Fine weather element prediction and early warning method for road transportation of launch vehicle

Abstract

The application relates to a refined meteorological element prediction and early warning method for road transportation of a carrier rocket, which comprises the following steps: S10, collecting meteorological prediction data, meteorological live data and road network data; S20, constructing a training sample set based on the meteorological prediction data and the meteorological live data, and training a prediction correction model based on the training sample set; S30, performing meteorological prediction on a target point on a transportation route by using the prediction correction model, and performing time scale reduction processing on a prediction result to obtain a meteorological element prediction result of a specified time resolution of the transportation route; performing space scale reduction processing based on the meteorological prediction data and geographical information of the target point to obtain a meteorological element prediction result of a specified space resolution of the transportation route; and S40, generating a prediction and early warning product of the transportation route based on the meteorological element prediction result. The application can perform refined meteorological element prediction and early warning on a target point in a transportation route.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, and in particular to a refined meteorological element forecasting and early warning method for road transport of launch vehicles. Background Technology

[0002] With the expansion of the space launch market and the explosive growth in the types of launch vehicles, launch vehicles are increasingly relying on road transport, moving away from primarily rail and sea transport. Road transport offers advantages such as low cost, high efficiency, high mobility, and a wide range of choices; however, compared to rail and sea transport, it is more susceptible to weather factors and secondary disasters caused by weather. Launch vehicles are high-value, oversized items with extremely stringent environmental requirements, making weather risk a key constraint in road transport. Therefore, it is necessary to plan road transport routes for launch vehicles, conduct detailed forecasts of key meteorological elements such as temperature, precipitation, wind speed, visibility, and road surface temperature, and provide accurate early warnings for severe weather to minimize adverse impacts during transport.

[0003] In terms of intelligent grid forecasting, my country has initially established a national-level seamless intelligent grid forecasting operational system, providing hourly digital grid forecast products with a resolution of 5 kilometers for 0-30 days nationwide, and improving near-ground resolution to 1 kilometer. However, this grid is a fixed, regular grid, which is disconnected from the actual geometric alignment of roads, while rocket transport routes are dynamically changing sets of irregular geometric lines in space. Therefore, existing intelligent grid products are difficult to directly convert into accurate forecasts for road segments down to the kilometer level. Summary of the Invention

[0004] The purpose of this invention is to provide a refined meteorological element forecasting and early warning method for road transportation of launch vehicles, so as to solve the problem that existing forecast products cannot be directly adapted to the needs of road transportation of launch vehicles.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A refined meteorological element forecasting and early warning method for road transportation of launch vehicles includes the following steps:

[0007] S10 collects weather forecast data, real-time weather data, and road network data, and obtains the geographical information of each target point in the carrier rocket transportation route based on the road network data.

[0008] S20, construct a training sample set based on the meteorological forecast data and the actual meteorological data, and train a forecast correction model based on the training sample set;

[0009] S30, the forecast correction model is used to make meteorological forecasts for target locations on the transportation route, and the forecast results are downscaled over time to obtain meteorological element forecast results for the transportation route at a specified time resolution; based on the meteorological forecast data and the geographic information of the target locations, spatial downscaling is performed to obtain meteorological element forecast results for the transportation route at a specified spatial resolution.

[0010] S40, Generate a forecast and early warning product for transportation routes based on the forecast results of the meteorological elements.

[0011] As an example, in S30, the spatial downscaling processing for temperature includes: firstly, selecting the temperature forecast data of the nearest smart grid point in the smart grid feature forecast product. The temperature data at the target location is used as the basis for calculation, and then the temperature data at the target location is corrected according to the terrain elevation. , The corrected temperature value for the target location. The rate of change of temperature with altitude, determined based on historical data. The terrain elevation of the target point. The terrain elevation of the nearest smart grid point.

[0012] As an example, in S30, the time downscaling processing of temperature includes: comparing the multi-model integrated temperature forecast results output by the forecast correction model with the diurnal variation curve of the CMA-MESO refined hourly model;

[0013] If the two trends are consistent, the 3-hour temperature integrated forecast is split into hourly forecasts by calculating the proportional relationship coefficient between the temperature at each 3-hour interval of the integrated temperature forecast and the corresponding temperature, the highest temperature and the lowest temperature of the time period in the refined hourly model.

[0014] If the two trends are inconsistent, the 3-hour temperature integrated forecast is broken down into hourly forecasts by approximating the previous day's actual rise and fall trend.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects:

[0016] 1. Filling a Technological Gap: This invention, based on real-time weather station data, intelligent grid forecasting, and multi-mode integrated forecasting products, produces a refined hourly meteorological element forecast product covering five types of severe weather for the next 72 hours, with a spatial resolution of 1 kilometer and a temporal resolution of 1 hour. This high-precision and high-resolution forecasting capability significantly improves the early warning and response capabilities for extreme weather events during the road transportation of launch vehicles. It solves the problem of how to apply low-resolution (e.g., 5km / 3h) intelligent grid element forecasting products to the 1km / 1h road segment forecasting required for rocket transport vehicles, and how to establish a quantifiable, multi-element comprehensive risk early warning system specifically applicable to rocket transportation safety management.

[0017] 2. Generation of Refined Early Warning Products: Based on the most important meteorological risk factors and geographical factors during the road transportation of launch vehicles, a scientific early warning level and early warning products were constructed. Different spatial and temporal downscaling methods were adopted for different meteorological elements, improving the precision of risk early warning. This not only enhanced the meteorological risk early warning capability during the road transportation of launch vehicles but also provided a referable technical solution for other similar scenarios (such as special cargo transportation and emergency rescue). It also solved the problem of how to eliminate the nonlinear impact of complex terrain (such as mountains, wind gaps, bridges, and tunnels) on local meteorological elements (such as wind and visibility).

