to reflect the stress on the conductor under actual operating conditions.

[0027] Topographic data includes spatial information such as terrain undulations, altitude, and landform distribution along the route provided by the geographic information system, which is used to describe the impact of terrain conditions on ice distribution and accessibility for de-icing operations.

[0028] The aforementioned multi-source data undergoes spatiotemporal alignment and fusion processing, and is stored in a unified decision feature library as a common data foundation for subsequent offline training and online inference.

[0029] An evaluation model for de-icing effectiveness was constructed, and the ensemble learning model was trained using historical case data, specifically as follows: Historical case data comes from de-icing operation records in different regions and under different climatic conditions, covering application scenarios, parameters, and effect feedback of various de-icing methods.

[0030] The effectiveness quantification scoring dimensions include safety, cost-effectiveness, effectiveness, and timeliness, with safety accounting for 30%, cost-effectiveness for 25%, effectiveness for 30%, and timeliness for 15%.

[0031] The comprehensive performance score of each de-icing method is calculated by weighted summation to form a standardized training target dataset for training the ensemble learning model.

[0032] A model versioning storage mechanism is implemented, which saves the evaluated and qualified models as specific versions, and supports the dynamic recall of the optimal model version according to actual needs during online inference.

[0033] The real-time decision features are input into the trained de-icing effect evaluation model to calculate the expected effect score of each candidate de-icing method, specifically: The system extracts the latest real-time data from the decision feature database at preset intervals and organizes it into a standard format feature vector that meets the model input requirements.

[0034] The optimal version of the model is retrieved from storage. The model combines historical experience patterns learned during training to predict the individual scores of safety, economy, effectiveness, and timeliness of each candidate de-icing method in the current scenario.

[0035] The overall expected performance score of each candidate method is calculated according to the preset weight ratio. At the same time, the calculation logic is dynamically adjusted according to the special circumstances of the current scenario to ensure that the score can truly reflect the actual application potential of each method.

[0036] Based on the expected effect score, the optimal de-icing solution is automatically selected and output, specifically as follows: The system sorts the candidate de-icing methods from highest to lowest based on their comprehensive expected performance score.

[0037] A priority adjustment mechanism is set up so that when the score difference between multiple methods is within a preset threshold range, a secondary selection is performed based on the constraints of the current scenario.

[0038] The final output includes the optimal de-icing solution, the second-best solution, and alternative solutions. The applicable conditions and precautions for each solution are clearly defined and presented in a visual form on the operation and maintenance decision-making platform. The content includes core information such as solution name, overall score, individual scores for each dimension, and implementation suggestions.

[0039] The decision-making basis of the optimal de-icing scheme is quantitatively explained through the interpretability analysis module, and the prediction confidence interval is given, specifically: Interpretability analysis was performed using the SHAP analysis tool, and the marginal contribution of each input feature to the expected score of the optimal ice-melting solution was calculated based on game theory. The calculation formula is as follows:

[0040] In the formula: Features in the feature set S The SHAP value; S represents no feature included. Any subset of features; N represents the set of all features; Represents the number of characteristics in set S; The contribution of the feature set S to the model's predicted output; Representative includes features Feature set Contribution to the predicted output.

[0041] Based on the calculated SHAP value, the system establishes an additive interpretation model for a single prediction, visually displaying the influence weights of each feature on the decision outcome. See Figure 5.

[0042] By combining the prediction error distribution of similar historical scenarios, the confidence interval of the current prediction result is calculated through statistical analysis. A narrow confidence interval indicates that the model prediction is highly reliable, while a wide confidence interval suggests that the current scenario may have insufficient data or high environmental complexity.

[0043] Generate a decision explanation report, presenting the SHAP analysis results and confidence intervals in the form of charts and text descriptions to enhance the credibility of the model's decisions.

[0044] Perform manual review and closed-loop feedback, and update the case library to achieve continuous model iteration, specifically as follows: Operations and maintenance personnel can view the visual display of the solution and the decision explanation report on the operations and maintenance decision platform, and conduct manual review based on practical experience to confirm the final de-icing solution to be implemented.

