Road icing multi-model feature importance screening and forecasting method based on optical fibers

Abstract

The invention discloses a road icing multi-model feature importance screening and forecasting method based on optical fibers, and belongs to the technical field of road icing early warning in plateau regions. The method specifically comprises the following steps: S1, constructing a road icing prediction data set of a high-altitude area, obtaining a data set including ground temperature actual measurement data, meteorological data and numerical weather forecast data, and transmitting the obtained data set to a subsequent processing link through an optical fiber; s2, data preprocessing: carrying out standardization processing on numerical features in the historical meteorological observation data to ensure that the data fall in a unified interval, and keeping the integrity of the data by adopting a missing value filling strategy; by automatically screening key features in meteorological data, interference of redundant information on the model is effectively reduced, and the generalization ability and prediction accuracy of the model are improved; a multi-model integration strategy is combined, the advantages of different models are fully utilized, and the stability and reliability of a forecasting result are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of road icing early warning in plateau areas, and in particular to an optical fiber-based road icing multi-model feature importance screening and forecasting method. Background Art

[0002] High-altitude regions experience cold climates and rapid weather changes, leading to frequent icing of roads, posing significant risks to traffic safety. With the continuous advancement of meteorological monitoring technology, existing icing prediction methods mostly rely on real-time monitoring of meteorological data or road sensors. These methods are poorly adaptable to complex terrain and sudden meteorological changes, lacking accuracy and real-time performance, making them unable to meet the demand for accurate icing predictions in high-altitude regions. Forecasts from single models suffer from certain biases and deficiencies, making them inadequate for high-precision warnings of severe weather. Traditional models often rely on a single algorithm or fixed parameters and have limitations in feature selection and model fusion, which compromises overall forecast effectiveness. In recent years, deep learning and ensemble algorithms have shown great potential in meteorological data processing, but they still require feature importance screening and multi-model integration strategies to improve forecast performance. Therefore, there is an urgent need for technologies that can combine multiple data sources, particularly those that utilize multi-model feature screening and integration techniques to provide fundamental data support for icing prediction.

[0003] To achieve this goal, the present invention proposes a meteorological forecasting method based on the integration of meteorological observation data and statistical methods to provide feature importance screening and multi-model integration. By automatically extracting and screening key features, key icing factors are summarized, which can provide important reference data for icing risks, effectively improve the accuracy and robustness of the model, and provide more reliable technical support for meteorological monitoring and early warning. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide an optical fiber-based road icing multi-model feature importance screening and forecasting method; it can solve the problem that traditional models mostly rely on a single algorithm or fixed parameters, and have limitations in feature selection and model fusion, which affects the overall forecast effect.

[0005] Technical solution: To solve the above technical problems, according to one aspect of the present invention, more specifically, a method for screening and predicting road icing using multiple models based on optical fiber features importance, specifically comprising the following steps:

[0006] S1. Construct a road icing prediction dataset for high-altitude areas, obtain a dataset including ground temperature measurement data, meteorological data, and numerical weather forecast data, and transmit the obtained dataset to subsequent processing links via optical fiber;

[0007] S2. Data preprocessing: standardize the numerical features in historical meteorological observation data to ensure that the data falls within a uniform interval and use missing value filling strategies to maintain data integrity;

[0008] S3. Screen key features. By setting thresholds, screen out features with higher importance and eliminate minor features.

[0009] S4. Multi-model integration: train multiple models on the filtered data to obtain the accuracy data of each model, select the most influential features based on the model's importance index, and optimize the model structure;

[0010] S5, fusion strategy, uses weighted average or voting algorithm to fuse the prediction results of multiple models to obtain the final prediction report.

[0011] Furthermore, in step S1, the measured ground temperature data is the actual road surface temperature collected by a road surface sensor or a mobile device; the meteorological data includes meteorological elements such as temperature, humidity, wind speed, and precipitation; and the numerical weather forecast data is meteorological forecast data related to road icing.

