A method for attributing extreme low-flow changes and assessing the potential of low-flow water resources

Abstract

The present invention discloses a method for attributing extreme low-flow changes and evaluating the potential of low-flow water resources, belonging to the technical field of water resources calculation. Grid data of runoff and its driving factors in the study area are obtained and data preprocessing is carried out. The non-parametric kernel density estimation method is used to derive the frequency distribution of runoff depth for each grid cell, and the proportion of grids meeting the runoff depth threshold of the guarantee rate in each year is statistically calculated, so as to determine the typical low-flow years under each guarantee rate. A non-linear regression simulation model coupled with optimized deep learning is used to carry out regression simulation research on extremely low-flow runoff depth under different combinations of driving factors, and the key combination of driving factors affecting the change of low-flow runoff depth is mined based on the results of regression simulation accuracy. Based on the results of attribution analysis, the constraint line derivation technology is used to analyze the constraint effect and possible thresholds of each key driving factor on extreme low-flow, and the potential of low-flow water resources in typical extreme low-flow years is evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of water resource calculation, and in particular to a method for attributing extreme low water changes and evaluating low water resource potential. Background Art

[0002] The attribution of extreme dry runoff changes and the assessment and analysis of dry runoff resource potential play an important role and significance in deeply understanding the causes of extreme events, promoting ecological and environmental protection, and enhancing disaster prevention and mitigation capabilities. In-depth analysis of the key driving factors of extreme dry runoff, scientific assessment of dry runoff resource potential, optimization of water resource allocation, and improvement of water resource utilization efficiency. Extreme dry runoff is an extreme state of river runoff. At present, there are several unresolved issues in the study of extreme dry runoff: First, traditional runoff attribution methods such as statistical analysis, hydrological models, and climate elasticity coefficients have high requirements for the temporal continuity of hydrological sequences, but extreme dry runoff sequences are discontinuous time series. Taking the annual scale as an example, extreme dry years are often separated by several years or even more than ten years. Traditional runoff attribution methods cannot meet the attribution needs of extreme dry runoff changes. Second, due to spatial limitations, traditional runoff change attribution is still based on runoff data from one or several hydrological stations, failing to consider the spatial heterogeneity of runoff changes and their influencing factors, reducing the reliability of attribution results. Third, there is currently a lack of evaluation technology that can evaluate the potential of dry runoff resources in a basin when extreme dry runoff scenarios occur. Therefore, it is urgently necessary to adopt scientific and efficient methods to explore the key driving factors affecting the spatial changes of extreme droughts from the perspective of spatial differentiation, attribute the changes in extreme droughts, and improve the accuracy of extreme drought simulation and drought resource potential assessment. Summary of the invention

[0003] The purpose of the present invention is to solve the problems existing in the prior art and to propose a method for attributing extreme low water changes and evaluating low water resource potential.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for attributing extreme low water changes and assessing low water resource potential comprises the following steps:

[0006] S1. Obtain raster data of runoff and its driving factors in the study area and perform data preprocessing;

[0007] S2. Based on the runoff depth grid data corresponding to step S1 in the study area, the frequency distribution of runoff depth of each grid cell is derived using the non-parametric kernel density estimation method, and the proportion of grids that meet the runoff depth threshold of the guarantee rate in each year is counted to the total grids, and then the typical dry year under each guarantee rate is determined;

[0008] S3. After obtaining the typical dry-year runoff corresponding to the study area in step S2, extract the runoff depth raster data and driving factor raster data corresponding to the typical dry year, and combine the fishnet generation technology and the random forest model to judge the importance degree of each driving factor to the extreme low-flow runoff depth;

[0009] S4. Based on the results of the driving factor importance judgment, use a non-linear regression simulation model that couples and optimizes deep learning to conduct a regression simulation analysis of the extreme low-flow runoff depth under different combinations of driving factors, and identify and mine the key driving factor combinations that affect the change of the low-flow runoff depth based on the accuracy of the regression simulation;

[0010] S5. Based on the results of the attribution analysis, use the constraint line derivation technology to analyze the constraint effect of each key driving factor on the extreme low flow and judge whether there is an optimal driving factor threshold, and evaluate the low-flow water resource potential of the extreme low-flow typical year.

[0011] Preferably, the preprocessing of the runoff and its driving factor raster data in step S1 includes:

[0012] Step 11, conduct missing value test, outlier test, and white noise processing on each type of data;

[0013] Step 12, conduct reprojection and resampling processing on each type of data to solve the problems of inconsistent resolution and unaligned pixels existing in each type of data.