[0045] The de-icing plan was implemented on-site, and data such as the de-icing effect, operation time, and consumption costs were collected during the actual implementation process.

[0046] The collected execution effect data and corresponding decision feature data are packaged together and updated into the historical case library as new cases.

[0047] The system periodically uses the updated case library to trigger the offline model training branch, retraining and updating the model to achieve a closed-loop iteration for optimizing the ice melting strategy.

[0048] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

Claims

1. A data-driven intelligent optimization method for de-icing of ultra-high voltage power lines, characterized in that, include: Multi-source heterogeneous data fusion processing, specifically: Data from different sources differ significantly in sampling frequency, time scale, and spatial coordinate representation. After collection, spatiotemporal alignment and preprocessing of multi-source data are required. In the time dimension, by constructing a unified time axis, data from different sampling periods are resampled and interpolated to map to the same time scale. In the spatial dimension, the spatial coordinate projection method is used to associate and map meteorological network data with the location of power line towers and sections; To reduce the impact of measurement noise on subsequent feature extraction, a Kalman filter is introduced to denoise and estimate the state of the continuous time series. The state update process is represented as follows: ； in, For the first The state vector at each moment; This is the state transition matrix; This is process noise; For observation vectors; The observation matrix; To observe noise; Provided that the data quality meets the requirements, the multi-source features are normalized and dimensionality reduced to eliminate the influence of differences in dimensions and scales on the fusion results. The normalization process is as follows: ； in, These are the normalized eigenvalues; These are the original eigenvalues; This represents the mean of the original feature samples; is the standard deviation of the original feature samples.

2. The intelligent optimization method for a data-driven ultra-high voltage line de-icing method according to claim 1, characterized in that, Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional feature space. Eigenvalue decomposition is performed using the covariance matrix, and principal components with satisfactory cumulative contribution rates are selected. This approach preserves key information while reducing feature redundancy. The covariance matrix is ​​as follows: ； in, The characteristic covariance matrix; The number of samples; For the first Feature vectors of each sample; It is the mean vector of all sample feature vectors; The fusion process is performed using a weighted mapping method as follows: ； in, This is the final fused feature vector; These are weighting coefficients used to measure the weighting factor of the first... The relative importance of class features in the fusion process; For the first Class feature vectors; After feature fusion is completed, the discriminative ability of the fused features is evaluated to verify its ability to distinguish between different icing states and the effectiveness of de-icing schemes. This is measured by the ratio of inter-class divergence to intra-class divergence. The discriminant index is as follows: ； in, As an indicator for feature discrimination capability; The projection direction vector; The inter-class scatter matrix; It is the intra-class scatter matrix.

3. The intelligent optimization method for a data-driven ultra-high voltage line de-icing method according to claim 2, characterized in that, It also includes effect evaluation and model optimization, specifically: Using known de-icing performance evaluation results from historical samples as supervision signals, model training is completed by minimizing the error between the predicted result and the actual evaluation result. The overall optimization objective of the model consists of a loss function and a regularization term, and the objective function is as follows: ； in, The objective function is... This represents the actual data value. These are the model's predicted values; This is the loss function for single-sample prediction error, used to measure the deviation between the model output and the actual evaluation results; This is a lateralization function used to constrain the structural complexity of the base learner; The loss function guides the model to continuously approximate the actual ice-melting effect evaluation results, enabling the model to have high fitting accuracy. The regularization function, by penalizing the model complexity, suppresses the occurrence of excessive invalid splits or overly deep structures. The regularization function is expressed as: ； in, The regularization coefficient is used. The number of leaf nodes in the base learner; These are the weight parameters corresponding to the leaf nodes.

4. The intelligent optimization method for a data-driven ultra-high voltage line de-icing method according to claim 3, characterized in that: The historical sample data is divided into several non-overlapping subsets. In each training round, one subset is selected as the validation set, and the remaining subsets are used as the training set. This process is repeated multiple times to complete the training and validation rounds. The overall validation error is then: ； in, The overall validation error of cross-validation; The number of folds for cross-validation; For the first The single-round verification error of round verification.

Images