[0012] Furthermore, the step S1 specifically includes the following steps:

[0013] S11. Download historical meteorological data of the traffic meteorological station at the same time of the target road section by accessing the local traffic meteorological service system;

[0014] S12. Select typical road sections at different altitudes and terrains in the target area. Study the meteorological data of the target road sections during icing periods over the years. Collect data every hour over a period of many years. Establish a low-temperature road surface data set for the target area. Select typical ground temperatures in different regions and analyze their spatiotemporal distribution characteristics to ensure the temporal continuity and representativeness of the samples. This will form a ground temperature risk zoning for the target area.

[0015] S13. Store the data in a CSV file. The record fields include the display number, site number, collection time, and the values ​​of various meteorological indicators. Study the impact of meteorological factors such as ground temperature, air pressure, atmospheric temperature, atmospheric humidity, rainfall, and slipperiness on road icing, and establish a preliminary road icing warning model.

[0016] Furthermore, the step S2 specifically includes the following steps:

[0017] S21. Clean the road icing prediction dataset for the target area. Perform a comprehensive null value count on each piece of raw data. For each feature column, calculate the missing value ratio in all samples. Columns that are null in more than 50% of the samples are considered to have insufficient information contribution and are removed.

[0018] S22. Use quality assurance to fill in the remaining features: forward filling or flattening is used for continuous numerical indicators, and "Unknown" or other categories are used to represent categorical variables or strings, ensuring that the meteorological values ​​at each time point are complete and continuous;

[0019] S23, adopt The formula standardizes continuous numerical features, subtracts the mean, divides by the standard deviation, and transforms each continuous feature;

[0020] S24. During the road icing model tuning phase, rainfall characteristics are gradually integrated, and the correlation modeling between this indicator and icing time is strengthened through a dynamic weight adjustment mechanism.

[0021] Furthermore, in step S3, when screening key features, a time window is set to extract meteorological information within a continuous time period. The window time is selected as the analysis unit according to actual needs. On this basis, multiple types of features are constructed using the original time series, including:

[0022] Lag characteristics: use the mean, maximum, and minimum values ​​of meteorological indicators over the past 1 hour, 2 hours, and until the end of the window;

[0023] Rate of change characteristics: Calculate the time series difference of continuous indicators;

[0024] Sliding window statistics: Within the set window time range, the sliding average, standard deviation, maximum and minimum values ​​of the indicator are calculated to capture important information about recent weather fluctuations.

[0025] Furthermore, in step S3, the feature aggregation process is a statistical calculation including mean, standard deviation, and extreme value. According to the definition of historical freezing events, the time period with continuous low temperature or meeting freezing conditions is marked as a positive sample. For the continuous meteorological data in these time periods, the feature vector is marked as "freezing", and the non-freezing time period is randomly selected as a negative sample.

[0026] Furthermore, in step S4, when performing multi-model training, all samples are randomly divided into a training set, a validation set, and a test set, and multi-model training is performed using gradient boosting tree, random forest, and Bayesian classifiers. The training process includes: fitting model parameters using training data, adjusting hyperparameters such as tree depth and learning rate in combination with cross-validation, and after the model training is completed, using the validation set to evaluate its classification effect on unseen data, including classification accuracy, F1 score, and confusion matrix, and then screening the most influential features based on the model's importance index.

[0027] According to another aspect of the present invention, a fiber-optic-based road icing multi-model feature importance screening and forecasting system is provided. The system is used to implement the above-mentioned fiber-optic-based road icing multi-model feature importance screening and forecasting method, including: a data preprocessing module, a feature importance analysis module, a multi-model training module and a fusion output module. Each module implements the content of the above steps and works together to form a complete forecasting process.

