Search

CN-121997740-A - Runoff simulation method and system for coupling SWAT model and machine learning algorithm

CN121997740ACN 121997740 ACN121997740 ACN 121997740ACN-121997740-A

Abstract

The invention provides a runoff simulation method and a system for coupling SWAT and a machine learning algorithm, which comprise the steps of collecting and preprocessing river basin space data, meteorological data and actually measured runoff data, taking intermediate physical variables and meteorological data output by a SWAT model as characteristic variables of ML, learning a directional converging topological relation among sub-river basins output by the SWAT model by a graphic annotation force network (Graph Attention Networks, GAT), capturing hydrological time sequence dependence by a Long Short-Term Memory (LSTM), optimizing super parameters by a grid search method, building SL (SWAT+LSTM) and SGL (SWAT+GAT+LSTM) coupling models, and evaluating the value of the coupling models for improving the runoff simulation precision. The invention combines the advantages of physical mechanism and data processing, solves the problems of low simulation precision of the traditional model and incapacity of the spatial association description of the traditional coupling model, realizes the accurate simulation of extreme flood events, and can provide scientific support for the optimal configuration of water resources and the prevention and control of flood disasters.

Inventors

  • ZHANG YUKUN
  • ZHANG QUAN

Assignees

  • 武汉大学
  • 张雨坤

Dates

Publication Date
20260508
Application Date
20260126

Claims (10)