[0018] 3. After adopting this method, for a complex mountainous transportation section in Southwest China, the mean absolute error (MAE) of the 24-hour temperature forecast decreased from 2.1℃ in the smart grid forecast product to 0.8℃; the wind speed forecast error decreased by 35%; and the hit rate of fog warnings for visibility below 500m reached 92%, with the false alarm rate reduced to 15%. Compared with traditional methods, the accuracy of identifying high-risk road sections improved by 40%. Attached Figure Description

[0019] Figure 1 The flowchart illustrates a refined meteorological element forecasting and early warning method for road transportation of launch vehicles, as exemplified in this embodiment. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This embodiment provides a refined meteorological element forecasting and early warning method for road transportation of launch vehicles, comprising the following steps:

[0022] S10 collects weather forecast data, real-time weather data, and road network data, and obtains the geographical information of each target point in the carrier rocket transportation route based on the road network data.

[0023] S20, Construct a training sample set based on meteorological forecast data and actual meteorological data, and train a forecast correction model based on the training sample set;

[0024] S30, the forecast correction model is used to make meteorological forecasts for target locations on the transportation route, and the forecast results are downscaled over time to obtain meteorological element forecast results for the transportation route at a specified time resolution; based on the meteorological forecast data and the geographic information of the target locations, spatial downscaling is performed to obtain meteorological element forecast results for the transportation route at a specified spatial resolution.

[0025] S40, Generate a forecast and early warning product for transportation routes based on the forecast results of the meteorological elements.

[0026] In this embodiment, the meteorological forecast data in step S10 is a smart grid element forecast product (referred to as the coarse-resolution product in this document). This product is obtained from a national meteorological operational platform, with a spatial resolution of 0.05° (approximately 5 kilometers), a lead time of 72 hours, and a temporal resolution of 3-hour increments. The forecast elements involved include temperature, cumulative precipitation, 10-meter wind speed, and relative humidity. Precipitation forecasts can be corrected using a subjective-objective fusion technique. This involves adding the forecaster's subjective precipitation area forecast to the model's quantitative precipitation forecast to correct the precipitation area, and also incorporating topographic correction.

[0027] The meteorological data consists of hourly real-time observation data from conventional ground meteorological stations (national reference stations, basic stations) and traffic meteorological stations along the transportation route, as well as hourly real-time analysis products from the China Smart Grid (the coarse-resolution products mentioned in this article), including air temperature, precipitation, wind speed, visibility, relative humidity, snowfall, road surface temperature, etc.

[0028] Collecting highway network data involves using ArcGIS software to read highway alignment topology vector data (shapefile format) and segmenting the rocket transport route at intervals of, for example, 3 kilometers, to obtain the geographic information of each segment point. Specifically, the route centerline is extracted, and a segment point is set every 3 kilometers, for example. Each segment is associated with the following attributes: segment number, starting station number, ending station number, midpoint latitude and longitude, altitude, administrative division (province / city / county level coding), highway alignment (straight / curve, curve radius), slope type (flat / moderate / steep), and special location markers (bridge / tunnel / elevated / windy section). Curved sections are defined as those with a curvature radius ≤ 500m, and steep slopes are defined as those with a slope ≥ 5%. Special location types such as bridges, viaducts, tunnel exits, and culverts are marked separately.

[0029] In step S20 above, a training sample set is constructed for each meteorological element along each segment of the launch vehicle's transportation route. A corresponding forecast correction model is then trained based on this training sample set. The method for constructing the training sample set for each meteorological element along each segment and the training process for the corresponding forecast correction model are the same.

[0030] Taking a specific road segment as an example, the construction of the training sample set first involves selecting continuous forecast sequences and corresponding real-time sequences for that road segment within the past M (e.g., 30) days, with the same start time (e.g., 08:00 Beijing time) and the same forecast lead time (e.g., forecasting the next 24 hours). The forecast sequence refers to a data sequence composed of forecast values, and the real-time sequence refers to a data sequence composed of real-time values. Then, model training is performed based on the initial sample set, using an adaptive sequence length algorithm to determine the optimal sequence length: starting from the past day, the sequence length is increased daily (e.g., 1 day, 3 days, 5 days, 7 days…), and the training error and validation error under each sequence length are calculated. The sequence length value with the smallest loss index is selected as the optimal sequence length N for model training. Finally, continuous forecast sequences and corresponding real-time sequences for that road segment within the past N (N < M) days, with the same start time (e.g., 08:00 Beijing time) and the same forecast lead time (e.g., forecasting the next 24 hours) are selected to construct the training sample set, which is used to train the forecast correction model to determine the model parameters. The forecast and actual values ​​of a point at an interval along the launch vehicle's transport route constitute an important part of a training sample (the training sample also includes the geographical information of that point; introducing geographical information can make the trained forecast correction model more suitable for forecasting meteorological elements at target points along the transport route, thereby improving forecast accuracy).

[0031] For determining the forecast and actual values ​​at each point along the launch vehicle's transport route, the actual values ​​are the measured values ​​from meteorological stations or smart grid real-time products, while the forecast values ​​are the forecast values ​​from smart grid element forecast products.

[0032] Specifically, the system queries a nearby weather station table (created from real-time weather data and highway network data for rocket transport routes) to find weather stations whose geometric orthogonal distance to the road segment is less than a set distance threshold (e.g., 1 kilometer). If there is only one nearby weather station and its data is valid and complete, then that weather station's data is used directly as the real-time value for the road segment. If multiple weather stations are matched, then the inverse distance weighted average of the geometric orthogonal distances between the weather stations and the road segment is used as the real-time value for the road segment.

[0033] If a road segment has no nearby weather station, the road network proximity matching algorithm and the hourly product data of the smart grid real-time analysis are used to calculate the real-time value of the road segment. The operation of the road network proximity matching algorithm is as follows: with the midpoint of the road segment as the center, match the four grid real-time product grid points surrounding the midpoint; for temperature and relative humidity, the weighted average of the four grid real-time product grid points is calculated using the inverse distance weighting method as the real-time value of the road segment; for precipitation and wind speed, the nearest point is calculated using the great circle distance, and the real-time value of the nearest point is used as the real-time value of the road segment.

[0034] As an example, assume points A, B, C, and D are the locations of the meteorological elements required along the launch vehicle's transport route; P1, P2, P3, and P4 are forecast grid points; T1, T2, T3, and T4 are real-time data from the smart grid; and S1, S2, S3, S4, and S5 are automatic weather stations. Therefore:

[0035] (1) There are four forecast grid points around point A: P1, P2, P3, and P4. Therefore, the temperature and relative humidity of point A are calculated using the inverse distance weighting method to obtain the weighted average of the four points as the forecast value of the road segment. The precipitation and wind speed of point A are obtained by using the forecast value of the grid point P4 that is closest to it as the forecast value of the road segment.

