An atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model

Abstract

The present invention provides an atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model, including: Step 1: Obtain GNSS atmospheric water vapor data and obtain ground meteorological data; Step 2: Perform quality control on the GNSS atmospheric water vapor data and the ground meteorological data; Step 3: Perform data preprocessing on the quality-controlled atmospheric water vapor data and the ground meteorological data to obtain the preprocessed atmospheric water vapor data and the ground meteorological data; Step 4: According to the preprocessed atmospheric water vapor data and the ground meteorological data, in combination with the observation date, time, and station altitude, design and train a multi-channel hybrid neural network atmospheric water vapor estimation model; Step 5: Use the preprocessed ground meteorological data, in combination with the observation date, time, and station altitude, based on the trained multi-channel hybrid neural network atmospheric water vapor estimation model, obtain the estimated value of the atmospheric water vapor, and evaluate the estimation accuracy of the multi-channel hybrid neural network atmospheric water vapor.

Description

Technical Field

[0001] The present invention relates to the field of atmospheric water vapor observation in atmospheric sounding, and particularly to an atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model. Background Art

[0002] Water vapor is an important component of the atmosphere. Although its content is small, it is the main driving force for weather changes and climate evolution, and is also a key factor affecting the formation and development of disastrous weather. However, its changes are rapid and the spatio-temporal distribution differences are complex. As a direct observation method, radiosonde is limited by reasons such as sparse station network distribution and fewer observation times, and the spatio-temporal resolution of the data is low. The emergence of GNSS / MET water vapor observation technology has improved the spatio-temporal resolution of atmospheric water vapor observation to a certain extent, but in most areas, it is still impossible to directly detect atmospheric water vapor data. Therefore, using ground meteorological data with a denser station network to estimate atmospheric water vapor has become an alternative technical approach.

[0003] Research shows that there is an obvious correlation between ground meteorological data and atmospheric water vapor, and ground meteorological data can be used to estimate atmospheric water vapor. However, existing atmospheric water vapor estimation methods generally use simple statistical models such as linear fitting or quadratic polynomials between ground water vapor pressure and atmospheric water vapor, which only reflect the relationship between a single ground meteorological element and atmospheric water vapor, and do not consider the temporal correlation between ground meteorological elements and atmospheric water vapor. Therefore, the estimation accuracy is low, and the root mean square error exceeds 4 mm. In recent years, with the rapid development of modern information technologies such as machine learning, some researchers have introduced methods such as neural networks to perform non-linear modeling and estimation of atmospheric water vapor using ground meteorological data, which has significantly improved the estimation accuracy of atmospheric water vapor, but only considered the relationship between ground meteorological elements and atmospheric water vapor, and did not consider the influence of non-meteorological elements such as observation date, time, and station altitude on atmospheric water vapor. Summary of the Invention

[0004] Embodiments of the present invention provide an atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model, which takes into account the temporal relationship between meteorological elements and atmospheric water vapor, and takes into account the influence of observation date, time, and station altitude on atmospheric water vapor, further improving the accuracy of estimating atmospheric water vapor based on ground meteorological data.

[0005] An atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model, characterized by comprising:

[0006] Step 1: Obtain GNSS atmospheric water vapor data and obtain ground meteorological data;

[0007] Step 2: Perform quality control on the GNSS atmospheric water vapor data and the ground meteorological data to obtain the quality-controlled atmospheric water vapor data and ground meteorological data;

[0008] Step 3: Perform data preprocessing on the quality-controlled atmospheric water vapor data and ground meteorological data to obtain the preprocessed atmospheric water vapor data and ground meteorological data;

[0009] Step 4: Design and train a multi-channel hybrid neural network atmospheric water vapor estimation model based on the preprocessed atmospheric water vapor data and ground meteorological data;

[0010] Step 5: Use the preprocessed ground meteorological data, combined with the observation date, time, and station altitude, and based on the trained multi-channel hybrid neural network atmospheric water vapor estimation model, obtain the estimated value of atmospheric water vapor, and evaluate the estimation accuracy of the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0011] As can be seen from the technical solutions provided by the embodiments of the present invention above, in the embodiments of the present invention, based on the trained multi-channel hybrid neural network atmospheric water vapor estimation model, the estimated value of atmospheric water vapor can be obtained, which can improve the estimation accuracy.

