Drought risk prediction method based on artificial intelligence and computer readable medium

Through artificial intelligence-based drought risk prediction methods, combined with meteorological and hydrological and socio-economic data, a variety of machine learning models and regression models are constructed to predict future drought risks, and the multivariate characteristics and non-translational characteristics of drought events cannot be accurately portrayed in the existing technology. The problem of consistency characteristics provides detailed assessment and early warning capabilities for future drought risks.

CN118966414APending Publication Date: 2024-11-15CHINA THREE GORGES UNIV
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202410991251.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to accurately characterize the multivariate characteristic attributes and non-consistent characteristics of drought events, cannot effectively reflect future drought risk changes, and fail to fully consider the potential risks of the socio-economic system.

Method used

Using an artificial intelligence-based drought risk prediction method, the relative humidity is calculated by combining meteorological and hydrological reanalysis data, land water reserve inversion samples and socioeconomic samples, and the Clausius-Clapperon thermodynamic equation and specific humidity formula. The spatial sliding window method and random forest model select key factors, build long and short-term memory networks, logistic regression models and support vector machines, evaluate the simulation accuracy of the machine learning model, establish a GI regression model, predict future land water reserves and socio-economic risks, and A joint probability distribution model is constructed based on run theory and Copula function.

Benefits of technology

It has realized the characterization of multivariate characteristic attributes of drought events and accurate assessment of future risks, providing important reference for global and regional water resource risk assessment and early warning under climate change scenarios, and enhancing the scientificity and engineering nature of responding to climate disasters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966414A_ABST
    Figure CN118966414A_ABST
Patent Text Reader

Abstract

The invention discloses a drought risk prediction method based on artificial intelligence and a computer readable medium. The method comprises the following steps: selecting a drought risk prediction area, obtaining meteorological and hydrological reanalysis data, a land water reserve inversion sample, a climate mode prediction sample and a social and economic sample, and calculating relative humidity, specific humidity and wet bulb temperature; the method comprises the following steps: firstly, optimizing key factors influencing land water reserves in each grid point by combining a space sliding window method and a random forest model, predicting a future land water reserve data set and GI data, establishing a joint probability distribution function based on a'and 'return period under a non-consistency condition, and introducing a GI index to predict social and economic risks caused by future drought; the method not only can be applied to drought risk assessment and disaster prevention and reduction, but also can provide an important and highly operable reference basis for risk assessment and early warning of global and regional water resources under a climate change scene.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction and processing, and in particular to an artificial intelligence-based drought risk prediction method and a computer-readable medium. Background Art

[0002] Drought events have complex causes, long time spans, and strong destructive power. They are important factors that restrict the sustainable development of natural ecosystems and social economy. They are often divided into meteorological drought, hydrological drought, agricultural drought, and socioeconomic drought. Among them, meteorological drought and hydrological drought have significant impacts on water resource management and various water-related activities, and are important categories of drought events that have attracted much attention. Meteorological drought mainly refers to the phenomenon of less precipitation, which is mostly caused by abnormal atmospheric circulation; meteorological drought is the cause of hydrological drought. Less precipitation and higher temperatures may cause drought in soil water, river and lake runoff, and groundwater, which further triggers hydrological drought. Severe drought events are often the result of the gradual development of meteorological drought and hydrological drought. There are many factors affecting drought, including hydrology, meteorology, and vegetation, and they are often closely correlated. Therefore, in order to quantitatively characterize the degree of water deficit, domestic and foreign scholars have proposed a large number of single-factor and multi-factor comprehensive drought indicators such as standardized precipitation index, standardized precipitation evapotranspiration index, Palmer drought index, and standardized runoff index. Although the research objects and physical processes of concern of various drought indices are different, they mainly consider one or more meteorological and hydrological elements such as precipitation, evapotranspiration, runoff and soil moisture content, and cannot fully characterize the intrinsic physical characteristics of drought events.

[0003] In March 2002, the Gravity Recovery and Climate Experiment (GRACE) satellite was successfully launched, providing a continuous, high-precision direct observation method for obtaining global large-scale migration of Earth's surface materials. The gravity field model based on the GRACE satellite signal solution can extract the change information of the Earth's lunar gravity field at a spatial scale of 300km×300km, deduct the influence of factors such as crustal material movement, atmospheric movement, ocean currents and tides, and can effectively reflect the gravity changes caused by ice and snow, surface water, soil water, groundwater and human factors, and comprehensively monitor the terrestrial water storage anomaly (TWSA) signal. In May 2018, after the GRACE gravity satellite stopped working for a year, the GRACE-FO (GRACE Follow-On) gravity satellite was successfully launched, continuing the scientific observation mission of the GRACE satellite. The GRACE / GRACE-FO gravity satellite effectively solved the problems of shallow ground observation range, uneven spatial distribution, and difficulty in data acquisition, and showed great potential in tracking regional and global drought events.