  1. 1. A method of radial flow simulation coupling a SWAT model with a machine learning algorithm, comprising: Collecting and preprocessing meteorological data, actual measurement daily runoff data of hydrologic stations, digital elevation data of DEM, land utilization data and soil type data; And inputting the pretreated DEM digital elevation data into a Soil and Water Assessment Tool (SWAT) model, automatically extracting a research basin boundary vector file, dividing sub-basins and calculating a confluence relation among the sub-basins. The preprocessed land utilization data and the preprocessed soil type data are cut by the research river basin boundary vector file and then input into a SWAT model together with gradient data, so that hydrological response unit (Hydrological Response Unit, HRU) division is completed; Driving a SWAT model built in a research area by using the meteorological data, and calibrating the SWAT model by using the real-time daily runoff data of the hydrologic station at regular dividing rate and a checking period to obtain an optimal parameter combination suitable for a research river basin; and obtaining an intermediate physical variable output by the SWAT model and simulating a daily runoff process. Taking the flood peak flow as an evaluation index, and acquiring a simulated typical flood process in a research period; Taking the meteorological data and the intermediate physical variable output by the SWAT model as machine learning characteristic variables, taking the measured daily runoff data of the hydrologic site as target variables, and GAT learning the converging relation among the sub-waterbasins and adaptively distributing weights. Hydrologic process timing dependencies are captured by LSTM. Constructing a SL (SWAT+LSTM) model and an SGL (SWAT+GAT+LSTM) model; The data is divided into training sets and test sets. Determining super parameters of the SL model and the SGL model by a grid search method, and obtaining a daily runoff process output by a coupling model SL and an SGL constructed in a research river basin and a simulated typical flood process of the coupling model in a research period; the Nash efficiency coefficient NSE, the root mean square error RMSE and the relative root mean square error RRMSE are used as evaluation indexes. And evaluating the runoff simulation precision of the hydrological model and the coupling model constructed in the research river basin.
  2. 2. The runoff simulation model of claim 1, wherein the acquiring and preprocessing of meteorological data, hydrographic site measured daily runoff data, DEM digital elevation data, land use data, soil type data comprises: The DEM digital elevation data adopts SRTMDEM data sets, has the spatial resolution of 90m multiplied by 90m, and is subjected to mosaic, projection and depression filling pretreatment. The land utilization data is a Chinese land utilization remote sensing monitoring data set, the resolution is 30m multiplied by 30m, and the data is preprocessed by merging, projecting, cutting, reclassifying and constructing an index table. The soil data is derived from HWSD v 1.2.2 databases, the resolution is 1000m multiplied by 1000m, and the data is preprocessed by projection, cutting, reclassification, parameter supplementing calculation and database importing. The meteorological data is derived from a daily value data set V3.0 of the Chinese ground climate data, and comprises daily precipitation, air temperature, relative humidity, solar radiation and wind speed in a research period, and is subjected to format standardization, quality control and index table construction pretreatment. The actual measurement daily runoff data of the hydrologic station are derived from hydrologic yearbook of the people's republic of China and are subjected to outlier correction pretreatment.
  3. 3. The runoff simulation model for coupling a SWAT model and a machine learning algorithm according to claim 1, wherein the preprocessed DEM digital elevation data is input into the SWAT model, and a research basin boundary vector file is automatically extracted, sub-basins are divided, and a confluence relation among the sub-basins is calculated. Inputting the preprocessed land use data and the preprocessed soil type data into a SWAT model together with gradient data after being cut by the research river basin boundary vector file, and completing hydrologic response unit (Hydrological Response Unit, HRU) division, comprising: And (3) inputting DEM data and setting a minimum water collecting area threshold, extracting a river channel network through a WATERSHED DELINEATION tool, finishing the division of the sub-river basin of the research river basin, and defining the upstream afflux relation of the sub-river basin. Dividing the river basin gradient into two groups, setting area thresholds of land utilization type, soil type and gradient grouping by combining the preprocessed land utilization data and the preprocessed soil data, and dividing the land utilization type, the soil type and the gradient grouping into hydrological response units.
  4. 4. The runoff simulation model of coupling a SWAT model and a machine learning algorithm according to claim 1, wherein the meteorological data is used to drive the SWAT model, the division rate is periodic and the inspection period, the SWAT model is calibrated by the hydrologic site actual measurement daily runoff data, and an optimal parameter combination suitable for a study basin is obtained, and the method comprises the following steps: The basin sensitivity parameters are firstly screened and researched through SUFI-2 algorithm. Setting an initial range of parameters, adopting an SCE-UA composite evolution algorithm, maximizing NSE at regular intervals to obtain an objective function, and setting a convergence threshold value of 0.001 to obtain a group of optimal parameters of the SWAT model in a research river basin.
  5. 5. The runoff simulation model of claim 1, wherein the intermediate physical variables output by the SWAT model are obtained to simulate a daily runoff process. Taking the flood peak flow as an evaluation index, obtaining a simulated typical flood process in a research period, comprising: and extracting typical flood processes in a research period from actual measurement daily runoff data by taking flood peak flow as a sequencing basis, wherein the extracted and verified SWAT model simulates the runoff of the flood processes, the type of flood peaks and the peak time.
  6. 6. The runoff simulation model for coupling a SWAT model and a machine learning algorithm according to claim 1, wherein the meteorological data and intermediate physical variables output by the SWAT model are used as machine learning characteristic variables, the measured daily runoff data of the hydrologic site are used as target variables, and the GAT learns the converging relationship among the sub-waterbasins, so that weights are distributed in a self-adaptive manner. Hydrologic process timing dependencies are captured by LSTM. Constructing a SL (SWAT+LSTM) model and an SGL (SWAT+GAT+LSTM) model, comprising: The feature variable and the target variable used by ML are normalized to the [0,1] interval by Min-Max Scaling. The training set and the testing set are divided according to a principle of no disturbance according to time sequence, the training set is divided into a training subset and a verification subset according to a ratio of 9:1, the SL model is constructed based on PyTorch frames and comprises an input layer, two LSTM hidden layers, a BatchNorm d layer and an output layer, a loss function is MSE, and an optimizer selects Adam. The SGL model introduces a GAT layer on the basis of the SL model, generalizes the sub-watershed into directed graph nodes, the converging direction is directed connection among the nodes, and calculates the differential contribution weight of the upstream sub-watershed through the GAT.
  7. 7. The runoff simulation model of claim 1, wherein the data is divided into a training set and a test set. Determining super parameters of the SL model and the SGL model through a grid search method, and obtaining a daily runoff process output by a coupling model SL and an SGL constructed in a research river basin and a simulated typical flood process of the coupling model in a research period, wherein the method comprises the following steps: The super-parametric candidate range includes LSTM hidden layer neuron number [32, 64, 128], learning rate [0.0001, 0.001, 0.01], dropout rate [0.1, 0.2, 0.3], batch size [16, 32, 64], GAT hidden layer dimension [32, 64, 128], GAT attention header number [4, 8, 16]. Traversing the super-parameter combination through GRIDSEARCHCV tools to verify that the subset NSE is the maximum optimal target, and determining the optimal super-parameter combination. The model training adopts an early-stopping strategy and a learning rate adjustment strategy, and a daily runoff simulation value and four-flood simulation process data are output after training is completed.
  8. 8. The runoff simulation model for coupling a SWAT model and a machine learning algorithm according to claim 1, wherein the step of evaluating the simulation capability of the hydrologic model SWAT and the coupling models SL and SGL to the daily runoff process and the typical flood process of the han river basin by selecting an appropriate hydrologic evaluation index comprises: NSE, RMSE, RRMSE, peak time errors are used as hydrologic evaluation indexes. In the formula, Representing the actual observed radial flow for the ith time step, The model simulation runoff quantity representing the ith time step, N is the total number of time steps in the study period.
  9. 9. A runoff simulation system coupling a SWAT model and a machine learning algorithm, comprising: the data preprocessing module is used for acquiring and preprocessing DEM data, land utilization data, soil type data, meteorological data and actual measurement runoff data; the SWAT modeling module is used for constructing a SWAT model in a research river basin and acquiring intermediate physical variables of the hydrological model for the coupling model; the coupling model building module is used for simulating runoffs by taking the intermediate physical variable of the hydrological model and the meteorological data as characteristic variables and adopting a machine learning algorithm, and meanwhile, considering the converging relation among sub-waterbasins to obtain coupling models SL and SGL; the model output evaluation module is used for evaluating the effect of constructing a hydrologic model and a machine learning algorithm on improving the runoff simulation effect of the research river basin.
  10. 10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements a radial flow simulation method of coupling a SWAT model physical mechanism with a machine learning algorithm as claimed in any one of claims 1 to 8 when the program is executed by the processor.