[0012] Additional aspects and advantages of the present invention will be given in part in the following description, which will become apparent from the following description, or can be understood through the practice of the present invention. Brief Description of the Drawings

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0014] Figure 1 It is a schematic diagram of an atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model of the present invention;

[0015] Figure 2 It is a method logic flow block diagram of the application scenario of the present invention;

[0016] Figure 3 It is a structural diagram of the multi-channel hybrid neural network atmospheric water vapor estimation model of the present invention. Detailed Embodiments

[0017] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] For the convenience of understanding the embodiments of the present invention, the following will further explain and illustrate by taking several specific embodiments as examples in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0019] As Figure 1 shown, an atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model of the present invention includes:

[0020] Step 1: Obtain GNSS atmospheric water vapor data and obtain ground meteorological data; the step of obtaining GNSS atmospheric water vapor data includes:

[0021] Step 1-1: Solve the GNSS observation data to obtain the total zenith delay of the troposphere; and perform averaging processing in hours to obtain the hourly total zenith delay ZTD;

[0022] Step 1-2: Use the ground pressure, temperature and altitude to calculate the zenith hydrostatic delay ZHD based on the Hopfield model;

[0023] Step 1-3: Obtain the zenith wet delay ZWD from ZWD = ZTD - ZHD;

[0024] Step 1-4: Calculate the GNSS atmospheric water vapor data PWV from PWV = Π × ZWD; where the water vapor conversion coefficient Π is a dimensionless coefficient calculated based on temperature.

[0025] The step of obtaining the ground meteorological data is specifically: the information obtained by observing the ground meteorological station is averaged and quantified in hours.

[0026] Step 2: Perform quality control on the GNSS atmospheric water vapor data and the ground meteorological data to obtain the atmospheric water vapor data and the ground meteorological data after quality control;

[0027] Step 3: Perform data preprocessing on the atmospheric water vapor data and the ground meteorological data after quality control to obtain the atmospheric water vapor data and the ground meteorological data after preprocessing;

[0028] Step 4: Design and train a multi-channel hybrid neural network atmospheric water vapor estimation model based on the preprocessed atmospheric water vapor data and surface meteorological data; and obtain the estimated value of atmospheric water vapor based on the trained multi-channel hybrid neural network atmospheric water vapor estimation model using the preprocessed atmospheric water vapor data and surface meteorological data.

[0029] Step 5: Calculate the root mean square error, mean absolute error, and mean absolute percentage error between the estimated value of atmospheric water vapor and the GNSS atmospheric water vapor data in the test sample set to obtain an evaluation value of the accuracy of the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0030] Among them, Step 2 includes:

[0031] Step 2-1: Perform gross error processing on the atmospheric water vapor data and surface meteorological data; specifically, Step 2-1 is: detect and delete using the 3-sigma rule.

[0032] Step 2-2: Perform missing data processing on the atmospheric water vapor data and surface meteorological data after gross error processing. Specifically, Step 2-2 is: use the Kalman smoothing algorithm to interpolate the missing data to form hourly data corresponding to the start and end times.

[0033] The said Step 3 includes:

[0034] Step 3-1: Generate surface meteorological data characteristic factors based on GNSS atmospheric water vapor data and surface meteorological data; Step 3-1 includes: conduct a correlation test on GNSS atmospheric water vapor data and surface meteorological data, calculate and generate a correlation coefficient; select the surface meteorological elements x i (t) corresponding to the correlation coefficients whose absolute values are greater than a predetermined value to form a feature vector X t =[x i (t)] T as the characteristic factors for input to the estimation model; where x i (t) represents the data of the i-th surface meteorological element at the t-th time.

[0035] Step 3-2: Standardize the GNSS atmospheric water vapor data and surface meteorological data characteristic factors to generate preprocessed atmospheric water vapor data and surface meteorological data.