[0004] The world is currently experiencing climate change dominated by "warming". The response of precipitation to climate warming has significant temporal and spatial heterogeneity, and its changes are affected by atmospheric water vapor capacity, relative humidity and atmospheric stability. Land water storage is affected by the synergistic effect of precipitation and evapotranspiration, and its response to global warming is more complex. It is not clear how drought will evolve in the future. Although GRACE / GRACE-FO satellites have begun to be used for drought monitoring and assessment, relevant research on predicting future drought based on land water storage anomalies has just started abroad, and there are few reports in China. At the same time, these documents usually only focus on changes in single characteristic attributes such as drought duration or intensity, and do not consider the inconsistency characteristics of drought. It is difficult to accurately characterize the multivariate characteristic attributes of drought events and cannot effectively reflect future changes in drought risks. At present, it has become a difficult problem to scientifically quantify multidimensional drought events under inconsistency conditions in the future, and how to quantitatively assess the potential risks of drought risk changes to the social and economic system also needs to be explored. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a drought risk prediction method based on artificial intelligence and a computer-readable medium to solve the problems raised in the background technology.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a drought risk prediction method based on artificial intelligence, comprising the following steps:

[0007] Step 1: Select a drought risk prediction area and obtain meteorological and hydrological reanalysis data, terrestrial water storage inversion samples, climate model prediction samples, and socio-economic samples within the drought risk prediction area;

[0008] Step 2: Input the 2-meter air temperature, dew point temperature, and air pressure of the meteorological and hydrological reanalysis data at each longitude and latitude coordinate point in each month of the reconstruction period into the Clausius-Clapeyron thermodynamic equation and specific humidity formula to calculate the relative humidity, specific humidity, and wet-bulb temperature at each longitude and latitude coordinate point in each month of the historical period;

[0009] Step 3: For each grid point in the drought risk prediction area, the key factors affecting terrestrial water storage in each grid point are selected by combining the spatial sliding window method and the random forest model; based on the selected key factors and the terrestrial water storage inversion samples, the optimized water storage long short-term memory network, the optimized water storage logistic regression model, and the optimized water storage support vector machine are obtained in turn by optimizing the training through the gradient descent method;

[0010] Step 4: For each grid point in the drought risk prediction area, evaluate the simulation accuracy of the machine learning model, and use the monthly multi-model weighted average model to optimize the weight of each machine learning model in each month;

[0011] Step 5: For the drought risk prediction area, establish a GI regression model of historical GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption, and deduce the parameters of the model;

[0012] Step 6: Input the future meteorological and hydrological variables predicted by the global climate model into the machine learning model and the monthly multi-model weighted average model established in Steps 4 and 5 to obtain the future terrestrial water storage data set; input the future industrial and agricultural water use and socio-economic data into the established GI regression model to predict future GI data;

[0013] Step 7: For the terrestrial water storage dataset under future scenarios, drought events are determined based on the run theory, and the annual average wet-bulb temperature is used as a covariate. A joint probability distribution model based on the “and” return period and the same frequency combination under non-consistent conditions is established, and the GI index is introduced to predict the socioeconomic risks caused by future droughts.

[0014] Preferably, the step 3 is specifically:

[0015] For each grid point, set the spatial sliding window threshold, slide it in sequence according to the interval range, and use the data of all grid points in the spatial range as the input of the machine learning model;

[0016] The random forest algorithm was used to construct a relationship model between the driving factors and the terrestrial water storage samples within the above spatial range. The driving factors included the average monthly temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity, and wet-bulb temperature. The time-lag effect of each driving factor on terrestrial water storage was considered. For each month, the driving factors of that month and the previous 1-3 months were selected as the input of the random forest model.

[0017] Based on the above random forest model, the important factors affecting terrestrial water storage were selected, 50% was set as the threshold, and the top 50% of the variables were selected as the key factors of water storage;

[0018] The above key factors of water reserves and historical terrestrial water reserves are input as samples into the long short-term memory network, and the long short-term memory network error loss function model is constructed. The optimized water reserve long short-term memory network is obtained by optimizing the training through the gradient descent method, and the long short-term memory simulated water reserve is obtained.

[0019] The above key factors of water reserves and historical terrestrial water reserves are input as samples into the logistic regression model, and the error loss function model of the logistic regression model is constructed. The optimized water reserve logistic regression model is obtained by optimizing the training through the gradient descent method, and the water reserve simulation of the logistic regression model is obtained;

[0020] The above-mentioned key factors of water reserves and historical terrestrial water reserves are input into the support vector machine as samples, and the support vector machine error loss function model is constructed. The optimized water reserve support vector machine is obtained by optimization training through the gradient descent method, and the support vector machine simulated water reserves are obtained.

[0021] Preferably, the step 4 is specifically:

[0022] The water storage data simulated by the long short-term memory network, the water storage data simulated by the logistic regression model, and the water storage data simulated by the support vector machine in the historical period are input into the monthly multi-mode weighted average model for calculation, and the water storage weight parameters of the optimized long short-term memory network, the optimized logistic regression model, and the optimized support vector machine in each month are obtained as follows:

[0023] For each month, the weight parameters of the water storage artificial intelligence model satisfy:

[0024]

[0025] Where i represents the i-th water storage artificial intelligence model, w tws represents the weight parameter of the combined scenario; LM represents the number of monthly runoff artificial intelligence models;

[0026] The weights are calculated using the following formula:

[0027]

[0028] Among them, w o (i) represents the weight parameter of the combined scenario, RBi represents the relative deviation between the terrestrial water storage simulated by the i-th water storage artificial intelligence model and the terrestrial water storage sample; AB i The absolute deviation between the terrestrial water storage simulated by the i-th water storage AI model and the terrestrial water storage sample.

[0029] Preferably, the step 5 is specifically:

[0030] Thiessen polygons were used to obtain the spatially averaged data on potential residential water consumption, potential electricity water consumption, potential irrigation water consumption, potential livestock water consumption, potential industrial water consumption and crop yield at an annual scale.

