Power transmission line icing prediction correction method fusing physical mechanism and artificial intelligence

Abstract

The invention discloses a power transmission line icing prediction correction method fusing a physical mechanism and artificial intelligence, relates to the technical field of power system power transmission line safety protection, and effectively reduces the problems of weather forecast error transmission and extreme event missing report by fusing the physical mechanism and the artificial intelligence and adopting a hierarchical modeling strategy. The model structure is integrated with clear physical mechanisms such as microtopography correction, sudden performance drop of a pure data driven model in a sample sparse region or in a weather-free mode is avoided, the adopted hierarchical modeling framework and the adaptive loss function can effectively balance contributions of samples of different grades, and the method is suitable for large-scale popularization and application. And the fitting capability of the model on a re-icing sample and the robustness on an abnormal value are enhanced. The output result is the graded icing thickness, and the potential risk level of the line is visually reflected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system transmission line safety protection, and particularly relates to an icing prediction correction method for a transmission line combining physical mechanisms and artificial intelligence. BACKGROUND

[0002] Transmission line icing is a major natural disaster that threatens the safe operation of power grids, especially in regions with complex geographical and climatic characteristics such as Xinjiang. Xinjiang is vast and has complex terrain, including the Tianshan Mountains, the Junggar Basin, the Tarim Basin, and other special landforms, as well as typical continental arid climate characteristics. The winter is cold and dry, and cold waves occur frequently. This special geographical and climatic condition makes the formation mechanism of transmission line icing more complex, posing a great challenge to icing prediction.

[0003] At present, domestic and foreign scholars have carried out a lot of research on transmission line icing prediction. Traditional prediction methods are mainly divided into two categories: numerical models based on physical mechanisms and statistical models based on data-driven methods. In terms of physical models, Makkonen model and Jones model are widely used, which simulate the icing growth process by solving thermodynamic equilibrium equations. However, these models have very high requirements for the accuracy of input meteorological data, and there are significant errors in the prediction of meteorological elements in Xinjiang under complex terrain, resulting in poor prediction results of physical models. In terms of statistical models, machine learning methods such as support vector machine and neural network are used to establish the mapping relationship between meteorological elements and icing thickness, but these models often lack physical mechanism support and have poor generalization ability under extreme weather conditions.

[0004] The existing research has the following outstanding problems: First, meteorological prediction errors are continuously accumulated and amplified in the icing prediction process. Numerical weather prediction models have systematic deviations in the prediction of key meteorological elements such as temperature and wind speed in Xinjiang under complex terrain, such as WRF model, which has a wind speed prediction error of more than 30% in mountainous areas. Second, the influence of micro-topographic features is seriously ignored. The altitude in Xinjiang changes dramatically (such as the altitude difference in the Tianshan Mountains exceeding 3000 meters), and special terrain such as passes and valleys can significantly change local meteorological conditions, but traditional models use uniform grid data and cannot reflect these micro-topographic effects. Third, the existing models have insufficient prediction ability for heavy icing events. The icing thickness in Xinjiang during the cold wave period can reach more than 50mm, and the existing model parameter threshold is difficult to identify heavy icing events in Xinjiang, with a high omission rate of up to 40%. Finally, the model has poor generalization ability and cannot adapt to the unique alternating cold-warm climate characteristics in Xinjiang, with a significant decrease in 48-hour prediction accuracy.

[0005] In recent years, some scholars have begun to explore fusion methods to improve icing prediction. Zhang Zhengda proposed an icing grade prediction method based on meteorological forecast revision and multi-feature fusion, which selects key features through gray correlation analysis, but does not fully consider the influence of micro-topography. Zhou Shaohui studied meteorological revision and model optimization under multi-source fusion and proposed the MCIF model, but its research area is mainly aimed at the humid climate in the south. Zhou Xueming et al. discussed the icing thickness revision method in the micro-topography area and proposed an altitude and windward-slip revision formula, which provides an important reference for micro-topography revision in Xinjiang region. However, the existing methods still cannot effectively solve the problem of icing prediction under special geographical and climatic conditions in Xinjiang, and it is urgent to develop a targeted prediction revision method.