[0014] Preferably, the step S2 includes:

[0015] Step 21, first based on the annual runoff depth raster data of the study area over the years, use the non-parametric kernel density estimation method to derive the frequency distribution of the annual runoff depth of each grid cell; first use the kernel function to derive the probability density function, and then derive the marginal distribution function of each duration rainstorm sample of the grid cell according to the kernel estimation probability density function; assume that the m-year runoff depth sample data y 1 , y 2 , …, y m of a certain grid are from a univariate continuous population distribution Y, the overall probability density function is f(y), and its kernel density estimation function can be expressed as:

[0016]

[0017] In the formula, is the kernel density estimation function of f(y); K(·) is the kernel function, and h is called the window width or bandwidth (h > 0);

[0018] Step 22: Determine the low-flow runoff depth thresholds under various guarantee rates; based on the frequency distribution results of Step 21, given a specific low-flow guarantee rate P, the low-flow runoff depth threshold RP corresponding to the low-flow guarantee rate P can be derived.

[0019] Step 23: Determine the low-flow typical years under various guarantee rates; annually count and calculate the proportion of grid cells that meet the runoff depth threshold of the guarantee rate in each year, and select the year with the largest proportion as the typical year; taking three low-flow guarantee rates of 75%, 90%, and 95% as examples: the 75% low-flow runoff depth refers to the runoff depth that is less than the runoff depth threshold of the 75% guarantee rate and greater than the runoff depth threshold of the 90% guarantee rate; the 90% low-flow runoff depth refers to the runoff depth that is less than the runoff depth threshold of the 90% guarantee rate and greater than the runoff depth threshold of the 95% guarantee rate; the 95% low-flow runoff depth refers to the runoff depth that is less than the runoff depth threshold of the 95% guarantee rate.

[0020] Preferably, Step S3 includes:

[0021] Step 31: After obtaining the low-flow typical years corresponding to Step S2 in the study area, extract the runoff depth grid data and driving factor grid data corresponding to the low-flow typical years.

[0022] Step 32: Adopt the fishnet construction and sampling techniques to collect the spatially continuous extreme low-flow runoff depth and its driving factor combination data in the study area within the typical year.

[0023] Step 33: Use the random forest model to identify the importance degree of the influence of each alternative driving factor on the change of the extreme low-flow runoff depth in the typical year; the calculation formula for the importance degree of the random forest model is

[0024]

[0025] In the formula, VI(X j ) represents the importance score of variable X j in the random forest; N OOB is the out-of-bag sample data; f(X i ) is the i-th observation value in the out-of-bag sample data; f n (X i ) and f n (X i ′) are the fitted values corresponding to the i-th observation value. On the n-th tree of the out-of-bag data, the former is before the observation value of the randomly permuted variable X j , and the latter is after the observation value of the randomly permuted variable X j ; I[f(X i ) = f n (X i )] is the discriminant function, that is, when f(X i ) = f n (X i) is 1 when the condition is met, and 0 otherwise; I[f(X i ) = f n (X i ′)] is also a discriminant function, that is, when f(X i ) = f n (X i ′), it is 1, and 0 otherwise; N t represents the number of trees in the random forest model.

[0026] Preferably, the step S4 includes:

[0027] Step 41, first, use a non - linear regression model to perform regression analysis on the extreme low - flow runoff depth under different combinations of driving factors, and record the residuals under each combination of driving factors;

[0028] Taking three variables as an example, the basic form of the non - linear regression model is as follows

[0029] Y = a0 + X1 * a1 + X2 * a2 + X3 * a3 + X1 * X2 * a4 + X1 * X3 * a5 + X2 * X3 * a6 + X1 * X1 * a7 + X2 * X2 * a8 + X3 * X3 * a9

[0030] In the formula, Y is the dependent variable of the non - linear regression model, X1 - X3 are the independent variables of the non - linear regression model, a0 is the constant term of the non - linear regression model, a1 - a3 are the linear term coefficients of the non - linear regression model, a4 - a6 are the cross - term coefficients of the non - linear regression model, and a7 - a9 are the quadratic term coefficients of the non - linear regression model; it can be seen from the basic form that the non - linear regression model adopted fully considers the influence of the constant term, linear term, cross - term, and quadratic term on the regression process;

[0031] Taking the first 70% of the time series as the calibration period of the regression model and the last 30% as the verification period of the regression model; then the residuals of the regression model in the calibration period and the verification period are respectively

[0032] e c,o = R c,s - R c,o

[0033] e v,o = R v,s - R v,o

[0034] In the formula, e c,o and e v,o are the residuals of the non - linear regression model in the calibration period and the verification period respectively; R c,s and R c,o are the simulated values and true values of the non - linear regression model in the calibration period respectively; R v,s and R v,oThey are the simulated value and the true value of the non-linear regression model during the validation period, respectively;

[0035] Step 42: Use the LSTM deep learning model optimized by differential creativity search to simulate the residual of the non-linear regression model;

[0036] The calculation steps of the LSTM model are as follows:

[0037] f k = σ[W f (Δe k-1 , R k ) + b f

[0038] i k = σ[W i (Δe k-1 , R k ) + b i

[0039]

[0040] o k = σ[W 0 (Δe k-1 , R k ) + b 0

[0041] Δe k = o k ⊙ tanh(C k )

[0042] In the formula, Δe k-1 is the error of the residual simulation result; f k , i k , o k are the forget gate, input gate, and output gate of the LSTM model, respectively; is the memory cell state; W f represents the weight of the forget gate; W i represents the weight of the input gate; W c is the weight; W 0 is the weight of the output gate; b f represents the bias term of the forget gate; b i is the bias term of the input gate; b 0 is the bias term of the output gate; and b c is the offset vector; C k is the state variable of the model, R k is the recursive input; σ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; ​​​

[0043] The calibration period and verification period of the LSTM deep learning model are consistent with those of the non - linear regression model; obtain the simulated residual sequences e c,s and e v,s ; then use the simulated residuals to correct the simulation results of the original non - linear regression model, that is, obtain the simulation results of the non - linear regression simulation model based on coupled optimization deep learning;

[0044] R1 c,s =R c,s +e c,s

[0045] R1 v,s =R v,s +e v,s

[0046] In the formula, R1 c,s and R1 v,s are the final simulation results of the non - linear regression simulation model based on coupled optimization deep learning in the calibration period and verification period;

[0047] Step 43, use the Nash efficiency coefficient NSE, R 2 , RMSE evaluation indicators to evaluate the simulation accuracy of the non - linear regression simulation model based on coupled optimization deep learning for extreme low - flow runoff under each combination of driving factors, and take the combination of driving factors with the best accuracy as the key driving factor combination affecting the change of extreme low - flow runoff depth.

[0048] Preferably, the step S5 includes:

[0049] Step 51, based on the identified key driving factor combination, use the constraint line derivation technique to obtain the constraint lines of each driving factor in the key driving factor combination on the extreme low - flow runoff depth, and reveal the constraint effect of each driving factor on the extreme low - flow runoff depth; the steps of constraint line derivation are as follows:

[0050] Take each driving factor as a limiting variable and the extreme low - flow runoff depth as the response variable to draw a two - dimensional scatter plot;

[0051] Divide the value range of the limiting variable into several groups, and select the 99.9% quantile point of the extreme low - flow runoff depth value in each group as the boundary point;

[0052] Perform an optimal fit of a polynomial function on all boundary points to extract the constraint line and determine its function equation;

[0053] Step 52, based on the constraint line results, judge whether there is an optimal driving factor threshold among various types of driving factors;

[0054] Step 53, based on the recognition result of the driving factor threshold, evaluate the potential of the dry water resources that the region may reach in the corresponding typical extremely dry years under the condition of regulating the driving factors; expressed by the average dry water runoff depth of the basin, the calculation formula for the potential of dry water resources is

[0055] R T =R S -R C

[0056] R S =f c (x s )

[0057] In the formula, R T is the potential of the dry water resources in the study area under the condition of the typical dry year, mm; R C represents the actual average dry water runoff depth of the basin in the typical year, mm; f c (x s ) is the corresponding constraint line equation; x s is the optimal driving factor threshold; R S represents the value of the dry water runoff depth that the basin can reach when the driving factor is infinitely close to the optimal driving factor threshold.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. The present invention provides a method for attributing extreme dry water changes and evaluating the potential of dry water resources. First, the non-parametric kernel density estimation method is used to derive the frequency distribution of the runoff depth of each grid cell, and then the typical dry years of the study area under each guarantee rate are determined. Secondly, after determining the typical years, the importance degree of each driving factor on the extreme dry water runoff depth is judged by combining the fishing net generation technology and the random forest model. Then, a non-linear regression simulation model coupled with optimized deep learning is used to carry out the regression simulation study of the extremely dry water runoff depth under different driving factor combinations, and the key driving factor combinations affecting the change of the dry water runoff depth are mined based on the regression simulation accuracy results. Finally, based on the results of the attribution analysis, the constraint line derivation technology is used to analyze the constraint effect of each key driving factor on the extreme dry water and the possible thresholds, and the potential of the dry water resources in the typical extreme dry years is evaluated.

