Drought forecasting method and system based on artificial intelligence and atmospheric circulation mechanism

Abstract

The present application provides a kind of drought forecasting method and system based on artificial intelligence and atmospheric circulation mechanism, comprising: S1, obtaining alternative precipitation runoff influence factor;S2, the importance of alternative precipitation runoff influence factor is sorted and the final precipitation runoff influence factor and precipitation runoff state number k value are determined;S3, with the Bayesian-NHMM model of minimum precipitation runoff prediction error as the best precipitation runoff prediction model, obtain precipitation and runoff prediction result;S4, realize the classification of land water storage forecast factor;S5, train machine learning model, and use the trained machine learning model as land water storage forecast model;S6, realize drought forecasting and early warning based on land water storage forecast model.The present application makes full use of the advantages of Bayesian-NHMM model and current machine learning model, makes up for the problem of low accuracy of traditional drought forecasting, and can deduce future drought risk and carry out drought risk early warning.

Description

Technical Field

[0001] This invention belongs to the field of drought prediction technology, and more specifically, relates to a drought forecasting method and system based on artificial intelligence and atmospheric circulation mechanisms. Background Technology

[0002] Drought events are complex in origin, span long periods, and are highly destructive, posing a significant challenge to the sustainable development of natural ecosystems and socio-economic systems. They are often categorized into meteorological drought, hydrological drought, agricultural drought, and socio-economic drought. Meteorological drought primarily refers to insufficient precipitation, often triggered by abnormal atmospheric circulation. Meteorological drought is a contributing factor to hydrological drought; insufficient precipitation and higher temperatures can lead to drought in soil water, river and lake runoff, and groundwater, further triggering hydrological drought. Drought is influenced by numerous factors, including hydrological, meteorological, and vegetation factors, which are often closely correlated. Therefore, to quantitatively characterize the degree of water deficit, scholars both domestically and internationally have proposed numerous single-factor and multi-factor comprehensive drought indices, such as the standardized precipitation index, standardized precipitation evapotranspiration index, Palmer drought index, and standardized runoff index. Although the research objects and physical processes focused on by each drought index differ, they primarily consider one or more meteorological and hydrological elements such as precipitation, evapotranspiration, runoff, and soil moisture content, thus failing to comprehensively characterize the intrinsic physical features of drought events.

[0003] In March 2002, the Gravity Recovery and Climate Experiment (GRACE) satellite was successfully launched, providing a continuous and high-precision direct observation method for obtaining large-scale global observations of material migration on the Earth's surface. Based on the gravity field model calculated from GRACE satellite signals, it can extract information on the variation of the Earth's monthly gravity field on a 300km × 300km spatial scale. After deducting the influence of crustal material movement, atmospheric motion, ocean currents, and tides, it can effectively reflect gravity changes caused by ice and snow, surface water, soil water, groundwater, and anthropogenic factors. It comprehensively monitors Terrestrial Water Storage Anomaly (TWSA) signals and shows great potential in tracking regional and global drought events.

[0004] Although the GRACE satellite has begun to be used for drought monitoring and assessment, research on drought assessment based on terrestrial water storage anomalies is still in its early stages abroad, and there are very few reports in China. Precipitation processes are a crucial factor affecting terrestrial water storage, driven not only by teleconnections of large-scale climatic factors but also closely related to local hydrological and meteorological factors and underlying surface conditions. One of the challenges in conducting high-precision drought event forecasting is how to fully consider the impact of large-scale and local factors on regional precipitation processes, construct reasonable and reliable precipitation forecasting models, and couple them with regional terrestrial water storage forecasting models. Furthermore, another major challenge in drought forecasting is how to comprehensively consider the multi-factor impacts under changing environments, downscale local hydrological and meteorological factors and underlying surface cover conditions into drought forecasts, obtain watershed drought forecasting information based on physical mechanism analysis, extend the flood forecast lead time, improve forecast accuracy, and construct a drought early warning mechanism. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of existing technologies, the present invention provides a drought forecasting method and system based on artificial intelligence and atmospheric circulation mechanisms. Its purpose is to solve the problems of short forecast period and low forecast accuracy of floods in complex environments.

[0006] Based on the above problems, the present invention adopts the following technical solution:

[0007] A drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms includes the following steps:

[0008] S1. Acquire data including large-scale climate factor data, hydro-meteorological factor data, and underlying surface impact factor data, and deduce relative humidity and specific humidity as alternative precipitation runoff impact factors;

[0009] S2. The importance of the candidate precipitation runoff influencing factors was initially ranked. Principal component analysis was used to reduce the dimensionality of the precipitation runoff influence. Finally, k-fold cross-validation was used to determine the final precipitation runoff influencing factors. The number of precipitation runoff states k in the Bayesian-NHMM precipitation probability forecasting model was determined by the BIC function.