[0036] The said Step 4 includes:

[0037] Step 4-1: Design of the multi-channel hybrid neural network atmospheric water vapor estimation model, specifically as Figure 3 shown. The multi-channel hybrid neural network atmospheric water vapor estimation model is specifically: adopt a multi-channel hybrid structure, and the model input is time series features and non-time series features;

[0038] Branch 1 is a time series feature extraction branch, which is used to extract time series features from characteristic factors of ground meteorological data. Branch 1 is a structure of 2 GRU neural network layers combined with 2 discard layers.

[0039] Branch 2 is a non-time series feature extraction branch, the input is the observation date, time and station altitude, and branch 2 is a 2-layer fully connected neural network layer sequential connection structure;

[0040] The time series features and non-time series features extracted by branch 1 and branch 2 are connected through a fully connected layer to estimate the GNSS water vapor data as label data.

[0041] Step 4-2: Determine the time window of the multi-channel hybrid neural network atmospheric water vapor estimation model, specifically: calculate the time correlation ρ of the GNSS atmospheric water vapor time series y(t) using the autocorrelation function Δt ; Where y(t) and y(t+Δt) are the GNSS atmospheric water vapor values ​​at time t and t+Δt, respectively, Cov(·) and σ(·) are the covariance and variance of the GNSS atmospheric water vapor values, respectively; the time interval Δt where the autocorrelation coefficient is greater than the predetermined value is selected as the time window W of the estimation model;

[0042] Step 4-3: Dataset reorganization, specifically: for the characteristic factors of ground meteorological data, use the time window to generate time series input samples seq_input as the input of the time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model; the length of each seq_input is equal to the time window W, that is, seq_input = {X t-W+1 ,X t-W+2 ΛX t The station altitude and the observation date and time corresponding to time t constitute the additional information input sample add_input, which is used as the input of the non-time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model. The input of the two branches and the GNSS atmospheric water vapor data y corresponding to time t constitute the data sequence sample data = {[seq_input, add_input], y}.

[0043] Step 4-4: Divide the training set and the test set; specifically: aggregate all the data sequence samples to obtain the model sample set data_set, divide the data_set into the training sample set train_set and the test sample set test_set according to the ratio r, which are used for model training and testing respectively;

[0044] Step 4-5: Training of the multi-path hybrid neural network atmospheric water vapor estimation model. For the neural network structure of the multi-path hybrid neural network atmospheric water vapor estimation model, set the number of neurons in the first GRU layer, the second GRU layer, the first fully connected layer, and the second fully connected layer, set the dropout rates of the first dropout layer and the second dropout layer. The activation functions of the GRU layer and the fully connected layer are tanh and linear respectively. Set the initial value of the learning rate to 1, set the number of training samples input into the model each time, and set the maximum number of training iterations to a predetermined number. Select RMSProp as the model optimizer, and the loss function is the mean square error between the estimated atmospheric water vapor value of the model and the GNSS atmospheric water vapor data label. Use the backpropagation algorithm to train the neural network model, calculate the gradient of the neural network parameters according to the loss function, and update the neural network parameters. One calculation of all data in the training sample set train_set is one iteration. When the number of iterations reaches the predetermined number of times or the mean square error is less than the predetermined value, the training of the multi-path hybrid neural network atmospheric water vapor estimation model is completed.

[0045] The said step 5 includes:

[0046] Step 5-1: Using the trained multi-path hybrid neural network atmospheric water vapor estimation model, take the seq_input and add_input in the test sample set test_set as the model input to obtain the atmospheric water vapor estimation value at the corresponding time.

[0047] Step 5-2: Calculate the root mean square error, mean absolute error, and mean absolute percentage error between the atmospheric water vapor estimation value and the GNSS atmospheric water vapor data in the test sample set to obtain the evaluation value of the atmospheric water vapor estimation accuracy of the multi-path hybrid neural network atmospheric water vapor estimation model.

[0048] The present invention takes into account the temporal relationship between meteorological elements and atmospheric water vapor, and takes into account the influence of the observation date, time, and station height on atmospheric water vapor, further improving the accuracy of estimating atmospheric water vapor based on ground meteorological data.

