Gust prediction method based on deep learning

Abstract

The present application provides a gust forecast method based on deep learning, which belongs to the cross technical field of meteorological numerical forecast and artificial intelligence. The present application takes the multi-dimensional environmental elements of ECMWF numerical forecast as the model input, takes the ERA5 reanalysis 10-meter instantaneous gust as the training true value, and constructs a high-dimensional feature vector containing three types of features of wind correlation, turbulence convection and underlying surface. By intercepting the elements of 3x3 grid points around the station and combining the bilinear interpolation to fuse the spatial circulation information, the data set is divided in time sequence and stratified to avoid data leakage. Gradient boosting tree, fully connected neural network and LSTM time series network are provided as three types of optional regression models, and a gust amplitude segmented weighted loss function is designed to strengthen the fitting of disaster-causing strong wind samples. After Adam optimization, learning rate decay and early stopping strategy, the model training is completed, and the business end reuses the preprocessing process to output multi-time wind forecast for each station, which is suitable for wind forecast and warning work in multiple fields such as transportation, power, aviation and marine disaster prevention.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of meteorological numerical forecasting and artificial intelligence, and in particular to a deep learning-based method for gust forecasting. Background Technology

[0002] Gusts are instantaneous maximum wind speeds that occur within a short period. Sudden, strong gusts can threaten road traffic, overhead power lines, aircraft takeoffs and landings, maritime shipping, and outdoor construction safety, making them a key target for meteorological disaster prevention and early warning. However, gust formation is influenced by multiple factors, including near-surface average wind, boundary layer turbulence, downward transmission of convection, topographic disturbance, and differences in the land-sea surface, resulting in complex physical mechanisms and high forecasting difficulty. Currently, there are three mainstream gust forecasting methods in the industry, all of which have significant drawbacks: First, the fixed gust coefficient method, which calculates gusts by multiplying the average wind speed by a fixed empirical coefficient. This coefficient cannot be adapted to special scenarios such as mountainous areas, coastal areas, and strong convection, leading to significant forecasting errors. Second, numerical models directly output gust diagnostic values. The diagnostic results are constrained by the model's built-in boundary layer parameterization scheme, resulting in systematic overestimation or underestimation under different weather conditions, making it difficult to reflect the impact of local small-scale topography on local gusts. Third, manual correction by forecasters, which relies on individual experience, leads to inconsistent outputs from different personnel, and the process cannot be automated or standardized, making it difficult to adapt to modern intelligent meteorological operational pipelines.

[0003] Currently, ECMWF numerical weather prediction can output spatiotemporally continuous and dimensionally comprehensive environmental meteorological grid fields, while ERA5 reanalysis data provides a high-precision historical ground truth dataset containing 10-meter instantaneous gusts. Both provide data support for data-driven intelligent gust forecasting. However, existing technologies have not fully explored the application potential of these two types of data, leaving several technological gaps. Most existing conventional modeling schemes only select a single wind field element, failing to simultaneously integrate the three key features of wind, turbulent convection, and underlying surface. They also ignore the spatial information of grid circulation around the station, making it impossible to fully characterize the evolution of gusts under the combined influence of multiple factors. Ordinary regression models assign equal loss weights to samples of light winds and disastrous gusts, while the proportion of historical samples of strong gusts is too low, making it easy for strong wind values ​​to be flattened by the model and for strong gusts to be underestimated and missed. Conventional random partitioning of datasets can cause leakage of the same weather process across training and test sets, resulting in insufficient model generalization ability. Existing solutions also lack a complete closed-loop process integrating data preprocessing, model training, automated forecasting, objective verification, model iteration, and strong wind warnings, making operational implementation difficult.