[0060] 2. This method can effectively solve the problem that the traditional runoff attribution method is not applicable to non-continuous dry water time series, and realizes the coupling of the non-linear regression model and the deep learning theory. It not only effectively eliminates the interference of irrelevant factors on the attribution, but also improves the accuracy of the attribution. At the same time, based on the results of the attribution analysis and combined with the constraint line derivation technology, it provides a new and effective path for the region to evaluate the potential of the dry water resources in the extreme dry years. Description of the Drawings

[0061] Figure 1 Schematic diagram of the method process of the present invention;

[0062] Figure 2 Spatial distribution results of the 75% low-flow runoff depth threshold;

[0063] Figure 3 Spatial distribution results of the 90% low-flow runoff depth threshold;

[0064] Figure 4 Spatial distribution results of the 95% low-flow runoff depth threshold;

[0065] Figure 5 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 1960s;

[0066] Figure 6 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 1970s;

[0067] Figure 7 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 1980s;

[0068] Figure 8 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 1990s;

[0069] Figure 9 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 2000s;

[0070] Figure 10 Results of the proportion of low-flow grid numbers to total grid numbers under various guarantee rate conditions in the 2010s;

[0071] Figure 11 Importance degree of each driving factor to the extreme low-flow runoff depth;

[0072] Figure 12 Process diagram of the simulation results of the extreme low-flow runoff depth by the non-linear regression simulation model based on coupled optimization deep learning with unoptimized influencing factors during the verification period;

[0073] Figure 13 Evaluation of the simulation accuracy of the extreme low-flow runoff depth by the non-linear regression simulation model based on coupled optimization deep learning with unoptimized influencing factors during the verification period;

[0074] Figure 14 Process diagram of the simulation results of the extreme low-flow runoff depth by the non-linear regression simulation model based on coupled optimization deep learning with optimized influencing factors during the verification period;

[0075] Figure 15Evaluation results of the simulation accuracy of the non - linear regression simulation model based on coupled optimization deep learning for extreme low - flow runoff depth during the verification period after optimizing influencing factors;

[0076] Figure 16 The constraint effect of rainfall on extreme low - flow runoff depth;

[0077] Figure 17 The constraint effect of air temperature on extreme low - flow runoff depth;

[0078] Figure 18 The constraint effect of potential evaporation on extreme low - flow runoff depth;

[0079] Figure 19 The constraint effect of NDVI on extreme low - flow runoff depth;

[0080] Figure 20 The constraint effect of SPEI on extreme low - flow runoff depth;

[0081] Figure 21 The constraint effect of irrigation water on extreme low - flow runoff depth. Specific implementation manners

[0082] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0083] Embodiment, referring to Figures 1 to 21 , a method for attributing extreme low - flow changes and evaluating the potential of low - flow water resources, comprising the following steps:

[0084] S1. Obtain the raster data of runoff and its driving factors in the study area, and perform data pre - processing;

[0085] S2. Based on the runoff - depth raster data corresponding to step S1 in the study area, use the non - parametric kernel density estimation method to derive the frequency distribution of runoff depth for each grid cell, count the proportion of grid cells that meet the runoff - depth threshold of the guarantee rate in each year to the total grid cells, and then determine the typical low - flow years under each guarantee rate;

[0086] S3. After obtaining the typical low - flow years corresponding to step S2 in the study area, extract the runoff - depth raster data and driving - factor raster data corresponding to the typical low - flow years, and combine the fishnet generation technology and the random forest model to judge the importance degree of each driving factor on extreme low - flow runoff depth;

[0087] S4. Based on the results of the importance judgment of driving factors, a non-linear regression simulation model that couples and optimizes deep learning is used to conduct a regression simulation analysis of the extreme low-flow runoff depth under different combinations of driving factors. Based on the accuracy of the regression simulation, the key driving factor combinations that affect the change of low-flow runoff depth are identified and mined;

[0088] S5. Based on the results of the attribution analysis, the constraint line derivation technique is used to analyze the constraint effects of each key driving factor on extreme low flow and to determine whether there is an optimal driving factor threshold, and the potential of low-flow water resources in typical years of extreme low flow is evaluated.

[0089] Taking the main stream area of the middle and lower reaches of the Yangtze River as an example. The grid runoff depth data collected the monthly runoff depth grid data of China from 1961 to 2018, with a data resolution of 0.25°. Through the ACRGIS software, the monthly runoff depth grid data of the main stream area of the middle and lower reaches of the Yangtze River from 1961 to 2018 was cropped, and then the annual runoff depth grid data of the main stream area of the middle and lower reaches of the Yangtze River from 1961 to 2018 was derived. For the driving factor data, precipitation data, temperature data, potential evapotranspiration data, NDVI data, elevation data, slope data, meteorological drought index SPEI, ecological drought index TVDI, GDP, population density, human activity index (HI), domestic water withdrawal, and agricultural irrigation water withdrawal in the main stream area of the middle and lower reaches of the Yangtze River were collected.

[0090] After obtaining the annual runoff depth grid data of the main stream area of the middle and lower reaches of the Yangtze River from 1961 to 2018, through the fishnet construction technology and spatial sampling technology, the annual runoff depth data of each grid in the main stream area of the middle and lower reaches of the Yangtze River from 1961 to 2018 was collected. Then, the non-parametric kernel density estimation method was used to derive the frequency distribution of the runoff depth of each grid cell, and the low-flow runoff depth thresholds under 75%, 90%, and 95% guarantees in the main stream area of the middle and lower reaches of the Yangtze River were derived. The results are as Figures 2 to 4 shown.