Description

Runoff simulation method and system for coupling SWAT model and machine learning algorithm Technical Field The invention relates to the fields of machine learning algorithms, hydrologic models, runoff simulation and the like, in particular to a runoff simulation method and system for coupling a SWAT model and a machine learning algorithm. Background The vulnerability of water resources is continuously improved, and the shortage of water resources and extreme drought and waterlogging disasters become core challenges for restricting the sustainable development of areas. The runoff is used as a key quantization index in the water circulation process, the accurate runoff simulation is a key means for understanding a land water circulation mechanism and predicting a future hydrologic evolution rule, a quantization basis can be provided for scientific regulation and optimization configuration of water resources of various watercourses, and the runoff simulation is used as a key technical support for establishing a high-efficiency disaster prevention and reduction system and guaranteeing stable operation of society and economy, and has important strategic significance for supporting sustainable development of watercourses. The existing runoff simulation technology is mainly divided into a physical hydrological model and a machine learning (MACHINE LEARNING, ML) model, but has the obvious defects that the physical hydrological model simulates a basin rainfall runoff process through a coupling structure and parameters, the simulation precision depends on parameter calibration and data quality, and the process of parameter optimization consumes more time and is difficult to find a globally optimal solution. Furthermore, the physical model has insufficient simulation ability for the runoff process. The strong nonlinear fitting capability of ML can flexibly adapt to runoff simulation requirements under different hydrologic conditions. Long Short-Term Memory (LSTM) networks are widely used in runoff simulation tasks in the hydrologic field by virtue of their advantages in processing time series data. Existing researchers have considered the spatiotemporal relevance of the runoff process and applied machine learning algorithms that couple LSTM with reflecting topological relationships to the hydrologic field. ML lacks physical mechanism support and has poor interpretability. Coupling the physical mechanisms of hydrologic models with machine learning has become a central research hotspot for hydrologic simulation. At present, the existing researches generally have the following problems that (1) most models adopt an undirected graph algorithm to describe spatial association, cannot adapt to the directional confluence characteristic of the upstream and downstream of a river basin, and (2) part of models are not fully integrated into physical process variables, and still take meteorological data as main input, so that the relevance of physical logic and simulation results is insufficient. In general, existing studies fail to adequately couple the simulation capabilities of hydrologic models on physical processes and the processing advantages of machine learning algorithms on spatiotemporal data. Disclosure of Invention The invention provides a runoff simulation method and a runoff simulation system for coupling a SWAT model and a machine learning algorithm, which are used for solving the problem that the simulation capability of a hydrologic model on an actual production converging physical process and the processing advantage of the machine learning algorithm on space-time data are insufficient in the prior art, so that the simulation capability on a daily runoff process and a flood process is insufficient. In a first aspect, the present invention provides a method of radial flow simulation coupling a SWAT model with a machine learning algorithm, comprising: And acquiring space data of a study river basin, meteorological data in a study period and actual measurement runoff data of the hydrological station. Wherein the spatial data includes digital elevation data (Digital Elevation Model, DEM), land use data, and soil type data. The spatial data input to the SWAT model is projected to the same coordinate system. The DEM data are longitude and latitude block data, and the preprocessed DEM data are obtained after three steps of mosaic, projection and depression filling. The land utilization data are needed to be combined, projected, reclassified and made into a land utilization type index table according to provinces and cities, and the preprocessed land utilization data are obtained. And the soil data is projected and cut. And merging soil groups according to the SU_SYM90 field in the HWSD soil database in the research flow field, perfecting the soil database, and establishing a soil type index table to obtain the preprocessed soil data. Inputting the pretreated DEM data into a SWAT model to extract a river channel and divid