[0004] Existing research on machine learning-based gust forecasting also has significant limitations. Most studies only model single-point wind speed elements, failing to incorporate key influencing factors such as convective available potential energy, vertical wind shear, and surface roughness. They lack spatial characteristics of the surrounding circulation and do not design weighted loss mechanisms for strong gusts, resulting in limited improvement in wind forecasting. Furthermore, they lack LSTM temporal networks to capture short-term abrupt changes in gust patterns, making it difficult to handle scenarios with rapid wind speed changes caused by frontal passages and strong convective outbreaks. Overall, their practicality and operational adaptability are poor. Therefore, there is an urgent need in this field for a deep learning-based gust forecasting method that integrates multi-dimensional meteorological elements, considers both spatial circulation and temporal evolution, enhances the fitting effect of disastrous winds, automates the entire process, and allows for long-term iterative optimization. Summary of the Invention

[0005] In view of this, the present invention provides a deep learning-based gust forecasting method to solve the problems of traditional gust forecasting, which uses fixed gust coefficients, model diagnostic quantities or manual correction methods, which seriously underestimate strong gusts in mountainous, coastal and severe convective scenarios, have poor forecast accuracy and weak generalization ability, and do not distinguish the importance of gust amplitude and ignore the spatial circulation and temporal evolution patterns around the site.

[0006] The technical solution proposed in this invention is: a deep learning-based gust forecasting method, specifically including: S1. Data preparation steps: Delineate the set of target meteorological stations across the entire region, collect ECMWF numerical forecast grid element fields as the model input data source, and collect 10-meter instantaneous gusts from ERA5 reanalysis grid fields as ground truth labels for model training; extract three types of core input features for each target station: wind-related, turbulent convection intensity, and underlying surface properties. S2. Sample construction steps: Training samples are constructed using "single site + single forecast time" as units. Bilinear interpolation is performed on ECMWF grid elements and ERA5 array style points to restore site elements. All features are standardized to eliminate dimensional differences. Training set, validation set and independent test set are divided into hierarchical groups according to natural year time series to avoid leakage of the same weather process across datasets. S3. Deep learning regression model training steps: Build a nonlinear regression model, input standardized site feature vectors into the model, and fit the mapping relationship between environmental elements and gust values; construct a gust amplitude weighted loss function, and configure higher loss weights for large gust samples; monitor the fitting effect with the validation set, and complete the training and save the optimal model by combining learning rate decay, regularization constraints, and early stopping strategies. S4. Automated operational gust forecasting steps: Real-time acquisition of ECMWF future forecast element field under operational time-of-care, reuse of S2 interpolation and standardized process to generate feature vectors for each site, input into the optimal model to output gust forecast values ​​for each site corresponding to the forecast time-of-care. S5. Quantitative verification steps for forecast performance: Match and compare the station gust forecast results output by the model with the actual gust conditions measured by the regional automatic weather stations station by station, calculate RMSE, MAE, wind threshold hit rate, false alarm rate, TS score and other multi-dimensional evaluation indicators, and complete the forecast performance evaluation under different terrain and weather types.

[0007] Further, the wind-related elements in step S1 include the 10-meter U-wind component, the 10-meter V-wind component, the 10-meter average wind speed, the 100-meter wind speed, the 850-hPa wind speed and direction, and the 925-hPa wind speed and direction; the turbulent convection intensity elements include the boundary layer height, convective effective potential energy (CAPE), vertical wind shear from the ground to 500-hPa, geopotential height, surface air pressure, 2-meter air temperature, and the temperature difference between 850-hPa and the ground; the underlying surface attribute elements include the site topographic height, surface roughness, land-sea identification attributes, and the variance of the topographic relief of the 3×3 grid points around the site.

[0008] Furthermore, the interpolation process in step S2 is as follows: extract all elements of the grid point where the target station is located and the surrounding 3×3 neighboring grid points. On the one hand, calculate the value of the station center element through bilinear interpolation. On the other hand, stitch the original elements of all surrounding grid points to the feature vector, and simultaneously retain the single-point environmental value and the surrounding circulation spatial information.

[0009] Furthermore, the dataset partitioning rule in step S2 is as follows: it is divided according to the natural year time sequence, with the data of the first 70% of the historical years as the training set, the data of the middle 20% of the years as the validation set, and the remaining 10% of the latest years as the test set, to ensure that the training, validation, and test datasets do not overlap in weather processes, and to prevent the model from overfitting due to leakage of time series data.