[0010] S3. Based on the final precipitation runoff influencing factor and the number of precipitation runoff states k, construct a Bayesian-NHMM model with combinations of the final precipitation runoff influencing factor and different numbers of precipitation runoff states k. Use the Bayesian-NHMM model with the smallest precipitation runoff forecasting error as the best precipitation runoff forecasting model to obtain precipitation and runoff forecasting results.

[0011] S4. Considering the combined impact of climate change and underlying surface changes on runoff, relevant hydro-meteorological elements, SIF, LAI and Bayesian-NHMM model-predicted precipitation values ​​are selected as clustering variables, and K-means clustering method is used to classify and grade the terrestrial water storage forecasting factors.

[0012] S5. Train the machine learning model using N types of training sub-samples to obtain N trained machine learning models as land water storage prediction models.

[0013] S6. Drought forecasting and early warning based on terrestrial water storage forecasting models.

[0014] Furthermore, step 2 includes the following sub-steps:

[0015] S2.1: Calculate and rank the correlations between candidate precipitation runoff influencing factors and precipitation amount;

[0016] S2.2: Determine the number of precipitation states in the Bayesian-NHMM precipitation probability forecast model based on the BIC function;

[0017] S2.3: The importance of runoff influencing factors is initially ranked based on linear correlation coefficient and nonlinear correlation mutual information index. Principal component analysis is used to reduce the dimensionality of the influencing factors, and k-fold cross-validation is used to determine the final runoff influencing factors. The number of runoff states in Bayesian-NHMM is determined by the BIC function.

[0018] Further, step S2.1 specifically includes:

[0019] (1) Calculate the correlation coefficient between each candidate precipitation forecast factor and precipitation amount: linear correlation is calculated using Pearson correlation coefficient; nonlinear correlation is calculated using mutual information index. Perform linear and nonlinear analyses on all candidate precipitation forecast factors and precipitation amounts respectively, and determine whether they belong to linear / nonlinear relationships based on the significance level.

[0020] (2) Optimize precipitation influencing factors based on principal component analysis: Select precipitation forecasting factors with correlation coefficients exceeding the preset threshold, and perform dimensionality reduction processing on the selected precipitation forecasting factors through principal component analysis;

[0021] (3) Determine the final precipitation impact factor through k-fold cross-validation: Perform k-fold cross-validation on the processed precipitation forecast factor. If the validation fails, adjust the preset threshold and return to step (2). If the validation passes, the processed precipitation forecast factor is the final precipitation impact factor.

[0022] Further, step S2.2 specifically includes:

[0023] The BIC function value is calculated as follows:

[0024] BIC = 2P - klog(T)

[0025] In the formula: P is the maximum likelihood estimate, which is related to the model; k is the number of precipitation states in the model; T is the number of days of data;

[0026] Maximum likelihood estimation of precipitation probability distribution based on hidden state (Z) P(R) t |X,Z,δ,θ) can be calculated using the following formula:

[0027]

[0028]

[0029] In the formula: t = 1, 2, ..., T represents time points; R t Let X be the observed precipitation at time t. t =X t,1 ,X t,2 ,…X t,p Let p be the precipitation forecast factors at time t, θ be the unknown parameter related to the probability matrix, δ be the unknown parameter related to the transformation probability, and ρ be the unknown parameter related to the transformation probability. j To be with X t The coefficients in the relevant probability distribution, let δ = δ i,j =(ρ j ,ε i,j );

[0030] Using the above formula, the BIC function values ​​corresponding to different k values ​​of the Bayesian-NHMM model are calculated, thereby establishing a correlation curve between the k value and the BIC function value. The k value corresponding to each inflection point in the correlation curve is used as the candidate value. If there is no inflection point, the k value corresponding to the minimum BIC value is selected.

[0031] Further, step S3 includes:

[0032] Based on the final precipitation runoff influencing factors and the number of precipitation runoff states k determined in step S2, Bayesian-NHMM models with different combinations of precipitation runoff forecasting factors are constructed. The root mean square error offset coefficient (CVRMSE) of the simulation results for different precipitation forecasting factors is calculated, i.e., the root mean square error offset coefficient (CVRMSE) between the precipitation runoff prediction results and the observed results of the Bayesian-NHMM model when k takes different alternative values. Different performances of multiple models are obtained. The CVRMSE calculation formula is as follows:

[0033]

[0034] In the formula: np is the number of forecast days; S iR represents the predicted value of precipitation or runoff on day i; i This represents the observed value of precipitation or runoff on day i.

[0035] For precipitation and runoff, the model with the smallest CVRMSE was selected as the optimal precipitation and runoff prediction model. Based on the optimal precipitation and runoff prediction model, precipitation and runoff were predicted using the final precipitation and runoff influencing factors.