[0049] The following describes the application scenarios of the present invention.

[0050] The present invention proposes an atmospheric water vapor estimation method based on a multi-path hybrid neural network atmospheric water vapor estimation model. As Figure 2 shown, the technical solution of the present invention includes the following steps:

[0051] Step 1: Acquisition of GNSS atmospheric water vapor data and ground meteorological data.

[0052] Step 1-1: Acquisition of GNSS atmospheric water vapor data. The tropospheric zenith total delay is solved from GNSS observation data and averaged hourly to obtain the hourly zenith total delay ZTD.

[0053] Using ground pressure, temperature, and altitude, the zenith hydrostatic delay is calculated based on the Hopfield model, and the calculation formula is as follows:

[0054] ZHD = 15.52 × P s × [40.082 + 0.14898 × (T s - 273.16) - H s / T s ⑴

[0055] Where P s , T s and H s are the ground pressure, ground temperature, and station altitude.

[0056] The zenith wet delay ZWD is obtained from ZWD = ZTD - ZHD. Based on this, the GNSS atmospheric water vapor data is calculated from PWV = Π × ZWD, where the water vapor conversion coefficient Π is a dimensionless coefficient calculated based on temperature.

[0057] Step 1 - 2: Acquisition of ground meteorological data. The ground meteorological data is obtained by averaging and quantifying the information observed by the ground meteorological station in hours.

[0058] Step 2: Perform quality control on the GNSS atmospheric water vapor data and ground meteorological data.

[0059] Step 2 - 1: Processing of gross errors in data. Due to anomalies in the observation instruments or processes such as data coding, processing, transmission, and storage, there are gross errors in the GNSS atmospheric water vapor data and ground meteorological data, and the method of the 3σ criterion is used for detection and deletion.

[0060] Step 2 - 2: Processing of missing data. Due to the influence of power supply, network and other failures and the processing of gross errors, there are a small number of missing measurements in the GNSS atmospheric water vapor data and ground meteorological data. The Kalman smoothing algorithm is used to interpolate the missing data to form hourly data corresponding to the start and end times.

[0061] Step 3: Data preprocessing.

[0062] Step 3 - 1: Selection of characteristic factors of ground meteorological data. Perform a correlation test on the GNSS atmospheric water vapor data and ground meteorological data, and calculate the correlation coefficient. Select the ground meteorological element x i (t) with an absolute value of the correlation coefficient greater than 0.6 with the GNSS atmospheric water vapor to form the characteristic vector X t = [x i (t)] T as the characteristic factor for input to the estimation model. Where x i(t) represents the data of the i-th surface meteorological element at the t-th time step. In this embodiment, air temperature, air pressure, dew point temperature, and water vapor pressure are selected as the characteristic factors for input to the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0063] Step 3-2: Data standardization. Standardize the GNSS atmospheric water vapor data obtained in Step 2 and the ground meteorological data characteristic factors obtained in Step 3-1. Use the zero-mean normalization method to transform the data distribution into a Gaussian distribution with a mean of 0 and a variance of 1.

[0064] Step 4: Train the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0065] Step 4-1: Design of the multi-channel hybrid neural network atmospheric water vapor estimation model. The design of the multi-channel hybrid neural network atmospheric water vapor estimation model is as Figure 2 shown. It adopts a multi-channel hybrid structure to extract the time series features and non-time series features of the model input respectively. Branch 1 is the time series feature extraction branch, which is used to extract the time series features of the ground meteorological characteristic factors obtained in Step 3. It is a structure composed of 2 layers of GRU neural network layers combined with 2 layers of dropout layers; Branch 2 is the non-time series feature extraction branch, and its input is the observation date, time, and station elevation. It is a sequential connection structure of 2 layers of fully connected neural network layers. The time series features and non-time series features extracted by Branch 1 and Branch 2 are connected through 1 fully connected layer to estimate the GNSS water vapor data used as the label data.

