CN115859823A - Time-varying parameter hydrological simulation method based on sliding window sub-period division - Google Patents

Abstract

The invention discloses a time-varying parameter hydrological simulation method based on sliding window subperiod division, which comprises the steps of collecting hydrological, meteorological and underlying surface data in a research flow domain; constructing a distributed hydrological model; reconstructing natural runoff of the drainage basin by using the model; setting a sliding window, and extending and dividing sub-periods forwards and backwards based on underlying surface data; calibrating optimal model parameters in stages, inputting corresponding meteorological data and LULC data as models, driving the models, and reconstructing a natural runoff sequence; screening sensitive parameters and selecting model parameters related to the height of the underlying surface by combining the physical significance of the parameters; fitting a complex nonlinear relationship between the model parameters and the watershed subsurface characteristic factors. The invention overcomes the defects that the prior art is limited by small sample number, weak sample continuity, neglecting the underlying surface condition of the drainage basin, only can be applied to a lumped model with simple structure and relatively few parameters, but can not be applied to a distributed hydrological model with finer model structure and relatively more parameters, and the like.

Description

Time-varying parameter hydrological simulation method based on sliding window sub-period division

Technical Field

The invention relates to the technical field of hydrological models, in particular to a time-varying parameter hydrological simulation method based on sliding window sub-period division.

Background

In conventional hydrologic simulation, it is often required that the hydrologic data have consistency so that the data or samples statistically satisfy the same distribution condition, and further, the variation of model parameters is often ignored in the hydrologic simulation, i.e., the model parameters are assumed to be invariant over time within a given period of time. However, the hydrological features of the region under the changing background are affected by three factors of the characteristics of the drainage basin, climate change and human activities, and the hydrologic consistency condition is destroyed, so that the assumption that the parameters of the hydrologic model are unchanged is no longer applicable. Although the model parameters change along with the climate change and the watershed characteristic change caused by human activities with the underlying surface change as the main characteristic, the understanding and quantitative description of the complex relationship between the model parameters and the change environment are relatively lacked at present, and the development of the time-varying parameter hydrological simulation method is more consistent with the current watershed situation in the unsteady state environment, thereby being beneficial to improving the hydrological simulation precision in the change environment and perfecting the cognition on the water circulation process.

The current time-varying parameter hydrological simulation method mainly comprises a data assimilation method, a model structure diagnosis method, a segmentation rate determination method and the like, for example, a time-varying parameter hydrological simulation based on the data assimilation method is proposed in a document modeling time-variant parameters of a two-parameter transparent water balance model [ J ]. Journal of Hydrology,2019,573, 918-936 (the doi number of the document SCI is 10.1016/j.jhydro.2019.04.027), such as Deng C. and the like, but the method is only applied to a lumped hydrological model which is relatively difficult to use and is applied to a distributed hydrological model which has more parameters and stronger physical properties because the lumped model parameters are relatively less and the structure is simple. Zhou L, et al, in the literature of diagnosis structural details of a hydraulic model by time-varying parameters [ J ] Journal of Hydrology,2022,605,127305 (the document SCI Doi No. 10.1016/J. Hydroyl. 2021.127305) propose a method for Diagnosing and improving a model structure using time-varying parameters, and since the model improved by this method has no need for input of an underlying surface, it is impossible to use a flow field underlying surface condition as an input factor of the model. Merz R. et al, in the references Time stability of probability model parameters for evaluating effects and analyses [ J ]. Water Resources research.2011,47 (2), W02531. (Doi No. 10.1029/2010WR009505 of SCI of the reference), propose methods of segmenting rate model parameters, but have the problems of limited data sequence length, sample segmentation parameters only representing the average state of the Time period, and different continuity degrees caused by different Time period divisions.

Therefore, it is urgently needed to develop a time-varying parameter hydrological simulation method which is not limited by the number of model parameters and the number of calibration samples, can be continuously applied and comprehensively considers the dynamic change process of the underlying surface of the basin, and provides a new technical support for realizing high-precision hydrological simulation and characterization of the basin under a changing environment and quantifying the relationship between parameters and the underlying surface.

Disclosure of Invention

The invention aims to overcome the defects that the prior art is limited by a small number of samples, low sample continuity, neglect of basin underlying surface conditions, only can be applied to a lumped model with a simple structure and relatively few parameters, but cannot be applied to a distributed hydrological model with a finer model structure and relatively many parameters, and the like, and provides a time-varying parameter hydrological simulation method based on sliding window sub-period division.