[0031] A GI regression model was established for the historical period GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption:

[0032] GI y =c1·PDWU y +c2·PEWU y +c3 PIGWU y +c4·PPWE y+c5·PMWE y +c6·CY y +c7·(POP y+ GDP y ) 2

[0033] Where: GI y is the standardized GI index for year y; PDWU y is the potential residential water consumption in year y, PEWU y is the potential electricity water consumption in year y, PIGWU y is the potential irrigation water consumption in year y, PPWE y is the potential livestock water consumption in year y, PMWE y is the potential industrial water consumption in year y, CY y is the crop yield in year y; POP y and GDP y are the population and GDP data for year y respectively; c1, c2, …, c7 are model parameters;

[0034] The GI regression model is calibrated using the least squares method to obtain model parameters c1, c2, ..., c7.

[0035] Preferably, the step 6 is specifically:

[0036] Under different shared socioeconomic pathways, the monthly average temperature and relative humidity of each global climate model are used to calculate the wet-bulb temperature. The monthly average temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity and wet-bulb temperature data for the future period predicted by each global climate model and the WaterGAP2-2e global hydrological model are then input into the machine learning model considering the spatial sliding window established in step 3 to predict the future terrestrial water storage series. Based on the monthly weights obtained in step 4, the terrestrial water storage prediction data sets for the historical period and future period of each global climate model are obtained, as follows:

[0037] Q(j)=ω k (j) ·Q k (j)

[0038] Where: Q(j) is the predicted land water storage data for the jth month after weighted average of various global climate models under a shared socioeconomic path; ω k (j) is the weight of the k-th water storage artificial intelligence model in the j-th month; Q k(j) is the simulated water storage of the kth artificial intelligence model in the jth month, k∈[1,3], if k=1, it means the optimized water storage long short-term memory network, if k=2, it means the optimized water storage random forest, if k=3, it means the optimized water storage support vector machine;

[0039] Under different shared socioeconomic pathways, the meteorological, hydrological and socioeconomic indices for future periods predicted by the global climate models, the WaterGAP2-2e global hydrological model and the LPJmL global vegetation model, as well as future population and GDP data, are input into the GI regression model established in step 5 to predict the GI index for future periods.

[0040] Preferably, the step 7 is specifically:

[0041] The TWS-DSI index is used to measure the dryness and wetness of land. A negative value of TWS-DSI indicates that the land water storage is lower than the average level of the study period, which is used to characterize the degree of drought. Similarly, a positive value can be used to measure the wetness of land. The calculation formula of the TWS-DSI series is as follows:

[0042]

[0043] Where: TWSA i,j represents the anomaly of TWS data in the jth month of the i-th year, and σ j are the mean and standard deviation of the TWS anomaly in the jth month during the study period; the reference period of the TWS anomaly is 2004-2009;

[0044] Based on the run theory, drought events under climate change are extracted. For each global climate model data under each SSP, the TWS-DSI index under climate change is first derived. Then, based on the run theory, the TWS-DSI index less than -0.8 is taken as the threshold to extract the drought duration D and drought intensity S in the historical period and future period respectively.

[0045] For any scenario under the climate scenario, the joint probability distribution function of drought duration and drought severity is constructed based on the Gumbel Copula function:

[0046]

[0047] in, is the Copula joint distribution function, the parameters of the Copula function The range is (1,∞); u t , v t are the probability density functions of the marginal distribution of drought duration D and drought severity S respectively;

[0048] Based on the definition of Copula function, the non-uniform two-variable Copula function is expressed as:

[0049]

[0050] Among them, F t (d t ,s t ) represents the time-varying joint distribution function of D and S; and Represent the time-varying marginal distribution function and time-varying parameters of the D and S variables respectively; further, the parameters of the time-varying Copula function are expressed by the covariate w as follows:

[0051]

[0052] Among them, g c represents the link function of the copula function, when hour, b0, b1 are the parameters of the model; w t is the average annual wet-bulb temperature in year t;

[0053] The same frequency combination model is:

[0054]

[0055] Where (d*, s*) represents a joint return period T AND is the most likely combination scenario of drought duration d and drought intensity s; μ is the average interval between drought events; where T AND is the "and" return period.

[0056] Solving the formula, we can obtain a return period T for each shared socioeconomic path and each global climate model historical period. h Corresponding drought duration and intensity (D h , S h ); further, construct the time-varying marginal distribution and Copula function of the future period, and convert (D h , S h ) is substituted into the time-varying distribution function of the kth sliding window in the future period, and the new return period T is calculated. f (k);

[0057] After obtaining the return period under each combination scenario, the weighted average method is used to obtain the average return period of the kth sliding window under M combination scenarios.

[0058]

[0059] Among them, i represents the combined scenario; wo (i) represents the weight parameter of the combined scenario; represents the return period of the kth sliding window under combination scenario i;

[0060] like It indicates that the drought risk in the kth window increases, otherwise it decreases; for the kth time window, the parameters of the marginal distribution and joint distribution are derived using the data of 15 years before and after the center point; the socioeconomic risk caused by drought in the future period is measured by the following formula:

[0061]

[0062]

[0063] Among them, E pop and E GDP The population and GDP risks affected by drought risk are represented by POP k and GDP k are the population and GDP of the kth year respectively; I(·) is the indicator function, The value is 1 when the time is right, otherwise it is 0; N1 and N2 represent the start and end years of the study period respectively; GI(k) is the GI index of the kth year predicted in step 6.

[0064] In addition, the present invention also discloses a computer-readable medium, which stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the above-mentioned artificial intelligence-based drought risk prediction method.