[0010] Furthermore, in step S3, the nonlinear regression model can be any of the following: 1) Use XGBoost or LightGBM gradient boosting tree model, configured with L1 and L2 regularization constraints to suppress overfitting; 2) A multi-layer fully connected deep neural network, with the network layers consisting of a fully connected layer, a batch normalized (BN) layer, a ReLU activation layer, a Dropout layer, and a single-node regression output layer at the end; 3) The LSTM temporal recurrent neural network stacks the features of multiple forecast stations into a temporal input, captures the short-term evolution of gusts and outputs multiple time-effect gust sequences; the model training modes are divided into unified single model training across the entire domain and grouped model training in the same terrain area. When training in groups, stations with similar terrain share a set of model parameters.

[0011] Furthermore, in step S3, the gust amplitude weighted loss function is constructed as follows: the absolute error or mean square error between the predicted gust and the true value of ERA5 is used as the basic loss, and the loss weight increases linearly with the amplitude of the true value of the gust, or a segmented step weight is set; when the gust amplitude exceeds the preset disaster-causing wind threshold, the weight is doubled to enhance the model's ability to fit strong gust samples.

[0012] Furthermore, the hyperparameters for neural network training in step S3 are configured as follows: the Adam optimizer is used, and the initial learning rate is set to 1×10⁻⁶. -3 When the validation set error does not decrease for 5 consecutive rounds, the learning rate is multiplied by 0.5 to achieve decay; the batch size is 64~512; dual regularization constraints of weight decay and Dropout random deactivation are set; when the validation set error does not reach the minimum value for 8 consecutive rounds, the early stopping mechanism is triggered to retain the model weight file with the lowest validation loss.

[0013] Furthermore, after outputting the gust forecast value in step S4, a sub-step for graded early warning output is added: based on the gust forecast value of the station and matching the wind force level threshold, four-level early warning labels of blue, yellow, orange and red for strong winds are automatically generated, and risk warnings for multiple industries such as transportation, power, aviation and construction sites corresponding to the station are linked in a synchronous manner. The output results can be directly connected to the automated meteorological early warning release system.

[0014] Furthermore, after completing the quantitative verification of the forecast effect in step S5, an adaptive iterative update step is added: the latest measured gust data from automatic weather stations is collected as new ground truth samples, the dataset is expanded monthly, and the training process of step S3 is reused to perform incremental fine-tuning on the original optimal model to continuously adapt to the differences in gust distribution caused by seasonal changes and interannual climate change.

[0015] The deep learning-based gust forecasting method provided by this invention has the following beneficial effects: 1. This invention integrates three dimensions of multi-source meteorological elements: wind field, turbulent convection, and underlying surface. It also incorporates the spatial circulation characteristics and temporal evolution characteristics of 3×3 grid points around the station, breaking through the limitation of traditional gust forecasts that rely solely on single-point wind speed coefficients. It is suitable for complex weather scenarios such as mountains, coastal areas, and severe convection. 2. This invention sets up a gust amplitude weighted loss function to increase the training weight of strong disaster-causing gust samples, effectively improving the defect of serious underestimation of strong gust forecasts in traditional schemes, and significantly improving the hit rate and TS score of wind-related forecast indicators. 3. This invention avoids the problem of leakage of time-series meteorological data by using a time-series hierarchical dataset partitioning strategy, ensuring that the model test results are consistent with real business scenarios and that the model has stronger generalization ability. 4. This invention is compatible with various deep learning models such as gradient boosting trees, fully connected networks, and LSTM, and supports two modes: global unified training and terrain-grouped training. It can be flexibly adapted according to the complexity of regional terrain. 5. This invention integrates business early warning output with a monthly incremental adaptive update mechanism, which can automatically generate graded gale warnings and continuously use newly added measured data to iteratively optimize the model, thereby continuously improving the stability of long-term forecasts and enabling direct implementation of automated meteorological operations.

[0016] When in use, this invention replaces the traditional fixed coefficient, model diagnosis, and manual correction methods for gust forecasting by extracting multi-source meteorological grid elements, fusing spatial circulation features, fitting weighted loss deep learning, outputting operational early warnings, and iterating monthly incrementally. This enables high-precision, automated, and continuously updated gust forecasts for meteorological stations across the entire region, significantly reducing the problems of missed or underestimated strong gusts, and providing reliable data support for meteorological disaster early warning and transportation and power safety assurance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0018] The accompanying drawings described below are only related to some embodiments of the invention and are not intended to limit the invention.