[0036] (2) There is an S1 automatic weather station within 1km of point B. Therefore, the temperature, relative humidity, precipitation and wind speed of point B are measured by the S1 automatic weather station as the actual values ​​of the road section.

[0037] (3) There are 4 automatic weather stations S2, S3, S4 and S5 within 1km of point C. Therefore, the temperature, relative humidity, precipitation and wind speed of point C are calculated by weighting the inverse distance of the geometric orthogonal distance between the station and the road segment as the actual value of the road segment.

[0038] (4) There is no automatic weather station within 1km of point D. Therefore, the temperature, relative humidity, precipitation, and wind speed at point D are calculated using the road network proximity matching algorithm and the intelligent grid real-time products (T1, T2, T3, T4). Specifically, the temperature and relative humidity at point D are calculated using the inverse distance weighting method to obtain the weighted average of the four points (T1, T2, T3, T4) as the real-time value of the road segment; the precipitation and wind speed at point D are calculated using the real-time value of the closest grid T3 as the real-time value of the road segment.

[0039] The forecast correction model preferably employs an artificial intelligence algorithm model, and is an integrated fusion of multiple artificial intelligence algorithm models. In this embodiment, XGBoost, LightGBM, and CatBoost are used as the base models, combined with a stack to integrate the forecast correction model. XGBoost, LightGBM, and CatBoost use feature variables including numerical features and categorical features, have low external data dependence, do not require the collection of a large amount of training data in advance, and compared with purely numerical input methods such as SVM, can more flexibly reflect changes in altitude and differences in local regional features, making them more suitable for correction processing of points in complex terrain based on forecasts.

[0040] More specifically, the input feature variables for the forecast correction model include administrative region code, altitude, longitude, latitude, and forecast and actual values ​​of meteorological elements (smart grid forecast values ​​for temperature, relative humidity, wind speed, etc.). In other words, each sample in the training sample set contains not only paired forecast and actual values, but also four types of geographic information: administrative region code, altitude, longitude, and latitude. Specifically, the administrative region code is converted into a categorical variable using scikit-learn's LabelEncoder, and longitude and latitude are converted into periodically varying sin and cosine components. The regressor is the forecast relative error, i.e., the difference between the actual and forecast values.

[0041] The constructed training sample set was used to train three boosting decision tree models: XGBoost, LightGBM, and CatBoost. Appropriate parameters were set for XGBoost, LightGBM, and CatBoost, primarily adjusted through cross-validation. After training, the outputs of the three base models were weighted and fused using a Stacking ensemble strategy. Specifically, StackingCVRegressor (a cross-validation stacked regression model that generates predictions from base models through cross-validation and trains a meta-model for fusion) from the mlxtend library was used, with linear regression as the meta-learner and 5-fold cross-validation. The outputs of the three base models were used as the second-layer input features, and the meta-learner learned the optimal combination weights. This involved combining the outputs of the three base models with various weights as new features, constructing a training set for training the meta-learner, and finally obtaining the weighted and fused prediction results.

[0042] After the forecast correction model is trained, it can continue to be trained, incrementally updated daily with the latest forecast-real-time paired data (geographical information is fixed and does not change over time), maintaining the optimal training sequence length. For example, after adding data for the 11th, the data for the 1st is removed and replaced with data for the 11th (the optimal sequence length is 10 days). Simultaneously, it checks for significant changes in the forecast bias mode (e.g., from systematically overestimating to systematically underestimating). If so, training continues to redetermine the optimal sequence length, adjusting it to shorten the response lag time. Continuous training ensures optimal training sequence lengths are maintained across different seasonal environments. Significant changes in the forecast bias mode are detected, for example, using standardized values ​​obtained from historical average wind speeds and standard deviations. A significant deviation from the standardized value indicates a significant change in the forecast bias mode.

[0043] In step S30 above, during the downscaling process, it is necessary to consider the forecasting effects of different interpolation analysis methods on different models in terms of spatial downscaling of continuous and discontinuous variables. Simultaneously, it is necessary to consider the impact of refined geographic information on different elements during the interpolation analysis process. Based on the spatiotemporal continuity characteristics of meteorological elements, forecast elements are divided into two main categories: continuous elements (such as temperature and relative humidity) and discontinuous elements (such as precipitation), and different spatial and temporal downscaling methods are applied to each category.

[0044] (a) Spatial downscaling method for continuous variable air temperature (relative humidity and road surface temperature are treated in the same way as air temperature).

[0045] The coarse-resolution (e.g., 5km) smart grid temperature forecast data is interpolated to a 1km fine-resolution grid point or a given fixed point (collectively referred to as the target point). In other words, the temperature forecast data of the grid point closest to the target point in the smart grid feature forecast product is selected as the target point temperature data along the transportation route. Target locations on a transportation route refer to physical locations along the transportation route where meteorological forecasts are to be conducted. For example, the distance between adjacent target locations is 1 km (high resolution).

[0046] It is easy to understand that coarse resolution and high resolution (or fine grid) in this article are relative concepts, not absolute concepts. For example, existing smart grid feature forecast products are coarse resolution models with a spatial interval of 5 km, while the fine forecasts described in this article require a spatial interval of 1 km, which is a high resolution model.

[0047] Considering the variation of air temperature with altitude, the following "temperature-topography" model is established to obtain the final air temperature data for the target location. : ,in, It is the temperature data of the target point obtained by interpolation from a coarse-resolution mode to a 1km fine grid or a given fixed point location. The rate of change of temperature with altitude, determined based on historical data. It is the terrain elevation of the refined target location. It is the terrain height of the nearest smart grid point in the coarse resolution mode.

[0048] (b) Spatial downscaling method for discontinuous variable precipitation (snowfall).