The invention adopts the following technical scheme:

the time-varying parameter hydrological simulation method based on sliding window sub-period division comprises the following steps:

the method comprises the following steps: collecting hydrology, meteorology, terrain and underlying surface data in a research basin, sorting station data according to a time sequence, and cutting spatial data according to a basin vector range;

step two: constructing a SWAT distributed hydrological model suitable for a research area based on the data information collected in the step one;

step three: reconstructing and researching natural runoff of the basin by using the hydrological model built in the step two;

the method comprises the following steps: carrying out hydrological long sequence non-consistency inspection by using a mutation inspection method, utilizing meteorological data before (namely a natural period) a mutation point to calibrate a hydrological model in the step two, and utilizing the calibrated hydrological model and meteorological data observed after the mutation point to reconstruct natural runoff in the whole period;

step four: selecting lambda year data to divide sub-period (SP) according to the length of continuous underlying surface data and meteorological data, extending the data before and after the underlying surface data, determining the start and stop of the sub-period sequence, sequentially sliding, sliding and constructing a SWAT model of each sub-period, and regarding SP _i (i =1,2,3 \8230;, α), taking 1 year as a preheating period, 0.7 (λ -1) year as a rate period, and 0.3 (λ -1) year as a verification period, and corresponding underlying surface data and sliding meteorological data to establish model input of α continuous sub-periods;

a distributed hydrological model with stronger physical property is taken as a research tool, continuously underlay LULC data for many years is taken as dynamic input of the underlay condition of the model, reasonable sliding years are set, and the LULC and meteorological data input are correspondingly connected to serve as model driving elements in multi-period hydrological simulation.

Step five: aiming at the alpha sub-periods, inputting meteorological data and land utilization/land coverage data according to the sequence start and stop divided in the fourth step, driving a model, automatically optimizing model parameters, simulating runoff corresponding to each sub-period in the third step, namely a natural runoff sequence reconstructed by the model, calibrating and verifying the model parameters respectively according to an optimal principle of an objective function, and obtaining an alpha group of optimal parameter sets of the model with the best simulation performance;

and on the basis of the sub-period divided in the step four, in each sub-period, preferably selecting the model parameters of the period, and simulating the runoff of each corresponding sub-period, namely simulating the natural runoff in the step three. The meteorological data and land utilization/land coverage data are input into the model, the output of the model is runoff, the historical hydrological data (historical observation runoff data) is used for calculating an objective function, namely, the runoff value simulated by the model is compared with the actual measurement runoff data and is calculated according to a certain objective function calculation formula, and the closer the simulated value is to the observed value, the better the representative parameter calibration result is and the better the model performance is.

Step six: analyzing the sensitivity of the parameters aiming at the parameters obtained in the step five, and screening out a plurality of parameters with high correlation with the characteristic factors of the underlying surface of the drainage basin in the sensitive parameters by combining the physical significance of the parameters;

step seven: and (5) fitting the relationship between the model parameters screened in the step six and the characteristic factors of the underlying surface of the drainage basin, and selecting an algorithm most suitable for representing the relationship between the model parameters and the characteristic factors of the underlying surface of the drainage basin through precision test, so as to establish a complex nonlinear relationship between the model parameters and the characteristic factors of the underlying surface of the drainage basin and obtain the optimal representation form of the parameters of the drainage basin.

Further, in the seventh step, comprehensively considering a plurality of remote sensing data sources of land utilization/land coverage and vegetation indexes, selecting a river basin underlying surface characteristic factor, considering that the relation between the model parameter and the underlying surface characteristic factor is difficult to accurately describe by a simple linear relation, fitting and verifying the relation between the model parameter and the river basin underlying surface characteristic factor by adopting a method combining a plurality of machine learning algorithms capable of describing a complex nonlinear relation, wherein the selected machine learning algorithm comprises the following steps: multivariate linear regression, random forest, artificial neural network, support vector machine, extreme learning machine, etc.

Preferably, in step one, the hydrological, meteorological, topographic, underlying surface data includes precipitation, wind speed, maximum/minimum air temperature, solar radiation, relative humidity, digital elevation, land use/coverage, soil data, runoff, vegetation index and leaf area index.

Specifically, in step four, if the sequence of the LULC data of the underlying surface is LULC ₁ -LULC _α The meteorological data sequence is Meteo ₁ -Meteo _β (wherein α is<β∩λ<β), then for the 1 st sub-period SP ₁ The weather input is Meteo ₁ -Meteo _λ And using LULC ₁ The land utilization/land coverage data is subjected to sliding construction of a SWAT model in the 1 st sub-period, 1 year is taken as a preheating period, 0.7 (lambda-1) year is taken as a rate period, and 0.3 (lambda-1) year is taken as a verification period; for the 2 nd sub-period SP ₂ The weather input is Meteo ₂ -Meteo _λ+1 Using LULC ₂ Land utilization/land cover, and so on, for the alpha sub-period SP _α The weather input is Meteo _β-λ -Meteo _β And using LULC _α Land use/land cover, thus correlating the LULC data with the sliding meteorological data, to establish a model input of alpha successive sub-periods.

