Deep learning-based flood runoff forecasting method and system

Through deep learning methods, combined with multiple data types, the CNN-GRU hybrid network structure is used to predict flood runoff, which solves the problem of inaccurate flood runoff prediction in the prior art and improves prediction accuracy and generalization capabilities.

CN118917483APending Publication Date: 2024-11-08XIAN UNIV OF TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411035008.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict flood runoff, and a single model cannot fully consider the correlation between multiple factors, resulting in inaccurate prediction results.

Method used

Using a deep learning-based method, a runoff forecast model is constructed through a convolutional neural network and a gated cycle unit. Combined rainfall, evaporation, terrain, land use and soil cover data, it is trained using a CNN-GRU hybrid network structure to obtain flood runoff prediction. variable.

Benefits of technology

The generalization ability and prediction accuracy of the model are improved, and the mapping relationship between rainfall, evaporation, terrain, land use and soil cover data and flood flow is established, effectively solving the optimization and search problems in runoff forecasting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917483A_ABST
    Figure CN118917483A_ABST
Patent Text Reader

Abstract

The invention provides a flood runoff forecasting method and system based on deep learning, and the method comprises the steps: collecting original data, carrying out the preprocessing of the original data, and constructing model input data; the model input data comprises rainfall data, evaporation data, flood flow data, topographic data, land utilization data and soil coverage data; constructing a runoff forecasting model by adopting a convolutional neural network and a gating circulation unit, and training the runoff forecasting model; and inputting the model input data into the trained runoff forecasting model, and obtaining a flood runoff forecasting variable based on the runoff forecasting model. According to the method, the strong feature extraction and nonlinear fitting capability of deep learning is utilized, the mapping relation between rainfall data, evaporation data, topographic data, land utilization data, soil coverage data and the flood flow of the target hydrometric station is established, and the generalization capability and precision of a flood flow prediction model are achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of river runoff forecasting, and in particular to a flood runoff forecasting method and system based on deep learning. Background Art

[0002] As a precious freshwater resource, floods will generate huge benefits if the runoff process can be obtained in advance and reasonably dispatched. Runoff forecasting is the basis for scientific planning, reasonable allocation, and adaptive utilization of water resources. Through accurate runoff forecasting, river flow in the future can be predicted, so as to reasonably formulate water resource allocation and dispatching plans. This is of great significance for improving the utilization efficiency of water resources, reducing water resource waste, and meeting the water resource needs of different regions and industries.

[0003] However, the formation and evolution of floods are affected by many factors, including atmospheric circulation, ocean tides, geophysical factors, etc. The interaction between these factors makes the flood system very complex. A single model may not be able to fully consider the impact of these factors and the relationship between various factors, resulting in inaccurate prediction results. Summary of the invention

[0004] In view of this, the present invention proposes a flood runoff forecasting method and system based on deep learning to solve the problems existing in the above-mentioned prior art.

[0005] On the one hand, to achieve the above-mentioned purpose, the present invention proposes a flood runoff forecasting method based on deep learning, comprising the following steps:

[0006] Collecting raw data and preprocessing the raw data to construct model input data; the model input data includes rainfall data, evaporation data, flood flow data, terrain data, land use data, and soil cover data;

[0007] A runoff forecast model is constructed by using a convolutional neural network and a gated recurrent unit, and the runoff forecast model is trained;

[0008] The model input data is input into the trained runoff forecast model, and flood runoff prediction variables are obtained based on the runoff forecast model.

[0009] Optionally, the construction process of the rainfall data includes:

[0010] The observation data of several rainfall stations are obtained, and the length n of the observation data is averaged according to the statistical period, and filled into the n time periods before the observation deadline.

[0011] Optionally, the process of constructing the evaporation data includes:

[0012] The daily average evaporation observation values ​​of several stations are obtained, and the daily average evaporation observation values ​​are averaged to obtain the evaporation data.

[0013] Optionally, the process of constructing the flood flow data includes:

[0014] Acquire flood element excerpt data in several years, use SG noise reduction method to reduce the noise of the flood element excerpt data, use cubic spline interpolation method to interpolate the flood element excerpt data, and obtain flood process data;

[0015] The SG denoising method is used to denoise the flood process data, and cubic spline interpolation is used to interpolate the denoised data to obtain the flow process lines of each flood.

[0016] Optionally, the process of constructing the terrain data includes:

[0017] The digital elevation model data of the study area is obtained, the digital elevation model data is processed by using GDAL, and elevation mapping is performed to generate a grid image, and the grid image is used as terrain data.