[0049] Nearest neighbor interpolation and downscaling methods based on precipitation variations with height and aspect are employed. Coarse-resolution model precipitation data are interpolated to finer grid points or fixed locations based on given geographic information using nearest neighbor interpolation. Considering the significant impact of different terrain heights and slope aspects (windward and leeward slopes) on precipitation magnitude or presence, precipitation is categorized into different levels (light rain (x1), moderate rain (x2), and heavy rain (x3)) based on fine-grained geographic information and historical precipitation observation data during the interpolation process. A training period of a certain length (e.g., at least one year of historical precipitation) is selected to establish ternary regression equations for different levels. Based on the training period data, the values ​​of regression coefficients a, b, and c are determined, and this relationship is applied to the terrain height and slope correction in the interpolation analysis from coarse-resolution model precipitation forecast data to fine-grid points or arbitrary required locations.

[0050] First, interpolation is performed using the nearest neighbor data. For a target location, its nearest neighbor grid point in the coarse-resolution data is found, and then the predicted precipitation for that grid point is... By combining the ternary regression equation and assigning fine grid points or target points, the baseline precipitation at the target point is obtained. .

[0051] Then, the baseline precipitation is corrected based on the relationship between precipitation and topographic height and slope aspect. The relationship between precipitation P and height H and slope aspect (windward or leeward) is an empirical exponential relationship:

[0052] ;

[0053] in, This is the baseline precipitation, which here is the precipitation after interpolating the coarse-resolution model forecast precipitation data using the nearest neighbor method; This indicates the impact of terrain altitude on precipitation; precipitation increases exponentially with increasing altitude. This reflects the impact of slope on precipitation. Parameter , , It is determined by fitting actual data to adapt to specific geographical environments.

[0054] (c) Spatial downscaling method for wind speed, a variable with both continuous and discontinuous characteristics.

[0055] Wind direction and speed are controlled by both weather and climate background and significantly influenced by topography. Therefore, wind direction and speed exhibit both continuity and discontinuity (abrupt changes). When downscaling in space, the influence of topographic height, slope, and aspect must be considered. Due to the complexity of wind speed and direction changes, the inverse distance interpolation method is adopted. During the interpolation process, based on refined geographic information and historical wind observation data, a training period of a certain length is selected to establish a multiple regression equation. The regression coefficients are determined during the training period, and the regression relationship is applied to the correction of topographic height, slope, and aspect in the interpolation analysis of wind forecast data from coarse-resolution models to 1km grid points or actual fixed location points.

[0056] Specifically, firstly, coarse-resolution wind speed (V) and wind direction (D) forecast data (intelligent grid feature forecast product) are collected, along with terrain height (H), slope (S), and aspect (A) data. For each target point location (x, y) to be forecasted, its distance to surrounding coarse-resolution grid points is calculated. Then, interpolation is performed using the inverse distance weighting formula, assigning greater weight to closer grid points:

[0057] , ;

[0058] Where is the distance between the l-th adjacent grid point and the target point in coarse resolution mode, which is the power of the distance (here, we can take the square).

[0059] Based on historical observation data and geographical information, a training period is selected, and a multiple linear regression equation for wind speed and wind direction is established:

[0060] ;

[0061] ;

[0062] in, , , , , , , , Both are regression coefficients, ε and This is the error term.

[0063] First, the least squares method is used to fit the selected training data to solve for the regression coefficients of the regression equation. Then, the coarse-resolution wind speed is calculated. and wind direction By substituting grid forecast data into a multiple linear regression equation, the wind speed at the high-resolution target point can be calculated. and wind direction :

[0064] ;

[0065] .

[0066] , This involves using wind speed and direction data at 1km grid points or actual fixed locations to perform downscaling analysis.

[0067] Visibility spatial downscaling employs a bilinear interpolation method. Bilinear interpolation is a linear interpolation extension of the interpolation function of two variables. Its core idea is to perform linear interpolation once in each of the two directions.

[0068] (d) Time downscaling method for continuous variable temperature.

[0069] A method is adopted to approximate the trend of the CMA-MESO integrated refined hourly model forecast with the actual rising and falling trend of the previous day. The multi-model integrated temperature forecast results output by the forecast correction model are compared with the diurnal variation curve of the CMA-MESO refined hourly model. If the trends of the two are consistent, it is considered that the refined hourly forecast trend can well reflect the rising and falling trend of the diurnal temperature variation. The proportional relationship coefficients are calculated by using the temperature at each of the 3-hour intervals in the integrated temperature forecast and the corresponding temperature, maximum temperature, and minimum temperature in the refined hourly model, and the 3-hour integrated temperature forecast is broken down into hourly forecasts.

[0070] Let the 3-hour temperature forecast value of the multi-model integrated temperature forecast be , and the corresponding 3-hour temperatures of the CMA-MESO refined hourly model be , and the highest temperature during this period be . The lowest temperature is The decomposition is performed at each time step to satisfy the overall trend and extreme value conditions. The proportional relationship can be expressed as follows: ;

[0071] in, This is the ratio coefficient at time j. At this point, the hourly temperatures after decomposition satisfy T. max and T min If the requirements are met, Then increase accordingly. The weight, if Then reduce accordingly. The weights are then assigned. Then, based on the proportional coefficients, T is allocated to each hourly segment: ,in, Let be the temperature forecast value at the j-th time after splitting.

[0072] If the multi-model integrated temperature forecast results are inconsistent with the diurnal variation curve trend of the refined model, it is considered that the hourly forecast trend of the refined model cannot well reflect the rising and falling trends in the diurnal temperature variation. In this case, the 3-hour integrated temperature forecast is broken down into hourly forecasts by approximating the actual rising and falling trends of the previous day. For example, assuming the integrated forecast temperature from 3 PM to 5 PM is 24℃, and the actual temperature from 3 PM to 5 PM is 10%, 5%, and 5% higher than the actual temperature from 3 PM to 5 PM the previous day, then the temperature in the first hour is 24℃ × 1.1 = 26.4℃, the temperature in the second hour is 24℃ × 1.05 = 25.2℃, and the temperature in the third hour is 24℃ × 0.95 = 22.8℃.