[0019] In the attached diagram: Figure 1 The overall flowchart of the site gust forecasting method of the present invention is shown; Figure 2 This diagram illustrates the site feature extraction and training label structure of the present invention. Figure 3 A schematic diagram of the site-by-site gust regression model of the present invention is shown; Figure 4 A schematic diagram of the model training process of the present invention is shown; Figure 5 A schematic diagram illustrating the verification of the site forecast results of this invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the described embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] I. Basic Implementation Conditions and Data Source Configuration: 1.1 Test Scope and Site Basic Information: This complete implementation case selected 2,000 national and regional automatic weather stations across East my country as forecast targets. The stations cover four typical underlying surfaces: plains, hills, mountains, and coastal islands. Each station has a unique pre-stored static attribute: latitude and longitude coordinates, topographic elevation, surface roughness coefficient, land-sea binary identifier (land=0, ocean=1), and topographic relief variance of the 3×3 grid points around the station.

[0022] 1.2 Complete Specifications for Three Core Data Sources: The model input data source (ECMWF forecast field) adopts the 0.25° horizontal resolution numerical forecast grid data from the European Centre for Medium-Range Weather Forecasts (ECMWF), with daily forecasts starting at 00 UTC and 12 UTC, a forecast lead time covering 0 to 72 hours, and an output interval of 3 hours. The grid field fully contains all dynamic features of wind, turbulent convection, and thermo-baric stratification, and is stored in NetCDF format.

[0023] The data source for model training ground truth (ERA5 reanalysis field) is ERA5 hourly reanalysis gridded data with a resolution of 0.25°. The 10-meter instantaneous gust of wind is extracted from the gridded field as the regression training label. The historical time span is selected from 15 consecutive years of hourly time series data from 2010 to 2024 as the baseline ground truth for model training.

[0024] The data source for real-world verification and incremental updates is hourly measured 10m gust observation data from automatic weather stations across the country. This data is used for offline performance verification of the model and construction of monthly incremental samples; it is also used to objectively compare the forecast accuracy of traditional schemes with that of this invention.

[0025] 1.3 Complete list of three main categories of input features: (1) Wind-related dynamic elements (8 items): 10-meter U-wind component, 10-meter V-wind component, 10-meter average wind speed, 100-meter height wind speed, 850 hPa wind speed, 850 hPa wind direction, 925 hPa wind speed, 925 hPa wind direction.

[0026] (2) Dynamic elements of turbulence and convection intensity (8 items): Boundary layer height, convective available potential energy (CAPE), vertical wind shear from the surface to 500 hPa, geopotential height at 500 hPa, surface-sea level pressure, near-surface air temperature at 2 m, air temperature at 850 hPa, and temperature difference between 850 hPa and 2 m (characterizing atmospheric stratification stability).

[0027] (3) Static fixing elements of the underlying surface (4 items): The data includes station topographic elevation, surface roughness coefficient, land-sea binary identifier, and variance of topographic relief at 3×3 grid points around the station. Static elements are read only once and are not updated with each forecast time; dynamic elements such as wind and convection are extracted in real time for each forecast time.

[0028] II. Detailed Implementation Guidelines for the Entire Sample Construction Process: 2.13×3 Grid Extraction and Bilinear Interpolation Operations: Centered on the latitude and longitude of a single target station, extract 9 neighboring ECMWF and ERA5 grid points in a 3×3 spatial grid. The standard bilinear interpolation algorithm was used to calculate the single-point element values ​​of the station's precise coordinates using 9 grid meteorological elements, which were then used as the station's local environmental characteristics. All the original uninterpolated features of the nine grid points are stitched together into the feature vector, simultaneously preserving the local information of a single point and the surrounding large-scale circulation spatial information, thus solving the problems of missing circulation background for single-point features and large prediction deviations in complex terrain.

[0029] 2.2 Feature Standardization Processing Details: All dynamic meteorological features are Z-Score standardized: , This is the average value of this feature over the entire historical period from 2010 to 2024. The standard deviation is the total historical period. Static underlying surface features are not standardized; raw values ​​are directly input. During the training phase, a complete set of mean and standard deviation parameters are calculated and stored. The same set of parameters is reused during the operational forecast inference phase to ensure a consistent distribution of training and inference data.