Specifically, in step five, the objective function is the nash efficiency coefficient NSE and the water volume error PBIAS, and the formula is as follows:

in the formula, t ₁ And t _n Respectively the start and end time of the runoff sequence;

and &>

Respectively reconstructed natural runoff at the time t and runoff quantity, m, of model simulation ³ /s；/>

Is the average value of natural runoff, m ³ /s；

The parameter set corresponding to the maximum NSE value and PBIAS <20% is used as the optimal parameter set for model simulation in this time period.

Further, in order to prevent the distributed model from different participation and common effects and expand the number of samples, a threshold θ of NSE is set, and a plurality of sets of parameters meeting the threshold condition of the objective function are selected, wherein the formula is as follows:

Para _max (j)＝{Para _i |NSE _i ＝NSE _max }(1≤i≤epochs,1≤j≤α)

Para _Group (j)＝{Para _i |NSE _i ≥θNSE _max }(1≤i≤epochs,1≤j≤α,0≤θ≤1)

in the formula, para _max (j) And Para _Group (j) Respectively referring to the optimal parameter set of the model in the jth sub-period and the conditional parameter set meeting a certain objective function threshold; para _i Is the parameter value when the parameter is preferred for the jth sub-period; NSE _i The corresponding objective function value when the parameter is optimized for i times is obtained; NSE _max The maximum objective function value reached in the parameter optimization process is referred to; θ is a threshold setting; epochs represent the number of parameter preferences.

Preferably, a Global Sensitivity parameter Sensitivity Analysis (GSA) method is used for analyzing the parameter Sensitivity of the distributed hydrological model, relatively sensitive parameters are screened out, and a plurality of parameters with high correlation with the characteristic factors of the drainage basin underlying surface in the sensitive parameters are selected out according to the physical significance of the parameters.

Preferably, considering that the remote sensing data has a certain error in acquisition and interpretation, especially the error is relatively larger when the remote sensing data are distributed in a staggered manner for multiple vegetation types, integrating multiple remote sensing data information sources, and selecting 5 under-lying surface characteristic factors in the basin in the seventh step, including: cultivated land area ratio, forest land area ratio, grassland area ratio, vegetation index NDVI and leaf area index LAI.

Further, the objective function for the accuracy test of the relation between the model parameters and the underlying surface characteristic factors in the seventh step comprises a standard deviation STDs, a root mean square error RMSE and a pearson correlation coefficient r, and the formula is as follows:

in the formula (I), the compound is shown in the specification,

and &>

Respectively setting the rate constant value and the machine learning algorithm analog value of the nth parameter in the jth sub-period; a represents the parameter value of the model calibration or the machine learning fitting, namely Para or psi; />

Is the model rate constant value or the machine learning algorithm analog value of the nth parameter in the jth sub-period; />

The average rate fixed value of the nth parameter and the average value of the machine learning algorithm simulation value are respectively set; />

Represents->

Or->

The invention has the beneficial effects that:

the method applies meteorological data and underlying surface data to the establishment of the basin SWAT distributed hydrological model, inputs the data of the underlying surface which is continuous for many years into the driving model, firstly reduces the actual measurement runoff into the reconstructed natural runoff through reduction calculation, combines the dynamic underlying surface data after the mutation point, and then simulates the reconstructed natural runoff sequence, thereby overcoming the defects that the prior art is limited by a small number of samples, weak sample continuity and neglected underlying surface conditions of the basin. The prior art is as follows, climatic change reduces influence [ D ] of water input and hydrological model parameters on hydrological simulation, zhejiang university, 2017, and three discontinuous time periods are selected in the text: 1992-1995,1965-1968,2003-2006 respectively represent different climate condition calibration parameters, considering the non-stationarity of hydrology, a distributed SWAT model is used, but the input of a fluctuating mat surface is not considered, LAND utilization data comes from GLOBAL LAND COVER 2000, the mat surface data of 2000 is used as a reference to replace the data in different periods, so that the influence of different climate states on the model parameters is only considered, the change of the mat surface data is not considered fully, and only three discontinuous period parameters are adopted, the sample size is small, and the influence of the climate conditions is large.

The time-varying attribute of the hydrological model parameter is fully considered, the problems of insufficient parameter samples and weak parameter continuity in the prior art are solved by constructing methods such as sliding window segmentation and threshold parameter selection, the parameter sample capacity is ensured to be sufficient by the calibration of the sliding window and the division of the threshold parameter group, the continuity exists in time, the sample representativeness and effectiveness are improved, the result reliability is enhanced, and the uncertainty is reduced.