[0073] Here, the correlation coefficient is used to determine whether the diurnal variation trends of the two models are consistent. For ease of description, the diurnal variation curve of the multi-model integrated temperature forecast and the diurnal variation curve of the CMA-MESO refined hourly model are defined as A and B, respectively. First, the average values ​​of curves A and B are calculated. , Then calculate the covariance. Then calculate the standard deviations of curves A and B respectively. , Finally, the correlation coefficient is calculated. .when A correlation coefficient close to 1 indicates a high degree of consistency between the two sets of data. In this meteorological data analysis, a correlation coefficient of 1 is considered to indicate a high degree of consistency between the two sets of data. When the two sets of temperature data show the same trend, they show different trends when they show different trends.

[0074] (e) Temporal downscaling method for discontinuous variable precipitation.

[0075] For the temporal downscaling of discontinuous precipitation variables, a fusion algorithm combining CMA-MESO refined model trend approximation and relative humidity approximation is employed. Specifically, when precipitation is predicted in both the 3-hour integrated precipitation forecast period and the corresponding refined hourly model, the 3-hour integrated precipitation forecast is broken down into hourly precipitation forecasts using the proportional relationship coefficient between the hourly model precipitation distribution curve and the precipitation amount. For example, if the hourly precipitation forecasts for the 3 hours are 4mm, 2mm, and 3mm respectively, with a forecast ratio of 4:2:3, and the total precipitation predicted for the 3 hours in the integrated forecast is 5mm, then the precipitation is allocated proportionally: 2.2mm for the first hour, 1.1mm for the second hour, and 1.7mm for the third hour. When the 3-hour integrated precipitation forecast predicts precipitation, but the hourly model predicts no precipitation for the corresponding time period, the precipitation is broken down according to the relative humidity trend. The total precipitation from the integrated forecast is allocated proportionally to each hour, maintaining consistency in hourly precipitation, relative humidity, and weather phenomena. For example, if the relative humidity for the 3 hours is 90%, 95%, and 80% (a ratio of 90:95:80), and the integrated forecast predicts a total precipitation of 5 mm for the 3 hours, then the precipitation is allocated proportionally: 1.7 mm for the first hour, 1.8 mm for the second hour, and 1.5 mm for the third hour. When the integrated forecast predicts no precipitation, but the hourly model predicts either precipitation or no precipitation, the integrated forecast takes precedence, predicting no precipitation.

[0076] (f) Time downscaling method for wind speed and wind direction, which have both continuous and discontinuous characteristics.

[0077] Wind speed time downscaling employs a method of approximating the rising and falling trends from high-resolution numerical models (such as hourly forecasts fused from GRAPES_MESO and NCEP_1H). First, the multi-model wind speed ensemble forecast results (every 3 hours) are compared with the daily variation curves of the high-resolution model trend forecast. If the trends are consistent, it is considered that the high-resolution model trend forecast effectively reflects the rising and falling trends in daily wind speed variation. The proportional relationship coefficient is then calculated using the 3-hour wind speed ensemble forecast results and the corresponding wind speeds from the high-resolution model trend forecasts, thus splitting the 3-hour wind speed ensemble forecast into hourly forecasts. If the trends of the multi-model wind speed ensemble forecast results and the daily variation curves of the high-resolution trend forecast are inconsistent, it is considered that the high-resolution trend forecast cannot effectively reflect the rising and falling trends in daily wind speed variation. In this case, a linear interpolation method is used to split the 3-hour multi-model wind speed ensemble forecast into hourly forecasts.

[0078] The wind direction time downscaling uses an algorithm that combines linear interpolation with high-resolution numerical model results. First, the integrated wind direction forecast (every 3 hours) is downscaled to hourly using a linear interpolation algorithm. Then, the interpolation result is compared with the high-resolution numerical model result at the corresponding time. If the difference between the two is large (e.g., greater than 22.5°), the two are weighted and summed to obtain the final hourly forecast result; otherwise, the linear interpolation result is used as the final hourly forecast result.

[0079] The most important weather risk warning information during the road transportation of launch vehicles includes: visibility, road icing, rainfall, snowfall, wind, and other risk warning levels, impact time, and affected road sections. To effectively differentiate risk levels, weather risk warnings are divided into three levels (see Table 1): Level I, Level II, and Level III, with Level I being the highest. These are indicated by numbers: 1 represents high risk, 2 represents relatively high risk, 3 represents moderate risk, and 0 represents no risk. The temporal resolution is 72 hours in increments of 1 hour, and the spatial resolution is 1 kilometer.

[0080] Table 1: Weather Risk Warning Levels for Traffic Safety Control

[0081]

[0082] In this embodiment, an indicator system is constructed based on the multi-dimensional influencing factors of meteorology and transportation routes. Using composite meteorological elements as the core indicator and highway alignment and special locations that characterize the disaster-prone environment as correction indicators, a risk warning indicator system with threshold matching for five types of severe weather (seven meteorological elements), three types of highway alignment, and two types of special locations is established.

[0083] Weather type is used as a primary indicator within the core indicator (weather). The main weather types include fog, rain, snow, wind, and road icing. Related meteorological elements under different weather types are used as secondary indicators. These meteorological elements include visibility, precipitation, snowfall, wind speed, air temperature, road surface temperature, and relative humidity. Based on actual conditions, highway alignment elements include curvature and slope. Curvature can be straight or curved, and slope can be flat, moderate, or steep. Therefore, the three types of highway alignment are: ordinary straight, moderate curvature, moderate slope, moderate curvature and slope, sharp curve, steep slope, moderate slope and sharp curve, moderate curve and steep slope, sharp curve and steep slope, and continuous downhill. In terms of highway location, besides ordinary road sections, bridges, viaducts, interchanges, and culverts are more susceptible to severe weather. Based on the impact of severe weather on different types of highway locations, special location types are divided into two categories: bridges (crossing rivers, lakes, and seas), viaducts, interchanges, culverts, tunnel exits (for viaducts), and windy road sections in mountainous areas.

[0084] As examples, Table 2 provides the indicator thresholds and examples for determining the risk warning level of a certain road section in foggy weather; Table 3 provides the indicator thresholds and examples for determining the risk warning level of a certain road section in windy weather; Table 4 provides the indicator thresholds and examples for determining the risk warning level of a certain road section in rainy weather; Table 5 provides the indicator thresholds and examples for determining the risk warning level of a certain road section in snowy weather; and Table 6 provides the indicator thresholds and examples for determining the risk warning level of a certain road surface icing.