[0036] Furthermore, step S5 specifically includes the following sub-steps:

[0037] S5.1: After preliminary screening of physical factors affecting the formation of terrestrial water storage and obtaining the final influencing factors, the physical factors are downscaled to simulated terrestrial water storage based on artificial neural network (ANN), support vector machine (SVM), random forest model (RF), and long short-term memory model (LSTM), and the future terrestrial water storage forecast is carried out respectively.

[0038] S5.2: Construct the probability density function of land water storage for each model based on Bayes' law of total probability;

[0039] S5.3: Determine the corresponding weights based on the relative contributions of the prediction effects of Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) to establish a Bayesian pattern average correction model.

[0040] Furthermore, in step S5.2, the probability density function of S is obtained according to Bayes' law of total probability as follows:

[0041]

[0042] In the formula: p k (S|f k R) represents the Kth machine learning model f k Given data R, the probability density function of the predicted value S; p(f k |R) is the posterior probability density function of the k-th prediction value of the machine learning model given the training data R, S is the simulated terrestrial water storage, R=[D,O] represents the model input data, where D is the input series of the machine learning model, O is the terrestrial water storage series retrieved by gravity satellite, and f=[f1,f2,…,f K ] represents the output results of K different machine learning models.

[0043] Further, step S5.3 specifically includes:

[0044] The Box-Cox function is used to perform a normal transformation on the gravity satellite inversion series and the simulation series obtained from various machine learning models at each grid point. Then, a weighted average is performed on the estimation results of multiple models based on the assumption of normal linear distribution.

[0045]

[0046] In the formula: The mean is f k The variance is The function follows a normal distribution; E represents the expected value of the function, w k Let be the weights of the k-th machine learning model.

[0047] Furthermore, in step 6, based on the land water storage forecasting model, the land water storage forecast value is obtained. Based on the land water storage forecast value and the derived TWSA-DSI index, early warning of drought events and their levels is achieved.

[0048] On the other hand, the present invention also provides a system for drought forecasting based on artificial intelligence and atmospheric circulation mechanisms, comprising:

[0049] The database construction module is used to collect various terrestrial water storage-related data in the study area as alternative precipitation runoff influencing factors.

[0050] The first parameter estimation module is used to calculate the correlation coefficient between each candidate influencing factor and precipitation runoff, screen out influencing factors whose correlation coefficient exceeds a preset threshold, and then process the screened precipitation runoff influencing factors through principal component analysis to obtain the final precipitation runoff influencing factors; construct a Bayesian-NHMM model and determine several candidate values ​​for the number of precipitation runoff states k in the Bayesian-NHMM model.

[0051] The precipitation runoff forecasting model building module is used to input the final precipitation runoff influencing factor k into the Bayesian-NHMM model with different alternative values, and to forecast precipitation runoff. The Bayesian-NHMM model with the smallest precipitation runoff forecasting error is taken as the best precipitation runoff forecasting model.

[0052] The second parameter estimation module is used to construct training samples based on a series of precipitation runoff values ​​predicted by the best precipitation runoff forecasting model, and to classify the training samples to obtain N types of training sub-samples.

[0053] The forecast model training module is used to train the machine learning model using N types of training subsamples, resulting in N trained machine learning models that serve as land water storage forecast models.

[0054] The joint forecasting module enables drought forecasting and early warning based on this terrestrial water storage forecasting model.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. This invention uses a Bayesian-NHMM model to predict precipitation runoff, and then uses the predicted values ​​as training data to train a machine learning model to predict terrestrial water storage. It fully utilizes the advantages of the Bayesian-NHMM mathematical statistical model and current machine learning models, making up for the problem of low accuracy in traditional drought forecasting. It can infer future drought risks and carry out drought risk early warning, providing a basis for disaster prevention and mitigation departments to initiate emergency responses.

[0057] 2. This invention makes full use of measured hydrological and meteorological data in the study area, especially considering the impact of current climate change and underlying surface changes (focusing on the influence of vegetation photosynthesis and respiration) on the watershed runoff generation and confluence mechanism. It couples the Bayesian-NHMM precipitation-runoff forecasting model with the runoff machine learning model, which has a strong statistical basis and can objectively reflect the regional physical characteristics. Attached Figure Description

[0058] Figure 1 This is a flowchart of a drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to an embodiment of the present invention;

[0059] Figure 2 This is a flowchart of the screening process for precipitation influencing factors according to an embodiment of the present invention;

[0060] Figure 3 This is a flowchart of the Bayesian-NHMM model in an embodiment of the present invention. Detailed Implementation

[0061] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] This invention is based on measured large-scale climate factors and hydrological and meteorological data of the study area, as well as long series of LAI and SIF indices. It fully considers the influence of various large-scale and local factors on precipitation and runoff. Principal component analysis and k-fold cross-validation are used to determine the influencing factors for precipitation and runoff forecasting. A Bayesian-NHMM forecasting model is constructed using the BIC criterion to forecast precipitation and runoff. The K-means clustering method is used to classify and grade the terrestrial water storage forecasting factors. Various terrestrial water storage forecasting factors are input into artificial neural networks (ANN), support vector machines (SVM), random forest models (RF), and long short-term memory models (LSTM) to obtain regional forecast terrestrial water storage combinations under the influence of different physical factors. The forecast results of each model are averaged using the Bayesian method to obtain the final terrestrial water storage forecast results for the study area and to predict drought risk warning results.