[0091] Based on the low-flow runoff depth threshold results of each grid cell under each guarantee rate, the proportion of grids that meet the low-flow conditions under each guarantee rate in each year can be derived ( Figures 5 to 10 ). In the present invention, the 75% low-flow runoff depth refers to the runoff depth that is less than the 75% guarantee rate runoff depth threshold and greater than the 90% guarantee rate runoff depth threshold; the 90% low-flow runoff depth refers to the runoff depth that is less than the 90% guarantee rate runoff depth threshold and greater than the 95% guarantee rate runoff depth threshold; the 95% low-flow runoff depth refers to the runoff depth that is less than the 95% guarantee rate runoff depth threshold. Figures 5 to 10It can be seen that in terms of the 75% low-flow runoff depth, the proportion of the grid numbers belonging to the 75% low-flow runoff depth in 1966 to the total grid numbers is the largest, reaching 40%. Thus, 1966 is determined as the typical year of 75% low flow. In terms of the 75% low-flow runoff depth, the proportion of the grid numbers belonging to the 75% low-flow runoff depth in 1966 to the total grid numbers is the largest, reaching 40%. Therefore, 1966 is determined as the typical year of 75% low flow in the main stream area of the middle and lower reaches of the Yangtze River. In terms of the 90% low-flow runoff depth, the proportion of the grid numbers belonging to the 90% low-flow runoff depth in 1978 to the total grid numbers is the largest, reaching 49%. Thus, 1978 is determined as the typical year of 90% low flow in the main stream area of the middle and lower reaches of the Yangtze River. In terms of the 95% low-flow runoff depth, the proportion of the grid numbers belonging to the 95% low-flow runoff depth in 2011 to the total grid numbers is the largest, reaching 64%. Therefore, 2011 is determined as the typical year of 95% low flow in the main stream area of the middle and lower reaches of the Yangtze River. In summary, the typical years of regional extreme low flow determined by the present invention are reasonable and reliable.

[0092] After determining the typical low-flow years, the attribution of extreme low-flow changes, simulation, and evaluation of low-flow water resource potential can be carried out. The present invention takes the extreme low flow with a 95% guarantee rate as an example. From the determination results of the typical extreme low-flow years, it can be seen that the typical year of extreme low flow with a 95% guarantee rate in the middle and lower reaches of the Yangtze River is 2011. Therefore, the grid data of the runoff depth in 2011 is obtained as the extreme low-flow runoff depth data. In addition, the corresponding driving factor data in 2011 is also obtained, including: rainfall data, temperature data, potential evapotranspiration data, NDVI data, altitude data, slope data, meteorological drought index SPEI, ecological drought index TVDI, GDP, population density, human activity index (HI), domestic water withdrawal, and agricultural irrigation water withdrawal. On this basis, the random forest model is used to judge the importance degree of each driving factor to the extreme low-flow runoff depth, and the results are as Figure 11 shown. It can be seen from Figure 11 that rainfall is the most important factor affecting the change of the 95% extreme low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River, followed by temperature and the meteorological drought index SPEI. Among the five indicators representing the impact of human activities, irrigation water is the most important factor affecting the 95% extreme low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River, followed by population density and domestic water withdrawal. The ecological drought index TVDI, slope, and HI are unimportant influencing factors.

[0093] Then, a probability regression simulation analysis of the extreme low-flow runoff depth is carried out using a non-linear regression simulation model that couples and optimizes deep learning, and the results are as Figures 12 to 15 shown. Among them, Figure 12 and Figure 13 are the regression process and accuracy evaluation results when the influencing factors are not optimized. When the influencing factors are not optimized, the performance of the regression model is poor due to the interference of unimportant factors, and R 2The Nash-Sutcliffe efficiency coefficient NSE are only 0.7149 and 0.6239 respectively, and the error value exceeds ±100 mm. Unimportant influencing factors are eliminated one by one from the least important to the most important, and regression analysis is carried out. The regression accuracy results are recorded and compared for each eliminated unimportant factor. It is found that when the ecological drought index TVDI, slope, human activity index (HI), altitude, GDP, domestic water consumption, and population density PD are successively eliminated, the accuracy evaluation indicators R 2 and the Nash-Sutcliffe efficiency coefficient NSE generally show an upward trend. After eliminating other influencing factors, the accuracy evaluation indicator R 2 and the Nash-Sutcliffe efficiency coefficient NSE continue to decrease. Therefore, the optimal combination of influencing factors is selected as rainfall, temperature, potential evapotranspiration, NDVI, meteorological drought index SPEI, and agricultural irrigation water consumption. The regression process and accuracy evaluation results are as Figure 14 and Figure 15 shown. After the influencing factors are optimized, the interference of unimportant factors is effectively excluded, and the performance of the regression model is significantly improved. R 2 and the Nash-Sutcliffe efficiency coefficient NSE reach 0.8470 and 0.8433 respectively, and the error value is within ±100 mm.