[0006] In summary, the existing icing prediction method for transmission lines in Xinjiang region faces severe challenges, and it is urgent to develop a new prediction method that can consider key issues such as meteorological forecast error revision, micro-topography effect modeling, and heavy icing event identification. SUMMARY

[0007] In order to overcome the shortcomings of the above technology, a method is provided for predicting icing on transmission lines in Xinjiang region under complex terrain and extreme climate conditions, which combines physical mechanism modeling with artificial intelligence spatiotemporal feature learning to achieve accurate prediction and revision of icing thickness.

[0008] The technical solution adopted by the present application to overcome the technical problems is: A fusion of physical mechanism and artificial intelligence for transmission line icing prediction revision method, comprising: S1. Obtain historical observation data, meteorological forecast data and micro-topography data, the historical observation data including icing thickness and wind direction observed by transmission tower in Xinjiang region, icing thickness is a three-dimensional tensor, meteorological forecast data includes: WRF mode output meteorological element set , predicted icing thickness , WRF mode output reference altitude , the micro-topography data includes altitude , slope , slope , meteorological elements , predicted icing thickness are all three-dimensional tensors; S2. Data alignment and interpolation are performed on the meteorological element set to obtain the processed meteorological element set ; S3. The processed meteorological element set standardized feature set ; S4. Utilize the reference elevation , elevation to correct the icing thickness , and obtain the actual elevation of the icing thickness ; S5. Correct the icing thickness on the windward and leeward slopes, and obtain the icing thickness corrected by the wind speed ; S6. Use the grey correlation analysis method to screen elements with a correlation degree greater than a set threshold from the standardized feature set , and obtain a feature set ; S7. Construct a feature set according to the feature set and the icing thickness corrected by the wind speed ; S8. Input the feature set into a prediction model composed of a space-time feature extraction module, a feature fusion and dimension reduction module, and a grading output module, and output a prediction value ; S9. Calculate a loss function , train the prediction model through the loss function , and obtain an optimized prediction model.

[0009] Further, the dimensions of the icing thickness in step S1 are longitude lon, latitude lat, and time t; and the dimensions of the meteorological elements and the predicted icing thickness are longitude lon, latitude lat, and time t, respectively.

[0010] Further, step S2 includes the following steps: S2-1. Use bilinear interpolation to interpolate the meteorological element set to the tower coordinate position, and obtain a meteorological element set at the tower coordinate position ; S2-2. Repair missing data in the meteorological element set using linear interpolation, and obtain a processed meteorological element set .

[0011] Further, in step S3, the processed meteorological element set , the slope , and the aspect are subjected to Min-Max standardization, and a standardized meteorological element set standardized slope standardized aspect building a standardized feature set , .

[0012] Further, the actual altitude is calculated in step S4 by the formula . .

[0013] Further, the wind speed corrected ice thickness is calculated in step S5 by the formula , wherein is a parameter, and the angle difference is calculated by the formula When the angle difference is greater than or equal to 0° and less than or equal to 60° or the angle difference is greater than or equal to 300° and less than or equal to 360°, it is determined that the tower position is located on the windward slope, with a value of 1.1-1.3; when the angle difference is greater than or equal to 120° and less than or equal to 240°, it is determined that the tower position is located on the leeward slope, with a value of 0.7-0.9; when the angle difference is greater than 60° and less than 120° or the angle difference is greater than 240° and less than 300°, it is determined that the tower position is located in the no significant influence zone, with a value of 1.

[0014] Preferably, the threshold value is set to 0.7-0.8 in step S6.

[0015] Further, step S7 includes the following steps: S7-1. If the ice thickness in the feature set is greater than or equal to 0 mm and less than 5 mm, it is determined that there is no risk, the risk level is 1, and the feature set is placed in the feature set . S7-2. If the ice thickness in the feature set is greater than or equal to 5 mm and less than 15 mm, it is determined that there is a warning level, the risk level is 2, and the feature set is placed in the feature set In step S7-1, the ice thickness is calculated according to the formula: S7-3. If the ice thickness is greater than or equal to 15 mm and less than 30 mm, the value of the same longitude lon, the same latitude lat, and the same time t is placed in the feature set In step S7-2, the ice thickness is calculated according to the formula: When the value of the same longitude lon, the same latitude lat, and the same time t is greater than or equal to 15 mm and less than 30 mm, this is a pre-warning level, the risk level is 3, and the feature set In step S7-3, the value of each of the same longitude lon, the same latitude lat, and the same time t is placed in the feature set In step S7-4, the ice thickness is calculated according to the formula: S7-4. If the ice thickness is greater than or equal to 30 mm, the value of the same longitude lon, the same latitude lat, and the same time t is placed in the feature set In step S7-2, the ice thickness is calculated according to the formula: When the value of the same longitude lon, the same latitude lat, and the same time t is greater than or equal to 30 mm, this is a warning level, the risk level is 4, and the feature set In step S7-4, the value of each of the same longitude lon, the same latitude lat, and the same time t is placed in the feature set In step S7-5, the feature set S7-5. Obtain the feature set , .

