Area rainfall forecasting method based on Beidou GNSS water vapor

Abstract

The invention discloses an area rainfall forecasting method based on Beidou GNSS water vapor, and the method specifically comprises the following steps: firstly, collecting original observation data and ephemeris data through GNSS equipment; surface water vapor distribution is obtained; and finally, a space-time prediction model of the area rainfall is constructed and trained, and for rainfall prediction, in order to suppress the predicted rainfall output by mistake of the model when no rainfall exists and improve the capability of the model to accurately output rainfall, a fusion prediction layer based on classification prediction is designed in a decoder part. According to the areal rainfall forecasting method based on Beidou GNSS water vapor, the areal rainfall condition in a certain area can be monitored and predicted in a high-precision mode, and the areal rainfall forecasting method has the advantages of being all-weather, free of influences of weather and environment, high in data reliability and low in equipment cost; therefore, the technology has wide application prospects in small and medium-sized drainage basins with limited resources and economically underdeveloped areas.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a surface rainfall forecasting method based on Beidou GNSS water vapor. Background Art

[0002] Surface rainfall refers to the average rainfall in a specific area or river basin within a certain period of time. It is one of the most important physical quantities in hydrological and meteorological operations and scientific research. It is also an important basis for flood and drought prevention, water resources assessment and estimation, engineering design, mountain torrent and geological disaster prevention and reduction in the river basin.

[0003] The traditional method of calculating surface rainfall mainly relies on the interpolation method to interpolate the point rainfall measured by the ground rain gauge station to obtain the surface rainfall of the region. At present, the meteorological and hydrological departments widely use the arithmetic mean method, Thiessen polygon method and Kriging method. These methods are easy to calculate, but they also have obvious disadvantages: they do not take into account the impact of uneven precipitation distribution and are also affected by the density of rain gauge stations. Sparse stations will lead to huge measurement errors; the surface rainfall measured by satellites and radars has the problem of poor temporal or spatial resolution, and is expensive and has high maintenance costs. It is not suitable for surface rainfall monitoring in small and medium-sized basins or economically underdeveloped areas.

[0004] With the rapid development of the new generation of global navigation satellite systems, GNSS-based atmospheric water vapor telemetry technology has the advantages of high precision, all-weather operation, long-term stability, and high temporal and spatial resolution, providing important data support for atmospheric water vapor monitoring. Among the various meteorological elements that affect weather changes, atmospheric water vapor is one of the most important parameters. Effective monitoring and rational use of it are important prerequisites and keys to achieving accurate forecasts of various types of disastrous weather events. However, traditional rainfall forecasts based on GNSS water vapor mainly use the accumulated water vapor above the site detected by the GNSS site to predict whether it will rain or the amount of rainfall, and do not make sufficient use of the regional water vapor distribution information within the detection range of the GNSS site. Summary of the invention

[0005] The purpose of the present invention is to provide a surface rainfall forecasting method based on Beidou GNSS water vapor, which is a new type of surface rainfall forecasting method, which can realize high-precision monitoring and prediction of surface rainfall conditions in a certain area, has all-weather, is not affected by weather and environment, has high data reliability, low equipment cost, and can also realize overall rainfall monitoring in an area, avoiding the limitations of single-point measurement of traditional rainfall monitoring instruments. Compared with traditional methods, it can achieve more accurate and real-time predictions, and is particularly suitable for application scenarios with high requirements for spatiotemporal resolution. In addition, the low cost and easy deployment characteristics make this technology have broad application prospects in small and medium-sized river basins with limited resources and economically underdeveloped areas. In the future, with the further improvement of the Beidou system and the development of deep learning technology, the prediction accuracy and application scope of this patented technology will be further expanded.

[0006] To achieve the above object, the present invention provides a surface rainfall forecasting method based on Beidou GNSS water vapor, which specifically includes the following steps:

[0007] S1, data collection;

[0008] S2, surface water vapor distribution acquisition;

[0009] S3. Build and train a spatiotemporal rainfall prediction model.

[0010] Preferably, in step S1, original observation data and ephemeris data are collected by GNSS equipment, and the oblique tropospheric delay and the corresponding satellite absolute position in the signal path of each satellite above the site are calculated by a precise single-point positioning algorithm.