[0065] Beneficial effects of the invention: The invention uses a variety of artificial intelligence models to characterize the meteorological and hydrological information that affects terrestrial water reserves, takes into account the resilience index of future social and economic development, and provides an important and highly operational reference basis for global and regional water resources risk assessment and early warning under climate change scenarios, and provides engineering reference value for responding to future climate disasters and scientifically formulating emission reduction strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 : A flow chart of a method provided by an embodiment of the present invention;

[0067] Figure 2 : Schematic diagram of the spatial sliding window method;

[0068] Figure 3 : A schematic diagram of the run theory provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0070] Figure 1 1 is a flow chart of a drought risk prediction method based on artificial intelligence provided by an embodiment of the present invention, wherein the method comprises the following steps:

[0071] Step 1: Select a drought risk prediction area and obtain meteorological and hydrological reanalysis data, terrestrial water storage inversion samples, climate model prediction samples, and socio-economic samples within the drought risk prediction area;

[0072] Step 2: Input the 2-meter air temperature, dew point temperature, and air pressure of the meteorological and hydrological reanalysis data at each longitude and latitude coordinate point in each month of the reconstruction period into the Clausius-Clapeyron thermodynamic equation and specific humidity formula to calculate the relative humidity, specific humidity, and wet-bulb temperature at each longitude and latitude coordinate point in each month of the historical period;

[0073] Step 3: For each grid point in the drought risk prediction area, the key factors affecting terrestrial water storage in each grid point are selected by combining the spatial sliding window method and the random forest model; based on the selected key factors and the terrestrial water storage inversion samples, the optimized water storage long short-term memory network, the optimized water storage logistic regression model, and the optimized water storage support vector machine are obtained in turn by optimizing the training through the gradient descent method;

[0074] Step 4: For each grid point in the drought risk prediction area, evaluate the simulation accuracy of the machine learning model, and use the monthly multi-model weighted average model to optimize the weight of each machine learning model in each month;

[0075] Step 5: For the drought risk prediction area, establish a GI regression model of historical GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption, and deduce the parameters of the model;

[0076] Step 6: Input the future meteorological and hydrological variables predicted by the global climate model into the machine learning model and the monthly multi-model weighted average model established in Steps 4 and 5 to obtain the future terrestrial water storage data set; input the future industrial and agricultural water use and socio-economic data into the established GI regression model to predict future GI data;

[0077] Step 7, for the terrestrial water storage dataset under the future scenario, drought events are determined based on the run theory, the annual average wet-bulb temperature is used as a covariate, a joint probability distribution model based on the "and" return period and the same frequency combination under inconsistent conditions is established, and the GI index is introduced to predict the socioeconomic risks caused by future droughts.

[0078] Step 1 is specifically as follows:

[0079] The 0.25° terrestrial water storage data of each grid point in the drought risk prediction area during the period of 1940-2014 were obtained from the terrestrial water storage data set (GTWS-MLrec data set) from 1940 to 2022 reconstructed based on machine learning technology by institutions such as Wuhan University, Oxford University and Columbia University. The data set includes the inversion results of the benchmark data of three different gravity satellite institutions. In this embodiment, the weighted average series of the benchmark data of the three institutions is used as the terrestrial water storage sample.

[0080] ERA5 is the fifth generation of atmospheric reanalysis dataset of the European Centre for Medium-Range Weather Forecasts, with a spatial resolution of 0.25°, providing hourly meteorological data covering the globe since 1940. The present invention obtains hourly precipitation, air pressure, 2m air temperature, 2m dew point temperature, shortwave radiation and runoff depth data of the ERA5 dataset in the study area during 1940-2014, and obtains monthly series after time scale conversion;

[0081] The M6A-LR global climate model, the GFDL-ESM4 global climate model, the MPI-ESM1-2-HR global climate model, the MRI-ESM2-0 global climate model, and the UKESM1-0-LL global climate model were selected; the monthly average temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, and shortwave radiation intensity of each latitude and longitude coordinate point in the historical period and the future period under each climate model were obtained; the future period was 2030-2100, and the present invention adopted the prediction results of three different shared socioeconomic pathways (SSP126, SSP370, and SSP585);

[0082] The datasets obtained by driving the above five global climate models using the WaterGAP2-2e global hydrological model under the ISIMIP3b framework were used to obtain runoff depth, potential residential water consumption, potential electricity water consumption, potential irrigation water consumption, potential livestock water consumption, and potential industrial water consumption under three different shared socioeconomic pathways for each grid point in the historical and future periods. The datasets obtained by driving the above five global climate models using the LPJmL global vegetation model under the ISIMIP3b framework were used to obtain annual crop yield data for each grid point in the historical and future periods.

[0083] To further obtain the population and GDP data of the shared socio-economic path dataset, the embodiment of the present invention uses a certain population policy estimation dataset released by the Collaborative Innovation Center for Meteorological Disaster Forecasting and Assessment of the School of Geography of a certain university; this dataset takes into account the results of previous domestic population and economic censuses, as well as the annual statistical yearbooks. The economic data is predicted by using the Cobb-Douglas model and the Population-Environment-Development (PED) model to predict the domestic socio-economic index from 2010 to 2100. This dataset has been widely used to assess the socio-economic risks of extreme hydrological events;

[0084] The population and GDP data of the drought risk prediction area from 1996 to 2014 were obtained through the socio-economic yearbook; the Governance Indicator (GI) from 1996 to 2014 was obtained from the World Bank. The index can characterize the adaptability to climate disasters and contains data in six dimensions, generally ranging from -2.5 to 2.5. This embodiment uses a standardized method to first map the GI index of each dimension to the interval [-1, 1], and finally take the arithmetic average of the GI index of the six dimensions, thereby obtaining the standardized GI index from 1996 to 2014.