[0085] Table 2: Example of Fog Risk Warning Level Determination

[0086]

[0087] Table 3: Example of Determining the Level of Wind Risk Warning

[0088]

[0089] Table 4: Example of Determining Rainy Weather Risk Warning Level

[0090]

[0091] Table 5: Example of Snow Weather Risk Warning Level Determination

[0092]

[0093] Table 6: Example of Determining the Risk Warning Level of Road Icing

[0094]

[0095] In step S40 above, after obtaining the forecast results of meteorological elements for each kilometer and hourly along the transportation route, the hourly risk level is determined according to Tables 2-6 above, and then the recommended measures are determined according to Table 1 above. Finally, a refined meteorological forecast and warning product covering the entire transportation route for the next 72 hours, kilometer by kilometer and hour by hour, is generated. The information items included in the forecast and warning product may include: time period, risk type, alert level, influencing factors, and recommended measures, as shown in Table 7.

[0096] The above method is based on real-time meteorological monitoring station data, intelligent grid forecasts, and multi-model integrated forecast products (forecast correction model output). It uses spatiotemporal downscaling technology to finely interpolate and correct forecast elements (temperature, precipitation, wind, relative humidity, visibility, etc.) with coarse temporal (forecast interval >= 1 hour) and coarse spatial (spatial resolution >= 3 km) resolutions to highway markers. Combined with risk warning indicators, it produces a highway weather risk warning product for 72 hours (hourly interval) for five types of severe weather, with each marker per kilometer.

[0097] Application Examples

[0098] (I) Overview of Application Scenarios

[0099] A certain type of launch vehicle needs to be transported by road from a manufacturing plant in a city in Southwest China to a launch site in South China. The planned route is approximately 850 kilometers long, passing through four provinces (regions): Sichuan, Guizhou, Guangxi, and Guangdong. The terrain along the route is complex, including hilly and mountainous areas (elevation 400-1200 meters), river valleys and basins, as well as sections of bridges, tunnels, and elevated roads. The transportation mission is scheduled for mid-November of that year. During this period, special attention needs to be paid to meteorological risks along the route, such as low visibility due to autumn fog, crosswinds, and road icing caused by sudden drops in temperature.

[0100] (II) Data Preparation and Preprocessing

[0101] The following preparation steps should be initiated 72 hours prior to the start of the transportation mission:

[0102] (1) Forecast data acquisition: Intelligent grid element forecast products were acquired from the national meteorological operational platform, with a spatial resolution of 0.05° (approximately 5 km), a lead time of 72 hours, and a temporal resolution of 3 hours. The forecast elements involved included air temperature, cumulative precipitation, 10-meter wind speed, and relative humidity. Precipitation forecasts were corrected using subjective-objective fusion technology, superimposing forecasters' subjective corrections of precipitation areas onto the model forecasts.

[0103] (2) Real-time data collection: Hourly real-time observation data from conventional ground meteorological stations (national reference stations and basic stations) and traffic meteorological stations along the transportation route are acquired, including temperature, precipitation, wind speed, visibility, relative humidity, etc. At the same time, the hourly real-time analysis product of China Smart Grid is acquired as a supplementary data source for model training and forecast effect verification.

[0104] (3) Road network information processing: The road alignment topology vector data (shapefile format) was read using ArcGIS software, and the transportation route was segmented at 3-kilometer intervals. First, the centerline of the route was extracted, and a segmentation point was set every 3 kilometers, generating approximately 283 road segments and their corresponding midpoint latitude and longitude coordinates. Each road segment was associated with the following attribute information: road segment number (RL-001 to RL-283), starting station number, ending station number, midpoint latitude and longitude, altitude, administrative division (province / city / county level coding), road alignment (straight line / curve, curve radius), slope type (flat slope / general slope / steep slope), and special location markers (bridge / tunnel / elevated / windy section). Among them, curved road segments were defined as road segments with a curvature radius ≤ 500m, steep slopes were defined as road segments with a slope ≥ 5%, and special location types such as bridges, viaducts, tunnel exits, and culverts were marked separately.

[0105] (III) Calculation of Actual Road Segment Values ​​and Construction of Training Sequences

[0106] Take section RL-103 (located in a mountainous area of ​​Guizhou Province, chainage K323-K326, altitude 1050 meters, a winding road section) as an example.

[0107] First, the road network's neighboring station table is used to query meteorological stations with a geometric orthogonal distance of less than 1 kilometer from RL-103. The query reveals one traffic meteorological station (station number GZ-1038) within a 500-meter orthogonal distance of this road segment, providing real-time temperature, relative humidity, and wind speed data. Since this station's data is valid and without missing measurements, it is directly used as the actual value for RL-103. If a road segment has no neighboring stations, the road network's neighbor matching algorithm and intelligent grid real-time data products are used to calculate the road segment's actual value. Specifically, four grid real-time data product points surrounding the midpoint of the road segment are matched, and the weighted average of these four points is calculated using the inverse distance weighting method as the road segment's actual value. For road segments with multiple matched stations, the inverse distance weighted average is calculated using the geometric orthogonal distance between the stations and the road segment as the road segment's observed value.

[0108] For model training, the temperature forecast correction model for the RL-103 road section is taken as an example. Continuous forecast sequences and corresponding real-time sequences for the past 30 days with the same start time (e.g., 08:00 Beijing time) and forecast lead time (e.g., forecasting the next 24 hours) are selected to construct an initial training sample set. An adaptive sequence length algorithm is used to determine the optimal sequence length: starting from the past day, the sequence length is increased daily (1 day, 3 days, 5 days, 7 days…), and the training error and validation error are calculated for each length. The length value with the smallest loss index is selected as the model training sequence length. Calculations show that the temperature forecast correction model for this road section obtains the minimum root mean square error when the sequence length is 10 days; therefore, the forecast-real-time paired data from the past 10 days are used as part of the training samples.

[0109] (iv) Training of forecast correction model

[0110] Taking the temperature forecast correction for the RL-103 section as an example, the forecast correction model is constructed according to the following steps.

