Flood forecasting method based on GNSS water vapor chromatography and deep learning

Abstract

The invention provides a GNSS (Global Navigation Satellite System) water vapor chromatography and deep learning-based flood forecasting method. The method comprises the following steps: S1, collecting GNSS, meteorological, flood and geographic data; s2, the PWV is estimated through a PPP equation and a Saastamoinen model; s3, grid water vapor density is calculated based on the three-dimensional water vapor chromatography technology; s4, constructing a Transform model, and performing rainfall prediction by taking PWV, water vapor density and the like as input; and S5, constructing a CNN-LSTM model based on rainfall prediction data to carry out flood forecasting. According to the rainfall and flood forecasting method established by the invention, GNSS multi-source and multi-mode navigation positioning signals are effectively utilized, the rainfall and flood forecasting method is not influenced by cloud and rain, all-weather observation can be realized, the water vapor gathering and conveying process in a three-dimensional space before rainfall can be carefully captured, the rainfall inversion process is optimized, the flood forecasting precision is improved, and the rainfall and flood forecasting period is prolonged; and a new thought is provided for flood control early warning of the drainage basin.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hydrological forecasting, and particularly relates to a flood forecasting method based on GNSS water vapor tomography and deep learning. BACKGROUND

[0002] Under the background of global climate change, extreme weather events (such as heavy rain, typhoon) occur frequently, which poses a serious challenge to social and economic development and life and property safety. The existing rain and flood forecasting technology mainly relies on statistical forecasting method, numerical weather prediction and radar echo extrapolation method, which has the problems of low inversion accuracy, low time resolution and short rain forecasting period.

[0003] In recent years, GNSS water vapor inversion technology has been introduced into the field of hydrology due to its continuous operation, all-weather, high precision, high temporal and spatial resolution and other advantages. It provides new data support for rainfall inversion through solving atmospheric precipitable water (PWV) and three-dimensional water vapor tomography technology. However, the existing technology still has the following bottlenecks:

[0004] Two-dimensional water vapor inversion limitation: traditional GNSS PWV only reflects the overall water vapor content of the region, and cannot reveal the dynamic changes of three-dimensional water vapor distribution, making it difficult to accurately locate the key water vapor accumulation area triggered by rainfall;

[0005] Insufficient model fusion: there is a lack of method for deep fusion of GNSS three-dimensional water vapor data with meteorological and geographical data, and the traditional machine learning model has limited ability to capture the complex rainfall-flood process of time and space correlation, resulting in insufficient prediction accuracy and stability. SUMMARY

[0006] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a flood forecasting method based on GNSS water vapor tomography and deep learning, which is not affected by cloud and rain, can be observed all-weather, can finely capture the water vapor accumulation and transportation process in three-dimensional space before rainfall, can optimize the rainfall inversion process, can improve the flood forecasting accuracy, and can prolong the rain and flood forecasting period.

[0007] A flood forecasting method based on GNSS water vapor tomography and deep learning, comprising the following steps:

[0008] Step S1: collect GNSS data, meteorological data, flood data and geographical data;

[0009] Step S2: based on the GNSS data collected in step S1, obtain zenith troposphere delay ZTD through real-time PPP observation equation, calculate zenith troposphere dry delay ZHD by using Saastamoinen model, and calculate zenith wet delay ZWD by the difference between ZTD and ZHD, and convert ZWD into atmospheric precipitable water PWV;

[0010] Step S3: Based on the GNSS data and geographic data collected in Step S1, the target area is discretized into grids, a set of GNSS water vapor tomography equations is constructed, and the water vapor density value of each grid is obtained by solving the set of GNSS water vapor tomography equations.

[0011] Step S4: Using the PWV obtained in Step S2 and the water vapor density value obtained in Step S3, combined with the meteorological data and geographical data in Step S1 as input, construct a Transformer deep learning model, train the model and output rainfall prediction data.

[0012] Step S5: Using the rainfall prediction data obtained in step S4 as input, construct a CNN-LSTM deep learning model, train the model, and output cross-sectional flood forecast data.

[0013] Furthermore, in step S1, the GNSS data includes raw GNSS observation data and IGS precise orbital clock difference data; the meteorological data includes in-situ synchronized temperature, air pressure, humidity, and rainfall; the flood data includes flood level and flood flow at the control section; and the geographic data includes topographic elevation, latitude and longitude, and section location.