[0016] Further, step S8 includes the following steps: S8-1. The spatio-temporal feature extraction module of the prediction model is composed of a ConvLSTM model, and the multi-dimensional meteorological and micro-topographic feature time series data are input features of the ConvLSTM model, and are labels, the high-latitude spatio-temporal feature tensor is extracted by the ConvLSTM model, and a three-dimensional tensor is output, the dimensions of the three-dimensional tensor are: time t, feature map height h, feature map width w x channel number p; S8-2. The feature fusion and dimension reduction module of the prediction model is composed of a global average pooling layer and a fully connected layer in sequence, and the three-dimensional tensor is input into the feature fusion and dimension reduction module of the prediction model, and a shared feature vector is output. S8-3. The graded output module of the prediction model is composed of a fully connected layer, and the shared feature vector is input into the graded output module of the prediction model, and a prediction value of the level is output.

[0017] Further, step S9 includes the following steps: S9-1. The loss is calculated by the formula , wherein is the icing thickness is the number of values, is the icing thickness is the number of values, is the number of values, , is the predicted value is the number of values, is the number of values is the number of values, is the smoothing threshold, is 1-5mm, is the adaptive weight, , is the hyperparameter, is 0.9-0.999, is the icing thickness is the number of values, is the number of values, is the number of values, S9-2. is calculated by the formula , wherein is the uncertainty parameter, is the icing thickness is 0.1-2.0; S9-3. The prediction model is trained using the AdamW optimizer and the loss function , wherein the initial learning rate is set to , the weight decay coefficient is set to , the batch size is set to 32-64, and the training round is set to 200.

[0018] The beneficial effects of the present application are: 1. The prediction accuracy is significantly improved: by fusing physical mechanisms and artificial intelligence, and adopting a grading modeling strategy, the present application effectively reduces the meteorological forecast error transmission and the extreme event missing report problem. Tests in typical heavy icing scenes in Xinjiang region show that for the prediction of icing thickness (30mm), the average absolute error (MAE) is reduced by about 20% compared with traditional physical models (such as Makkonen model) and single AI models (such as standard LSTM), and the critical success index (CSI) of heavy icing event is improved by about 15%, and the result is statistically significant p<0.05).

[0019] ​2. Physical interpretability and enhanced generalization capability: This method is not a "black box" model. Its feature selection is based on the physical correlation between meteorological elements and icing, and the model structure incorporates explicit physical mechanisms such as micro-topographic correction. This physical constraint makes the model show stronger generalization capability and stability when facing the complex and variable geographical and climatic conditions in Xinjiang region (such as the Tianshan Mountain Pass, the Bosten Lake effect area, etc.), avoiding the performance drop of pure data-driven models in sparse sample areas or unobserved weather patterns.

[0020] 3. Optimized response capability to extreme events: In view of the rapid growth of icing during the cold wave period in Xinjiang, the adopted hierarchical modeling framework and adaptive loss function can effectively balance the contribution of samples at different levels and enhance the model's fitting capability to heavy icing samples and robustness to outliers. This makes the model more sensitive to the sudden change signals of icing thickness caused by freezing rain, strong temperature drop, etc.

[0021] 4. High engineering practical value: The final output is the icing thickness in different levels, which directly reflects the potential risk level of the line (e.g., 0-5mm is no risk, 5-15mm is attention level, 15-30mm is warning level, 30-50mm is alarm level). This output form is convenient for power grid operation and maintenance personnel to quickly understand the risk level and start differentiated anti-icing and de-icing emergency plans (such as adjusting de-icing current or load transfer strategy for different levels), which improves the decision-making efficiency and safety. DETAILED DESCRIPTION

[0022] The present application will be further described below.