[0111] Step 1: Feature Variable Construction. The following feature variables are selected as model inputs: administrative region code (coded as Guizhou Province → 52), altitude (1050 meters), longitude (105.67°E) and latitude (26.33°N) are converted into sin and cos periodic components respectively, the original temperature value of the smart grid forecast (T=12.2℃), the original relative humidity forecast value (RH=86%), the temperature change trend over the past 6 hours (-1.5℃ / 6h), and the orthogonal distance between the station and the midpoint of the road segment (0.3 km).

[0112] Step 2: Definition of Regression Quantity. The regression quantity is defined as the forecast relative error, which is the difference ΔT between the actual temperature and the forecast temperature. This forecast relative error is used as the model's prediction target.

[0113] Step 3: Ensemble Model Training. The constructed training sample set (approximately 80 samples, forecasted every 3 hours over 10 days) is input into three boosting decision tree models: XGBoost, LightGBM, and CatBoost, for training. Taking XGBoost as an example, the maximum tree depth is set to 6, the learning rate to 0.05, and the subsampling ratio to 0.7. LightGBM and CatBoost are configured with appropriate parameters, primarily adjusted through cross-validation. After training, the outputs of the three models are weighted and fused using a Stacking ensemble strategy. Specifically, StackingCVRegressor (a cross-validated stacked regression model that generates predictions from base models through cross-validation and trains a meta-model for fusion) from the mlxtend library is used, with linear regression as the meta-learner and 5-fold cross-validation. The outputs of the three base models are used as the second-layer input features. The meta-learner learns the optimal combination weights, meaning the outputs of the three base models are combined with various weights as new features to construct a training set for training the meta-learner, ultimately yielding the weighted and fused prediction results. Calculations show that the average absolute error of the RL-103 section after integration is reduced by approximately 0.3°C compared to the optimal value of the single model.

[0114] Step 4: Adaptive Update Mechanism. After model training is complete, incremental updates are performed daily using the latest forecast-real-world paired data, maintaining the optimal training sequence length (10 days). For example, after adding data for the 11th, data for the 1st is removed and replaced with data for the 11th. When the forecast bias mode changes significantly (e.g., from systematically high to systematically low), the adaptive algorithm can adjust the sequence length to shorten the response lag time.

[0115] (v) Spatiotemporal downscaling

[0116] Taking the temperature data of the RL-103 road section as an example, this illustrates the downscaling process of continuous variables. The temperature data obtained from the smart grid forecast is in 3-hour intervals and needs to be downscaled to hourly forecasts. First, the hourly temperature forecast from the CMA-MESO refined hourly model is obtained as a high temporal resolution trend reference. The multi-model integrated temperature forecast results (i.e., the 3-hourly forecast results predicted by the forecast correction model) are compared with the diurnal variation curve of the CMA-MESO refined hourly model. If the diurnal variation trends of the two are consistent, it is considered that the hourly model can well reflect the temperature rise and fall trends, and then the proportional relationship coefficient is calculated: using the initial temperature T0 and the final temperature T3 within the 3-hour forecast period, combined with the highest and lowest temperatures of the corresponding period in the hourly model, the temperature change trend is distributed to the hourly nodes according to the weighting coefficient. If the trends of the two are inconsistent, the rise and fall ratio of the actual temperature at the same time of the previous day is used to split the 3-hour integrated forecast into hourly forecasts.

[0117] For the RL-103 section, after comparison, the daily variation trend of the multi-model integrated forecast and the CMA-MESO hourly model is basically consistent (correlation coefficient r=0.92). Using the trend approximation method, the temperature from 11:00 to 13:00 on November 12 (forecast period) is downscaled from 3-hour resolution to hourly. The multi-model integrated forecast for the 3-hour period is 12.9℃, and the hourly temperature forecast of the CMA-MESO hourly model is 12.2℃ at 11:00, 12.8℃ at 12:00, and 13.2℃ at 13:00. At this time, the proportional relationship coefficients can be obtained as 0.32, 0.34, and 0.34, and finally broken down into 12.3℃ at 11:00, 13.1℃ at 12:00, and 13.1℃ at 13:00.

[0118] For non-continuous variables (such as precipitation), the time downscaling adopts a fusion algorithm of CMA-MESO refined model trend approximation and relative humidity approximation. When the 3-hour integrated precipitation forecast indicates precipitation but the corresponding hourly model shows no precipitation, the data is split according to the relative humidity trend coordination and consistency algorithm to ensure the logical consistency of hourly precipitation, relative humidity, and weather phenomena.

[0119] (vi) Generation of risk warning products

[0120] Take the generation of early warning products for the RL-103 section during the forecast period (0:00 to the next 72 hours) on November 12 as an example.

[0121] (1) Crosswind and wind risk warning: Based on the wind speed forecast, the wind speed of the RL-103 section will reach 22.5 m / s (level 9) between 15:00 and 17:00 on the 12th. The section is located in the valley between two mountains and is a windy section. According to Table 3 (Wind Risk Warning Level Judgment Table), the wind speed is 20.8 m / s ≤ Ws ≤ 28.5 m / s. Therefore, the wind risk warning level for this period is determined to be Level I (high risk). As shown in Table 1, the corresponding measure is to prohibit rocket carrier vehicles from passing through this section during this period.

[0122] (2) Road icing risk warning: Based on the forecast of air temperature, relative humidity, precipitation and road surface temperature, the air temperature of RL-103 section from 06:00 to 09:00 on the 12th was 1.5℃, the road surface temperature was 2.2℃, the relative humidity was 92% and the precipitation was 0.6mm. Referring to Table 6 (Road Icing Risk Warning Level Determination Table), this section is a curved and sloping road alignment and includes a bridge section (chainage K323+500 to K326+300), which meets the criteria of "curved and sloping road alignment and / or bridge and tunnel section". According to the table, the road icing risk warning level is determined to be Level II (higher risk). As shown in Table 1, the corresponding measure is to prohibit rocket carrier vehicles from passing through this section during this period.

[0123] (vii) Early warning product output

[0124] Based on the above calculations, taking road segment RL-103 as an example, the following comprehensive forecast and early warning products were generated for different time periods on November 12, as shown in Table 7.