[0063] The technical solution of the present invention will be further described in detail below through embodiments and with reference to the accompanying drawings.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment provides a drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms, including the following steps:

[0066] S1. Acquire data including large-scale climate factor data, hydro-meteorological factor data, and underlying surface impact factor data, and deduce relative humidity and specific humidity as alternative precipitation runoff impact factors;

[0067] In this embodiment, ENSO index, atmospheric circulation index, Indian Ocean Dipole, etc., are obtained from large-scale climate factors to construct a large-scale climate factor database.

[0068] The study area was defined, and meteorological and hydrological data, including precipitation, air pressure, 2-meter air temperature, 2-meter dew point temperature, shortwave radiation, longwave radiation, runoff, and soil moisture, were obtained from the ERA5 reanalysis dataset to construct a hydro-meteorological factor database. ERA5 is the fifth-generation 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 1979.

[0069] To acquire terrestrial water storage anomaly data retrieved from the GRACE gravity satellite, international institutions such as the Potsdam Geoscience Center in Germany, the Jet Propulsion Laboratory (JPL) at Caltech, the Center for Space Research at the University of Texas at Austin (CSR), and NASA's Goddard Space Flight Center (GSFC) are primarily responsible for data interpretation and releasing monthly-scale global gravity field output data. This invention utilizes the latest sixth-generation products from JPL, CSR, and GSFC, interpolating these three GRACE gravity satellite datasets to a 0.25°×0.25° spatial raster (making the data consistent with the spatial resolution of ERA5). The average value of each product is taken at each time step, ultimately yielding a TWSA monthly dataset covering the study area from 2002 to 2022.

[0070] In addition, leaf area index (LAI, an important parameter for measuring the intensity of energy and material exchange between the ecosystem and the atmosphere) and sunlight-induced chlorophyll fluorescence (SIF, an important parameter for indicating vegetation photosynthesis and physiological status) were collected in the study area to construct a database of underlying surface influencing factors.

[0071] Relative humidity and specific humidity are derived from meteorological data based on the ERA5 dataset.

[0072] The Clausius-Clapeyron thermodynamic equation can quantitatively describe the saturated water vapor pressure e. sat Nonlinear relationship with temperature T:

[0073]

[0074] Among them, T0 and e s0 Let L be the first and second integration constants, taken as 273.16 K and 611 Pa respectively; v Let be the latent heat of vaporization constant, taken as 2.5 × 10⁻⁶. 6 J kg -1 ;R v Let be the water vapor gas constant, taken as 461 J kg. -1 K -1 .

[0075] Dew point temperature characterizes the temperature at which air, under constant water vapor content and pressure, cools to water vapor saturation. Substituting this into the Clausius-Clapeyron equation, it can measure the actual water vapor pressure. ERA5 2m air temperature (T) 2m ) and dew point temperature (T dew Substituting these values ​​into formula (1), we can derive the near-ground relative humidity RH = e sat (T dew ) / e sat (T 2m ).

[0076] Specific humidity q is the ratio of water vapor mass to the total mass of the air mass, calculated using ERA5 near-surface pressure p and dew point temperature, as shown in the following formula:

[0077]

[0078] S2. The importance of the candidate precipitation runoff influencing factors was initially ranked. Principal component analysis was used to reduce the dimensionality of the precipitation runoff influence. Finally, k-fold cross-validation was used to determine the final precipitation runoff influencing factors. The number of precipitation runoff states k in the Bayesian-NHMM precipitation probability forecasting model was determined by the BIC function.

[0079] Taking precipitation as an example, such as Figure 2As shown, the importance of precipitation influencing factors was initially ranked based on linear correlation coefficients and nonlinear correlation mutual information indices. Principal component analysis was used to reduce the dimensionality of the influencing factors, and k-fold cross-validation was used to determine the final precipitation influencing factors. The number of precipitation states in the Bayesian Hidden Markov Precipitation Probability Prediction Model (Bayesian-NHMM) was determined using the BIC function. Using the above methods, the influencing factors and the number of states of runoff were obtained.

[0080] Specifically, step S2 includes:

[0081] S2.1: Calculate and rank the correlations between candidate precipitation influencing factors and precipitation amount.