[0028] Beneficial effects: By automatically screening key features in meteorological data, the interference of redundant information on the model is effectively reduced, and the generalization ability and prediction accuracy of the model are improved; combined with multi-model integration strategies, the advantages of different models are fully utilized to significantly enhance the stability and reliability of forecast results. It is not only simple, efficient, and highly adaptable, but can also be widely used in meteorological fields such as short-term weather forecasting and disaster monitoring to improve meteorological warning levels and emergency response capabilities. Standardization processing transforms each continuous feature and controls the range within a reasonable range, which helps the gradient descent method to converge quickly, reduce the impact of dimensional differences of different indicators, ensure feature equivalence during model training, and avoid dimensional differences affecting model effects.

[0029] Categorical features are filled with categories or missing values ​​are marked. This process consolidates meteorological monitoring data from multiple sites and time periods into a complete, continuous database. Detailed preprocessing of the dataset addresses issues such as missing values ​​and outliers, and data standardization improves data quality and optimizes model performance.

[0030] During the model tuning stage, rainfall characteristics are gradually integrated, and the correlation modeling between this indicator and icing events is strengthened through a dynamic weight adjustment mechanism. This staged feature fusion strategy effectively alleviates the problems of data sparsity and distribution skewness, allowing the actual impact of rainfall to be accurately quantified.

[0031] The feature aggregation process forms a set of positive and negative samples, which helps the model distinguish between icing and non-icing weather. By independently sampling data from multiple monitoring stations and using Bootstrap resampling technology to balance the category distribution, the sample size is expanded from the initial 11 groups to 73 groups, achieving a significant improvement in data representation capabilities.

[0032] The data set is divided to ensure that the sample distribution between different sets is consistent. After the model training is completed, the validation set is used to evaluate its classification effect on unseen data. Then, the most influential features are selected based on the model's importance index to optimize the model structure.

[0033] The entire technical solution effectively improves the prediction accuracy of icing events through large-scale, multi-scale, and multi-angle meteorological feature extraction, combined with advanced machine learning technology, providing a scientific basis for meteorological warning and disaster management, and can be widely promoted and applied in related meteorological forecasting fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is the flowchart of Tibet's high-altitude road icing warning;

[0035] Figure 2 、 3 It is a historical meteorological element map collected by weather station monitoring;

[0036] Figure 4 It is an hourly granularity trend chart;

[0037] Figure 5 It is a daily granularity trend chart;

[0038] Figure 6 、 7 It is an important feature screening graph;

[0039] Figure 8 It is a confusion matrix diagram of multi-model prediction accuracy. DETAILED DESCRIPTION

[0040] In order to make the technical solution of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] Step 1: Construct a road icing prediction dataset for the high-altitude areas of Tibet from 2021 to 2024.

[0043] The first step is data collection, constructing a dataset of road icing dates in Tibet's high-altitude areas. This involves accessing the local traffic meteorological service system to download historical meteorological data from traffic meteorological stations for the target road sections during the same period. Typical sections of National Highways 109 and 318, with varying altitudes and terrain, were selected to study meteorological data from 10km and 50km sections of National Highways 109 and 318 during periods of road icing from 2021 to 2024. Data were collected hourly, spanning many years of actual observations. A dataset of road surface geothermal temperatures in Tibet covering the past 10 years was established. Typical geothermal temperatures in different regions were selected and their spatiotemporal distribution characteristics were analyzed to ensure the temporal continuity and representativeness of the samples, thereby forming a risk zoning system for road surface geothermal temperatures in Tibet. The data is stored in a CSV file, with record fields including the station number (STATION_ID), the collection time (BJ_DATETIME), and the values ​​of various meteorological indicators. The impact of meteorological factors such as geothermal temperature, air pressure, atmospheric temperature, atmospheric humidity, rainfall, and slipperiness on road icing was studied, and a preliminary road icing warning model was established.