[0023] Xinjiang has complex terrain (such as the Tianshan Mountains and the Junggar Basin), and numerical weather prediction (WRF, etc.) has significant temperature and wind speed prediction bias due to initial field error and physical parameterization scheme not suitable for local dry climate. For example, wind speed fluctuates greatly in winter in Xinjiang, and WRF model may overestimate wind speed in the basin and underestimate wind speed in the mountains, with icing thickness prediction bias exceeding 30% after the error is transmitted through the icing model (such as Makkonen model, Jones model). Traditional linear correction methods (such as MOS) cannot capture the nonlinear characteristics of strong weather processes in Xinjiang, especially during the cold wave period when error accumulation is more obvious.

[0024] Ignoring micro-topographic features: Xinjiang has various micro-topographic types (such as Tianshan Mountain Pass, valley wind channel, and basin edge), and traditional models do not integrate factors such as altitude jump and slope convergence. For example, temperature decreases by 0.6°C for every 100m increase in altitude, but traditional models do not correct the altitude gradient, which can lead to underestimation of icing in mountainous areas; humidity can increase by 20% within 2.5km around lakes (such as Bosten Lake) in Xinjiang, and traditional models do not consider the enhancement of fog by water bodies.

[0025] Serious icing missing: Xinjiang winter cold events occur frequently, and the traditional physical model parameter threshold (such as the Jones model freeze rain discrimination condition) is not suitable for the local dry and cold climate, resulting in missing reports of heavy icing events.

[0026] Poor model generalization: Traditional statistical models (such as ARIMA) respond slowly to the sudden cold-warm alternating process of icing in Xinjiang, with a 48-hour forecast CSI falling below 0.2; training data is mostly from the humid climate in the middle and eastern regions, and the sparse icing sample in Xinjiang dry and cold climate leads to poor performance of machine learning models (such as CNN-LSTM) in identifying extreme values, with error increasing sharply when icing is greater than 30mm.

[0027] Therefore, the present application proposes a power line icing prediction correction method combining physical mechanism and artificial intelligence, comprising: S1. Obtain historical observation data, meteorological forecast data and microtopography data, the historical observation data including icing thickness and wind direction observed by power transmission towers in Xinjiang region, the icing thickness is a three-dimensional tensor, the meteorological forecast data includes: a set of meteorological elements output by WRF model, predicted icing thickness , and reference elevation output by WRF model, the microtopography data includes elevation , slope , and aspect extracted based on high-precision DEM, the meteorological elements , and predicted icing thickness are all three-dimensional tensors.

[0028] S2. Align and interpolate the set of meteorological elements to obtain the processed set of meteorological elements .

[0029] S3. Standardize the processed set of meteorological elements to obtain the standardized feature set .

[0030] S4. Use the reference elevation and the elevation to correct the icing thickness to obtain the icing thickness at the actual elevation .

[0031] S5. Correct the icing thickness for windward and leeward slopes to obtain the icing thickness corrected by wind speed.

[0032] S6. Screening elements with correlation degree greater than a set threshold from the standardized feature set . .

[0033] S7. Constructing a feature set according to the feature set and the icing thickness corrected by wind speed . .

[0034] S8. Inputting the feature set into a prediction model composed of a space-time feature extraction module, a feature fusion and dimension reduction module and a grading output module to output a prediction value.

[0035] S9. Calculating a loss function , training the prediction model through the loss function and obtaining an optimized prediction model.

[0036] By combining physical mechanism modeling and artificial intelligence space-time feature learning, the grading accurate prediction and correction of icing thickness are realized.

[0037] In an embodiment of the present application, the dimensions of the icing thickness in step S1 are respectively longitude lon, latitude lat and time t, and the dimensions of the meteorological elements and the predicted icing thickness are respectively longitude lon, latitude lat and time t.

[0038] In an embodiment of the present application, step S2 includes the following steps: S2-1. Using bilinear interpolation to interpolate the meteorological element set to the tower coordinate position to obtain a meteorological element set of the tower position, so as to realize the space-time synchronization of meteorological data and icing observation.