[0094] Finally, the constraint effect of rainfall, temperature, potential evapotranspiration, NDVI data, meteorological drought index SPEI, and agricultural irrigation water consumption on the extreme low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River is analyzed using the constraint line technique, the threshold of the most suitable driving factor is determined, and the low-flow water resource potential in the typical low-flow years in the main stream area of the middle and lower reaches of the Yangtze River is evaluated. The results of the constraint effect of precipitation, temperature, potential evapotranspiration, NDVI data, meteorological drought index SPEI, and agricultural irrigation water consumption on the extreme low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River are as Figures 16 to 21 shown. As shown by Figures 16 to 21It can be seen that the constraint ability of precipitation on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River decreases with the increase of precipitation, and there is no optimal rainfall threshold; before the annual average temperature reaches 16°C, the constraint ability of temperature on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River decreases with the increase of temperature. After the temperature reaches 16°C, the constraint ability of temperature on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River increases with the increase of temperature, and 16°C is the optimal temperature threshold; the constraint ability of potential evapotranspiration on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River increases with the increase of potential evapotranspiration, and there is no optimal evapotranspiration threshold; the NDVI in the main stream area of the middle and lower reaches of the Yangtze River is mainly distributed between 0.5 and 0.95. Before the NDVI reaches 0.8, the constraint ability of NDVI on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River decreases with the increase of NDVI. After the NDVI reaches 0.8, the constraint ability of NDVI on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River increases with the increase of NDVI, and 0.8 is the optimal NDVI threshold; the constraint ability of SPEI on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River decreases with the increase of the SPEI value, and there is no optimal SPEI threshold; the constraint ability of irrigation water use on the extremely low-flow runoff depth in the main stream area of the middle and lower reaches of the Yangtze River increases with the increase of irrigation water use, and there is no optimal irrigation water use threshold.

[0095] Based on the results of the constraint effect and threshold analysis, it can be seen that there are optimal thresholds for temperature and NDVI. By regulation, the temperature and NDVI can reach the thresholds to obtain the best low-flow runoff depth. Among them, temperature is related to carbon emissions, while NDVI is related to the underlying surface vegetation and can be regulated through ecological construction projects. After obtaining the optimal temperature and NDVI thresholds, substituting the optimal thresholds into the corresponding constraint line equations, the low-flow runoff depth under the threshold conditions can be obtained. Comparing the low-flow runoff depth under the threshold conditions with the actual low-flow runoff depth in the typical year, the difference is the low-flow water resource potential of the typical low-flow year. The evaluation results of the low-flow water resource potential of the 95% low-flow typical year (2011) in the main stream area of the middle and lower reaches of the Yangtze River are shown in Table 1. As can be seen from the table, the low-flow runoff depth of the 95% low-flow typical year (2011) in the main stream area of the middle and lower reaches of the Yangtze River under the optimal temperature threshold condition is 1096 mm, and the corresponding low-flow water resource potential is 375.52. The low-flow runoff depth of the 95% low-flow typical year (2011) in the main stream area of the middle and lower reaches of the Yangtze River under the optimal NDVI threshold condition is 1136 mm, and the corresponding low-flow water resource potential is 415.52.

[0096] Table 1 Evaluation results of low-flow water resource potential of 95% guarantee rate extremely low-flow typical years

[0097]

[0098] The present invention provides a method for attributing extreme low-flow changes and evaluating the potential of low-flow water resources. First, the non-parametric kernel density estimation method is used to derive the frequency distribution of runoff depth for each grid cell, and then the typical low-flow years under different guarantee rates in the study area are determined. Secondly, after determining the typical years, the importance degree of each driving factor on the extreme low-flow runoff depth is judged by combining the fishing net generation technology and the random forest model. Then, a non-linear regression simulation model coupled with optimized deep learning is used to carry out the regression simulation study of the extremely low-flow runoff depth under different combinations of driving factors, and the key driving factor combinations affecting the change of low-flow runoff depth are mined based on the regression simulation accuracy results. Finally, based on the results of the attribution analysis, the constraint line derivation technology is used to analyze the constraint effect and possible thresholds of each key driving factor on the extreme low-flow, and the potential of low-flow water resources in the typical extreme low-flow years is evaluated.

[0099] This method can effectively solve the problem that traditional runoff attribution methods are not applicable to discontinuous low-flow time series, and realizes the coupling of non-linear regression models and deep learning theory, which not only effectively eliminates the interference of irrelevant factors on attribution, but also improves the accuracy of attribution. At the same time, based on the results of the attribution analysis and combined with the constraint line derivation technology, a new and effective path is provided for regional evaluation of the potential of low-flow water resources in extreme low-flow years.