[0014] Furthermore, step S2 specifically includes the following steps:

[0015] Step S2-1: Obtain the Zenith Tropospheric Delay (ZTD) using the real-time PPP observation equation;

[0016] The real-time PPP observation equation is expressed as follows:

[0017] ;

[0018] ;

[0019] By incorporating the IGS precise satellite ephemeris and clock bias, the final positioning error equation is obtained, expressed as:

[0020] ;

[0021] ;

[0022] Of the above formulas, , These are pseudorange and phase observations. The superscript ' indicates the corresponding observations after precise satellite ephemeris and clock error correction. dt represents the geometric distance from the satellite to the station; dT represents the satellite clock error, and dT represents the receiver clock error. Represents tropospheric parameters; Represents multipath effects; N represents the integer ambiguity parameter of the ionosphere-free combination. The random error representing the observed value; the ppp positioning error term calculated using the positioning error equation is used as the observation to estimate the zenith tropospheric delay (ZTD);

[0023] Step S2-2: Calculate the zenith tropospheric dry delay ZHD using the Saastamoinen model. The expression is:

[0024] ;

[0025] ;

[0026] Among them, P s Here, φ represents atmospheric pressure, φ represents latitude, and H represents the altitude of the station.

[0027] Step S2-3: Calculate the zenith tropospheric wet delay ZWD and convert it to PWV; the zenith wet delay ZWD equals the total zenith delay ZTD minus the zenith dry delay ZHD. Multiply the zenith wet delay ZWD by the conversion factor. The atmospheric precipitation potential (PWV) is expressed as follows:

[0028] ;

[0029] Zenith wet delay (ZWD) is converted into total rainable volume (PWV) through calculation, using the conversion factor in the formula. The expression is:

[0030] ;

[0031] Where, ρ w R represents water vapor density. v M represents the water vapor gas constant. w M d K1, K2, and K3 represent the molar masses of water vapor and dry air molecules, respectively, and represent the atmospheric refractive indices: K1 = 77.6890 K / hPa, K2 = 71.2952 K / hPa, and K3 = 375463 K. 2 / hPa;

[0032] Representing the weighted average temperature of the atmosphere, the expression is:

[0033] ;

[0034] Where e represents water vapor pressure, T represents absolute temperature, and T d Represents the dew point temperature, e s e represents the saturated vapor pressure. s0The saturated water vapor pressure at 0 degrees Celsius is 6.11 hPa. For water surfaces, a0 = 7.5 and b0 = 273.3. For ice surfaces, a0 = 9.5 and b0 = 265.7.

[0035] It is a dimensionless numerical value, and its value is related to T. m The size is related, T m Using ground or radiosonde data to study T m Numerical integration is performed to calculate, and then the ground temperature is used to correlate with T. m Perform statistical modeling.

[0036] Furthermore, step S3 includes the following steps:

[0037] Step S3-1: Discretize the target area into a mesh to obtain a 3D voxel block: This includes setting the target area height, determining the horizontal and vertical resolutions; setting the heights of the bottom and top of the mesh to 0km and 11km respectively; the horizontal and vertical resolutions of the horizontal mesh division are determined according to application requirements.

[0038] Step S3-2: Construct the GNSS water vapor tomography equation set, the expression of which is:

[0039] ;

[0040] in, This represents the intercept vector of the GNSS signal within each voxel block. X is the vector of SWV observations; X is the vector of water vapor density. The coefficient matrix of the horizontal constraint follows the characteristics of a Gaussian weighted function; The coefficient matrix of the vertical constraint follows an exponentially decreasing characteristic.

[0041] Step S3-3: Perform three-dimensional tomography equation solving to obtain the water vapor density value for each grid; specifically, the SVD method is used for solving, and the expression is:

[0042] ;

[0043] In the formula, P is a matrix composed of adjustment constraints and observation information weighting factors; L is the constant matrix of the tomographic equation system, expressed as:

[0044] .

[0045] Furthermore, step S4 includes the following steps:

[0046] Step S4-1: Construct a Transformer neural network model using PWV, water vapor density value, meteorological data, and geographical data as input layers, and predicted rainfall as the output layer, and perform "time series embedding-encoding-decoding" calculations. First, use the Embedding time series embedding layer to project the input layer matrix into another dimension space; second, use a Transformer-based encoder to transform the input time series into a fixed-dimensional feature map; finally, use a Transformer-based decoder to decode the feature map and generate future time series.