[0018] Optionally, the process of constructing the land use data includes:

[0019] The land use data of the study area are obtained, the land use type of the study area is obtained based on the land use data, and a land use raster image is constructed based on the land use type and the land use data.

[0020] Optionally, the process of constructing the soil cover data includes:

[0021] A soil data set is obtained according to a soil database, soil types and soil distribution conditions of a study area are obtained based on the soil data set, and a soil coverage grid map of the study area is constructed based on the soil types and the soil distribution conditions.

[0022] Optionally, the runoff prediction model is a CNN-GRU hybrid network structure. During the training of the runoff prediction model, a TPE algorithm based on decision tree and Bayesian optimization is used to optimize the model hyperparameters of the CNN-GRU hybrid network structure. The training set MAE, MAPE, Loss, R 2 The evaluation index evaluates the model's ability to learn features, uses the test set loss as an indicator to evaluate the model's generalization ability, uses the ADAM optimizer to optimize the model's training, and uses early stopping and learning rate decay learning strategies to obtain the model with the highest fitting accuracy and the strongest generalization ability.

[0023] Optionally, the process of obtaining flood runoff prediction variables based on the runoff forecast model includes:

[0024] Label the data points in the rainfall data and evaporation data with the flood number and flood event number, and add the latitude and longitude information of the monitoring station;

[0025] Integrate and filter the labeled data to obtain rainfall and evaporation data at each station at each moment during the flood process, and retain the label information of rainfall and evaporation data as well as flood time and geographical location;

[0026] Inputting the rainfall data and evaporation data into the runoff forecast model, extracting the spatial features of the data based on a one-dimensional convolution kernel, generating a feature map, and downsampling the feature map through a pooling layer;

[0027] The downsampled and filtered feature map is input into the gated recurrent unit to generate time series features;

[0028] Based on another convolutional neural network, feature extraction is performed on the terrain data, land use data and soil cover data to generate one-dimensional feature data;

[0029] The time series features and the one-dimensional feature data are merged based on a fully connected layer to obtain the flood flow of the hydrological station, that is, the flood runoff prediction variable.

[0030] On the other hand, to achieve the above-mentioned purpose, the present invention proposes a flood runoff forecasting system based on deep learning, comprising a processor and a computer program stored in a memory and capable of running on the processor, wherein the processor implements the above-mentioned method when executing the computer program.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention utilizes the powerful feature extraction and nonlinear fitting capabilities of deep learning to establish a mapping relationship between rainfall data, evaporation data, terrain data, land use data, soil cover data and flood flow of the target hydrological station. The terrain data, soil cover data and land use data are used as inputs of the convolutional neural network in the form of pictures, and are integrated with other one-dimensional features (rainfall, evaporation data) in the fully connected layer, ultimately realizing the prediction of flood flow, effectively solving the optimization and search problems in runoff forecasting, and improving the generalization ability and prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0034] Figure 1 It is a schematic diagram of the flow of a flood runoff forecasting method based on deep learning in an embodiment of the present invention;

[0035] Figure 2 It is a statistical diagram of rainfall data during the process of the embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of process data of evaporation data in an embodiment of the present invention;

[0037] Figure 4 is a schematic diagram of flood flow interpolation results in an embodiment of the present invention;

[0038] Figure 5 is a terrain grid image in an embodiment of the present invention;

[0039] Figure 6 A land use distribution map in an embodiment of the present invention;

[0040] Figure 7 The land use grid image in the embodiment of the present invention;

[0041] Figure 8 A soil cover distribution map in an embodiment of the present invention;

[0042] Fig. 9 A soil cover grid map in an embodiment of the present invention;

[0043] Fig.10 Schematic diagram of the model structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to be able to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] This embodiment proposes a flood runoff forecasting method based on deep learning, such as Figure 1 As shown, the steps include:

[0046] Collect and pre-process the raw data to construct model input data; the model input data includes rainfall data, evaporation data, flood flow data, terrain data, land use data, and soil cover data;

[0047] A convolutional neural network and gated recurrent unit were used to construct a runoff forecast model and train the runoff forecast model.

[0048] The model input data is input into the trained runoff forecast model, and flood runoff prediction variables are obtained based on the runoff forecast model.

[0049] The specific implementation process of the above method is as follows:

[0050] Model Overview

[0051] Raw data:

[0052] GIS data (DEM, watershed, river and hydrological station, etc.), hydrological and meteorological data from 2007 to 2012 (water data from hydrological stations, rainfall data from rainfall stations and evaporation data).