The invention overcomes the defects that the prior art can only be applied to a lumped model with simple structure and relatively few parameters but can not be applied to a distributed hydrological model with finer model structure and relatively more parameters, for example, the application number is 2021101127230, the invention name is a hydrological model parameter time-varying form construction method, 1981-1985, 1986-1990 and 1986-1990 are respectively selected as model rate periods, a lumped model of GR4J is adopted, a basin is seen as a whole by the lumped model, the meteorological data can only be input into the model without considering basin space difference, dynamic underlying surface data can not be used as model driving data when the model is built, only NDVI of another data source is adopted to replace underlying surface characteristics, and a linear time-varying parameter form is built by the underlying surfaces represented by NDVI, potential precipitation PET and NDVI, but the relationship between model parameters and external factors is complex, and the linear relationship representation is not enough. The method quantitatively separates the influence of the change of the underlying surface on the hydrological parameters, quantifies the nonlinear relation between the model parameters and the characteristic factors of the underlying surface, and has certain theoretical value and practical significance for perfecting basin refined hydrological simulation under the changing background and predicting future water resources.

Drawings

FIG. 1 is a distribution diagram of the geographical locations and site locations of the Fenghe basin in example 1;

fig. 2 is a spatial distribution diagram of the vegetation index NDVI, the leaf area index LAI, and the land use LULC in example 1, where 2a is a year-round average NDVI distribution, 2b is an NDVI evolution trend, 2c is a year-round average LAI distribution, 2d is an LAI evolution trend, 2e is a land use pattern in 1982, and 2f is a land use pattern in 2015;

fig. 3 is a structure diagram of a SWAT model constructed in the river basin of the fen in embodiment 1;

fig. 4 is a natural radial flow graph of the multi-year precipitation radial flow distribution and SWAT model reconstruction of the fen river basin in example 1;

FIG. 5 is a hydrological simulation process based on a sliding window time-varying parameter in example 1;

FIG. 6 is the SWAT parameter sensitivity analysis of example 1, wherein a is the parameter sensitivity ranking for different time periods, the numbers in the figure are the sensitivity ranking, b is the result of significance test, "-" indicates passing significance test;

FIG. 7 is a graph of CN in example 1 based on the 5-machine learning regression analysis method ₂ CANMX and ESCO, where a is to CN ₂ B is a simulation result of CANMX, c is a simulation result of ESCO; wherein, point A represents a reference value, point B represents a multiple linear regression, point C represents a random forest, point D represents an artificial neural network, and point E representsRepresenting the support vector machine and point F representing the extreme learning machine.

Detailed Description

The technical solution of the present invention is further explained with reference to the accompanying drawings and specific embodiments.

Example 1

Fig. 1 shows the geographical location of the river basin and the distribution locations of the hydrological and meteorological stations, taking the river basin as an example, the area is located in the midstream of the yellow river basin, is the second major branch of the yellow river, and is 3.9 kilometers squared, which is an important agricultural production base of the yellow river basin. 16 meteorological stations are shared in the research area and the periphery, and the drainage basin control station is an river water station. The Fenriver basin is in a semiarid and semihumid area, the distribution is uneven in rainfall years, water resources are relatively deficient, main vegetation types in the basin are cultivated land, woodland and grassland, the evolution of the underlying surface is mainly changed among the three vegetation types under the dual effects of climate change and human activities, and the research on the time-varying hydrological effect of dynamic change of the underlying surface has important guiding significance for the development and utilization of regional water resources and the treatment of the flood and drought disasters.

The embodiment is based on a GLASS-GLC global annual land utilization data set provided by Qinghua university, a GIMMS NDVI normalized vegetation index data set provided by NASA (national space and flight administration) of the United states, a GLOBMAP LAI V3.0 leaf area index data set provided by the institute of geography, science and resources of China academy, and the space-time evolution characteristic of an underlying surface characteristic factor in a river basin is analyzed, then, the daily meteorological data provided by a China meteorological network, hydrological station runoff data provided by a China hydrological almanac and an annual land utilization data set are utilized to build a distributed hydrological model suitable for a Fenriver basin, an 11-year sliding window calibration method is adopted to build 34 sub-periods, a periodic rate and a verification period are respectively planned in each sub-period, and an optimal parameter set and an 80% threshold parameter set are obtained. And simulating the relationship between the parameter group and the underlying surface factor by adopting various machine learning methods, thereby constructing the functional relationship between the river basin underlying surface vegetation dynamic and the model parameter.