[0047] Step S4-2: Conduct neural network training to obtain the optimal model parameters and perform rainfall prediction; divide the training set and test set, use the sample data of the training set to train the neural network model, continuously adjust the model parameters, use the root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators, use the validation set data to evaluate the accuracy of the model's predicted rainfall data, and determine the optimal model parameters.

[0048] Furthermore, step S5 includes the following steps:

[0049] Step S5-1: Construct a CNN-LSTM neural network model, including an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, and an output layer. The input layer receives rainfall prediction data. The CNN layer extracts the spatial features of the rainfall prediction data, which are then used as the input to the LSTM layer. The LSTM layer recursively processes the spatial features of the rainfall prediction data step-by-step, outputting a feature vector containing time-series information. The attention mechanism layer performs weighted aggregation on the feature vectors output by the LSTM layer to obtain a context feature vector. The output layer converts the context feature vector output by the attention mechanism layer into the final prediction result, obtaining the flood change trend over a future period.

[0050] Step S5-2: Conduct neural network training to obtain optimal model parameters for flood forecasting; extract GNSS data, meteorological data, and measured flood data based on the latitude and longitude of the hydrological section; divide the training set and test set, use the training set sample data to train the neural network model, continuously adjust the model parameters, use the root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators, use the validation set data to evaluate the accuracy of the model's predicted rainfall data, and determine the optimal model parameters.

[0051] The present invention has the following beneficial effects:

[0052] 1. Breaking through the limitations of traditional observations to achieve high-precision three-dimensional dynamic monitoring of water vapor: In response to the problem that traditional methods cannot capture the three-dimensional distribution of water vapor before precipitation in all weather conditions and at high resolution, a high-precision inversion chain is constructed by combining GNSS multi-source data (raw observation data, IGS precise orbital clock difference) and three-dimensional water vapor tomography technology, from "zenith tropospheric delay (ZTD) → precipitationable volume (PWV) → grid water vapor density", which can finely characterize the spatiotemporal characteristics of water vapor accumulation and transport.

[0053] 2. Integrating multi-source data with deep learning models to improve the accuracy of rainfall and flood forecasts: To address the problems of traditional models' weak ability to integrate multimodal data (water vapor, meteorology, and topography) and insufficient capture of spatiotemporal correlations, a Transformer-based rainfall prediction model (integrating PWV, gridded water vapor density, meteorological, and topographic data) and a CNN-LSTM flood forecast model (extracting spatial features of rainfall and capturing time series dependencies) are designed to solve the problems of rainfall area discretization, large magnitude errors, and the lag of flood processes.

[0054] 3. Extend the forecast period for rainfall and floods and achieve all-weather applicability: In view of the shortcomings of traditional methods, such as being affected by clouds and rain and having a short forecast period, this method takes advantage of the fact that GNSS signals are not affected by weather. Combined with the rapid response capability of three-dimensional water vapor dynamic monitoring and deep learning models, the forecast timeliness of rainfall is improved to 30-60 minutes and the flood forecast period is extended, providing a scientific basis for watershed flood prevention and early warning. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a flood forecasting method based on GNSS water vapor tomography and deep learning, according to an embodiment of the present invention.

[0056] Figure 2 This is a detailed block diagram of the technical route of the present invention;

[0057] Figure 3 This is a schematic diagram of a neural network model according to an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram illustrating the relationship between model-predicted floods and measured floods during the forecast period;

[0059] Figure 5 This is a flowchart of the areal rainfall prediction model. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a flood forecasting method based on GNSS water vapor tomography and deep learning, including the following steps:

[0062] Step S1: Collect GNSS data, meteorological data, flood data, and geographic data, etc.; the GNSS data includes raw GNSS observation data and IGS precise orbital clock error data; the meteorological data includes in-situ synchronized temperature, air pressure, humidity, and rainfall; the flood data includes flood level and flood flow at the control section; the geographic data includes topographic elevation, latitude and longitude, and section location.

[0063] Step S2: Estimate the atmospheric water vapor precipitation potential (PWV) along the GNSS signal path: Obtain the zenith tropospheric delay (ZTD) using the real-time PPP observation equation; calculate the zenith dry tropospheric delay (ZHD) using the Saastamoinen model; calculate the zenith wet tropospheric delay (ZWD) and convert it to PWV. The specific implementation steps are as follows:

[0064] Step S2-1: Obtain the Zenith Tropospheric Delay (ZTD) using the real-time PPP observation equation;

[0065] The real-time PPP observation equation is expressed as follows:

[0066]

[0067] By incorporating the IGS precise satellite ephemeris and clock bias, the final positioning error equation (i.e., the atmospheric wet delay observation equation) can be obtained, which is expressed as follows:

[0068]

[0069]

[0070] Of the above formulas, , These are pseudorange and phase observations. The superscript ' indicates the corresponding observations after precise satellite ephemeris and clock error correction. dt represents the geometric distance from the satellite to the station; dT represents the satellite clock error, and dT represents the receiver clock error. Representative tropospheric parameters (composed of projection function and zenith tropospheric parameters); Represents multipath effects; N represents the integer ambiguity parameter of the ionosphere-free combination. The random error representing the observed value; the ppp positioning error term calculated using the positioning error equation is used as the observation to estimate the zenith tropospheric delay (ZTD);

[0071] Step S2-2: Calculate the zenith tropospheric dry delay ZHD using the Saastamoinen model. The expression is:

[0072]

[0073]

[0074] Among them, P s Here, φ represents atmospheric pressure, φ represents latitude, and H represents the altitude of the station.

[0075] Step S2-3: Calculate the zenith tropospheric wet delay ZWD and convert it to PWV; the zenith wet delay ZWD equals the total zenith delay ZTD minus the zenith dry delay ZHD. Multiply the zenith wet delay ZWD by the conversion factor. The atmospheric precipitation potential (PWV) is expressed as follows:

[0076]

[0077] Zenith wet delay (ZWD) is converted into total rainable volume (PWV) through calculation, using the conversion factor in the formula. The expression is:

[0078]

[0079] Where, ρ w R represents water vapor density. v M represents the water vapor gas constant. w M d K1, K2, and K3 represent the molar masses of water vapor and dry air molecules, respectively, and represent the atmospheric refractive indices: K1 = 77.6890 K / hPa, K2 = 71.2952 K / hPa, and K3 = 375463 K. 2 / hPa.

[0080] Representing the weighted average temperature of the atmosphere, the expression is:

[0081]

[0082] Where e represents water vapor pressure, T represents absolute temperature, and T d Represents the dew point temperature, e se represents the saturated vapor pressure. s0 The saturated water vapor pressure at 0 degrees Celsius is 6.11 hPa. For water surfaces, a0 = 7.5 and b0 = 273.3. For ice surfaces, a0 = 9.5 and b0 = 265.7.

[0083] It is a dimensionless numerical value, and its value is related to T. m The size is related, T m Using ground or radiosonde data to study T m Numerical integration is performed to calculate, and then the ground temperature is used to correlate with T. m Perform statistical modeling.

[0084] Step S3: Calculate the water vapor density value of the grid based on three-dimensional water vapor tomography: Discretize the target area into a grid to obtain three-dimensional voxel blocks; construct the GNSS water vapor tomography equation set; solve the three-dimensional tomography equations to obtain the water vapor density value of each grid. The specific implementation steps are as follows:

[0085] Step S3-1: Discretize the target area into a mesh to obtain a three-dimensional voxel block; this includes setting the target area height, determining the horizontal and vertical resolutions; setting the heights of the bottom and top of the mesh to 0km and 11km (tropospheric height) respectively; the horizontal and vertical resolutions of the horizontal mesh division are determined according to application requirements.

[0086] Step S3-2: Construct the GNSS water vapor tomography equation set, the expression of which is:

[0087]

[0088] in, This represents the intercept vector of the GNSS signal within each voxel block. X is the vector of SWV observations; X is the vector of water vapor density. The coefficient matrix of the horizontal constraint follows the characteristics of a Gaussian weighted function; The coefficient matrix of the vertical constraint follows an exponentially decreasing characteristic.

[0089] Step S3-3: Perform three-dimensional tomography equations to obtain the water vapor density value for each grid; typically, the SVD method is used for this calculation, and the expression is:

[0090]

[0091] In the formula, P is a matrix composed of adjustment constraints and observation information weighting factors; L is the constant matrix of the tomographic equation system, expressed as:

[0092]

[0093] Step S4: Construct a Transformer deep learning model and perform rainfall prediction: Using PWV, water vapor density, meteorological data, and topographic elevation data as input layers, and rainfall prediction as the output layer, construct a Transformer neural network model (e.g.,...). Figure 3 (As shown in the image); conduct neural network training to obtain the optimal model parameters for rainfall prediction. The specific implementation steps are as follows:

[0094] Step S4-1: Using PWV, water vapor density, meteorological data, and topographic elevation data as input layers, and predicting rainfall as the output layer, construct a Transformer neural network model and perform "time series embedding-encoding-decoding" calculations. First, use the Embedding time series embedding layer to project the input layer matrix into another dimension space. Second, use a Transformer-based encoder to transform the input time series into a fixed-dimensional feature map. Finally, use a Transformer-based decoder to decode the feature map and generate future time series.