[0044] Step 2: Data preprocessing, data cleaning is performed on the constructed road icing forecast dataset for the high altitude areas of Tibet from 2021 to 2024. Comprehensive null value missing statistics are performed on each piece of raw data. For each feature column, the missing proportion in all samples is calculated. If a column is empty in more than 50% of the samples, it is judged that its information contribution is insufficient and is eliminated to avoid introducing noise. For the remaining features, a quality assurance filling method is adopted: continuous numerical indicators are forward filled (bfill) or flattened, and categorical variables or strings are represented by "Unknown" or other categories to ensure that the meteorological values ​​at each time point are complete and continuous. Categorical features are filled with categories or marked with missing status. After this processing process, meteorological monitoring data from multiple sites and time periods are organized into a complete and continuous database. Through detailed preprocessing of the dataset, problems such as missing values ​​and outliers are solved, and the data is standardized to improve data quality and optimize model effects.

[0045] Standardization: Adopt The formula standardizes continuous numerical features, subtracts the mean, divides by the standard deviation, and transforms each continuous feature so that its value falls between [-1, 1]. The range is controlled within a reasonable interval, which helps the gradient descent method converge quickly, reduces the impact of dimensional differences between different indicators, ensures feature equivalence during model training, and avoids the impact of dimensional differences on model effects.

[0046] During the model tuning phase, rainfall features were gradually integrated, and the correlation between this indicator and icing events was strengthened through a dynamic weight adjustment mechanism. This phased feature integration strategy effectively alleviated data sparsity and distribution skewness, enabling precise quantification of the actual impact of rainfall.

[0047] Step 3: Filter key features. By setting a threshold such as importance ≥ 0.01, filter out features with higher importance and eliminate secondary features to reduce noise interference and improve model performance. When filtering key features, set a time window to extract meteorological information within a continuous time period. Select the window time as the analysis unit based on actual needs. On this basis, use the original time series to construct multiple types of features, including:

[0048] Lag characteristics: use the mean, maximum, and minimum values ​​of meteorological indicators over the past 1 hour, 2 hours, and until the end of the window;

[0049] Rate of change characteristics: Calculate the time series difference of continuous indicators;

[0050] Sliding window statistics: Within the set window time range, the sliding average, standard deviation, maximum and minimum values ​​of the indicator are calculated to capture important information about recent weather fluctuations.

[0051] The feature aggregation process includes the calculation of statistics such as mean, standard deviation, and extreme value. According to the definition of historical freezing events, time periods with continuous low temperatures or meeting freezing conditions are marked as positive samples. For continuous meteorological data within these time periods, the feature vectors are marked as "freezing". Non-freezing time periods are randomly sampled as negative samples. In this way, a set of positive and negative samples is formed, which helps the model distinguish between freezing and non-freezing weather. For example, if the data segment is intercepted with a time window of x days, the data from days x+1 to 2x before the freezing day are classified as negative samples, and the data from days x close to the freezing day are marked as positive samples. Through independent sampling of data from multiple monitoring stations, the Bootstrap resampling technology balances the category distribution and expands the sample size from the initial 11 groups to 73 groups, achieving a significant improvement in data characterization capabilities.

[0052] Step 4: Multi-model integration, train multiple models on the filtered data to obtain the accuracy data of each model, screen the most influential features according to the model's importance index, optimize the model structure, and randomly divide all samples into training set, validation set and test set during multi-model training. The division ratio can be set to 80%, 10%, and 10% to ensure consistent sample distribution between different sets. Use gradient boosting tree, random forest and Bayesian classifier multi-model training. The training process includes: fitting model parameters with training data, combining cross-validation to adjust hyperparameters such as tree depth and learning rate. After the model training is completed, use the validation set to evaluate its classification effect on unseen data, including classification accuracy, F1 score and confusion matrix, and then screen the most influential features according to the model's importance index.

[0053] The core idea of ​​the ensemble learning algorithm based on the gradient boosting framework is to improve prediction performance by iteratively training multiple weak learners (usually decision trees) and weighting the combined results to gradually optimize the prediction results through an additive model. This captures the nonlinear relationships and interactive features of features. Its prediction function is:

[0054]

[0055] in is the predicted value of the tth iteration, f t (x i ) is the output value of the t-th tree.