[0053] Model input and output:

[0054] The rainfall data, evaporation data, topographic map, soil distribution map and land use map are used as the input data of the model, and the flow of hydrological station B is used as the output of the model.

[0055] Model building:

[0056] This model uses part of the available data in the original data as the model input, and uses the SG denoising and multiple spline interpolation methods to smooth and denoise the rainfall and evaporation data and obtain the time process data of various data. At the same time, based on the GIS data in the original data, Python and GDAL are used to generate terrain, soil distribution and land use layers that can be used for convolutional neural network input. When dividing the data set, the relevant data from 2007 to 2010 are used to train the model, and the flood data from 2011 to 2012 are used to test the generalization ability of the model.

[0057] When designing the model infrastructure, a convolutional neural network (CNN) was first used to extract features from rainfall data from 45 rain gauges and evaporation data from 6 meteorological stations to adaptively learn the potential relationship between these data. The features extracted by the CNN were then used as input to the gated recurrent unit (GRU) time series prediction model. The GRU learns the changing characteristics in the time series by updating its gated units, and then the output of the GRU layer is concatenated with the convolution results of the terrain, soil distribution, and land use layers using a fully connected layer, and finally a regression model between the input data and the target variable (i.e., flood flow) is constructed.

[0058] In order to ensure that the model has the best simulation accuracy and good generalization ability, the TPE algorithm based on decision tree and Bayesian optimization is used to optimize the model hyperparameters of the CNN-GRU hybrid network structure.2 Model accuracy evaluation indicators such as sparse_ ...

[0059] First, the original data was preprocessed by noise reduction, interpolation and normalization to ensure that the data met the model input requirements. Secondly, the hyperparameter optimization technology and ADAM algorithm were used to optimize the model training process to improve the accuracy and generalization ability of the model in flood flow prediction. Then, the model was used to predict and analyze the flood event in 2013. Finally, the key influencing factors and influencing range of the hydrological station runoff were further determined by performing sensitivity analysis on the model input characteristics.

[0060] Model input and output data preprocessing

[0061] The input data of this flood prediction model mainly includes two categories: hydrological and meteorological data and GIS data. The hydrological and meteorological data selects the rainfall and evaporation process data after noise reduction and interpolation, and stores it in the form of Excel tables in "xlsx" format, and then uses Pandas1.4.0 to read and use the data. For geographic information data, Python is used as the development environment to read, write and convert such data with GDAL, and finally obtains the terrain, soil distribution, and land use layers suitable for the CNN-LSTM hybrid network model. Such data is stored in the "png" format with a resolution of "5760×3180".

[0062] The model uses a specific time step d as the length of the time series, rainfall and evaporation process data from td to t as the model input, and flow data at t+1 as the model output. The basic summary of input and output data is shown in Table 1.

[0063] Table 1

[0064]

[0065]

[0066] In addition to the given data, this modeling adds land use and soil distribution data, which are downloaded from the China National Geographic Information Network and the China Soil Information Network. In order to adapt to the input size of the CNN-LSTM model, a series of preprocessing operations are performed on the original data. The preprocessing of various types of data is now introduced.

[0067] Preprocessing of rainfall data

[0068] Deep learning has the ability to automatically extract features and can adapt to the automatic feature extraction of high-dimensional data. Therefore, in the modeling process here, the data of all 45 rainfall stations are used as the input of the model, and the model adaptively learns the important features of these data.

[0069] In order to construct the temporal process data of rainfall observations, the rainfall observations of each raingauge station from 2007 to 2013 were averaged according to the length of its statistical period n and filled into the n periods before the observation cutoff time, thus constructing the temporal process data of rainfall at each station from 2007 to 2013.

[0070] The rainfall observation process data statistics of each station are as follows: Figure 2 As shown, the peak rainfall occurred at 18:00 on July 26, 2007 at station 7, with a value of 74.5 mm. In addition, the rainfall was concentrated in the range of 0-20 mm at most times, with a small amount occurring in the range of 20-60 mm.

[0071] Preprocessing of evaporation data

[0072] The original data is the daily average evaporation observation value. To match the input of the rest of the data, the daily average value is averaged to obtain Figure 3 The evaporation data shown are process data. The evaporation amounts at stations 2, 3, and 4 are relatively close, with the peak evaporation data occurring at station 3, with a value of 0.83 mm. In addition, the evaporation amount at most times is concentrated between 0 and 0.6 mm, with a small amount occurring between 0.6 and 0.8 mm.