[0095] Specifically, this invention uses a Transformer-based areal rainfall prediction model to solve the GNSS areal rainfall prediction task. The first part is a time-series embedding layer, and the second part is a time-series encoder-decoder. The embedding-based time-series embedding layer projects the input matrix into another dimension space. Then, the Transformer-based encoder-decoder first transforms the input time series into a fixed-dimensional feature map, and then the decoder decodes the feature map to generate future time series.

[0096] Among them, the learning framework of the Transformer-based areal rainfall prediction model is as follows: Figure 5 As shown:

[0097] Specifically, the method used will be continuous The system takes a historical time series X with a step size as input and outputs a series of... The future value Y of the time step, where and In the first stage, an embedding-based time-series embedding layer is designed, which includes a scalar projection module, position encoding, and time encoding. To input the time-series data into the model, the data is first projected onto the required dimension using the scalar projection part. Then, position information is added to its position embedding. Finally, time-related embeddings are added to obtain the data M input to the model. emb .

[0098] In the second stage, a Transformer-based time-series encoder-decoder is designed. The time-series encoder-decoder consists of two parts: an encoder and a decoder. While multi-head attention mechanisms perform well in capturing long-range dependencies, they neglect the importance of capturing nearest-neighbor dependencies. Therefore, the model effectively captures both nearest-neighbor and long-range dependencies through convolutional attention modules and multi-head attention mechanisms. The encoder consists of a convolutional attention module and a downsampling module, while the decoder consists of a convolutional attention module, a multi-head attention mechanism, and fully connected layers. The input to the time-series encoder-decoder is M. emb The output is The model takes the output M of the time series embedding layer. emb This serves as the input to the time-series encoder-decoder. A convolutional attention module is used in the encoder to focus on the input M. emb Perform calculations to further obtain high-level features, while capturing nearest-neighbor dependencies, to obtain M. cbam Then M cbam The input is fed into the downsampling module, where it passes through a max-pooling layer to reduce parameters and filter features, yielding the encoder's output M. encoder The decoder's input is the encoder's output M. encoder The output is The model first uses a convolutional attention module in the decoder to further capture nearest-neighbor dependencies, while preventing the receptive field from growing exponentially with the number of layers, which could lead to attention being paid to unimportant features. Then, a multi-head attention mechanism is used to capture long-range dependencies, and finally, a fully connected layer is used to project the dimensions onto the original dimensions of the time series to obtain the output. In addition, in the time-series encoder-decoder, the model uses a convolutional attention module and a max-pooling layer in the downsampling module to reduce the GPU memory footprint during training.

[0099] Step S4-2: Conduct neural network training to obtain the optimal model parameters and perform rainfall prediction; divide the training set and test set, use the sample data of the training set to train the neural network model, continuously adjust the model parameters, use the validation set data to evaluate the accuracy of the model's predicted rainfall data (root mean square error RMSE, mean absolute error MAE), and determine the optimal model parameters.

[0100] Specifically, the GNSS water vapor density data is raster data, with a time resolution of 10 minutes, a spatial resolution of 1 km * 1 km, and a spatial range of 100 km. 2 The dataset is divided into training and testing sets. When training the model, the input data is GNSS water vapor density and other data, and the output data is the rainfall prediction data for the next period.

[0101] The model is validated using a test set, and metrics such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are calculated to evaluate its accuracy and reliability. Based on the validation results, the model is optimized, including adjusting model parameters and selecting more suitable features. Step S5: Based on GNSS rainfall prediction data, a CNN-LSTM deep learning model is constructed for flood forecasting: a CNN-LSTM neural network model is constructed using rainfall prediction data as the input layer and cross-sectional flood forecast data as the output layer; the neural network is trained to obtain the optimal model parameters for flood forecasting. The specific implementation steps are as follows:

[0102] Step S5-1: Construct a CNN-LSTM neural network model, including an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, and an output layer;

[0103] The input layer receives rainfall forecast data and converts it into a time series matrix format suitable for model processing.

[0104] The CNN layer is used to extract spatial features from rainfall prediction data and contains multiple convolutional and pooling layers. The convolutional layers perform convolution operations on the input data using appropriately sized kernels to abstractly represent the original data. The pooling layers perform downsampling operations on the convolutional output, preserving strong features and removing weak features, while also reducing the number of parameters to prevent overfitting.