[0085] Step 2 is as follows:

[0086] The Clausius-Clapeyron thermodynamic equation is defined as follows:

[0087]

[0088] Where T0 is the first integral constant, which is 273.16K, e s0 is the second integral constant, take 611Pa, L v is the latent heat of vaporization constant, taken as 2.5×10 6 J kg -1 , R v is the water vapor gas constant, take 461 Jkg -1 K -1 , T is the input variable of the Clausius-Clapeyron thermodynamic equation;

[0089] The specific humidity is calculated using the following formula:

[0090]

[0091] Among them, q j,g represents the specific humidity of the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period, p j,g represents the air pressure at the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period, T 2m (j, g) represents the 2-meter temperature of the meteorological and hydrological reanalysis data at the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period, T dew (j, g) represents the dew point temperature at the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period;

[0092] Use air temperature and relative humidity to derive wet bulb temperature:

[0093]

[0094] Among them, T w(jg) represents the wet-bulb temperature of the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period, T 2m (j, g) represents the 2-meter temperature of the meteorological sample at the g-th longitude and latitude coordinate point in the reconstruction period, atan is the inverse tangent function, RH j,g represents the ground relative humidity at the g-th latitude and longitude coordinate point in the j-th month of the reconstruction period, T dew (j, g) represents the dew point temperature at the g-th longitude and latitude coordinate point in the th month of the reconstruction period;

[0095] Step 3 is as follows:

[0096] like Figure 2 As shown in the figure, for each grid point, 5×5 is set as the spatial sliding window threshold, and the window is slid sequentially according to the interval range. The data of all grid points in the spatial range are used as the input of the machine learning model. This method can improve the robustness of the machine learning model.

[0097] A random forest algorithm was used to construct a relationship model between the driving factors and the terrestrial water storage samples from 1940 to 2014 within the above spatial range, wherein the driving factors included average monthly temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity, and wet-bulb temperature (a total of 8 driving factors). Considering the time lag effect of each driving factor on terrestrial water storage, for each month, the driving factors of that month and the previous 1-3 months were selected as inputs of the random forest model, and each grid point in this embodiment had 8×3=24 variables as model inputs.

[0098] The random forest model can rank the importance of each input variable to the simulation variable. Therefore, based on the above random forest model, the important factors affecting terrestrial water storage are selected. In this embodiment, 50% is set as the threshold, and the top 50% of the variables are selected as the key factors of water storage;

[0099] The above key factors of water reserves and historical terrestrial water reserves are input as samples into the long short-term memory network, and the long short-term memory network error loss function model is constructed. The optimized water reserve long short-term memory network is obtained by optimizing the training through the gradient descent method, and the long short-term memory simulated water reserve is obtained.

[0100] The above key factors of water reserves and historical terrestrial water reserves are input as samples into the logistic regression model, and the error loss function model of the logistic regression model is constructed. The optimized water reserve logistic regression model is obtained by optimizing the training through the gradient descent method, and the water reserve simulation of the logistic regression model is obtained;

[0101] The above key factors of water reserves and historical terrestrial water reserves are input as samples into the support vector machine, and the error loss function model of the support vector machine is constructed. The optimized water reserve support vector machine is obtained by optimizing the training through the gradient descent method, and the simulated water reserve of the support vector machine is obtained;

[0102] Step 4 is as follows:

[0103] The water storage data simulated by the long short-term memory network, the water storage data simulated by the logistic regression model, and the water storage data simulated by the support vector machine in the historical period are input into the monthly multi-mode weighted average model for calculation, and the water storage weight parameters of the optimized long short-term memory network, the optimized logistic regression model, and the optimized support vector machine in each month are obtained as follows:

[0104] There are January to December in a year. In the weight derivation scheme, the weight of each month in different years is the same, so the weights of January to December need to be derived separately; the weight parameters of the combined scenario are obtained by normalizing the independent weight parameters and skill weight parameters under the combined scenario, as follows:

[0105] For each month, the weight parameters of the water storage artificial intelligence model satisfy:

[0106]

[0107] Where i represents the i-th water storage artificial intelligence model established in step 3, and w tws represents the weight parameter of the combined scenario; LM represents the number of monthly runoff artificial intelligence models, which is 3;

[0108] The weights are calculated using the following formula:

[0109]

[0110] Among them, w o (i) represents the weight parameter of the combined scenario, RB i Characterizes the relative deviation between the terrestrial water storage simulated by the i-th water storage artificial intelligence model and the terrestrial water storage sample; AB i The absolute deviation between the terrestrial water storage simulated by the i-th water storage AI model and the terrestrial water storage sample;

[0111] Step 5 is as follows:

[0112] Based on the prediction results of the WaterGAP2-2e global hydrological model and the LPJmL global vegetation model under the ISIMIP3b framework in step 1, the Thiessen polygons were used to obtain the spatially averaged potential residential water consumption, potential electricity water consumption, potential irrigation water consumption, potential livestock water consumption, potential industrial water consumption and crop yield data on an annual scale during the period 1996-2014;

[0113] Based on the standardized GI data in step 1, a GI regression model was established between the historical GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption:

[0114] GI y =c1·PDWU y +c2·PEWU y +c3 PIGWU y +c4·PPWE y +c5·PMWE y +c6·CY y +c7·(POP y+ GDP y ) 2 (6)

[0115] Where: GI y is the standardized GI index for year y; PDWU y is the potential residential water consumption in year y, PEWU y is the potential electricity water consumption in year y, PIGWU y is the potential irrigation water consumption in year y, PPWE y is the potential livestock water consumption in year y, PMWE y is the potential industrial water consumption in year y, CY y is the crop yield in year y; POP y and GDP y are the population and GDP data for year y respectively; c1, c2, …, c7 are model parameters.