[0073] Preprocessing of flood flow

[0074] The original data only contains the flood feature extracts from 2007 to 2012. To obtain a sample data set that can be used for deep learning, SG denoising is first used to reduce the noise of the original data, and then cubic spline interpolation is used to interpolate the flood observation data to obtain the time process data of the flood process. After processing, a total of 16 flood process data are obtained.

[0075] SG was used to reduce the noise of each flood process. The absolute error (MAE) and relative error (MAPE) of the noise-reduced data compared with the original data are shown in Table 2. The relative errors are basically less than 5%, indicating that the noise reduction results are relatively consistent with the original data. Then, cubic spline interpolation was used to interpolate the noise-reduced data to obtain the flow process lines of each flood, as shown in Figure 4 shown.

[0076] Table 2

[0077]

[0078] Preprocessing of terrain data

[0079] Topographic data plays a key role in the accuracy and reliability of hydrological models. It affects the formation of surface runoff and is very important in runoff prediction.

[0080] The study area uses a digital elevation model (DEM), covering the range from 33.77°N to 34.50°N, and 109.73°E to 111.50°E, using the WGS84 geographic coordinate system. The resolution of the DEM is 250 meters, and the elevation values ​​range from 386 meters to 2595 meters, with significant terrain undulations. In order to make the data suitable for the convolutional neural network model, GDAL is used to process the original DEM data, and Matplotlib's 'terrain' is used for elevation mapping. Finally, a grid image with a resolution of 5760×3180 pixels is generated. For specific images, see Figure 2 .

[0081] Preprocessing of land use

[0082] Land use data plays a vital role in runoff simulation. Different land use types have significantly different effects on water infiltration, evaporation and surface runoff, which in turn affect the formation of runoff.

[0083] The land use data of the study area was obtained from CLCD (China Land Cover Dataset) with a resolution of 30×30m. The study area mainly involves 7 types of land use, and the classification of the types and their respective proportions in the entire area are shown in Table 3. The regional land use distribution is shown in Figure 6 shown.

[0084] Table 3

[0085]

[0086] Finally, the land use was made into Figure 7 The land use grid image shown will eventually serve as the input file of the CNN-LSTM hybrid network model.

[0087] Preprocessing of soil cover data

[0088] Soil type not only affects water infiltration rate and water retention capacity, but also affects evaporation and vegetation water use efficiency, thereby directly and indirectly affecting runoff volume and runoff characteristics.

[0089] A 1:4 million soil dataset was obtained from the China Soil Database, and the soil type statistics of the study area are shown in Table 4. The specific soil distribution is shown in Figure 8 shown.

[0090] Table 4

[0091]

[0092]

[0093] The data was processed using Python and GDAL to distinguish different soil cover types with different chromaticities, and finally a soil cover grid map that can be input into the model was obtained, such as Fig. 9 shown.

[0094] Modeling Platform Introduction

[0095] The construction, optimization, and testing of the flood flow prediction model were all carried out on the NVIDIA GeForce RTX 4090 server of the Sky Information cloud computing platform. The main Python tool libraries and software information involved are shown in Table 5.

[0096] Table 5

[0097]

[0098] Introduction to model structure

[0099] For the flood simulation here, a CNN-GRU hybrid neural network model was constructed, considering the complex interactions of various input variables in spatial and temporal dimensions, as well as the integration of multiple data types, aiming to improve the accuracy and reliability of flood flow prediction. Fig.10 As shown, the model is briefly introduced below.

[0100] First, the rainfall and evaporation data containing time series information were preprocessed. Each data point was labeled with its flood number and the sequence number of the data point in the current flood event. At the same time, in order to strengthen the model's understanding of the geographical relationship between sites, corresponding geographic information, that is, the longitude and latitude information of each monitoring site, was added to each data point. Then, these labeled one-dimensional data were integrated together, and through screening and processing, the rainfall and evaporation data of each site at each flood process were obtained, and the label information of these data and the flood time and geographical location was retained.

[0101] Afterwards, these integrated data will be input into a deep learning model. The input data is processed using a one-dimensional convolution kernel to generate multiple feature maps, thereby extracting the spatial features of the data and increasing the dimension of the data layer by layer to obtain more feature information. Afterwards, the feature maps are downsampled through the pooling layer (MaxPooling) to retain the most important features and reduce the amount of data calculations to prevent the model from being too complex. These feature maps will then be used as input to the gated recurrent unit (GRU) module to learn the time series features in the data, thereby helping the model capture the dynamic changes in the development of the flood.

[0102] At the same time, another convolutional network (Conv2D) is used to process the distribution images of terrain, land use, and soil types. These image data are extracted through the convolution layer and flattened into one-dimensional data through the Flatten layer.