[0105] The LSTM layer is used to learn the long-term dependencies and contextual information of the rainfall prediction input data, and to recursively process the spatial features of the rainfall prediction data step by step, outputting a feature vector containing time series information.

[0106] The attention mechanism layer is used to weight and converge the feature vectors output by the LSTM layer to obtain the context feature vector;

[0107] The output layer is used to convert the context feature vector output by the attention mechanism layer into the final prediction result, which yields the flood change trend over a future period of time.

[0108] Regarding model parameter settings, the CNN layer parameters are set as follows:

[0109] Kernel size (kernel_size): Select a 3x3 kernel size.

[0110] Number of convolutional kernels (filters / channels): 32 convolutional kernels are selected based on task complexity.

[0111] Activation function: Uses the ReLU model to learn complex features.

[0112] Pooling layer: Average pooling is used to reduce the size of the feature map, reduce the amount of computation, and retain important features.

[0113] The LSTM layer parameters are set as follows:

[0114] Hidden size: 64 hidden units are selected based on task complexity and dataset size.

[0115] Number of LSTM layers (num_layers): Select 2 LSTM layers.

[0116] Attention mechanism layer parameter settings:

[0117] Multi-head attention: When using a multi-head attention mechanism, the number of heads needs to be set.

[0118] Dimension scaling: When calculating attention weights, the input dimension needs to be scaled to prevent the gradient from vanishing in the softmax function.

[0119] Other parameter settings:

[0120] Learning rate: 0.001 was chosen as the learning rate.

[0121] Batch size (batch_size): Select 8 as the batch size.

[0122] Number of iterations (num_epochs): 100 is chosen as the number of iterations.

[0123] Optimizer: Select to use the Adam optimizer.

[0124] Loss function: Select the mean squared error loss function.

[0125] Step S5-2: Conduct neural network training to obtain optimal model parameters for flood forecasting; extract GNSS data, meteorological data, and measured flood data based on the latitude and longitude of the hydrological section; divide the training set and test set, use the training set sample data to train the neural network model, continuously adjust the model parameters, and use the validation set data to evaluate the accuracy of the model's predicted rainfall data (root mean square error RMSE, mean absolute error MAE) to determine the optimal model parameters.

[0126] Specifically, based on the latitude and longitude of the hydrological section, GNSS data, meteorological data, and measured flood data for that location are extracted.

[0127] GNSS rainfall prediction data is raster data, with a time resolution of 10 minutes, a spatial resolution of 1 km * 1 km, and a spatial range of 100 km. 2 The dataset is divided into training and testing sets. When training the model, the input data is GNSS rainfall prediction data, and the output data is flood prediction data for the next period.

[0128] The model is validated using a test set, and metrics such as root mean square error (RMSE) and mean absolute error (MAE) are calculated to assess its accuracy and reliability. Based on the validation results, the model is fine-tuned, including adjusting model parameters and selecting more suitable features.

[0129] The following is a specific example: Hourly GNSS data from four base stations in a CORS network of a province in China, collected from September 1, 2014 to August 31, 2015, were used for experiments. The method proposed in this invention was used for short-term rainfall prediction and flood forecasting, and the results were verified by combining hourly rainfall and flood runoff data from corresponding hydrological stations.

[0130] The analysis of the four rainfall events selected in this example shows the changes in the accuracy of the predicted rainfall area over the next hour, with each event occurring 6 minutes later. The prediction accuracy ranged from 68.6% to 80.9% at 30 minutes; however, it decreased to 40.2% to 69.8% at 60 minutes. Because clouds are constantly moving, deforming, splitting, and merging, their spatial dispersion becomes increasingly pronounced with increasing prediction time, making the prediction of the predicted rainfall area increasingly difficult.

[0131] Table 1. Changes in the accuracy of predicted rainfall areas

[0132]

[0133] The accuracy of the magnitude of four rainfall events was verified, and the relationship between the model-predicted rainfall and the measured rainfall during the forecast period was analyzed. As shown in Table 2, the correlation coefficient ranged from 0.21 to 0.54, the BIAS ranged from -7.57 to 2.12 mm, the RMSE ranged from 3.19 to 21.47 mm, the average absolute value of the relative error was 28.92%, and the overall magnitude accuracy was 71.08%.