The specific method of the embodiment is as follows:

the method comprises the following steps: downloading meteorological data of the inside and surrounding meteorological sites of the Fenhe basin, including precipitation, highest/lowest temperature, wind speed, relative humidity, solar radiation and the like, and performing statistical arrangement to obtain a daily scale sequence (1960-2019). And collecting monthly actual measurement runoff data of the Fenhezujin hydrological station (1960-2019). Downloading elevation DEM, soil data, GLASS-GLC land utilization/land coverage (LULC) data, GIMMS vegetation index and GLOBMAP leaf area index of the Fenriver basin, and cutting according to vector boundaries of the Fenriver basin, thereby obtaining annual land utilization data of the Fenriver basin (1982-2015), and simultaneously obtaining annual vegetation index NDVI and leaf area index LAI sequence of the Fenriver basin by utilizing maximum synthesis statistics (1982-2015).

Fig. 2 shows the spatial distribution of the annual average and evolution trend of the vegetation index NDVI and the leaf area index LAI of the fen river basin and the spatial distribution of the basin land cover in 1982 and 2015 through year-by-year data statistics. Covering the underlying surface of the Fenhe basin mainly comprises cultivated land, woodland and grassland, wherein the cultivated land and the grassland are mainly distributed along the dry flow, and the woodland mostly appears in the peripheral area of the basin. Spatially, about 80% of the watershed area shows a tendency to turn green, with the watershed NDVI increasing at a rate of 0.0257/10a on average, and the LAI increasing at a rate of 0.244/10a on average.

Step two: selecting a SWAT (Soil and Water Association Tool) distributed hydrological model as a research Tool, and building the SWAT distributed hydrological model suitable for the river basin by using the data collected and processed in the step one, wherein FIG. 3 is a conceptual basin diagram built based on the SWAT model.

Step three: and D, reconstructing natural runoff of the basin by using the built SWAT hydrological model of the Fenriver basin in the step two. Carrying out non-consistency test on hydrological long sequence data by using a Mann-Kendall (M-K) mutation test method to obtain that the runoff mutation point of the Fenriver basin appears near 1970, using meteorological data in 1960-1970 years (natural period) before the mutation point to calibrate a model and preferably obtain parameters, considering that the parameters are model parameters reflecting the natural period which is not influenced by human activities, keeping the fixed parameters unchanged, using the meteorological data observed in 1971-2019 as the model input into the well-calibrated hydrological model, and calculating a simulated runoff sequence, wherein the simulated runoff is the natural runoff reconstructed by the model, thereby reconstructing the monthly-scale natural runoff (the runoff assuming that climate and underlying surface do not change) in the Fenriver basin in 1960-2019.

In hydrology, it is generally considered that the time before the mutation point is a natural stage less affected by human activities, and the time after the mutation point is a stage more affected by human activities. In hydrologic simulation, adopt meteorological data drive and rate hydrologic model, what the simulation obtained is the change of runoff sequence, and meteorological data is the input promptly, and hydrologic data is output, and hydrologic data refers to the runoff in this application. The meteorological data used for model driving or parameter calibration in the application specifically comprises the collected precipitation, the highest/lowest air temperature, the wind speed, the relative humidity and the solar radiation data.

The hydrological data include runoff, evaporation, soil water and the like, but are limited by data observation, the actual evaporation is difficult to directly measure by evaporation, soil moisture content sites of the soil water are relatively few, the runoff is a relatively complete series in the current observation and can well represent the hydrological characteristics of a watershed or an area, and therefore the measured runoff data are generally adopted in the non-consistency inspection of a hydrological sequence. In addition, the meteorological data such as precipitation can also be subjected to non-consistency inspection, the runoff is only used because the runoff is highly related to the precipitation, and the inspection result according to the runoff sequence can indicate that the whole basin has sudden change in which time period.

The method adopts the model parameters which are calibrated at the natural period and the meteorological data driving model of 1971-2019 of the influence period to obtain the natural runoff reconstructed by the model in the year 1960-2019 of the Fenhe basin, and fig. 4 shows the rainfall, the actual measurement runoff sequence and the SWAT reconstructed natural runoff sequence in the basin in the year 1960-2019. In 1960-2019, annual average rainfall of a basin is 455.2mm, and the average natural runoff of a model reconstructed for many years is 42.04m ³ Measured runoff of 27.15m per year ³ And s. Wherein the average natural runoff of the model reconstruction in the years of 1971-2019 is 39.65m ³ Per s, runoff measured in many years is 20.83m ³ And s. The fitting condition between the average natural runoff and the measured runoff of the natural period SWAT reconstruction for many years is good, the NSE coefficient reaches 0.80, the relative error PBIAS is less than 5 percent, and the natural runoff reconstructed by the model is reliable. Natural period moldThe model effect is good, the constructed model or the calibrated model parameters have good applicability in the drainage basin, the runoff in the natural period can be simulated, the parameters are reliable, and the model and the parameters are proved to be reliable when being transferred to the influence period to calculate the natural runoff.