[0100] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A method for attributing extreme low water changes and assessing low water resource potential, characterized in that: The following steps are involved: S1. Obtain raster data of runoff and its driving factors in the study area and perform data preprocessing; S2. Based on the runoff depth grid data corresponding to step S1 in the study area, the frequency distribution of runoff depth of each grid cell is derived using the non-parametric kernel density estimation method, and the proportion of grids that meet the runoff depth threshold of the guarantee rate in each year is counted to the total grids, and then the typical dry year under each guarantee rate is determined; S3. After obtaining the typical dry year corresponding to step S2 of the study area, extract the runoff depth raster data and driving factor raster data corresponding to the typical dry year, combine the fishing net generation technology and the random forest model, and determine the importance of each driving factor to the extreme low water runoff depth; S4. Based on the results of the importance of driving factors, a nonlinear regression simulation model with coupled optimization deep learning is used to carry out regression simulation analysis of extreme low-flow runoff depth under different driving factor combinations. Based on the accuracy of regression simulation, the key driving factor combinations that affect the changes in low-flow runoff depth are identified and mined; S5. Based on the results of attribution analysis, the constraint line inference technology is used to analyze the constraint effect of each key driving factor on extreme drought and to determine whether there is an optimal driving factor threshold, so as to evaluate the dry water resource potential in a typical year of extreme drought.

2. The method for attributing extreme low water changes and evaluating low water resource potential according to claim 1, characterized in that: The preprocessing of the runoff and its driving factor raster data in step S1 includes: Step 11, perform missing value test, outlier test, and white noise processing on each type of data; Step 12, reprojecting and resampling each type of data to solve the problem of inconsistent resolution and non-strict pixel alignment among various types of data.

3. The method for attributing extreme low water changes and evaluating low water resource potential according to claim 1, characterized in that: The step S2 comprises: Step 21, firstly, based on the annual runoff depth grid data of the study area over the years, the non-parametric kernel density estimation method is used to deduce the frequency distribution of the annual runoff depth of each grid unit; firstly, the kernel function is used to deduce the probability density function, and then the marginal distribution function of the heavy rain samples of each duration of the grid unit is derived based on the kernel estimated probability density function; the m-year runoff depth sample data y1, y2, ..., y m It comes from a univariate continuous population distribution Y, the overall probability density function is f(y), and its kernel density estimation function can be expressed as: In the formula, is the kernel density estimation function of f(y); K(·) is the kernel function, and h is called the window width or bandwidth (h>0); Step 22, determining the low-flow runoff depth threshold value under each guarantee rate; based on the frequency distribution result of step 21, given a specific low-flow guarantee rate P, the low-flow runoff depth threshold value RP corresponding to the low-flow guarantee rate P can be deduced; Step 23, determine the typical dry year under each guarantee rate; calculate the proportion of grids that meet the guarantee rate runoff depth threshold in each year to the total grids, and select the year with the largest proportion as the typical year; 75%, 90%, and 95% three dry guarantee rates: 75% dry runoff depth refers to the runoff depth less than the 75% guarantee rate runoff depth threshold but greater than the 90% guarantee rate runoff depth threshold; 90% dry runoff depth refers to the runoff depth less than the 90% guarantee rate runoff depth threshold but greater than the 95% guarantee rate runoff depth threshold; 95% dry runoff depth refers to the runoff depth less than the 95% guarantee rate runoff depth threshold.

4. The method for attributing extreme low water changes and evaluating low water resource potential according to claim 1, characterized in that: The step S3 comprises: Step 31, after obtaining the typical dry year corresponding to step S2 of the study area, extracting the runoff depth raster data and driving factor raster data corresponding to the typical dry year; Step 32, using fishing net construction and sampling technology to collect spatially continuous extreme low-flow runoff depth and its driving factor combination data in the study area in a typical year; Step 33, using the random forest model, identify the importance of each candidate driving factor on the change of extreme low-flow runoff depth in a typical year; the importance calculation formula of the random forest model is: In the formula, VI(X j ) represents the variable X in the random forest j Importance score of OOB is the out-of-bag sample data; f(X i ) is the i-th observation value in the out-of-bag sample data; f n (X i ) and f n (X′ i ) is the fitted value corresponding to the i-th observation. On the n-th tree of out-of-bag data, the former is the fitted value corresponding to the random permutation variable X j The latter is before the observation of the random permutation variable X j After the observation value of i ) = f n (X i )] is the discriminant function, that is, when f(X i ) = f n (X i ), the value is 1, otherwise it is 0; I[f(X i ) = f n (X′ i )] is also the discriminant function, that is, when f(X i ) = f n (X′ i ), the value is 1, otherwise it is 0; N t Represents the number of trees in the random forest model.