[0082] (1) Calculate the correlation coefficients between each candidate precipitation forecast factor and precipitation amount: linear correlation is calculated using the Pearson correlation coefficient; nonlinear correlation is calculated using the mutual information index. Specifically, linear and nonlinear analyses are performed on all candidate precipitation forecast factors and precipitation amounts respectively, and the specific linear / nonlinear relationship is determined based on the significance level; the candidate precipitation forecast factors include: LAI, SIF, air pressure, 2m air temperature, 2m dew point temperature, shortwave radiation, longwave radiation, relative humidity and specific humidity, and other meteorological and underlying surface factors and large-scale climate factors. In addition, considering the lag effect of large-scale climate factors, LAI and SIF on regional precipitation, when calculating their correlation with regional precipitation, the correlation between each factor and precipitation for the previous 1-12 months is calculated first, and then the months with the largest correlation between each factor and precipitation are selected.

[0083] (2) Selection of precipitation influencing factors based on principal component analysis: Select precipitation forecasting factors with correlation coefficients exceeding the preset threshold, and perform dimensionality reduction processing on the selected precipitation forecasting factors through principal component analysis.

[0084] (3) Determine the final precipitation impact factor through k-fold cross-validation: Perform k-fold cross-validation on the processed precipitation forecast factor. If the validation fails, adjust the preset threshold and return to step (2); if the validation passes, the processed precipitation forecast factor is the final precipitation impact factor. In this embodiment, k = 0.6 is used for k-fold cross-validation, that is, calibration is performed with 0.6 times the data length and validation is performed with 0.4 times the data length to determine the final precipitation impact factor.

[0085] S2.2: Determine the number of precipitation states in the Bayesian Hidden Markov Precipitation Probability Prediction Model (Bayesian-NHMM) based on the BIC function.

[0086] like Figure 3As shown, the Bayesian-NHMM model is a model based on hidden state transformation. In the model, the precipitation of each day corresponds to a hidden state, and the state transformation is determined by the features of a first-order Markov chain. Figure 3 China Q t Let X be the probability transformation matrix considering the influencing variables at time t. t,b X t,a These represent different categories of exogenous influencing variables. The transition probabilities between different states change over time and are closely related to the input forecast factors. Therefore, the Bayesian-NHMM model can reflect the spatial correlation and temporal dependence between precipitation and forecast factors. The choice of the number of hidden states significantly impacts the model's simulation results; therefore, determining the number of precipitation states is a crucial step in model building. The number of states, k, can be determined by comparing the BIC function values ​​of models with different k values. The BIC function value is calculated as follows:

[0087] BIC=2P-klog(T) (3)

[0088] In the formula: P is the maximum likelihood estimate, which is related to the model; k is the number of precipitation states in the model; T is the number of days of data. Maximum likelihood estimate P(R) based on the hidden state (Z) precipitation probability distribution. t |X,Z,δ,θ) can be calculated using the following formula:

[0089]

[0090]

[0091] In the formula: t = 1, 2, ..., T represents time points; R t Let X be the observed precipitation at time t. t =X t,1 ,X t,2 ,…X t,p Let p be the precipitation forecast factors at time t, θ be the unknown parameter related to the probability matrix, δ be the unknown parameter related to the transformation probability, and ρ be the unknown parameter related to the transformation probability. j To be with X t The coefficients in the relevant probability distribution, let δ = δ i,j =(ρ j ,ε i,j ).

[0092] Using the above formula, the BIC function values ​​of the Bayesian-NHMM model for different k values ​​are calculated, thus establishing a correlation curve between the k value and the BIC function value. The k values ​​corresponding to each inflection point in this correlation curve are used as candidate values. If no inflection point exists, the k value corresponding to the minimum BIC value is selected.

[0093] S2.3: Based on the above method, the importance of runoff influencing factors is initially ranked based on linear correlation coefficient and nonlinear correlation mutual information index. Principal component analysis is used to reduce the dimensionality of the influencing factors. Finally, k-fold cross-validation is used to determine the final runoff influencing factors. The number of runoff states in Bayesian-NHMM is determined by the BIC function.

[0094] S3. Based on the final precipitation runoff influencing factor and the number of precipitation runoff states k, construct a Bayesian-NHMM model with combinations of the final precipitation runoff influencing factor and different numbers of precipitation runoff states k. Use the Bayesian-NHMM model with the smallest precipitation runoff forecasting error as the best precipitation runoff forecasting model to obtain precipitation and runoff forecasting results.