Step four: in the present embodiment, the 11-year window sliding method is adopted to divide the sub-periods, and in each sub-period, 1 year is a warm-up period, 7 years is a rate period, and 3 years is a verification period. Based on a SWAT distributed model and continuous underlying surface input (34-stage LULC data in 1982-2015), the data input is prolonged before and after the LULC data period, namely, meteorological data in 1976-2019 are divided into time periods every 11 years, such as 20 8230in 1976-1986, 20-1987, 8230in 2009-2019, 34 time periods in 2009-2019, the SWAT model of each stage is constructed in a sliding mode according to different time periods, wherein 1 year is selected as a preheating period, 7 years are selected as a rate period, and 3 years are selected as a verification period. For example, for the first sub-period SP ₁ 1976-1986,1976 for the preheat phase, 1977-1983 for the rate phase, 1984-1986 for the validation phase, and so on. The LULC data and the 11-year sliding meteorological data are thus correlated to establish a model input for 34 consecutive sub-periods.

Step five: and aiming at the starting time and the ending time of the 34 continuous sub-period sequences established by the fourth division, inputting corresponding meteorological data and LULC data as models, driving the models, adopting an SUFI2 algorithm to optimize parameters, and simulating the natural runoff sequences reconstructed by the models in the third step. Respectively calibrating and verifying model parameters according to an optimal principle of an objective function, and calculating an objective function Nash efficiency coefficient NSE and a water volume error PBIAS when the model is calibrated each time, wherein the formula is as follows:

in the formula, t ₁ And t _n Respectively the start and end time of the runoff sequence;

and &>

Respectively reconstructed natural runoff at the time t and runoff quantity, m, of model simulation ³ /s；/>

Is the average value of natural runoff, m ³ /s。

Thus, the optimal parameter set of the model with the best simulation performance of 34 sets corresponding to 34 sub-periods is obtained, and particularly, in order to prevent the distributed model abnormal participation phenomenon and expand the number of samples, the optimal parameter sets of the model satisfying the NSE threshold of the objective function 80% are counted, and the formula is as follows:

Para _max (j)＝{Para _i |NSE _i ＝NSE _max }(1≤i≤epochs,1≤j≤α)

Para _Group (j)＝{Para _i |NSE _i ≥0.8NSE _max }(1≤i≤epochs,1≤j≤α,0≤θ≤1)

in the formula, para _max (j) And Para _Group (j) Respectively referring to the optimal parameter set of the model in the jth sub-period and the conditional parameter set meeting a certain objective function threshold; para a _i Is the parameter value when the parameter is preferred for the jth sub-period; NSE _i The corresponding objective function value when the parameter is optimized for i times is obtained; NSE _max The maximum objective function value reached in the parameter optimization process is referred to; θ is a threshold setting; epochs represent the number of parameter preferences.

Fig. 5 is a time-varying hydrological simulation process line with a multi-year sliding window rating. The time-varying parameter hydrological simulation method based on the sliding window sub-period has a good runoff simulation result. In each sliding window sub-period of 11 years, the NSE calculated through parameter calibration is greater than 0.85, the PBIAS is less than 10% except for individual year, and the runoff process of basin simulation is well matched with natural runoff. The simulated time-varying runoff mostly lies below the natural runoff, which shows that the change of the underlying surface contributes to the runoff in a negative direction in the influence period. This is consistent with the evolution of vegetation greening in the Fenghe basin.

Step six: according to the parameter optimization for multiple times in the step five, a Global Sensitivity parameter Sensitivity Analysis (GSA) is utilized to analyze the parameter Sensitivity of the SWAT hydrological model and screen out relatively sensitive parameters, FIG. 6 is the SWAT parameter Sensitivity sequencing and significance Analysis in the embodiment, a total of 12 model parameters are selected for parameter calibration in the embodiment, wherein CN ₂ SOL _ AWC, CANMX, ALPHA _ BNK, ESCO are relatively sensitive and pass the significance test.

And finally selecting three parameters with high correlation with the characteristic factors of the underlying surface of the drainage basin in the sensitive parameters by combining the physical meanings of the parameters: CN ₂ 、CANMX、ESCO。

Step seven: fitting an empirical formula of the model parameters and the characteristic factors of the underlying surface of the basin according to the model continuous parameter sets obtained in the fifth step and the sensitive parameters screened in the sixth step, thereby respectively constructing basin parameters CN ₂ CANMX, ESCO and underlying surface characteristic factors. Selecting a plurality of watershed subsurface feature factors includes: cultivated land occupation ratio, forest land occupation ratio, grassland occupation ratio, NDVI and LAI. The method adopts a plurality of machine learning algorithms capable of depicting complex nonlinear relations to fit and verify the relation between the model parameters and the characteristic factors of the underlying surface of the watershed: multivariate linear regression, random forests, artificial neural networks, support vector machines, extreme learning machines, and the like.