[0066] Step 4-2: Determine the time window of the estimation model. Use the autocorrelation function to calculate the time correlation of the GNSS atmospheric water vapor time series y(t), and analyze the influence degree of the time interval on the current GNSS water vapor. The calculation formula is as follows:

[0067]

[0068] where y(t) and y(t + Δt) are the GNSS atmospheric water vapor values at times t and t + Δt respectively, and Cov(·) and σ(·) are the covariance and variance of the GNSS atmospheric water vapor values respectively. Select the time interval with an autocorrelation coefficient greater than 0.6 as the time window W of the estimation model. In this embodiment, 24 hours is selected as the time window of the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0069] Step 4-3: Dataset reorganization. For the ground meteorological data characteristic factors obtained in Step 3, use the time window obtained in Step 4-2 to generate the time series input sample seq_input, which is used as the input of the time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model. The length of each seq_input is equal to the time window W, that is, seq_input = {X t-W+1 , X t-W+2 ΛXt}; The station altitude and the observation date and time corresponding to time t constitute the additional information input sample add_input, which is used as the input of the non-time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model. The input of the two branches and the GNSS atmospheric water vapor data y corresponding to time t constitute the data sequence sample data = {[seq_input, add_input], y}.

[0070] Step 4-4: Divide the training set and the test set. Aggregate all the data sequence samples obtained in step 4-3 to obtain the model sample set data_set, and divide data_set into the training sample set train_set and the test sample set test_set according to the ratio r, which are used for model training and testing respectively. In this embodiment, r is 0.7.

[0071] Step 4-5: Training of the multi-channel hybrid neural network atmospheric water vapor estimation model. For the neural network structure of step 4-1, the number of neurons in the first GRU layer, the second GRU layer, the first fully connected layer, and the second fully connected layer are 100, 50, 20, and 10 respectively, the dropout rates of the first dropout layer and the second dropout layer are 0.2 and 0.1 respectively, the activation functions of the GRU layer and the fully connected layer are tanh and linear respectively, the initial value of the learning rate is 0.001, the number of training samples input into the model each time is 128, the maximum number of training iterations is 500, the model optimizer selects RMSProp, and the loss function is the mean square error of the model estimated atmospheric water vapor value and the GNSS atmospheric water vapor data label. The neural network model is trained using the back propagation algorithm, the neural network parameter gradient is calculated according to the loss function, and the neural network parameters are updated. One calculation for all the data in the training sample set train_set is considered an iteration, and the training of the multi-channel hybrid neural network atmospheric water vapor estimation model is completed when the number of iterations reaches 500 or the mean square error is less than 0.001.

[0072] Step 5: Estimate atmospheric water vapor based on the multi-channel hybrid neural network atmospheric water vapor estimation model.

[0073] Step 5-1: Estimation of atmospheric water vapor data. Using the multi-channel hybrid neural network atmospheric water vapor estimation model trained in step 4-5, the seq_input and add_input in the test sample set test_set obtained by 4-4 are used as model inputs to obtain the atmospheric water vapor estimation data of the corresponding time.

[0074] Step 5-2: Evaluation of the atmospheric water vapor estimation accuracy of the multi-channel hybrid neural network atmospheric water vapor estimation model. Calculate the root mean square error, mean absolute error, and mean absolute percentage error between the atmospheric water vapor estimation data and the GNSS atmospheric water vapor data of the test samples, and the accuracy of the multi-channel hybrid neural network atmospheric water vapor estimation model can be evaluated.

[0075] The present invention has the following technical effects:

[0076] Given the GNSS observation data (tropospheric zenith total delay) and ground meteorological data (temperature, pressure, dew point temperature, water vapor pressure, precipitation, 10-minute average wind speed, 10-minute average wind direction) in a certain area from April 1, 2016 to December 1, 2020. Perform quality control on the obtained GNSS atmospheric water vapor data and ground meteorological data according to the technical solution of the present invention. Select temperature, pressure, dew point temperature, and water vapor pressure with a correlation coefficient greater than 0.6 with the GNSS atmospheric water vapor data as the input characteristic factors of the estimation model, and select 24 hours as the time window of the estimation model. Reorganize the model sample set (data_set) with a total of 40,921 samples according to the 24-hour time window and combined with additional information, and divide it into a training sample set (train_set)

[0077] with 28,645 samples and a test sample set (test_set) with 12,276 samples according to a ratio of 0.7. After training the multi-channel hybrid neural network atmospheric water vapor estimation model based on the training sample set, use the corresponding time period of the test sample set as the time period to be estimated, estimate the atmospheric water vapor, and compare it with the GNSS atmospheric water vapor data. The root mean square error between the two is 1.05 mm, the average error is 0.64 mm, and the mean absolute percentage error is 4.37%.