[0095] Specifically, step S3 includes:

[0096] Based on the final precipitation runoff influencing factors and the number of precipitation runoff states k determined in step S2, Bayesian-NHMM models with different combinations of precipitation runoff forecasting factors are constructed. Then, the root mean square error (CVRMSE) of the simulation results for different precipitation forecasting factors is calculated; that is, the CVRMSE of the precipitation runoff prediction results and the observed results of the Bayesian-NHMM model when k takes different alternative values, obtaining different performances of multiple models. The CVRMSE calculation formula is as follows:

[0097]

[0098] In the formula: np is the number of forecast days; S i R represents the predicted value of precipitation or runoff on day i; i This represents the observed value of precipitation or runoff on day i.

[0099] For precipitation and runoff, the model with the smallest CVRMSE was selected as the optimal precipitation and runoff prediction model. Based on the optimal precipitation and runoff prediction model, precipitation and runoff were predicted using the final precipitation and runoff influencing factors.

[0100] S4. Considering the combined impact of climate change and underlying surface changes on runoff, relevant hydro-meteorological elements, SIF, LAI and Bayesian-NHMM model-predicted precipitation values ​​are selected as clustering variables, and K-means clustering method is used to classify and grade the terrestrial water storage forecasting factors.

[0101] Specifically, step S4 includes:

[0102] K-means clustering is a classic partitioning clustering method. Its advantages include simple principle, fast convergence speed, and good clustering results. Furthermore, its algorithm is highly interpretable, requiring adjustment only of the number of clusters. The basic principle of K-means is: given a sample set, divide the set into n clusters based on the distance between samples, with the goal of making the points within each cluster as close together as possible while maximizing the distance between clusters.

[0103] In this embodiment, considering that the main factors affecting terrestrial water storage are precipitation and runoff, the precipitation and runoff values ​​for the next period of time are predicted by the optimal precipitation and runoff forecasting model. Then, the obtained precipitation values ​​are divided into three categories: low, medium, and high, i.e., the number of clusters is initially determined to be 3. Then, the distance between other factors and the three categories is calculated according to the K-means clustering algorithm, thereby classifying all training samples and obtaining three sub-samples.

[0104] S5. Train the machine learning model using N types of training sub-samples to obtain N trained machine learning models as land water storage prediction models.

[0105] Specifically, step S5 includes:

[0106] S5.1: After preliminary screening of physical factors affecting the formation of terrestrial water storage and obtaining the final influencing factors, the physical factors are downscaled to simulate terrestrial water storage based on artificial neural network (ANN), support vector machine (SVM), random forest model (RF), and long short-term memory model (LSTM), and the future terrestrial water storage forecast is carried out respectively.

[0107] The LSTM, ANN, SVM and RF were trained using the three types of subsamples obtained in step S4, and the land water storage prediction results for each grid point were obtained.

[0108] S5.2: Construct the probability density function of each model to simulate the land water storage based on Bayes' law of total probability.

[0109] Specifically, let S represent the simulated terrestrial water storage, R = [D, O] characterize the model input data (where D is the input series of the machine learning model, and O is the terrestrial water storage series retrieved from gravity satellites), and f = [f1, f2, ..., f K Given the outputs of K different machine learning models, the probability density function of S, obtained using Bayes' law of total probability, is as follows:

[0110]

[0111] In the formula: p k (S|f k R) represents the Kth machine learning model f kGiven data R, the probability density function of the predicted value S; p(f k |R) is the posterior probability density function of the k-th machine learning model prediction value given training data R.

[0112] S5.3: Determine the corresponding weights based on the relative contributions of each machine learning model to the prediction performance, thereby establishing a Bayesian pattern average correction model;

[0113] Specifically, firstly, the gravity satellite inversion series and the simulation series obtained from various machine learning models at each grid point are normally transformed using the Box-Cox function, and then the estimation results of multiple models are weighted and averaged based on the assumption of normal linear distribution.

[0114]

[0115] In the formula: The mean is f k The variance is The function follows a normal distribution; E represents the expected value of the function, w k Let be the weights of the k-th machine learning model.

[0116] Furthermore, in this embodiment, K = 4.

[0117] For a certain sub-sample, based on the calculated weight w k Four pre-trained machine learning models were weighted to obtain the terrestrial water storage forecasting model corresponding to each sample type, thus yielding terrestrial water storage forecasting models for the three sample types. This model effectively considers the dynamic impacts of climate change and underlying surface factors on the physical mechanisms of the study area, and can effectively extend the forecast period and improve the accuracy of terrestrial water storage forecasts.

[0118] S6. Drought forecasting and early warning based on terrestrial water storage forecasting models.

[0119] First, relevant data of the study area are acquired in real time (specifically, real-time measured hydrological and meteorological factor data). The similarity between the relevant data and the three types of training sub-samples is calculated. The land water storage forecasting model corresponding to the training sub-sample with the highest similarity is selected. The relevant data is input into the land water storage forecasting model to obtain the forecasted land water storage.