And through precision inspection, selecting an algorithm which is most suitable for representing the relationship between the model parameters and the characteristic factors of the underlying surface of the basin, so as to establish a complex nonlinear relationship between the model parameters and the underlying surface of the basin and obtain the optimal representation form of the parameters of the basin. The objective function of the parameter simulation precision test comprises standard deviation STDs, root mean square error RMSE and Pearson correlation coefficient r, and the formula is as follows:

in the formula (I), the compound is shown in the specification,

and &>

Respectively setting the rate constant value and the machine learning algorithm analog value of the nth parameter in the jth sub-period; a represents the parameter value of the model calibration or the machine learning fitting, namely Para or psi; />

Is the model rate constant value or the machine learning algorithm analog value of the nth parameter in the jth sub-period; />

The average rate fixed value of the nth parameter and the average value of the machine learning algorithm simulation value are respectively set; />

Represents->

Or->

FIG. 7 is a Taylor diagram of the accuracy of parameter modeling based on the 5-machine learning regression analysis method in example 1. Support vector machine pair CN ₂ And the simulation result of ESCO is relatively good; the CANMX simulation result of the support vector machine and the extreme learning machine is relatively good, and the effect is similar. The simulation accuracy of the linear regression method is generally inferior to that of other methods, which shows that compared with a simple linear relation, the complex nonlinear relation represented by the machine learning algorithm is more suitable for parameters and underlying surface variablesAnd (5) describing quantitative relations.

Under the background of changing environment, the method provided by the invention can provide technical support for unsteady-state time-varying hydrological simulation, and provides a theoretical basis in the applications of water resource planning, early warning and prevention of flood and drought disasters and the like in future areas.

It should be noted that the above embodiments and examples are preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby, and any equivalent changes or equivalent modifications made on the basis of the technical solutions according to the technical ideas of the present invention still belong to the protection scope of the technical solutions of the present invention.

Claims (10)

1. The time-varying parameter hydrological simulation method based on sliding window sub-period division is characterized by comprising the following steps of:

the method comprises the following steps: collecting and organizing hydrology, meteorology, terrain and underlying surface data information in a research flow domain;

step two: constructing a SWAT distributed hydrological model suitable for a research basin based on the data information in the step one;

step three: reconstructing and researching natural runoff of the basin by using the hydrological model built in the step two;

step four: selecting data of lambda year to divide sub-period SP according to length of underlying surface data and meteorological data _i Extending the data before and after the underlying surface data, determining the start and the end of each sub-period sequence, sequentially sliding, constructing a SWAT model of each sub-period in a sliding manner, and aiming at each SP _i Wherein i =1,2,3 \8230, alpha, 1 year is taken as a preheating period, 0.7 (lambda-1) year is a rate period, 0.3 (lambda-1) year is a verification period, underlying surface data and sliding meteorological data are corresponded, and alpha model input of continuous sub-periods is established;

step five: aiming at the alpha sub-period, inputting meteorological data and underlying surface data into and driving a model according to the sequence start and stop divided in the fourth step, simulating a natural runoff sequence reconstructed by the model in the third step, and calibrating and verifying model parameters respectively according to an optimal principle of a target function to obtain an alpha group of model parameter sets with the best simulation performance;

step six: analyzing the sensitivity of the parameters aiming at the parameters obtained in the step five, and screening out the parameters with high correlation with the characteristic factors of the underlying surface of the drainage basin in the sensitive parameters by combining the physical significance of the parameters;

step seven: and (3) matching the relationship between the model parameters screened in the sixth step and the characteristic factors of the underlying surface of the basin by various machine learning algorithms, and selecting the algorithm most suitable for representing the relationship between the model parameters and the characteristic factors of the underlying surface of the basin by precision inspection so as to establish the relationship between the model parameters and the underlying surface of the basin.

2. The time-varying parameter hydrological simulation method based on sliding window subperiod division according to claim 1, wherein the specific method for reconstructing and researching natural runoff of a basin in the third step is as follows: and (3) carrying out non-consistency check on the hydrological long sequence data by using a mutation check method, utilizing the natural-period meteorological data before the mutation point to calibrate the hydrological model in the step two, and utilizing the calibrated hydrological model and the meteorological data observed after the mutation point to reconstruct the natural runoff in the whole period.