5. The method for attributing extreme low water changes and evaluating low water resource potential according to claim 1, characterized in that: The step S4 comprises: Step 41, firstly, a nonlinear regression model is used to perform regression analysis on the extreme low-flow runoff depth under different driving factor combinations, and the residual under each driving factor combination is recorded; Among them, the basic form of the three-variable nonlinear regression model is as follows Y=a0+X1*a1+X2*a2+X3*a3+X1*X2*a4+X1*X3*a5+X2*X3*a6+X1*X1*a7+X2*X2*a8+X3*X3*a9 In the formula, Y is the dependent variable of the nonlinear regression model, X1~X3 are the independent variables of the nonlinear regression model, a0 is the constant term of the nonlinear regression model, a1~a3 are the linear term coefficients of the nonlinear regression model, a4~a6 are the cross term coefficients of the nonlinear regression model, and a7~a9 are the square term coefficients of the nonlinear regression model; The first 70% of the time series is used as the rate period of the regression model, and the last 30% is used as the validation period of the regression model; the residuals of the regression model in the rate period and validation period are e c,o =R c,s -R c,o e v,o =R v,s -R v,o In the formula, e c,o and e v,o are the residuals of the nonlinear regression model in the rate period and the validation period; R c,s and R c,o are the simulated value and true value of the nonlinear regression model at the rate period; R v,s and R v,o are the simulated value and true value of the nonlinear regression model in the validation period, respectively; Step 42, using an LSTM deep learning model based on differential creative search optimization parameters to simulate the residual of the nonlinear regression model; The LSTM model calculation steps are as follows: f k =σ[W f (Δe k-1 ,R k )+b f ] i k =σ[W i (Δe k-1 ,R k )+b i ] the k =σ[W0(Δe k-1 ,R k )+b0] Δe k =o k ⊙tanh(C k ) In the formula, Δe k-1 is the error of the residual simulation result; f k ,i k , o k They are the forget gate, input gate and output gate of the LSTM model respectively; is the memory unit state; W f Represents the weight of the forget gate; W i Represents the weight of the input gate; W c for The weight of the output gate; W0 is the weight of the output gate; b f Characterizes the bias term of the forget gate; b i is the bias term of the input gate; b0 is the bias term of the output gate; and b c yes The offset vector of k is the state variable of the model; R k is the recursive input; σ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; The rate period and validation period of the LSTM deep learning model are consistent with the nonlinear regression model; the simulated residual sequence e of the training period and the test period is obtained c,s and e v,s ; Then, the simulation results of the original nonlinear regression model are corrected using the residuals obtained by simulation, that is, the simulation results of the nonlinear regression simulation model based on coupled optimization deep learning are obtained; R1 c,s =R c,s +e c,s R1 v,s =R v,s +e v,s Where R1 c,s and R1 v,s The final simulation results of the nonlinear regression simulation model based on coupled optimization deep learning in the rate period and the verification period; Step 43, using Nash efficiency coefficient NSE, R 2 , RMSE evaluation indicators, to evaluate the accuracy of extreme low-flow runoff simulation by the nonlinear regression simulation model based on coupled optimization deep learning under each group of driving factor combinations, and take the driving factor combination with the best accuracy as the key driving factor combination affecting the deep change of low-flow runoff.

6. The method for attributing extreme low water changes and evaluating low water resource potential according to claim 1, characterized in that: The step S5 comprises: Step 51, based on the identified key driving factor combination, the constraint line derivation technology is used to obtain the constraint line of each driving factor in the key driving factor combination on the extreme low-flow runoff depth, revealing the constraint effect of each driving factor on the extreme low-flow runoff depth; the constraint line derivation steps are as follows: A two-dimensional scatter plot was drawn with each driving factor as the limiting variable and the extreme low-flow runoff depth as the response variable; The value range of the limiting variable is divided into several groups, and the 99.9% quantile point of the extreme low-water runoff depth value is selected from each group as the boundary point; Perform optimal fitting of polynomial functions on all boundary points to extract constraint lines and determine their functional equations; Step 52, based on the constraint line result, judging whether there is an optimal driving factor threshold among various driving factors; Step 53, based on the identification result of the driving factor threshold, evaluate the low water resource potential of the region in the corresponding typical extreme dry year when the driving factor is regulated; expressed as the average low water runoff depth of the basin, the calculation formula of the low water resource potential is: R T =R S -R C R S =f c (x s ) In the formula, R T is the low water resource potential of the research area under the conditions of a typical dry year, mm; R C represents the actual average low-water runoff depth of the basin in a typical year, mm; f c (x s ) is the corresponding constraint line equation; x s is the optimal driving factor threshold; R S It indicates the low-flow depth that the basin can reach when the driving factor is infinitely close to the optimal driving factor threshold.

Images