[0030] 2.3 Time Series Dataset Partitioning Rules (To Avoid Time Series Data Leakage): Strictly adhering to the natural year timeline, the samples are not randomly shuffled throughout the entire process to prevent the leakage of the same weather event across datasets. Training set: All time samples from 2010 to 2020 (the first 70% of the years); Validation set: All time samples from 2021–2022 (the middle 20% of the years); Independent test set: Samples from 2023–2024 (the latest 10% of years); training, validation, and test sets have no overlapping years and no overlapping weather processes, objectively reproducing real business forecast scenarios and preventing the model test accuracy from being artificially high.

[0031] 2.4 Sample cleaning and storage specifications: Single sample definition: unique station number + reporting start time + forecast lead time; each sample consists of a standardized feature vector + ERA5 interpolated gust ground truth label, with a total feature dimension of 184 dimensions; Abnormal sample removal rules: Station records with gust values ​​exceeding the extreme range of 0~40m / s, large-area missing grid elements, or failed interpolation are directly deleted; Missing data filling rules: When a small number of features are missing at a grid point, the average value of the same feature at the surrounding 3×3 grid points shall be used to fill the missing data; samples with missing gusts measured at the site shall not be included in training and testing. Storage format: Training samples are uniformly stored in CSV format, while model weights, mean, and standard deviation files are stored in HDF5 format.

[0032] III. Model Building, Weighted Loss, and Complete Training Strategy: This embodiment also implements the three optional regression models recorded in the handover, which can be freely switched according to the complexity of the regional terrain; the experiment selects the LSTM time series model as the main business model, and simultaneously supports two modes: unified training across the entire domain and terrain-grouped training.

[0033] 3.1 Complete Structure of the Three Types of Regression Model: Option 1: Gradient Boosting Tree Model (LightGBM / XGBoost, simple shape optimization for plains) Input: 184-dimensional standardized site feature vectors; Hyperparameter configuration: maximum tree depth 6, minimum number of samples per leaf 30, base learning rate 0.01; add L1 regularization coefficient 0.05 and L2 regularization coefficient 0.1 to suppress overfitting; Output layer: Single-node linear regression, directly outputting the predicted gust values ​​for the stations; Applicable scenarios: Plain areas with small terrain undulations and weak convection activity, with fast training speed and low inference computing power overhead.

[0034] Option 2: Multi-layer fully connected deep neural network (general basic model, adaptable to all terrains) Fixed-level cascaded structure: Input layer (184-dimensional) → Fully connected layer (256 neurons) → Batch normalized layer (BN) → ReLU activation function → Dropout random deactivation layer (deactivation rate 0.3) → Fully connected layer (128 neurons) → Batch normalized layer (BN) → ReLU activation function → Dropout random deactivation layer (deactivation rate 0.2) → Output layer (single neuron, gust regression prediction).

[0035] Option 3: LSTM Long Short-Term Memory Temporal Network (preferred for scenarios with strong convection and sudden gusts) Temporal input construction: Stack the 184-dimensional features of 8 consecutive forecast times from the same site into a temporal tensor; Network structure: LSTM hidden layer (128 units) + two fully connected layers; Output: A continuous gust sequence with 3-hour intervals from 0 to 72 hours in the future, which can capture the short-term fluctuations of gusts, the passage of fronts, and the sudden changes in wind speed caused by convective outbreaks.

[0036] 3.2 Two model training modes: Mode 1: Unified single model training across the entire region, with 2000 sites sharing a single set of model weights, suitable for scenarios with minimal regional terrain differences; Mode 2: Independent training by terrain groups. Subgroups are divided into plains, hills, mountains, and coastal islands, and each group of stations trains its own dedicated model. In this embodiment, the group training mode is adopted, and the RMSE of gust forecasts at mountain and coastal stations is reduced by more than 22%.