[0116] The GI regression model is calibrated using the least squares method to obtain model parameters c1, c2, ..., c7.

[0117] Step 6 is as follows:

[0118] Under different shared socioeconomic pathways, the monthly average temperature and relative humidity of each global climate model are used to calculate the wet-bulb temperature. The monthly average temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity and wet-bulb temperature data for the future period predicted by each global climate model and the WaterGAP2-2e global hydrological model are then input into the three machine learning models considering spatial sliding windows established in step 3 to predict the future terrestrial water storage series. Based on the monthly weights obtained in step 4, the terrestrial water storage prediction data sets for the historical period (1985-2014) and the future period (2015-2100) of each global climate model are obtained, as follows:

[0119] Q(j)=ω k(j) ·Q k (j) (7)

[0120] Where: Q(j) is the predicted land water storage data for the month after weighted average of various global climate models under a shared socioeconomic path; ω k (j) is the weight of the k-th water reserve artificial intelligence model in the month; Q k (j) is the simulated water reserves of the kth artificial intelligence model in the jth month, k∈[1,3], if k=1, it represents the optimized water reserves long short-term memory network, if k=2, it represents the optimized water reserves random forest, and if k=3, it represents the optimized water reserves support vector machine.

[0121] Under different shared socioeconomic pathways, the meteorological, hydrological and socioeconomic indices for the future period (2015-2100) predicted by the global climate models, the WaterGAP2-2e global hydrological model and the LPJmL global vegetation model, as well as future population and GDP data, are input into the GI regression model established in step 5 to predict the GI index for the future period.

[0122] Step 7 is as follows:

[0123] The TWS-DSI index is used to measure the dryness and wetness of land. It is a dimensionless standardized water storage anomaly index that is spatially comparable between different hydrological and climatic regions. A negative value of TWS-DSI indicates that the land water storage is lower than the average level of the study period, which is used to characterize the degree of drought; similarly, a positive value can be used to measure the level of land moisture. The calculation formula of the TWS-DSI series is as follows:

[0124]

[0125] Where: TWSA i,j represents the anomaly of TWS data in the jth month of the i-th year, and σ j are the mean and standard deviation of the TWS anomaly in the first and second months of the study period respectively; the base period of the TWS anomaly is 2004-2009.

[0126] This paper selects the long series from 1985 to 2100 to calculate the monthly mean and standard deviation of TWSA, and divides the land dryness and wetness intensity into different levels based on the TWS-DSI index (see Table 1 for details).

[0127] Table 1 Dryness and wetness classification standards based on TWS-DSI index

[0128]

[0129] Based on the run theory, drought events under climate change are extracted. For each global climate model data under each SSP, the TWS-DSI index under climate change is first derived. Then, based on the run theory, the TWS-DSI index is less than -0.8 (preset drought threshold) as the threshold to extract the drought duration (D) and drought severity (S) in the historical period and future period. Figure 3 The schematic diagram of the run theory shown is that the run theory is usually used to identify drought events. When the drought index is lower than a certain threshold and the duration exceeds a certain length, a drought event is considered to have occurred.

[0130] For any scenario under the climate scenario, the joint probability distribution function of drought duration and drought severity is constructed based on the Gumbel Copula function:

[0131]

[0132] in, is the Copula joint distribution function, the parameter θ of the Copula function c t The range is (1,∞); u t , v t They are the probability density functions of the marginal distribution of drought duration D and drought severity S respectively.

[0133] Based on the definition of Copula function, the non-uniform two-variable Copula function can be expressed as:

[0134]

[0135] Among them, F t (d t ,s t ) represents the time-varying joint distribution function of D and S; and Denote the time-varying marginal distribution function and time-varying parameters of the D and S variables respectively. Further, the parameters of the time-varying Copula function are expressed by the covariate w as follows:

[0136]

[0137] Among them, g c represents the link function of the copula function, when When (for GH Copula), b0, b1 are the parameters of the model; w t is the average annual wet-bulb temperature in year t.

[0138] The same frequency combination model is:

[0139]

[0140] Where (d*, s*) represents a joint return period T AND is the most likely combination scenario of drought duration d and drought intensity s; μ is the average interval between drought events; where T AND is the "and" return period.

[0141] Solving formula (12), we can obtain the return period (T) of each shared socioeconomic path in each global climate model historical period (1985-2014). h ) corresponding to the drought duration and intensity (D h , S h ); further, using 30 years as a sliding window (consistent with the length of the historical period), we construct the time-varying marginal distribution and Copula function for the future period (2015-2100) and convert (D h , S h ) is substituted into the time-varying distribution function of the kth sliding window in the future period, and the new return period T is calculated. f (k).

[0142] After obtaining the return period under each combination scenario, the weighted average method is used to obtain the average return period of the kth sliding window under M combination scenarios.

[0143]

[0144] Among them, i represents the combined scenario; w o (i) represents the weight parameter of the combined scenario; represents the return period of the kth sliding window under combination scenario i.

[0145] like This indicates that the drought risk in the kth window increases, and vice versa. For the kth time window, the parameters of the marginal distribution and joint distribution are derived using the data of 15 years before and after the center point. The socioeconomic risk caused by drought in the future period is measured by the following formula:

[0146]

[0147]

[0148] Among them, E pop and E GDP The population and GDP risks affected by drought risk are represented by POP k and GDP k are the population and GDP of the kth year respectively; I(·) is the indicator function, The value is 1 when the time is right, otherwise it is 0; N1 and N2 represent the start and end years of the study period respectively; GI(k) is the GI index of the kth year predicted in step 6.