[0078] It can be seen from the above application experiments that the atmospheric water vapor estimation method using the above multi-channel hybrid neural network atmospheric water vapor estimation model can obtain a relatively accurate atmospheric water vapor estimation value based on ground meteorological data and non-meteorological data such as observation date, time, and station altitude, effectively making up for the deficiency of the low accuracy of the linear fitting or quadratic polynomial atmospheric water vapor estimation model, and providing higher-precision atmospheric water vapor data for disaster weather monitoring and early warning and meteorological disaster prevention and mitigation.

[0079] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An atmospheric water vapor estimation method based on a multi-channel hybrid neural network atmospheric water vapor estimation model, characterized in that, Including: Step 1: Obtain GNSS atmospheric water vapor data and obtain ground meteorological data; Step 2: Perform quality control on the GNSS atmospheric water vapor data and the ground meteorological data to obtain the quality-controlled atmospheric water vapor data and ground meteorological data; Step 3: Perform data preprocessing on the quality-controlled atmospheric water vapor data and ground meteorological data to obtain the preprocessed atmospheric water vapor data and ground meteorological data; Step 4: According to the preprocessed atmospheric water vapor data and ground meteorological data, combined with the observation date, time, and station elevation, design and train a multi-path hybrid neural network atmospheric water vapor estimation model; Step 5: Use the preprocessed ground meteorological data, combined with the observation date, time, and station elevation, based on the trained multi-path hybrid neural network atmospheric water vapor estimation model, to obtain the estimated value of atmospheric water vapor; And evaluate the estimation accuracy of the multi-path hybrid neural network for atmospheric water vapor; The said Step 3 includes: Step 3-1: Generate ground meteorological data characteristic factors according to GNSS atmospheric water vapor data and ground meteorological data; Step 3-2: Standardize the GNSS atmospheric water vapor data and ground meteorological data characteristic factors to generate the preprocessed atmospheric water vapor data and ground meteorological data; The said Step 3-1 includes: Perform a correlation test on GNSS atmospheric water vapor data and ground meteorological data, and calculate and generate a correlation coefficient; Select the surface meteorological element x corresponding to the correlation coefficient whose absolute value is greater than a predetermined value i (t) to form the feature vector X t =[x i (t)] T , as the feature factor input to the estimation model; where x i (t) represents the data of the i-th surface meteorological element at the t-th time The multi-path hybrid neural network atmospheric water vapor estimation model is specifically: adopt a multi-path hybrid neural network structure, and the model input is time series features and non-time series features; Branch 1 is a time series feature extraction branch, which is used to extract time series features from ground meteorological data characteristic factors. Branch 1 is a structure composed of 2 layers of GRU neural network layers combined with 2 layers of dropout layers; Branch 2 is a non-time series feature extraction branch, and the input is the observation date, time, and station elevation. Branch 2 is a sequential connection structure of 2 layers of fully connected neural network layers; The time series features and non-time series features extracted by Branch 1 and Branch 2 are connected by 1 fully connected layer to estimate the GNSS water vapor data used as label data.

2. The method according to claim 1, wherein The steps for obtaining GNSS atmospheric water vapor data include: Step 1-1: Solve GNSS observation data to obtain the total zenith delay of the troposphere; and perform averaging processing in hours to obtain the hourly total zenith delay ZTD; Step 1-2: Use ground pressure, temperature, and elevation to calculate the zenith hydrostatic delay ZHD based on the Hopfield model; Step 1-3: Obtain the zenith wet delay ZWD from ZWD = ZTD - ZHD; Step 1-4: Calculate the GNSS atmospheric water vapor data PWV from PWV = Π × ZWD; where the water vapor conversion coefficient Π is a dimensionless coefficient calculated based on temperature.