3. The sliding-window sub-period division-based time-varying parameter hydrological simulation method of claim 1, wherein data in step one specifically include precipitation, wind speed, maximum/minimum air temperature, solar radiation, relative humidity, digital elevation, land utilization/land cover, soil data, runoff, vegetation index and leaf area index.

4. The time-varying parametric hydrological modeling method based on sliding window subperiod partitioning of claim 1, wherein in step four, if the sequence of underlying LULC data is LULC ₁ -LULC _α The meteorological data sequence is Meteo ₁ -Meteo _β (wherein α is<β∩λ<β), then for the 1 st sub-period SP ₁ The weather input is Meteo ₁ -Meteo _λ And using LULC ₁ Land utilization/land coverage data, sliding to construct a SWAT model of the 1 st sub-period, taking 1 year as a preheating period, and taking 0.7The (lambda-1) year is the regular period, and the 0.3 (lambda-1) year is the verification period; for the 2 nd sub-period SP ₂ The weather input is Meteo ₂ -Meteo _λ+1 Using LULC ₂ Land utilization/land cover, and so on, for the alpha sub-period SP _α The weather input is Meteo _β-λ -Meteo _β And using LULC _α Land use/land cover, thus mapping the LULC data to the sliding meteorological data, together establishing a model input of a consecutive sub-periods.

5. The time-varying parameter hydrological simulation method based on sliding window subperiod division according to claim 1, wherein in step five, the objective function is nash efficiency coefficient NSE and water volume error PBIAS, and the formula is as follows:

/>

in the formula, t ₁ And t _n Respectively the start and end time of the runoff sequence;

and &>

Respectively reconstructed natural runoff at the time t and runoff quantity, m, of model simulation ³ /s；/>

Is the average value of natural runoff, m ³ /s；

The parameter set corresponding to the maximum NSE value and PBIAS <20% is used as the optimal parameter set for model simulation in this time period.

6. The time-varying parameter hydrological simulation method based on sliding window sub-period division according to claim 5, wherein in step five, a threshold θ of NSE is set, and a plurality of sets of parameters satisfying an objective function threshold condition are selected, wherein the formula is as follows:

Para _max (j)＝{Para _i |NSE _i ＝NSE _max }(1≤i≤epochs,1≤j≤α)

Para _Group (j)＝{Para _i |NSE _i ≥θNSE _max }(1≤i≤epochs,1≤j≤α,0≤θ≤1)

in the formula, para _max (j) And Para _Group (j) Respectively indicating the optimal parameter set of the model in the jth sub-period and the condition parameter set meeting a certain objective function threshold; para _i Is the parameter value at which the i-th parameter is preferred for the j-th sub-period; NSE _i The corresponding objective function value when the parameter is optimized for i times is obtained; NSE _max The maximum objective function value reached in the parameter optimization process is referred to; θ is a threshold setting; epochs represent the number of parameter preferences.

7. The time-varying parameter hydrological simulation method based on sliding window sub-period division according to claim 1, wherein in the sixth step, a global sensitivity parameter sensitivity analysis method is used for analyzing parameter sensitivity of the distributed hydrological model, screening out relatively sensitive parameters therein, and selecting out a plurality of parameters with high correlation with a basin underlying surface characteristic factor in the sensitive parameters in combination with physical significance of the parameters.

8. The time-varying parameter hydrological simulation method based on sliding window sub-period division according to claim 7, wherein the number of flow field underlying surface feature factors in the step seven is 5, and the method comprises the following steps: cultivated land area ratio, forest land area ratio, grassland area ratio, vegetation index NDVI and leaf area index LAI.

9. The time-varying parameter hydrological simulation method based on sliding window sub-epoch division according to claim 1, wherein the machine learning algorithm in step seven comprises: multivariate linear regression, random forests, artificial neural networks, support vector machines, and extreme learning machines.

10. The time-varying parameter hydrologic simulation method based on sliding window sub-period division according to claim 9, wherein the objective function of the accuracy test of the relation between the model parameters and the underlying surface characteristic factors in step seven comprises standard deviation STDs, root mean square error RMSE and pearson correlation coefficient r, and the formula is as follows:

in the formula (I), the compound is shown in the specification,

and &>

Respectively setting the rate constant value and the machine learning algorithm analog value of the nth parameter in the jth sub-period; a represents the parameter value of the model calibration or the machine learning fitting, namely Para or psi; />

Is the model rate constant value or the machine learning algorithm analog value of the nth parameter in the jth sub-period; />

The average rate fixed value of the nth parameter and the average value of the machine learning algorithm simulation value are respectively set; />

Represents->

Or->

/>

Images