[0011] Preferably, in step S2, based on the oblique tropospheric delays of different satellite paths and the corresponding satellite absolute positions calculated in step S1, a tomography algorithm is used to perform three-dimensional tomography on the data collected by multiple GNSS stations.

[0012] Preferably, the specific process of using a tomography algorithm to perform three-dimensional tomography on the data collected by multiple GNSS stations is as follows:

[0013] S21, dividing the regional three-dimensional grid into voxel blocks with exponentially increasing intervals in the vertical direction, and calculating the intercept matrix;

[0014] S22, setting horizontal and vertical constraints through the boundary constraint and initialization module, and obtaining a priori background field according to the numerical model;

[0015] S23. Through the water vapor density inversion module, the observation equation and constraint equation are established to construct the tomography model. The non-iterative singular value decomposition (SVD) is used to solve the tomography equation to obtain the three-dimensional tomography wet refractive index grid data. The wet refractive index of the Z axis is integrated and then calculated to obtain the two-dimensional surface water vapor grid data. The steps are repeated to obtain the final three-dimensional water vapor spatiotemporal distribution sequence.

[0016] Preferably, for any two-dimensional grid point in step S23, the total water vapor value PWV is calculated by the following formula for the wet refractive index integral of the Z-axis height axis:

[0017]

[0018] PWV (x,y) =Π·ZWD (x,y) ;

[0019] Among them, ZWD (x,y) is the total wet delay of the two-dimensional surface grid at row x and column y, H i is the height of the i-th point on the Z-axis height axis, H i+1 is the height of the i+1th point on the Z-axis height axis of this point, is the wet refractive index of the ith point on the height axis, is the wet refractive index of the i+1th point on the height axis, Π is the coefficient calculated based on temperature and air pressure, PWV (x,y) is the total water vapor value of the two-dimensional surface grid at the xth row and yth column.

[0020] Preferably, in step S3, the specific process of constructing and training the surface rainfall spatiotemporal prediction model is as follows:

[0021] S31. Training and testing data set collection:

[0022] Collect historical GNSS data and rainfall data, temperature and air pressure meteorological data, calculate the historical spatiotemporal data of surface water vapor distribution based on the GNSS data according to the steps of S2; divide them into model training and test data sets in a ratio of 3:1;

[0023] S32. Model building:

[0024] A Transformer-based surface rainfall spatiotemporal prediction model is constructed, wherein the prediction model includes a time series embedding layer part and a time series encoder-decoder part, wherein the time series embedding layer part projects the input matrix into another dimensional space, and then uses a Transformer-based encoder-decoder to first convert the input time series into a feature map of fixed dimensions by the encoder, and then decode the feature map to generate a future time series by the decoder, wherein the decoder includes a rainfall generation module and a level prediction module, and the final output of the surface rainfall spatiotemporal prediction model is the output result of the rainfall generation module multiplied by the level prediction module;

[0025] S33. Training model:

[0026] The constructed historical data set is used to pre-train the spatial-temporal rainfall prediction model, and the network structure and hyperparameters are adjusted by traversal, and the loss function is set to the improved root mean square error function;

[0027] The improved root mean square error loss function takes into account the rainfall events that are of main concern in surface rainfall prediction, and the greater the rainfall, the greater the weight of the loss function. The calculation formula is as follows:

[0028]

[0029] Among them, L is the loss function, n is the number of samples, y i is the i-th real rainfall value, is the corresponding predicted rainfall value, α is the weight hyperparameter;

[0030] S34. Fine-tune the model based on measured data:

[0031] According to the general and common meteorological characteristics, correlations and laws of surface water vapor and surface rainfall learned by the pre-trained model obtained by pre-training in step S33, in the fine-tuning stage, the pre-trained model is fine-tuned using the local meteorological data of the target area or station. During the fine-tuning process, all other layer weights are frozen, and the output layer weights are released. The neural network is trained using the enhanced data, and only the output layer weights are updated for fine-tuning.