[0149] A specific embodiment of the present invention also provides a computer-readable medium.

[0150] The computer readable medium is a server workstation;

[0151] The server workstation stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of drought risk prediction based on artificial intelligence in an embodiment of the present invention.

[0152] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A drought risk prediction method based on artificial intelligence, characterized in that: The following steps are involved: Step 1: Select a drought risk prediction area and obtain meteorological and hydrological reanalysis data, terrestrial water storage inversion samples, climate model prediction samples, and socio-economic samples within the drought risk prediction area; Step 2: Input the 2-meter air temperature, dew point temperature, and air pressure of the meteorological and hydrological reanalysis data at each longitude and latitude coordinate point in each month of the reconstruction period into the Clausius-Clapeyron thermodynamic equation and specific humidity formula to calculate the relative humidity, specific humidity, and wet-bulb temperature at each longitude and latitude coordinate point in each month of the historical period; Step 3: For each grid point in the drought risk prediction area, the key factors affecting terrestrial water storage in each grid point are selected by combining the spatial sliding window method and the random forest model; based on the selected key factors and the terrestrial water storage inversion samples, the optimized water storage long short-term memory network, the optimized water storage logistic regression model, and the optimized water storage support vector machine are obtained in turn by optimizing the training through the gradient descent method; Step 4: For each grid point in the drought risk prediction area, evaluate the simulation accuracy of the machine learning model, and use the monthly multi-model weighted average model to optimize the weight of each machine learning model in each month; Step 5: For the drought risk prediction area, establish a GI regression model of historical GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption, and deduce the parameters of the model; Step 6: Input the future meteorological and hydrological variables predicted by the global climate model into the machine learning model and the monthly multi-model weighted average model established in Steps 4 and 5 to obtain the future terrestrial water storage data set; input the future industrial and agricultural water use and socio-economic data into the established GI regression model to predict future GI data; Step 7: For the terrestrial water storage dataset under future scenarios, drought events are determined based on the run theory, and the annual average wet-bulb temperature is used as a covariate. A joint probability distribution model based on the "and" return period and the same frequency combination under non-consistent conditions is established, and the GI index is introduced to predict the socioeconomic risks caused by future droughts.

2. The drought risk prediction method based on artificial intelligence according to claim 1, characterized in that: The step 3 is specifically as follows: For each grid point, set the spatial sliding window threshold, slide it in sequence according to the interval range, and use the data of all grid points in the spatial range as the input of the machine learning model; The random forest algorithm was used to construct a relationship model between the driving factors and the terrestrial water storage samples within the above spatial range. The driving factors included the average monthly temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity, and wet-bulb temperature. The time-lag effect of each driving factor on terrestrial water storage was considered. For each month, the driving factors of that month and the previous 1-3 months were selected as the input of the random forest model. Based on the above random forest model, the important factors affecting terrestrial water storage were selected, 50% was set as the threshold, and the top 50% of the variables were selected as the key factors of water storage; The above key factors of water reserves and historical terrestrial water reserves are input as samples into the long short-term memory network, and the long short-term memory network error loss function model is constructed. The optimized water reserve long short-term memory network is obtained by optimizing the training through the gradient descent method, and the long short-term memory simulated water reserve is obtained. The above key factors of water reserves and historical terrestrial water reserves are input as samples into the logistic regression model, and the error loss function model of the logistic regression model is constructed. The optimized water reserve logistic regression model is obtained by optimizing the training through the gradient descent method, and the water reserve simulation of the logistic regression model is obtained; The above-mentioned key factors of water reserves and historical terrestrial water reserves are input into the support vector machine as samples, and the support vector machine error loss function model is constructed. The optimized water reserve support vector machine is obtained by optimization training through the gradient descent method, and the support vector machine simulated water reserves are obtained.

3. The drought risk prediction method based on artificial intelligence according to claim 1, characterized in that: The step 4 is specifically as follows: The water storage data simulated by the long short-term memory network, the water storage data simulated by the logistic regression model, and the water storage data simulated by the support vector machine in the historical period are input into the monthly multi-mode weighted average model for calculation, and the water storage weight parameters of the optimized long short-term memory network, the optimized logistic regression model, and the optimized support vector machine in each month are obtained as follows: For each month, the weight parameters of the water storage artificial intelligence model satisfy: Where i represents the i-th water storage artificial intelligence model, w tws represents the weight parameter of the combined scenario; LM represents the number of monthly runoff artificial intelligence models; The weights are calculated using the following formula: Among them, w o (i) represents the weight parameter of the combined scenario, RB i Characterizes the relative deviation between the terrestrial water storage simulated by the i-th water storage artificial intelligence model and the terrestrial water storage sample; AB i The absolute deviation between the terrestrial water storage simulated by the i-th water storage AI model and the terrestrial water storage sample.

4. The drought risk prediction method based on artificial intelligence according to claim 1, characterized in that: The step 5 is specifically as follows: Thiessen polygons were used to obtain the spatially averaged data on potential residential water consumption, potential electricity water consumption, potential irrigation water consumption, potential livestock water consumption, potential industrial water consumption and crop yield at an annual scale. A GI regression model was established for the historical period GI and socioeconomic indices such as crop yield, population, GDP, industrial water consumption, and agricultural water consumption: GI y =c1·PDWU y +c2·PEUW y +c3·PIGWU y +c4·PPWE y +c5·PMWE y +c6·CY y +c7·(POP y+ ·GDP y ) 2 Where: GI y is the standardized GI index for year y; PDWU y is the potential residential water consumption in year y, PEWU y is the potential electricity water consumption in year y, PIGWU y is the potential irrigation water consumption in year y, PPWE y is the potential livestock water consumption in year y, PMWE y is the potential industrial water consumption in year y, CY y is the crop yield in year y; POP y and GDP y are the population and GDP data for year y respectively; c1, c2, …, c7 are model parameters; The GI regression model is calibrated using the least squares method to obtain model parameters c1, c2, ..., c7.