[0039] S2-2. Repairing missing data in the meteorological element set by linear interpolation to obtain a processed meteorological element set .

[0040] In an embodiment of the present application, the processed meteorological element set , slope and slope direction in step S3 are respectively subjected to Min-Max standardization processing to ensure that different features are input into the model in a dimensionless form, and a standardized meteorological element set , standardized slope and standardized slope direction are obtained to construct a standardized feature set , .

[0041] In one embodiment of the present application, the actual altitude is calculated in step S4 by the formula and the ice thickness .

[0042] In one embodiment of the present application, the wind speed corrected ice thickness is calculated in step S5 by the formula , wherein is a parameter, and the angle difference is calculated by the formula When the angle difference is greater than or equal to 0° and less than or equal to 60° or the angle difference is greater than or equal to 300° and less than or equal to 360°, it is determined that the tower position is on the windward slope, with a value of 1.1-1.3; when the angle difference is greater than or equal to 120° and less than or equal to 240°, it is determined that the tower position is on the leeward slope, with a value of 0.7-0.9; when the angle difference is greater than 60° and less than 120° or the angle difference is greater than 240° and less than 300°, it is determined that the tower position is in the no significant influence zone, with a value of 1.

[0043] In one embodiment of the present application, the threshold value is set to 0.7-0.8 in step S6.

[0044] In one embodiment of the present application, step S7 includes the following steps: S7-1. If the ice thickness in the feature set is greater than or equal to 0 mm and less than 5 mm, it is determined that there is no risk, the risk level is 1, and the values of longitude lon, latitude lat, and time t in the feature set are placed in the feature set .

[0045] S7-2. If the ice thickness in the feature set is greater than or equal to 5 mm and less than 15 mm, it is determined that there is a warning level, the risk level is 2, and the values of longitude lon, latitude lat, and time t in the feature set are placed in the feature set .

[0046] S7-3. If the ice thickness is greater than or equal to 15 mm and less than 30 mm, a warning level is reached, and the risk level is 3. The feature set is placed in the feature set .

[0047] S7-4. If the ice thickness is greater than or equal to 30 mm, an alarm level is reached, and the risk level is 4. The feature set is placed in the feature set .

[0048] S7-5. The feature set is obtained. .

[0049] In an embodiment of the present application, step S8 comprises the following steps: S8-1. The spatio-temporal feature extraction module of the prediction model is composed of a ConvLSTM model. The multi-dimensional meteorological and micro-topographic feature time series data are taken as the input features of the ConvLSTM model, and the are taken as the labels to extract the high-dimensional spatio-temporal feature tensor through the ConvLSTM model, and the output three-dimensional tensor is obtained. The high-dimensional spatio-temporal feature tensor is extracted through the ConvLSTM model (including convolution structure and gating mechanism). The dimensions of the three-dimensional tensor are: time t, feature map height h, feature map width w x channel number p. The channel number corresponds to the feature dimension, which captures the spatio-temporal local correlation and time series dynamics. The embedded attention mechanism (CBA) further weights the key time steps, enhancing the model's ability to perceive key signals.

[0050] S8-2. The feature fusion and dimension reduction module of the prediction model is composed of a global average pooling layer and a fully connected layer in sequence. The three-dimensional tensor is input into the feature fusion and dimension reduction module of the prediction model. The spatio-temporal feature tensor is aggregated along the spatial and temporal dimensions through the global average pooling operation, and is converted into a one-dimensional feature vector. This vector retains the overall characteristics of the sequence. Subsequently, one or more fully connected layers are used for dimension reduction, compressing the feature vector to a fixed dimension, and obtaining a low-dimensional, dense shared feature vector .​

[0051] S8-3. The grading output module of the prediction model is composed of a full connection layer, and the shared feature vector is input into the grading output module of the prediction model, and a predicted value of each grade is output. The predicted value is a three-dimensional tensor, and the dimensions of the predicted value are longitude lon, latitude lat and time t, respectively. The prediction model outputs the respective predicted values of the four grades, which intuitively reflects the potential risk level of the line, and this output form is convenient for power grid operation and maintenance personnel to quickly understand the risk degree and start differentiated anti-icing and ice-melting emergency plans accordingly. In an embodiment of the present application, step S9 comprises the following steps:

[0052] S9-1. To solve the problems of uneven sample quantity of each grade and observation noise, the model introduces an adaptive loss function mechanism to improve the robustness of the model to abnormal values and balance the contribution of samples of different grades. Specifically, the loss is calculated by the formula , wherein n is the number of median values of the ice thickness, is the th value of the ice thickness, is the value in the predicted value corresponding to the th value of the ice thickness, lon is the longitude, lat is the latitude, t is the time, is a smoothing threshold, the value is 1-5 mm, is an adaptive weight, is a hyperparameter, the value is 0.9-0.999, is the number of values in the ice thickness of the grade corresponding to the feature set and the values of the same longitude, the same latitude and the same time.

[0053] S9-2. is calculated by the formula , wherein is the uncertainty parameter of the grade, and the value is 0.1-2.0.

[0054] ​​​​​​​​​​​​​​​​​​​​​​​​​​​​S9-3. Using the AdamW optimizer with the loss function The prediction model is trained to obtain the optimized prediction model. The initial learning rate during training is set to... The weight decay coefficient is set to The batch size is set to 32-64, and the number of training epochs is set to 200. In practice, this can be controlled using an early stopping mechanism. A cosine annealing algorithm is used to dynamically adjust the learning rate. The learning rate is halved if the validation set loss shows no improvement for 5 consecutive epochs. If the validation set loss shows no improvement for 15 consecutive epochs, training is terminated, and the optimal model parameters are saved. A gradient norm threshold of 1.0 can be set to prevent gradient explosion during training.

[0055] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and correcting icing on power transmission lines that integrates physical mechanisms and artificial intelligence, characterized in that, include: S1. Acquire historical observation data, meteorological forecast data, and micro-topographic data. The historical observation data includes the ice thickness observed on transmission towers in Xinjiang. and wind direction Ice thickness The weather forecast data is a three-dimensional tensor, including the set of meteorological elements output by the WRF model. Forecast ice thickness The baseline altitude output in WRF mode The micro-topographic data includes elevation data extracted based on a high-precision DEM. ,slope Slope aspect meteorological elements Forecast ice thickness All are three-dimensional tensors; S2. Set of meteorological elements Data alignment and interpolation are performed to obtain the processed meteorological element set. ; S3. Set the processed meteorological elements Standardization is performed to obtain a standardized feature set. ; S4. Utilizing the baseline elevation ,altitude Regarding ice thickness Altitude correction was performed to obtain the actual altitude. ice thickness ; S5. Regarding ice thickness Perform windward / leeward slope correction to obtain the icing thickness corrected for wind speed. ; S6. Using grey relational analysis to analyze standardized feature sets From each element in the dataset, elements with a correlation greater than a set threshold are selected to obtain the feature set. ; S7. Based on the feature set and icing thickness corrected for wind speed Constructing feature sets ; S8. Feature set The input is fed into a prediction model consisting of a spatiotemporal feature extraction module, a feature fusion and dimensionality reduction module, and a hierarchical output module, and the output is the predicted value. S9. Calculate the loss function Through the loss function Train the prediction model to obtain the optimized prediction model.

2. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence as described in claim 1, characterized in that: Ice thickness in step S1 The dimensions are: longitude (lon), latitude (lat), and time (t); meteorological elements. Forecast ice thickness The dimensions are: longitude lon, latitude lat, and time t.

3. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence as described in claim 1, characterized in that, Step S2 includes the following steps: S2-1. Using bilinear interpolation to integrate meteorological element sets Interpolate to the tower coordinates to obtain the meteorological element set for the tower location. ; S2-2. Repairing meteorological element sets using linear interpolation. The missing data in the dataset was used to obtain a processed set of meteorological elements. .

4. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence as described in claim 1, characterized in that: In step S3, the processed meteorological element set is processed respectively. ,slope Slope aspect Min-Max standardization was used to obtain a standardized set of meteorological elements. Standardized slope Standardized slope aspect Construct a standardized feature set , .

5. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence according to claim 1, characterized in that: In step S4, the formula is used. The actual altitude was calculated. ice thickness .

6. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence according to claim 1, characterized in that: In step S5, the formula is used. The calculated icing thickness after wind speed correction In the formula As parameters, through the formula The angle difference was calculated. When the included angle difference Greater than or equal to 0° and less than or equal to 60° or the difference in angle When the angle is greater than or equal to 300° and less than or equal to 360°, the tower is considered to be located on the windward slope. The value ranges from 1.1 to 1.3; when the angle difference... When the angle is greater than or equal to 120° and less than or equal to 240°, the tower is determined to be located on the leeward slope. The value ranges from 0.7 to 0.9; when the angle difference... greater than 60° and less than 120° or the difference in angle When the angle is greater than 240° and less than 300°, the tower location is determined to be in a region with no significant influence. The value is 1.

7. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence according to claim 1, characterized in that: In step S6, the threshold is set to 0.7-0.

8.

8. The method for predicting and correcting icing on transmission lines by integrating physical mechanisms and artificial intelligence as described in claim 2, characterized in that, Step S7 includes the following steps: S7-1. If the icing thickness China and the establishment of wind speed-corrected icing thickness When the values ​​of longitude (lon), latitude (lat), and time (t) are greater than or equal to 0 mm and less than 5 mm, there is no risk, and the risk level is 1. The feature set... The values ​​of longitude lon, latitude lat, and time t for each element are placed in the feature set. middle; S7-2. If the icing thickness China and the establishment of wind speed-corrected icing thickness When the values ​​of longitude (lon), latitude (lat), and time (t) are greater than or equal to 5 mm and less than 15 mm, the level is considered "attention level," with a risk level of 2. The feature set... The values ​​of longitude lon, latitude lat, and time t for each element are placed in the feature set. middle; S7-3. If the icing thickness China and the establishment of wind speed-corrected icing thickness When the values ​​of longitude (lon), latitude (lat), and time (t) are greater than or equal to 15 mm and less than 30 mm, it is considered a warning level with a risk level of 3. The feature set... The values ​​of each longitude (lon), latitude (lat), and time (t) are placed in the feature set. middle; S7-4. If the icing thickness China and the establishment of wind speed-corrected icing thickness When the values ​​of longitude (lon), latitude (lat), and time (t) are greater than or equal to 30 mm, it is an alarm level with a risk level of 4. The feature set... The values ​​of each longitude (lon), latitude (lat), and time (t) are placed in the feature set. middle; S7-5. Obtain the feature set , .

9. The method for predicting and correcting icing on transmission lines by integrating physical mechanisms and artificial intelligence as described in claim 8, characterized in that, Step S8 includes the following steps: S8-1. The spatiotemporal feature extraction module of the prediction model consists of a ConvLSTM model, which extracts multi-dimensional meteorological and micro-topographic feature time series data. As input features to the ConvLSTM model, The labels are used to extract high-dimensional spatiotemporal feature tensors through the ConvLSTM model, and the output is a three-dimensional tensor. 3D tensor The dimensions are: time t, feature map height h, and feature map width w × number of channels p; S8-2. The feature fusion and dimensionality reduction module of the prediction model consists of a global average pooling layer and a fully connected layer, which integrates the three-dimensional tensor. The input is fed into the feature fusion and dimensionality reduction module of the prediction model, and the output is a shared feature vector. ; S8-3. The hierarchical output module of the prediction model consists of fully connected layers that share feature vectors. The input is fed into the ranking output module of the prediction model, and the output is the ranking. Predicted value .

10. The transmission line icing prediction and correction method integrating physical mechanisms and artificial intelligence according to claim 9, characterized in that, Step S9 includes the following steps: S9-1. Through formula Calculate the loss In the formula Ice thickness The number of medians Ice thickness The Middle One value, , For predicted values Middle and the first Value Values ​​of the same longitude (lon), same latitude (lat), and same time (t). For the smoothing threshold, The value is 1-5mm. For adaptive weights, , For hyperparameters, The value ranges from 0.9 to 0.

999. For level Ice thickness In and feature set The number of values ​​with the same longitude (lon), same latitude (lat), and same time (t); S9-2. Through formula Calculated In the formula For level Uncertainty parameters, The value ranges from 0.1 to 2.0; S9-3. Using the AdamW optimizer with the loss function The prediction model is trained to obtain the optimized prediction model. The initial learning rate during training is set to... The weight decay coefficient is set to The batch size was set to 32-64, and the number of training rounds was set to 200.

Images