[0134] Table 2 Accuracy Analysis of Predicted Rainfall Levels

[0135]

[0136] Select a single rainfall event to calculate the cross-sectional flood, and analyze the relationship between the model-predicted flood and the measured flood during the forecast period. Figure 4 As shown, the model predictions are largely consistent with the measured values, with a correlation coefficient of 0.553.

[0137] This invention has the following features and effects:

[0138] 1. Multi-source data fusion and high-precision water vapor inversion: Combining GNSS data (including raw observation data and IGS precise orbital clock error), meteorological data, and geographic data, PWV is estimated through real-time PPP observation equations and the Saastamoinen model. Combined with three-dimensional water vapor tomography (discrete grid generation, tomography equation set construction, and SVD solution), the three-dimensional spatial water vapor accumulation and transport process before precipitation is precisely captured, providing high-precision input for precipitation inversion.

[0139] 2. Deep learning model optimizes forecast accuracy: The Transformer model is used to fuse PWV, water vapor density, meteorological and topographic data. The "time series embedding-encoding-decoding" mechanism is used to improve the spatiotemporal correlation of rainfall forecast. The CNN-LSTM model is combined to extract the spatial features of rainfall and capture the time series dependence. The attention mechanism is used to weight and converge key features, which effectively improves the accuracy of flood forecast (correlation coefficient reaches 0.553).

[0140] 3. Extended forecast period and all-weather applicability: Utilizing the all-weather observation advantage of GNSS signals, which are not affected by clouds and rain, and combining the rapid response capabilities of three-dimensional water vapor dynamic monitoring and deep learning, the forecast period for rain and flood is extended, providing a scientific basis for watershed flood prevention and early warning.

[0141] The above description is merely a specific embodiment 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 flood forecasting method based on GNSS water vapor tomography and deep learning, characterized in that... This includes the following steps: Step S1: Collect GNSS data, meteorological data, flood data, and geographic data; Step S2: Based on the GNSS data collected in step S1, the zenith tropospheric delay ZTD is obtained through the real-time PPP observation equation, the zenith dry tropospheric delay ZHD is calculated using the Saastamoinen model, and the zenith wet delay ZWD is calculated through the difference between ZTD and ZHD. The ZWD is then converted into atmospheric precipitation PWV. Step S3: Based on the GNSS data and geographic data collected in Step S1, the target area is discretized into grids, a set of GNSS water vapor tomography equations is constructed, and the water vapor density value of each grid is obtained by solving the set of GNSS water vapor tomography equations. Step S4: Using the PWV obtained in Step S2 and the water vapor density value obtained in Step S3, combined with the meteorological data and geographical data in Step S1 as input, construct a Transformer deep learning model, train the model and output rainfall prediction data. Step S5: Using the rainfall prediction data obtained in step S4 as input, construct a CNN-LSTM deep learning model, train the model, and output cross-sectional flood forecast data.

2. The flood forecasting method based on GNSS water vapor tomography and deep learning as described in claim 1, characterized in that, In step S1, the GNSS data includes raw GNSS observation data and IGS precise orbital clock error data; the meteorological data includes in-situ synchronized temperature, air pressure, humidity, and rainfall; the flood data includes flood level and flood flow at the control section; and the geographic data includes topographic elevation, latitude and longitude, and section location.

3. The flood forecasting method based on GNSS water vapor tomography and deep learning as described in claim 1, characterized in that, Step S2 specifically includes the following steps: Step S2-1: Obtain the Zenith Tropospheric Delay (ZTD) using the real-time PPP observation equation; The real-time PPP observation equation is expressed as follows: ； ； By incorporating the IGS precise satellite ephemeris and clock bias, the final positioning error equation is obtained, expressed as: ； ； In the above formulas, , These are pseudorange and phase observations. The superscript ' indicates the corresponding observations after precise satellite ephemeris and clock error correction. dt represents the geometric distance from the satellite to the station; dT represents the satellite clock error, and dT represents the receiver clock error. Represents tropospheric parameters; Represents multipath effects; N represents the integer ambiguity parameter of the ionosphere-free combination. The random error representing the observed value; the ppp positioning error term calculated using the positioning error equation is used as the observation to estimate the zenith tropospheric delay (ZTD); Step S2-2: Calculate the zenith tropospheric dry delay ZHD using the Saastamoinen model. The expression is: ； ； Among them, P s Here, φ represents atmospheric pressure, φ represents latitude, and H represents the altitude of the station. Step S2-3: Calculate the zenith tropospheric wet delay ZWD and convert it to PWV; the zenith wet delay ZWD equals the total zenith delay ZTD minus the zenith dry delay ZHD. Multiply the zenith wet delay ZWD by the conversion factor. The atmospheric precipitation potential (PWV) is expressed as follows: ； Zenith wet delay (ZWD) is converted into total rainable volume (PWV) through calculation, using the conversion factor in the formula. The expression is: ； Where, ρ w R represents water vapor density. v M represents the water vapor gas constant. w M d K1, K2, and K3 represent the molar masses of water vapor and dry air molecules, respectively, and represent the atmospheric refractive indices: K1 = 77.6890 K / hPa, K2 = 71.2952 K / hPa, and K3 = 375463 K. 2 / hPa; Representing the weighted average temperature of the atmosphere, the expression is: ； Where e represents water vapor pressure, T represents absolute temperature, and T d Represents the dew point temperature, e s e represents the saturated vapor pressure. s0 The saturated water vapor pressure at 0 degrees Celsius is 6.11 hPa. For water surfaces, a0 = 7.5 and b0 = 273.