[0037] 3.3 Piecewise weighted loss function for gust amplitude: The base loss is calculated using mean squared error (MSE). To address the shortcomings of conventional regression in smoothing out extreme wind speeds and generally underestimating severe, destructive gusts, a segmented, stepped weighting strategy is employed. True value gust < 6 m / s: weight w = 1; 6 m / s ≤ true gust < 10.8 m / s (critical threshold for gale force 6): weight w = 2; True value gust ≥ 10.8 m / s (disastrous wind): weight w = 5; single sample loss calculation formula: The total batch loss is the sum of the weighted losses of all samples. During backpropagation, the gradient of strong gust samples is amplified first, so that the model focuses on learning the characteristics of strong winds.

[0038] 3.4 Unified training hyperparameters, learning rate decay, and early stopping mechanism: Optimizer: Adam adaptive optimizer, with a fixed initial learning rate of 1×10⁻⁶. -3 ; Batchsize: Set to 256 uniformly. If the total number of samples is insufficient, it can be adaptively adjusted in the range of 64 to 512. Double regularization constraint: Neural network configuration weight decay (coefficient 1×10) -4 +Dropout random deactivation; tree model configured with L1 / L2 regularization terms; Learning rate decay rule: If the verification set-weighted loss does not decrease for 5 consecutive training rounds, the learning rate is decayed by multiplying it by 0.5. Early stopping rule: If the validation set loss does not refresh the historical minimum for 8 consecutive rounds, training is terminated immediately, the optimal model weight file with the lowest validation loss is automatically saved, and the overfitting weights in the later stages are discarded.

[0039] 3.5 Training Hardware Environment: Offline model training is accelerated using a single RTX 4090 graphics card; online business inference only requires a regular server CPU to complete batch site real-time forecasting.

[0040] IV. Automated business forecasting process + Level 4 gale warning output: 4.1 Complete steps of real-time inference: The latest starting grid element field of ECMWF is retrieved at 00:00 and 12:00 every day. For all 2000 target sites, the 3×3 grid interpolation, feature splicing, and Z-Score standardization processes in the sample construction phase were repeated, and the mean and standard deviation parameters stored in the training phase were strictly reused. The standardized feature vectors are input into the pre-trained optimal grouping model, and the gust prediction values ​​for each site are output for 24 forecast periods, ranging from 0 to 72 hours with 3-hour intervals. Store gust forecast results for each site and time period, and generate standardized JSON business messages.

[0041] 4.2 Automatic generation of gale level warnings (expanded implementation functions): Based on national meteorological wind force standards, a four-level warning threshold is set, and the warning level is automatically marked according to the forecast gusts and linked to industry risk warnings: Blue alert: Gusts of 10.8~13.8 m / s (level 6) indicate risks to mountain roads and small-scale outdoor construction. Yellow alert: Gusts of 13.9~17.1 m / s (level 7), indicating risks to ports, terminals, and low-altitude navigation; Orange alert: Gusts of 17.2~20.7 m / s (level 8) indicate risks to overhead power lines and tower crane operations; Red Alert: Gusts ≥ 20.8 m / s (level 9 and above), indicating risks to airports, maritime shipping, and large structures; the output message can be directly connected to the provincial meteorological automated early warning release system without the need for manual secondary editing.

[0042] V. Complete Implementation Rules for Quantitative Verification of Forecast Results: Data matching rules: The model output stations and time-leading gust forecast values ​​are accurately matched one-to-one with the actual measured gust conditions at the same stations and times by automatic weather stations; Batch calculation of basic error indicators: Mean Absolute Error (MAE), Root Mean Square Error (RMSE); Special evaluation indicators for gale (with gale force 6 as the judgment threshold): hit rate, false alarm rate, missed alarm rate, and TS score; Stratified group statistics: All indicators are statistically analyzed by grouping the four types of weather processes into plains / hills / mountains / coastal terrain and clear sky / light precipitation / strong convection / frontal winds. Monthly inspection reports are generated to quantitatively compare the error differences between this invention and the traditional fixed gust coefficient method and the original gust diagnosis of the model. The test results show that the TS score of disaster-causing winds of level 6 and above is improved by 35%, and the problem of underestimation of strong gusts is significantly improved.