[0032] Therefore, the present invention adopts the above-mentioned surface rainfall forecasting method based on Beidou GNSS water vapor, and the beneficial effects are as follows:

[0033] (1) The present invention can realize high-precision monitoring and prediction of surface rainfall conditions in a certain area, is all-weather and is not affected by weather and environment, has high data reliability and low equipment cost.

[0034] (2) The present invention can also monitor the rainfall in an area as a whole, avoiding the limitations of single-point measurement of traditional rainfall monitoring instruments. Compared with traditional methods, it can achieve more accurate and real-time predictions, and is particularly suitable for application scenarios with high requirements for temporal and spatial resolution.

[0035] (3) The low cost and easy deployment of the present invention make this technology have broad application prospects in small and medium-sized river basins with limited resources and economically underdeveloped areas. In the future, with the further improvement of the Beidou system and the development of deep learning technology, the prediction accuracy and application scope of the present invention will be further expanded.

[0036] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the overall process of an embodiment of a method for surface rainfall forecasting based on Beidou GNSS water vapor of the present invention;

[0038] Figure 2 It is a diagram of the spatial-temporal prediction model of surface rainfall according to an embodiment of a surface rainfall forecasting method based on Beidou GNSS water vapor of the present invention;

[0039] Figure 3 It is a schematic diagram of the equipment location in Hong Kong, my country and the Thiessen polygon result of the reference rainfall station in an embodiment of the surface rainfall forecasting method based on Beidou GNSS water vapor of the present invention;

[0040] Figure 4 It is an overview diagram of the hourly comparison of the predicted surface rainfall in Hong Kong GNN network in my country according to an embodiment of the surface rainfall forecasting method based on Beidou GNSS water vapor of the present invention;

[0041] Figure 5 It is a comparison diagram of the GNSS predicted surface rainfall at different times and the interpolation distribution of the reference rainfall station in an embodiment of the surface rainfall forecasting method based on Beidou GNSS water vapor of the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0043] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0044] like Figure 1 As shown, a surface rainfall forecasting method based on Beidou GNSS water vapor specifically includes the following steps:

[0045] S1, data collection;

[0046] The original observation data and ephemeris data are collected by GNSS equipment, and the oblique tropospheric delay and the corresponding satellite absolute position in the signal path of each satellite above the site are calculated using the precise single-point positioning algorithm.

[0047] S2, surface water vapor distribution acquisition;

[0048] According to the oblique tropospheric delays of different satellite paths and the corresponding satellite absolute positions obtained by step S1, a tomography algorithm is used to perform three-dimensional tomography on the data collected by multiple GNSS stations.

[0049] The specific process is as follows:

[0050] S21, dividing the regional three-dimensional grid into voxel blocks with exponentially increasing intervals in the vertical direction, and calculating the intercept matrix;

[0051] S22, setting horizontal and vertical constraints through the boundary constraint and initialization module, and obtaining a priori background field according to the numerical model;

[0052] S23. Through the water vapor density inversion module, the observation equation and constraint equation are established to construct the tomography model. The non-iterative singular value decomposition (SVD) is used to solve the tomography equation to obtain the three-dimensional tomography wet refractive index grid data. The wet refractive index of the Z axis is integrated and then calculated to obtain the two-dimensional surface water vapor grid data. The steps are repeated to obtain the final three-dimensional water vapor spatiotemporal distribution sequence.

[0053] Among them, for any two-dimensional grid point, the total water vapor value PWV (that is, the total water vapor result after integrating the three-dimensional grid on the z-axis) is calculated by the following formula for the wet refractive index integral of the z-axis, that is, the height axis:

[0054]

[0055] PWV (x,y) =Π·ZWD (x,y) ;

[0056] Among them, ZWD (x,y) is the total wet delay of the two-dimensional surface grid at row x and column y, H i is the height of the i-th point on the Z axis (height axis), H i+1 is the height of the i+1th point on the Z axis (height axis) of this point, is the wet refractive index of the ith point on the Z axis (height axis), is the wet refractive index of the i+1th point on the Z axis (height axis), Π is the coefficient calculated based on temperature and air pressure, PWV (x,y) is the total water vapor value of the two-dimensional surface grid at the xth row and yth column.