5. The drought risk prediction method based on artificial intelligence according to claim 2, characterized in that: The step 6 is specifically as follows: Under different shared socioeconomic pathways, the monthly average temperature and relative humidity of each global climate model are used to calculate the wet-bulb temperature. The monthly average temperature, snowfall, precipitation, relative humidity, near-ground wind speed, runoff depth, shortwave radiation intensity and wet-bulb temperature data for the future period predicted by each global climate model and the WaterGAP2-2e global hydrological model are then input into the machine learning model considering the spatial sliding window established in step 3 to predict the future terrestrial water storage series. Based on the monthly weights obtained in step 4, the terrestrial water storage prediction data sets for the historical period and future period of each global climate model are obtained, as follows: Q(j)=ω k (j)·Q k (j) Where: Q(j) is the predicted land water storage data for the jth month after weighted average of various global climate models under a shared socioeconomic path; ω k (j) is the weight of the k-th water storage artificial intelligence model in the j-th month; Q k (j) is the simulated water storage of the kth artificial intelligence model in the jth month, k∈[1,3], if k=1, it means the optimized water storage long short-term memory network, if k=2, it means the optimized water storage random forest, if k=3, it means the optimized water storage support vector machine; Under different shared socioeconomic pathways, the meteorological, hydrological and socioeconomic indices for future periods predicted by the global climate models, the WaterGAP2-2e global hydrological model and the LPJmL global vegetation model, as well as future population and GDP data, are input into the GI regression model established in step 5 to predict the GI index for future periods.

6. The drought risk prediction method based on artificial intelligence according to claim 1, characterized in that: The step 7 is specifically as follows: The TWS-DSI index is used to measure the dryness and wetness of land. A negative value of TWS-DSI indicates that the land water storage is lower than the average level of the study period, which is used to characterize the degree of drought. Similarly, a positive value can be used to measure the wetness of land. The calculation formula of the TWS-DSI series is as follows: Where: TWSA i,j represents the anomaly of TWS data in the jth month of the i-th year, and σ j are the mean and standard deviation of the TWS anomaly in the jth month during the study period; the reference period of the TWS anomaly is 2004-2009; Based on the run theory, drought events under climate change are extracted. For each global climate model data under each SSP, the TWS-DSI index under climate change is first derived. Then, based on the run theory, the TWS-DSI index less than -0.8 is taken as the threshold to extract the drought duration D and drought intensity S in the historical period and future period respectively. For any scenario under the climate scenario, the joint probability distribution function of drought duration and drought severity is constructed based on the Gumbel Copula function: in, is the Copula joint distribution function, the parameters of the Copula function The range is (1,∞); u t , v t are the probability density functions of the marginal distribution of drought duration D and drought severity S respectively; Based on the definition of Copula function, the non-uniform two-variable Copula function is expressed as: Among them, F t (d t ,s t ) represents the time-varying joint distribution function of D and S; and Represent the time-varying marginal distribution function and time-varying parameters of the D and S variables respectively; further, the parameters of the time-varying Copula function are expressed by the covariate w as follows: Among them, g c represents the link function of the copula function, when hour, b0, b1 are the parameters of the model; w t is the average annual wet-bulb temperature in year t; The same frequency combination model is: Where (d*, s*) represents a joint return period T AND is the most likely combination scenario of drought duration d and drought intensity s; μ is the average interval between drought events; where T AND is the return period of "and". Solving the formula, we can obtain a return period T for each shared socioeconomic path and each global climate model historical period. h Corresponding drought duration and intensity (D h , S h ); further, construct the time-varying marginal distribution and Copula function of the future period, and convert (D h , S h ) is substituted into the time-varying distribution function of the kth sliding window in the future period, and the new return period T is calculated. f (k); After obtaining the return period under each combination scenario, the weighted average method is used to obtain the average return period of the kth sliding window under M combination scenarios. Among them, i represents the combined scenario; w o (i) represents the weight parameter of the combined scenario; represents the return period of the kth sliding window under combination scenario i; like It indicates that the drought risk in the kth window increases, otherwise it decreases; for the kth time window, the parameters of the marginal distribution and joint distribution are derived using the data of 15 years before and after the center point; the socioeconomic risk caused by drought in the future period is measured by the following formula: Among them, E pop and E GDP The population and GDP risks affected by drought risk are represented by POP k and GDP k are the population and GDP of the kth year respectively; I(·) is the indicator function, The value is 1 when the time is right, otherwise it is 0; N1 and N2 represent the start and end years of the study period respectively; GI(k) is the GI index of the kth year predicted in step 6.

7. A computer-readable medium, characterized in that: It stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the drought risk prediction method based on artificial intelligence as described in any one of claims 1 to 6.

Citation Information

Cited By

  • Drought and flood sharp turning prediction method and device based on artificial intelligence and satellite remote sensing

    CN119202482A

  • A method and device for predicting sudden drought and flood transition based on artificial intelligence and satellite remote sensing

    CN119202482B

  • Regional drought toughness simulation analysis method and device

    CN119475813A

  • Non-consistent design flood estimation method and system considering climate and human activities

    CN119962158A

  • Land water reserve prediction method and system based on artificial intelligence, and electronic equipment

    CN120046798A