3. The method according to claim 1, characterized in that The steps for obtaining ground meteorological data are specifically: Obtained by averaging and quantifying the information observed by the ground meteorological station in hours.

4. The method according to claim 1, wherein The said Step 2 includes: Step 2-1: Perform gross error processing on the atmospheric water vapor data and ground meteorological data; Step 2-2: Perform missing data processing on the atmospheric water vapor data and surface meteorological data after gross error processing.

5. The method according to claim 4, wherein the specific operation of step 2-1 is: detecting and deleting by using the Pauta criterion; the specific operation of step 2-2 is: using the Kalman smoothing algorithm to interpolate the missing data to form hourly data corresponding to the start and end times.

6. The method according to claim 1, characterized in that The step 4 includes: Step 4-1: Determine the time window of the multi-channel hybrid neural network atmospheric water vapor estimation model, specifically: Calculate the temporal correlation ρ of the GNSS atmospheric water vapor time series y(t) using the autocorrelation function Δt ; where y(t) and y(t+Δt) are the GNSS atmospheric water vapor values at times t and t+Δt respectively, Cov(·) and σ(·) are the covariance and variance of the GNSS atmospheric water vapor values respectively; Select the time interval Δt with an autocorrelation coefficient greater than a predetermined value as the time window W of the estimation model; Step 4-2: Dataset recombination, specifically: for the characteristic factors of ground meteorological data, use the time window to generate a time series input sample seq_input, which is used as the input of the time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model; the length of each seq_input is equal to the time window W, that is, seq_input = {X t-W+1 , X t-W+2 … X t}; the station altitude and the observation date and time corresponding to the time t form an additional information input sample add_input, which is used as the input of the non-time series feature extraction branch of the multi-channel hybrid neural network atmospheric water vapor estimation model. The inputs of the two branches and the GNSS atmospheric water vapor data y corresponding to the time t form a data sequence sample data = {[seq_input, add_input], y}; Step 4-3: Divide the training set and the test set. Specifically, summarize all the data sequence samples to obtain the model sample set data_set, and divide data_set into the training sample set train_set and the test sample set test_set according to the ratio r, which are used for training and testing the model respectively; Step 4-4: Train the multi-channel hybrid neural network atmospheric water vapor estimation model. For the neural network structure of the multi-channel hybrid neural network atmospheric water vapor estimation model, set the number of neurons in the first GRU layer, the second GRU layer, the first fully connected layer, and the second fully connected layer, set the dropout rates of the first dropout layer and the second dropout layer, the activation functions of the GRU layer and the fully connected layer are tanh and linear respectively, set the initial value of the learning rate, set the number of training samples input into the model each time, the maximum number of training iterations is a predetermined number, the model optimizer selects RMSProp, and the loss function is the mean square error between the estimated atmospheric water vapor value of the model and the GNSS atmospheric water vapor data label; use the backpropagation algorithm to train the neural network model, calculate the gradient of the neural network parameters according to the loss function, and update the neural network parameters; one calculation of all the data in the training sample set train_set is one iteration, and when the number of iterations reaches the predetermined number of times or the mean square error is less than the predetermined value, the training of the multi-channel hybrid neural network atmospheric water vapor estimation model is completed.

7. The method according to claim 1, wherein It further includes: Step 5-1: Using the trained multi-channel hybrid neural network atmospheric water vapor estimation model, take the seq_input composed of the surface meteorological data in the test sample set test_set and the add_input composed of the observation date, time, and station altitude as the model input to obtain the estimated value of atmospheric water vapor at the corresponding time; Step 5-2: Calculate the root mean square error, mean absolute error, and mean absolute percentage error between the estimated value of the atmospheric water vapor and the GNSS atmospheric water vapor data in the test sample set to obtain the evaluation value of the accuracy of the multi-channel hybrid neural network atmospheric water vapor estimation model.

Images