[0057] S3. Build and train the spatial and temporal prediction model for surface rainfall:

[0058] According to the three-dimensional spatiotemporal distribution sequence of water vapor in the region, it is input into the spatiotemporal prediction model of surface rainfall to predict the spatiotemporal distribution sequence of surface rainfall in the current region in the next 1-3 hours.

[0059] The specific process of constructing and training the spatial-temporal prediction model of surface rainfall is as follows:

[0060] S31. Training and testing data set collection:

[0061] Collect historical GNSS data and rainfall data, temperature, air pressure and other meteorological data, and calculate the historical spatiotemporal data of surface water vapor distribution based on the GNSS data according to step S2.

[0062] like Figure 3 As shown, this embodiment conducts experiments in Hong Kong, China, and uses 8 GNSS stations to train and test the Beidou GNSS surface rainfall forecast model. 22 reference rainfall stations are used as reference values ​​to verify the accuracy of the Beidou GNSS surface rainfall forecast model. The temperature and air pressure data measured by the meteorological sensors at the Hong Kong GNSS station data and the stations are collected and divided into a 3:1 ratio as the model training and test data sets; GNSS data can be downloaded from the official website of the International GNSS Service (IGS), and ERA5 reanalysis data can be downloaded from the official website of the European Center for Medium-Term Weather Forecasts ECMWF.

[0063] S32. Model building:

[0064] like Figure 2 As shown in the figure, a Transformer-based surface rainfall spatiotemporal prediction model is constructed. The model input is the historical 6-hour water vapor spatiotemporal distribution sequence, and the output is the surface rainfall distribution sequence for the next 1-3 hours. Among them, the base model Transformer used performs well in processing spatiotemporal sequence data and can effectively capture the dynamic changes in time and space dimensions.

[0065] The prediction model consists of a time series embedding layer and a time series encoder-decoder part. The time series embedding layer projects the input matrix into another dimensional space, and then uses the Transformer-based encoder-decoder to first convert the input time series into a feature map of fixed dimensions by the encoder, and then decode the feature map through the decoder to generate the future time series.

[0066] Among them, the learning framework diagram of the Transformer-based surface rainfall prediction model is as follows:

[0067] The specific process of building a prediction model is as follows:

[0068] Step 1: Take the historical water vapor spatiotemporal sequence X of continuous α steps as input and output a series of future values ​​Y of β time steps, where and

[0069] Step 2: In the first stage, the time series embedding layer is designed, which includes the scalar projection module, position coding and time coding. First, the water vapor spatiotemporal sequence X needs to be projected to the required dimension in the scalar projection part, and then the position information is added. Finally, the time-related coding is added to obtain the data M input to the model. emb ;

[0070] Step 3. In the second stage, a Transformer-based time series encoder-decoder is designed. The time series encoder-decoder consists of two parts: the encoder and the decoder. Therefore, the model effectively captures neighbor dependencies and long-distance dependencies through convolutional attention modules and multi-head attention mechanisms.

[0071] The encoder consists of a convolutional attention module and a downsampling module, and the decoder consists of a convolutional attention module, a multi-head attention mechanism, and a fully connected layer. The input of the time series encoder-decoder is M emb , the output is The model embeds the time series into the output of the layer M emb As input to the time series encoder-decoder.

[0072] In the encoder, a convolutional attention module is used to pay attention to the encoder input M emb Calculate and further obtain advanced features while capturing neighbor dependencies to obtain M cbam ; Then M cbam Input to the downsampling module, through the maximum pooling layer, reduce parameters, filter features, and obtain the encoder output M encoder ;

[0073] The input of the decoder is the output M of the encoder encoder , the output is The model first uses a convolutional attention module in the decoder to further capture neighbor dependencies and prevent the receptive field from growing exponentially with the number of layers, which would lead to attention to unimportant features. Then use the multi-head attention mechanism to capture long-distance dependencies, and finally use the fully connected layer to project the dimension to the original dimension of the time series to get the output

[0074] In addition, in the time series encoder-decoder, the model uses a convolutional attention module and a maximum pooling layer of a downsampling module to reduce the GPU memory usage during training.