[0043] VI. Monthly Model Adaptive Incremental Iterative Update Mechanism: On the 1st of each month, the system automatically performs the entire incremental update process to adapt to the differences in gust distribution caused by seasonal and interannual climate changes. Collect real-time gust data from all automatic weather stations over the past month, combine them with historical gridded data from ECMWF and ERA5 during the same period, and generate new training samples according to the sample construction rules. New samples are incorporated into the existing training set, while the validation set and independent test set remain unchanged, and the original evaluation benchmark is not altered. Load the optimal model weights from the previous version of the business as pre-training parameters, and reuse the weighted loss and the full set of training hyperparameters to carry out incremental fine-tuning training. After fine-tuning, the new and old models were compared on an independent test set using a full set of evaluation metrics. The online business inference model was replaced only when the new model’s wind TS score improved by ≥3% and the RMSE decreased by ≥3%. If the accuracy did not improve or even deteriorated, the original model was retained to avoid abnormal observation data from damaging the model’s forecast performance.

[0044] This scheme abandons the three deficient schemes of traditional fixed gust coefficient, numerical model native gust diagnostic quantity, and manual subjective correction. It integrates three dimensions of multi-source meteorological elements: wind field, turbulent convection, and underlying surface, and introduces 3×3 grid point circulation spatial information around the station. Combined with the large gust weighted loss function, it significantly improves the accuracy of gust forecast in mountainous, coastal, and severe convective scenarios. The entire process operates in a fully automated closed loop, requiring no manual intervention from forecasters, and supports monthly incremental iterations to continuously optimize the model; the output products are equipped with graded gale warnings and industry risk alerts, which can be directly applied to meteorological disaster prevention scenarios such as transportation, power, marine, and construction. The multi-model compatible design allows for flexible switching between gradient boosting trees, fully connected networks, and LSTM time series models based on regional terrain complexity and server computing power, adapting to the business needs of different regions and forecast lead times.

[0045] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based gust forecasting method, characterized in that, Includes the following steps: S1. Data preparation: Determine the set of target meteorological stations across the entire region, collect ECMWF numerical forecast grid element fields as the model input data source, and collect 10-meter instantaneous gusts from the ERA5 reanalysis grid field as the model training ground value label data source. For each target meteorological station, three types of core input features are extracted, including wind-related elements, turbulent convection intensity elements, and underlying surface attribute elements. S2. Sample Construction: Independent training samples are constructed using a single site and a single forecast time. Bilinear interpolation is used to restore the site elements for ECMWF grid elements and ERA5 matrix grid elements respectively. Standardize all features to eliminate differences in units, and divide the training set, validation set, and independent test set into layers according to time sequence to avoid leakage of the same weather process across datasets. S3. Deep learning regression model training: Build a nonlinear regression model and input the standardized site feature vector into the model to fit the mapping relationship between environmental factors and gust values; construct a gust amplitude weighted loss function, configure higher loss weights for large gust samples, use a validation set to monitor the model fitting effect in real time, and train and save the optimal model with learning rate decay, regularization constraints, and early stop strategy. S4. Automated Operational Gust Forecast: Real-time acquisition of ECMWF future forecast element field during operation, reuse of S2 interpolation and standardization process to generate feature vectors for each target site, input into the trained optimal model, and output gust prediction values ​​for each site with corresponding forecast lead time. S5. Quantitative verification of forecast performance: The gust forecast results of the model output stations are matched and compared with the actual gust conditions of the regional automatic weather stations station by station. RMSE, MAE, wind threshold hit rate, false alarm rate, TS score and other multi-dimensional indicators are calculated to complete the forecast performance evaluation under different terrain and weather types.

2. The deep learning-based gust forecasting method according to claim 1, characterized in that, The wind-related elements mentioned in step S1 include: 10-meter U-wind component, 10-meter V-wind component, 10-meter average wind speed, 100-meter wind speed, 850 hPa wind speed and direction, and 925 hPa wind speed and direction; The turbulent convection intensity elements include: boundary layer height, convective effective potential energy (CAPE), vertical wind shear from the ground to 500 hPa, geopotential height, ground air pressure, air temperature at 2 m, and temperature difference between 850 hPa and the ground. The underlying surface attributes include: site topographic height, surface roughness, land-sea identification attributes, and variance of topographic relief at 3×3 grid points around the site.