[0120] The TWSA-DSI index, proposed by scholars both domestically and internationally in recent years, is used to measure the degree of land aridity and wetness. It is a dimensionless, standardized water storage anomaly index, spatially comparable across different hydroclimatic regions. A negative TWSA-DSI value indicates that land water storage is below the average level for the study period, characterizing the degree of drought; similarly, positive values ​​can be used to measure the level of land wetness. The calculation formulas for the TWSA-DSI series are as follows:

[0121]

[0122] In the formula: TWSA i,j represents the TWSA data for the i-th year and the j-th month, and σ j are the mean and standard deviation of TWSA for the j-th month during the study period, respectively.

[0123] In this paper, the monthly means and standard deviations of TWSA are calculated for the long series from 2002 to 2022, and the terrestrial dry and wet intensity is divided into different grades based on the TWSA-DSI index. For example, some river basins mainly focus on two types of drought events: moderate-severe drought (-1.6 < TWSA-DSI ≤ -0.8) and extreme-severe drought (TWSA-DSI ≤ -1.6).

[0124] Based on the predicted values of terrestrial water storage and the derived TWSA-DSI index, early warning of drought events and their grades can be achieved, providing decision-making references for relevant departments.

[0125] Embodiment 2

[0126] This embodiment provides a system for a drought forecasting method based on artificial intelligence and atmospheric circulation mechanism, including:

[0127] A database construction module, used to collect various terrestrial water storage-related data in the study area as alternative precipitation-runoff influencing factors;

[0128] A first parameter estimation module, used to calculate the correlation coefficients between each alternative influencing factor and precipitation runoff respectively, screen out the influencing factors whose correlation coefficients exceed the preset threshold, and then process the screened precipitation-runoff influencing factors through principal component analysis to obtain the final precipitation-runoff influencing factors; construct a Bayesian-NHMM model, and determine several alternative values of the number k of precipitation-runoff states in the Bayesian-NHMM model;

[0129] A precipitation-runoff forecasting model construction module, used to input the final precipitation-runoff influencing factors into the Bayesian-NHMM model when k takes different alternative values for precipitation-runoff forecasting, and use the Bayesian-NHMM model with the minimum precipitation-runoff forecasting error as the best precipitation-runoff forecasting model;

[0130] A second parameter estimation module, used to construct training samples based on a series of precipitation-runoff values predicted by the best precipitation-runoff forecasting model, and classify the training samples to obtain N types of training sub-samples;

[0131] A forecasting model training module, used to train a machine learning model through N types of training sub-samples respectively to obtain N trained machine learning models as terrestrial water storage forecasting models,

[0132] The joint forecasting module enables drought forecasting and early warning based on this terrestrial water storage forecasting model.

[0133] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms, characterized in that, Includes the following steps: S1. Acquire data including large-scale climate factor data, hydro-meteorological factor data and underlying surface impact factor data, and deduce relative humidity and specific humidity as alternative precipitation runoff impact factors; S2. The importance of candidate precipitation-runoff influencing factors was initially ranked. Principal component analysis was used to reduce the dimensionality of the precipitation-runoff impact. Finally, k-fold cross-validation was used to determine the final precipitation-runoff influencing factors. The number of precipitation-runoff states in the Bayesian-NHMM precipitation probability forecasting model was determined using the BIC function. k Value; including the following sub-steps: S2.1: Calculate and rank the correlations between candidate precipitation runoff influencing factors and precipitation amount; S2.2: Determine the number of precipitation states in the Bayesian-NHMM precipitation probability forecast model based on the BIC function; S2.3: The importance of runoff influencing factors is initially ranked based on linear correlation coefficient and nonlinear correlation mutual information index. Principal component analysis is used to reduce the dimensionality of the influencing factors, and k-fold cross-validation is used to determine the final runoff influencing factors. The number of runoff states in Bayesian-NHMM is determined by the BIC function. S3. Based on the final precipitation runoff influencing factors and the number of precipitation runoff states k Values ​​were used to construct the final precipitation runoff influencing factors and the number of different precipitation runoff states. k The Bayesian-NHMM model with the smallest precipitation-runoff forecast error was selected as the optimal precipitation-runoff forecast model to obtain precipitation and runoff forecast results; including: Based on the final precipitation runoff influencing factors and the number of precipitation runoff states determined in step S2 k We constructed Bayesian-NHMM models with different combinations of precipitation-runoff forecasting factors, and calculated the root mean square error offset coefficient (CVRMSE) of the simulation results for different precipitation forecasting factors. This is the CVRMSE of the precipitation-runoff prediction results and observed results of the Bayesian-NHMM model when k takes different alternative values. We obtained different performances of multiple models, and the CVRMSE calculation formula is as follows: In the formula: np Forecast number of days; S i For the first i Forecast values ​​for precipitation or runoff; R i For the first i Observations of precipitation or runoff; For precipitation and runoff, the model with the smallest CVRMSE was selected as the optimal precipitation and runoff prediction model. Based on the optimal precipitation and runoff prediction model, precipitation and runoff were predicted using the final precipitation and runoff influencing factors. S4. Considering the combined impact of climate change and underlying surface changes on runoff, relevant hydro-meteorological elements, SIF, LAI and Bayesian-NHMM model-predicted precipitation values ​​are selected as clustering variables, and K-means clustering method is used to classify and grade the terrestrial water storage forecasting factors. S5. Train the machine learning model using N types of training sub-samples to obtain N trained machine learning models as land water storage prediction models. S6. Drought forecasting and early warning based on terrestrial water storage forecasting models.