[0075] Among them, for rainfall prediction, in order to suppress the model's erroneous output of predicted rainfall when there is no rainfall and improve the model's ability to accurately output rainfall, a fusion prediction layer based on classification forecast is designed in the decoder part, including a rainfall generation module and a level prediction module. The rainfall generation module consists of a multi-layer perceptron with 1024 hidden layers, which outputs grid rainfall prediction results; the level prediction module consists of another multi-layer perceptron with 512 hidden layers and a Softmax layer, which predicts whether there will be rainfall in the next 3 hours and classifies it into different rainfall levels: no rain (0), rain (1). The two modules are trained separately, with root mean square error and cross entropy as loss functions respectively. After training, the final output of the surface rainfall spatiotemporal prediction model is the output result of the rainfall generation module multiplied by the model level prediction module.

[0076] S33. Training model:

[0077] The constructed historical data set (a data set consisting of water vapor data and rainfall data inverted from historical GNSS, and temperature and pressure meteorological data) is used to pre-train the spatial-temporal rainfall prediction model. The network structure and hyperparameters are adjusted by traversal, and the loss function of the rainfall generation module is set to the improved root mean square error function, and the loss function of the level prediction module is set to the cross entropy loss function. After 1000 rounds of training, the model with the smallest loss function is obtained, which is the optimal model.

[0078] Among them, the improved root mean square error loss function considers the rainfall events that are the main focus in surface rainfall prediction, and the greater the rainfall, the greater the weight of the loss function. The calculation formula is as follows:

[0079]

[0080] Among them, L is the loss function, n is the number of samples, y i is the i-th real rainfall value, is the corresponding predicted rainfall value, α is the weight hyperparameter;

[0081] S34. Fine-tune the model based on measured data:

[0082] According to the general and common meteorological characteristics, correlations and laws of surface water vapor and surface rainfall learned by the pre-trained model obtained by pre-training in step S33, in the fine-tuning stage, the pre-trained model is fine-tuned using the local meteorological data of the target area or station. During the fine-tuning process, all other layer weights are frozen, and the output layer weights are released. The neural network is trained using the enhanced data, and only the output layer weights are updated for fine-tuning.

[0083] like Figure 4The figure shows the comparison between the average surface rainfall forecast results of the BeiDou GNSS inverted water vapor surface rainfall forecast model in Hong Kong, my country and the average surface rainfall calculated by Thiessen polygons at the reference rainfall station. It can be seen that the 1-hour surface rainfall forecast results of the BeiDou GNSS surface rainfall forecast model at the 8 stations in Hong Kong, my country are excellent, basically consistent with the reference surface rainfall, and can effectively capture the entire rainfall process.

[0084] Through the above detailed implementation steps, the spatial-temporal prediction model of surface rainfall can make full use of Beidou GNSS water vapor data, as well as other meteorological data such as temperature and air pressure, to achieve high-precision prediction of surface rainfall in a specific area within the next 1-3 hours. Figure 5 As shown in Figure 1, Figure (a) is a comparison of the spatial distribution of rainfall predicted by the GNSS forecast model one hour later and the Kriging interpolation of the reference rainfall station, and Figure (b) is a comparison of the spatial distribution of rainfall predicted by the GNSS forecast model two hours later and the Kriging interpolation of the reference rainfall station. This method not only improves the accuracy and efficiency of rainfall forecasting, but also effectively reduces related costs, and has important application value in water resources management, disaster prevention, meteorological research and other fields.

[0085] Therefore, the present invention adopts the above-mentioned surface rainfall forecasting method based on Beidou GNSS water vapor, which can realize the overall rainfall monitoring in a region, avoiding the limitations of single-point measurement of traditional rainfall monitoring instruments. Compared with traditional methods, it can achieve more accurate and real-time predictions, with high data reliability, and is particularly suitable for application scenarios with high requirements for spatiotemporal resolution.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A surface rainfall forecasting method based on Beidou GNSS water vapor, characterized in that: The specific steps include: S1, data collection; S2, surface water vapor distribution acquisition; S3. Build and train a spatiotemporal rainfall prediction model.

2. The method for surface rainfall forecasting based on Beidou GNSS water vapor according to claim 1 is characterized in that: In step S1, the original observation data and ephemeris data are collected by GNSS equipment, and the oblique tropospheric delay and the corresponding satellite absolute position in the signal path of each satellite above the site are calculated by the precise single-point positioning algorithm.