[0103] Finally, the features from the rainfall evaporation data and the features from the geographic image data are combined in a fully connected layer, which fuses the data from different sources and outputs the target prediction variable, i.e., the flood flow at hydrological station B.

[0104] The present invention utilizes the feature extraction of convolution kernels and the time series processing capabilities of gating units to capture the dynamic changes of various types of data and their temporal and spatial relationships during the development of floods. The terrain data, soil cover data, and land use data are used as inputs of the convolutional neural network in the form of pictures, and are integrated with other one-dimensional features (rainfall, evaporation data) in the fully connected layer to ultimately achieve the prediction of flood flow.

[0105] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or perform equivalent replacements on some of the technical features thereof; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. They should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A flood runoff forecasting method based on deep learning, It is characterized in that The following steps are involved: Collecting raw data and preprocessing the raw data to construct model input data; the model input data includes rainfall data, evaporation data, flood flow data, terrain data, land use data, and soil cover data; A runoff forecast model is constructed by using a convolutional neural network and a gated recurrent unit, and the runoff forecast model is trained; The model input data is input into the trained runoff forecast model, and flood runoff prediction variables are obtained based on the runoff forecast model.

2. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The construction process of the rainfall data includes: The observation data of several rainfall stations are obtained, and the length n of the observation data is averaged according to the statistical period, and filled into the n time periods before the observation deadline.

3. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The process of constructing the evaporation data includes: The daily average evaporation observation values ​​of several stations are obtained, and the daily average evaporation observation values ​​are averaged to obtain the evaporation data.

4. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The construction process of the flood flow data includes: Acquire flood element excerpt data in several years, use SG noise reduction method to reduce the noise of the flood element excerpt data, use cubic spline interpolation method to interpolate the flood element excerpt data, and obtain flood process data; The SG denoising method is used to denoise the flood process data, and cubic spline interpolation is used to interpolate the denoised data to obtain the flow process lines of each flood.

5. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The construction process of the terrain data includes: The digital elevation model data of the study area is obtained, the digital elevation model data is processed by using GDAL, and elevation mapping is performed to generate a grid image, and the grid image is used as terrain data.

6. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The process of constructing the land use data includes: The land use data of the study area are obtained, the land use type of the study area is obtained based on the land use data, and a land use raster image is constructed based on the land use type and the land use data.

7. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The process of constructing the soil cover data includes: A soil data set is obtained according to a soil database, soil types and soil distribution conditions of a study area are obtained based on the soil data set, and a soil coverage grid map of the study area is constructed based on the soil types and the soil distribution conditions.

8. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The runoff prediction model is a CNN-GRU hybrid network structure. During the training process of the runoff prediction model, the TPE algorithm based on decision tree and Bayesian optimization is used to optimize the model hyperparameters of the CNN-GRU hybrid network structure. The training set MAE, MAPE, Loss, R 2 The evaluation index evaluates the model's ability to learn features, uses the test set loss as an indicator to evaluate the model's generalization ability, uses the ADAM optimizer to optimize the model's training, and uses early stopping and learning rate decay learning strategies to obtain the model with the highest fitting accuracy and the strongest generalization ability.

9. The flood runoff forecasting method based on deep learning according to claim 1, It is characterized in that The process of obtaining flood runoff prediction variables based on the runoff forecast model includes: Label the data points in the rainfall data and evaporation data with the flood number and flood event number, and add the latitude and longitude information of the monitoring station; Integrate and filter the labeled data to obtain rainfall and evaporation data at each station at each moment during the flood process, and retain the label information of rainfall and evaporation data as well as flood time and geographical location; Inputting the rainfall data and evaporation data into the runoff forecast model, extracting the spatial features of the data based on a one-dimensional convolution kernel, generating a feature map, and downsampling the feature map through a pooling layer; The downsampled and filtered feature map is input into the gated recurrent unit to generate time series features; Based on another convolutional neural network, feature extraction is performed on the terrain data, land use data and soil cover data to generate one-dimensional feature data; The time series features and the one-dimensional feature data are merged based on a fully connected layer to obtain the flood flow of the hydrological station, that is, the flood runoff prediction variable.

10. A flood runoff forecasting system based on deep learning, It is characterized in that The method comprises a processor and a computer program stored in a memory and capable of running on the processor, wherein the processor implements the method according to any one of claims 1 to 9 when executing the computer program.

Citation Information

Cited By

  • Runoff prediction method and system based on interpretable Bayesian gating circulation unit

    CN119599195A

  • Runoff prediction method based on space-time heterogeneous graph neural network and terminal

    CN119940609A