[0056] The objective function consists of a loss function and a regularization term, and the balance between the two represents the balance between bias and variance in machine learning.

[0057]

[0058] Loss function:

[0059] Mean Squared Error (MSE):

[0060] Cross Entropy Loss:

[0061] Regularization term Ω(f k ): Control model complexity and prevent overfitting.

[0062]

[0063] Where T is the leaf node tree of the tree, w is the weight of the leaf node, γ and λ are hyperparameters, which penalize the structural complexity and weight of the tree respectively.

[0064] Training method: Define the objective function (including the training loss term and the regularization term), and then optimize it (applicable to all supervised learning models). Each time, fix the learned tree and add a new tree. Each step of adding a tree should optimize the objective function. Objective function optimization (second-order Taylor expansion)

[0065] Perform a second-order Taylor expansion on the loss function, and the approximate objective function is:

[0066]

[0067] The first derivative (gradient) of the loss function;

[0068] The second derivative of the loss function (Hessian matrix);

[0069] Random Forest

[0070] Set parameters such as the number of trees (n_estimators = 100), the maximum depth (max_depth = None or tuned according to cross-validation), etc., and train the random forest model. Random forest integrates multiple decision trees and uses bagging and random feature selection to avoid overfitting. For each training subset, random forest builds a decision tree. At the splitting node of each tree, a subset of features is randomly selected for splitting instead of using all features. When each node splits, m candidate features (usually d features) are randomly selected from all d features. Or log2d), the node splitting indicator here uses information gain:

[0071]

[0072] It is suitable for saving computational costs and has a certain degree of model interpretability. Finally, the results of all decision trees are combined by voting (classification) or averaging (regression) to obtain the final prediction.

[0073]

[0074] Bayesian Classifier (Gaussian Process Classifier):

[0075] Equipped with appropriate kernel functions (such as RBF kernel, RBF()) and hyperparameters (length scale, noise parameter) for training. Gaussian process uses probability to infer probability and provides better uncertainty estimation, which is especially suitable for complex meteorological data with limited sample size or noise. For cases where the features are continuous, Gaussian Bayesian classifier can be used, which assumes that the features conform to Gaussian (normal) distribution under each category. The feature X under each category i The conditional probability is:

[0076]

[0077] After the model training is completed, the validation set is used to evaluate its classification effect on unseen data, including classification accuracy, F1 score and confusion matrix. Then, the most influential features are selected according to the model's importance index to optimize the model structure.

[0078] Step 5: Fusion strategy, using weighted average or voting algorithm to fuse the prediction results of multiple models to obtain the final prediction report.

[0079] Furthermore, by analyzing the importance of the features output by the model, the meteorological indicators that contribute most to icing warnings, such as temperature, humidity, and wind speed, are identified, with importance values ​​typically above 0.01. Ultimately, the trained model can be applied to actual meteorological monitoring systems to quickly generate icing risk warnings based on the latest meteorological observation data, enabling real-time and efficient meteorological disaster warnings.

[0080] The entire technical solution effectively improves the prediction accuracy of icing events through large-scale, multi-scale, and multi-angle meteorological feature extraction, combined with advanced machine learning technology, providing a scientific basis for meteorological warning and disaster management, and can be widely promoted and applied in related meteorological forecasting fields.

[0081] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. The optical fiber-based road icing multi-model feature importance screening prediction method is characterized by: The specific steps include: S1. Construct a road icing prediction dataset for high-altitude areas, obtain a dataset including ground temperature measurement data, meteorological data, and numerical weather forecast data, and transmit the obtained dataset to subsequent processing links via optical fiber; S2. Data preprocessing: standardize the numerical features in historical meteorological observation data to ensure that the data falls within a uniform interval and use missing value filling strategies to maintain data integrity; S3. Screen key features. By setting thresholds, screen out features with higher importance and eliminate minor features. S4. Multi-model integration: train multiple models on the filtered data to obtain the accuracy data of each model, select the most influential features based on the model's importance index, and optimize the model structure; S5, fusion strategy, uses weighted average or voting algorithm to fuse the prediction results of multiple models to obtain the final prediction report.

2. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: In step S1, the measured ground temperature data is the actual road surface temperature collected by a road surface sensor or a mobile device; the meteorological data includes meteorological elements such as temperature, humidity, wind speed, and precipitation; and the numerical weather forecast data is meteorological forecast data related to road icing.

3. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11. Download historical meteorological data of the traffic meteorological station at the same time of the target road section by accessing the local traffic meteorological service system; S12. Select typical road sections at different altitudes and terrains in the target area. Study the meteorological data of the target road sections during icing periods over the years. Collect data every hour over a period of many years. Establish a low-temperature road surface data set for the target area. Select typical ground temperatures in different regions and analyze their spatiotemporal distribution characteristics to ensure the temporal continuity and representativeness of the samples. This will form a ground temperature risk zoning for the target area. S13. Store the data in a CSV file. The record fields include the display number, site number, collection time, and the values ​​of various meteorological indicators. Study the impact of meteorological factors such as ground temperature, air pressure, atmospheric temperature, atmospheric humidity, rainfall, and slipperiness on road icing, and establish a preliminary road icing warning model.

4. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21. Clean the road icing prediction dataset for the target area. Perform a comprehensive null value count on each piece of raw data. For each feature column, calculate the missing value ratio in all samples. Columns that are null in more than 50% of the samples are considered to have insufficient information contribution and are removed. S22. Use quality assurance to fill in the remaining features: forward fill or flatten continuous numerical indicators, and use "Unknown" or other categories to represent categorical variables or strings, to ensure that the meteorological values ​​at each time point are complete and continuous; S23, use The formula standardizes continuous numerical features, subtracts the mean, divides by the standard deviation, and transforms each continuous feature; S24. During the road icing model tuning phase, rainfall characteristics are gradually integrated, and the correlation modeling between this indicator and icing time is strengthened through a dynamic weight adjustment mechanism.

5. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: In step S3, when screening key features, a time window is set to extract meteorological information within a continuous time period. The window time is selected as the analysis unit according to actual needs. On this basis, multiple types of features are constructed using the original time series, including: Lag characteristics: use the mean, maximum, and minimum values ​​of meteorological indicators over the past 1 hour, 2 hours, and until the end of the window; Rate of change characteristics: Calculate the time series difference of continuous indicators; Sliding window statistics: Within the set window time range, the sliding average, standard deviation, maximum and minimum values ​​of the indicator are calculated to capture important information about recent weather fluctuations.

6. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: In step S3, the feature aggregation process involves calculating statistics including mean, standard deviation, and extreme values. Based on the definition of historical freezing events, time periods with sustained low temperatures or meeting freezing conditions are marked as positive samples. For continuous meteorological data within these time periods, the feature vectors are marked as "freezing." Non-freezing time periods are randomly selected as negative samples.

7. The optical fiber-based road icing multi-model feature importance screening and prediction method according to claim 1, characterized in that: In step S4, when performing multi-model training, all samples are randomly divided into a training set, a validation set, and a test set, and multi-model training is performed using gradient boosting trees, random forests, and Bayesian classifiers. The training process includes: fitting model parameters using training data, adjusting hyperparameters such as tree depth and learning rate in combination with cross-validation, and after the model training is completed, using the validation set to evaluate its classification effect on unseen data, including classification accuracy, F1 score, and confusion matrix, and then screening the most influential features based on the model's importance index.

8. Fiber-optic-based road icing multi-model feature importance screening and forecasting system, characterized by: The system is used Implementing the optical fiber-based road icing multi-model feature importance screening and prediction method according to any one of claims 1 to 7, Including: data preprocessing module, feature importance analysis module, multi-model training module and fusion output module, Each module implements the above steps and works together to form a complete forecast process.

Images