3. The deep learning-based gust forecasting method according to claim 1, characterized in that, The interpolation process in step S2 is as follows: extract all features of the grid point where the target station is located and all 3×3 neighboring grid points. On the one hand, calculate the value of the station center feature through bilinear interpolation. On the other hand, stitch together the original features of all surrounding grid points into the feature vector, and simultaneously retain the single-point environmental value and the surrounding circulation spatial information.

4. The deep learning-based gust forecasting method according to claim 1, characterized in that, The dataset partitioning rule in step S2 is as follows: it is divided according to the natural year time sequence, with the first 70% of the historical data as the training set, the middle 20% of the data as the validation set, and the remaining 10% of the latest data as the test set. This ensures that the training, validation, and test datasets do not overlap with weather processes and prevents the model from overfitting due to time series data leakage.

5. The deep learning-based gust forecasting method according to claim 1, characterized in that, The nonlinear regression model mentioned in step S3 is one of the following three: Gradient boosting tree model, XGBoost or LightGBM, is selected, and L1 and L2 regularization constraints are introduced to suppress overfitting; A multi-layer fully connected deep neural network, with the network layer structure consisting of a fully connected layer, a batch normalized (BN) layer, a ReLU activation layer, a Dropout layer, and a single-node regression output layer at the end; The LSTM temporal recurrent neural network takes the features of multiple consecutive forecast times as a time series input, captures the short-term evolution pattern of gusts, and outputs future multi-time-effect gust sequences. The model training modes are divided into two types: unified single model training across the entire domain and grouped model training in the same terrain area. During group training, stations with similar terrain share the same set of model parameters.

6. The deep learning-based gust forecasting method according to claim 1, characterized in that, The gust amplitude weighted loss function described in step S3 is constructed as follows: the absolute error / mean square error between the predicted gust and the true value of ERA5 is used as the base loss, and the weight increases linearly with the amplitude of the true value of the gust, or a segmented step weight is set. When the gust amplitude exceeds the preset disaster-causing wind threshold, the weight is doubled to enhance the model's ability to fit strong gust samples.

7. The deep learning-based gust forecasting method according to claim 1, characterized in that, Step S3: Configure the neural network training hyperparameters as follows: Use the Adam optimizer, and set the initial learning rate to 1×10⁻⁶. -3 The learning rate is multiplied by 0.5 to decay when the validation set error does not decrease for 5 consecutive rounds; the batch size is set to 64~512; dual regularization of weight decay and Dropout random deactivation is set; early stopping is triggered when the validation set error does not reach the minimum value for 8 consecutive rounds, and the model weight file with the lowest validation loss is retained.

8. The deep learning-based gust forecasting method according to claim 1, characterized in that, After the business forecast is output in step S4, a sub-step for graded early warning output is added: based on the gust forecast value of the station, the wind force level threshold is matched, and four levels of warning labels for strong winds (blue, yellow, orange, and red) are automatically output. The risk warnings of the transportation, power, aviation, and construction industries to which the station belongs are simultaneously associated, and the system is directly connected to the automated meteorological early warning release system.

9. The deep learning-based gust forecasting method according to claim 1, characterized in that, After the quantitative verification in step S5 is completed, a new adaptive iterative update step is added: the latest measured gust data from the automatic weather station is used as the new ground truth sample, the dataset is expanded monthly, the training process in S3 is reused to make incremental fine-tuning of the original optimal model, and the gust distribution differences caused by seasonal and interannual climate changes are continuously adapted.

10. The deep learning-based gust forecasting method according to any one of claims 1-9, characterized in that, This method abandons the traditional fixed gust coefficient method, the original gust diagnostic quantity of numerical models, and the manual subjective correction. By integrating three dimensions of multi-source grid point environmental elements, namely wind, turbulent convection, and underlying surface, and combining gust weighted loss deep learning regression, it simultaneously utilizes the spatial information and temporal evolution characteristics of the surrounding grid point circulation to achieve automated, stable, and reproducible high-precision forecasting of gusts at stations under complex mountainous, coastal, and severe convective weather conditions, significantly reducing the underestimation bias of large gust forecasts.

Images