[0125] Table 7: Forecast and Early Warning Products for Section RL-103 from 0:00 to 24:00 on November 12

[0126]

[0127] The warning level for the entire route (RL-001 to RL-283) is calculated kilometer by kilometer and hour by hour using the same method, ultimately generating a refined meteorological warning product covering the entire transportation route for the next 72 hours, kilometer by kilometer and hour by hour. The warning results can be output in the form of a data table of "Route - Time - Risk Type - Warning Level - Recommended Measures" through a visualization platform, supporting rapid retrieval and sorting according to dimensions such as high-risk routes and high-impact periods, providing a scientific basis for rocket transportation scheduling decisions.

[0128] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A refined meteorological element forecast warning method for road transportation of launch vehicles, characterized by, Includes the following steps: S10 collects weather forecast data, real-time weather data, and road network data, and obtains the geographical information of each target point in the carrier rocket transportation route based on the road network data. S20, construct a training sample set based on the meteorological forecast data and the actual meteorological data, and train a forecast correction model based on the training sample set; S30, the forecast correction model is used to make meteorological forecasts for target locations on the transportation route, and the forecast results are downscaled over time to obtain meteorological element forecast results for the transportation route at a specified time resolution; based on the meteorological forecast data and the geographic information of the target locations, spatial downscaling is performed to obtain meteorological element forecast results for the transportation route at a specified spatial resolution. S40, Generate a forecast and early warning product for transportation routes based on the forecast results of the meteorological elements.

2. The refined weather element forecast warning method for road transportation of launch vehicles according to claim 1, characterized in that, In S20, the transportation route is divided into several segments according to the set distance interval. For each segment, the forecast values ​​of the segment in the past M days with the same start time and the same forecast time are selected from the meteorological forecast data and the meteorological real-time data to form a forecast sequence and the corresponding real-time values ​​to form a real-time sequence, thus constructing an initial training sample set. Then, the model is trained based on the initial sample set to determine the optimal sequence length N; finally, the forecast sequences and corresponding real-time sequences of the road segment in the past N days with the same reporting time and the same forecast lead time are selected to construct a training sample set, where N < M and all are positive integers.

3. The refined weather element forecast warning method for road transportation of launch vehicles according to claim 2, characterized in that, The process of determining the optimal sequence length N by training the model based on the initial sample set includes: increasing the sequence length daily starting from the past day, training the model based on training samples with different sequence lengths, calculating the training error and validation error for each sequence length, and selecting the sequence length corresponding to the minimum loss index as the optimal sequence length N.

4. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 2, characterized in that, For determining the forecast and actual values ​​at each interval point along the transportation route, the actual values ​​are the measured values ​​from meteorological stations or smart grid real-time products, and the forecast values ​​are the forecast values ​​from smart grid element forecast products. The method for determining the actual value is: query the meteorological stations whose geometric orthogonal distance to the road segment is less than the set distance threshold; If there is only one nearby weather station and the data from that weather station is valid and complete, then the data from that weather station shall be used directly as the actual value for that road segment. If multiple weather stations are matched, the inverse distance weighted average value is calculated based on the geometric orthogonal distance between the weather stations and the road segment and used as the actual value of the road segment. If there is no nearby meteorological station for a certain road segment, then the four grid points of real-time products surrounding the midpoint of the road segment are matched. For temperature and relative humidity, the weighted average of the four grid points of real-time products is calculated using the inverse distance weighting method as the real-time value of the road segment. For precipitation and wind speed, the nearest point is calculated by the great circle distance, and the real-time value of the nearest point is used as the real-time value of the road segment.

5. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 1, characterized in that, The forecast correction model is an integration of multiple artificial intelligence algorithm models.

6. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 5, characterized in that, The prediction correction model is trained in the following way: the constructed training sample set is input into the three models XGBoost, LightGBM and CatBoost respectively for training. After training, the outputs of the three models are weighted and fused by the Stacking ensemble strategy.

7. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles based on claim 1, characterized in that, In step S30, the spatial downscaling of temperature includes: firstly, selecting the temperature forecast data of the smart grid point closest to the target point in the smart grid feature forecast product. The temperature data at the target location is used as the basis for calculation, and then the temperature data at the target location is corrected according to the terrain elevation. , The corrected temperature value for the target location. The rate of change of temperature with altitude, determined based on historical data. The terrain elevation of the target point. The terrain elevation of the nearest smart grid point.

8. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 1, characterized in that, In step S30, the spatial downscaling of precipitation includes: firstly, selecting the precipitation data of the smart grid point closest to the target point in the smart grid feature forecast product. And combined with the impact of precipitation level Baseline precipitation for the target location , a, b, and c are regression coefficients, and x1, x2, and x3 represent different levels of precipitation. Then, the baseline precipitation at the target location is corrected based on topographic elevation and slope aspect. P represents the corrected precipitation at the target location. This is the baseline precipitation for the target location. Indicates the relationship between terrain elevation and precipitation; Parameters reflecting the impact of slope on precipitation , , Determined through fitting actual data, where H represents terrain height. Indicates slope direction.

9. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 1, characterized in that, In S30, the time downscaling processing for wind speed includes: First, the wind speed and direction at the target location in the coarse-resolution mode are obtained using the inverse distance weighting method. , , It is the distance between the l-th adjacent grid point and the target point. The distance is a power of itself. Then, based on historical observation data and geographical information, a training period is selected, and a multiple linear regression equation for wind speed and wind direction is established: , ,in, , , , , , , , Both are regression coefficients, ε and The error term is obtained by least squares fitting using the selected training data; Finally, and wind direction Substituting the values ​​into the multiple linear regression equation, the wind speed at the target location in the high-resolution mode is calculated. and wind direction : , .

10. The refined meteorological element forecasting and early warning method for road transportation of launch vehicles according to claim 1, characterized in that, In S30, the time downscaling processing of temperature includes: comparing the multi-model integrated temperature forecast results output by the forecast correction model with the diurnal variation curve of the CMA-MESO refined hourly model; If the two trends are consistent, the 3-hour temperature integrated forecast is split into hourly forecasts by calculating the proportional relationship coefficient between the temperature at each 3-hour interval of the integrated temperature forecast and the corresponding temperature, the highest temperature and the lowest temperature of the time period in the refined hourly model. If the two trends are inconsistent, the 3-hour temperature integrated forecast is broken down into hourly forecasts by approximating the previous day's actual rise and fall trend.

Images