3. For ice surfaces, a0 = 9.5 and b0 = 265.

7. It is a dimensionless numerical value, and its value is related to T. m The size is related, T m Using ground or radiosonde data to study T m Numerical integration is performed to calculate, and then the ground temperature is used to correlate with T. m Perform statistical modeling.

4. The flood forecasting method based on GNSS water vapor tomography and deep learning as described in claim 1, characterized in that, Step S3 includes the following steps: Step S3-1: Discretize the target area into a mesh to obtain a 3D voxel block: This includes setting the target area height, determining the horizontal and vertical resolutions; setting the heights of the bottom and top of the mesh to 0km and 11km respectively; the horizontal and vertical resolutions of the horizontal mesh division are determined according to application requirements. Step S3-2: Construct the GNSS water vapor tomography equation set, the expression of which is: ； in, This represents the intercept vector of the GNSS signal within each voxel block. X is the vector of SWV observations; X is the vector of water vapor density. The coefficient matrix of the horizontal constraint follows the characteristics of a Gaussian weighted function; The coefficient matrix of the vertical constraint follows an exponentially decreasing characteristic. Step S3-3: Perform three-dimensional tomography equation solving to obtain the water vapor density value for each grid; specifically, the SVD method is used for solving, and the expression is: ； In the formula, P is a matrix composed of adjustment constraints and observation information weighting factors; L is the constant matrix of the tomographic equation system, expressed as: 。 5. The flood forecasting method based on GNSS water vapor tomography and deep learning as described in claim 1, characterized in that, Step S4 includes the following steps: Step S4-1: Construct a Transformer neural network model using PWV, water vapor density value, meteorological data, and geographical data as input layers, and predicted rainfall as the output layer, and perform "time series embedding-encoding-decoding" calculations. First, use the Embedding time series embedding layer to project the input layer matrix into another dimension space; second, use a Transformer-based encoder to transform the input time series into a fixed-dimensional feature map; finally, use a Transformer-based decoder to decode the feature map and generate future time series. Step S4-2: Conduct neural network training, obtain the optimal model parameters, and perform rainfall prediction; The model is divided into training and testing sets. The neural network model is trained using the sample data in the training set, and the model parameters are continuously adjusted. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators. The accuracy of the model's prediction of rainfall data is evaluated using the validation set data, and the optimal model parameters are determined.

6. The flood forecasting method based on GNSS water vapor tomography and deep learning as described in claim 1, characterized in that, Step S5 includes the following steps: Step S5-1: Construct a CNN-LSTM neural network model, including an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, and an output layer. The input layer receives rainfall prediction data. The CNN layer extracts the spatial features of the rainfall prediction data, which are then used as the input to the LSTM layer. The LSTM layer recursively processes the spatial features of the rainfall prediction data step-by-step, outputting a feature vector containing time-series information. The attention mechanism layer performs weighted aggregation on the feature vectors output by the LSTM layer to obtain a context feature vector. The output layer converts the context feature vector output by the attention mechanism layer into the final prediction result, obtaining the flood change trend over a future period. Step S5-2: Conduct neural network training to obtain optimal model parameters for flood forecasting; extract GNSS data, meteorological data, and measured flood data based on the latitude and longitude of the hydrological section; divide the training set and test set, use the training set sample data to train the neural network model, continuously adjust the model parameters, use the root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators, use the validation set data to evaluate the accuracy of the model's predicted rainfall data, and determine the optimal model parameters.

Images