3. The method for surface rainfall forecasting based on Beidou GNSS water vapor according to claim 2 is characterized in that: In step S2, according to the oblique tropospheric delays of different satellite paths and the corresponding satellite absolute positions calculated in step S1, a tomography algorithm is used to perform three-dimensional tomography on the data collected by multiple GNSS stations.

4. The method for surface rainfall forecasting based on Beidou GNSS water vapor according to claim 3 is characterized in that: The specific process of using the tomography algorithm to perform three-dimensional tomography on the data collected by multiple GNSS stations is as follows: S21, dividing the regional three-dimensional grid into voxel blocks with exponentially increasing intervals in the vertical direction, and calculating the intercept matrix; S22, setting horizontal and vertical constraints through the boundary constraint and initialization module, and obtaining a priori background field according to the numerical model; S23. Through the water vapor density inversion module, the observation equation and constraint equation are established to construct the tomography model. The non-iterative singular value decomposition (SVD) is used to solve the tomography equation to obtain the three-dimensional tomography wet refractive index grid data. The wet refractive index of the Z axis is integrated and then calculated to obtain the two-dimensional surface water vapor grid data. The steps are repeated to obtain the final three-dimensional water vapor spatiotemporal distribution sequence.

5. The method for surface rainfall forecasting based on Beidou GNSS water vapor according to claim 4 is characterized in that: For any two-dimensional grid point in step S23, the total water vapor value PWV is calculated by the following formula for the wet refractive index integral of the Z-axis height axis: PWV (x,y) =Π·ZWD (x,y) ; Among them, ZWD (x,y) is the total wet delay of the two-dimensional surface grid at row x and column y, H i is the height of the i-th point on the Z-axis height axis, H i+1 is the height of the i+1th point on the Z-axis height axis of this point, is the wet refractive index of the ith point on the height axis, is the wet refractive index of the i+1th point on the height axis, Π is the coefficient calculated based on temperature and air pressure, PWV (x,y) is the total water vapor value of the two-dimensional surface grid at the xth row and yth column.

6. The method for surface rainfall forecasting based on Beidou GNSS water vapor according to claim 5 is characterized in that: In step S3, the specific process of constructing and training the spatial-temporal prediction model of surface rainfall is as follows: S31. Training and testing data set collection: Collect historical GNSS data and rainfall data, temperature and air pressure meteorological data, calculate the historical spatiotemporal data of surface water vapor distribution based on the GNSS data according to the steps of S2; divide them into model training and test data sets in a ratio of 3:1; S32. Model building: A Transformer-based surface rainfall spatiotemporal prediction model is constructed, wherein the prediction model includes a time series embedding layer part and a time series encoder-decoder part, wherein the time series embedding layer part projects the input matrix into another dimensional space, and then uses a Transformer-based encoder-decoder to first convert the input time series into a feature map of fixed dimensions by the encoder, and then decode the feature map to generate a future time series by the decoder, wherein the decoder includes a rainfall generation module and a level prediction module, and the final output of the surface rainfall spatiotemporal prediction model is the output result of the rainfall generation module multiplied by the level prediction module; S33. Training model: The constructed historical data set is used to pre-train the spatial-temporal rainfall prediction model, and the network structure and hyperparameters are adjusted by traversal, and the loss function is set to the improved root mean square error function; The improved root mean square error loss function takes into account the rainfall events that are of main concern in surface rainfall prediction, and the greater the rainfall, the greater the weight of the loss function. The calculation formula is as follows: Among them, L is the loss function, n is the number of samples, y i is the i-th real rainfall value, is the corresponding predicted rainfall value, α is the weight hyperparameter; S34. Fine-tune the model based on measured data: According to the general and common meteorological characteristics, correlations and laws of surface water vapor and surface rainfall learned by the pre-trained model obtained by pre-training in step S33, in the fine-tuning stage, the pre-trained model is fine-tuned using the local meteorological data of the target area or station. During the fine-tuning process, all other layer weights are frozen, and the output layer weights are released. The neural network is trained using the enhanced data, and only the output layer weights are updated for fine-tuning.

Images