2. The drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to claim 1, characterized in that: Step S2.1 specifically includes: (1) Calculate the correlation coefficient between each candidate precipitation forecast factor and precipitation amount: linear correlation is calculated using Pearson correlation coefficient; nonlinear correlation is calculated using mutual information index. Perform linear and nonlinear analyses on all candidate precipitation forecast factors and precipitation amounts respectively, and determine whether they belong to linear / nonlinear relationships based on the significance level. (2) Optimize precipitation influencing factors based on principal component analysis: Select precipitation forecasting factors with correlation coefficients exceeding the preset threshold, and perform dimensionality reduction on the selected precipitation forecasting factors through principal component analysis; (3) Through k Cross-validation was used to determine the final precipitation influencing factors: the processed precipitation forecast factors were analyzed. k Cross-validation is performed. If the validation fails, the preset threshold is adjusted and the process returns to step (2). If the validation passes, the processed precipitation forecast factor is the final precipitation impact factor.

3. The drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to claim 1, characterized in that: Step S5 specifically includes the following sub-steps: S5.1: After preliminary screening of physical factors affecting the formation of terrestrial water storage and obtaining the final influencing factors, the physical factors are downscaled to simulated terrestrial water storage based on artificial neural network (ANN), support vector machine (SVM), random forest model (RF), and long short-term memory model (LSTM), and the future terrestrial water storage forecast is carried out respectively. S5.2: Construct the probability density function of land water storage for each model based on Bayes' law of total probability; S5.3: Determine the corresponding weights based on the relative contributions of the prediction effects of Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) to establish a Bayesian pattern average correction model.

4. The drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to claim 3, characterized in that: In step S5.2, the probability density function of S is obtained according to Bayes' law of total probability as follows: In the formula: For the Kth machine learning model Given data R, predict the probability density function of the value S; Let S be the posterior probability density function of the k-th prediction value of the machine learning model given the training data R, where S is the simulated terrestrial water storage, and R = [D, O] represents the model input data, where D is the input series of the machine learning model, and O is the terrestrial water storage series retrieved from gravity satellites. Let K be the outputs of different machine learning models.

5. The drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to claim 3, characterized in that: Step S5.3 specifically includes: The Box-Cox function is used to perform a normal transformation on the gravity satellite inversion series and the simulation series obtained from various machine learning models at each grid point. Then, a weighted average is performed on the estimation results of multiple models based on the assumption of normal linear distribution. In the formula: The mean is The variance is The function follows a normal distribution; E represents the expected value of the function. For the first k The weights of each machine learning model.

6. The drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to claim 1, characterized in that: In step 6, based on the land water storage forecasting model, the land water storage forecast value is obtained. Based on the land water storage forecast value and the derived TWSA-DSI index, early warning of drought events and their levels is achieved.

7. A system for a drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms, used to perform the steps in the drought forecasting method based on artificial intelligence and atmospheric circulation mechanisms according to any one of claims 1-6, characterized in that, include: The database construction module is used to collect various terrestrial water storage-related data in the study area as alternative precipitation runoff influencing factors. The first parameter estimation module is used to calculate the correlation coefficient between each candidate influencing factor and precipitation runoff, screen out influencing factors whose correlation coefficient exceeds a preset threshold, and then process the screened precipitation runoff influencing factors through principal component analysis to obtain the final precipitation runoff influencing factors; construct a Bayesian-NHMM model and determine several candidate values ​​for the number of precipitation runoff states k in the Bayesian-NHMM model. The precipitation runoff forecasting model building module is used to input the final precipitation runoff influencing factor k into the Bayesian-NHMM model with different alternative values, and to forecast precipitation runoff. The Bayesian-NHMM model with the smallest precipitation runoff forecasting error is taken as the best precipitation runoff forecasting model. The second parameter estimation module is used to construct training samples based on a series of precipitation runoff values ​​predicted by the best precipitation runoff forecasting model, and to classify the training samples to obtain N types of training sub-samples. The forecast model training module is used to train the machine learning model using N types of training subsamples, resulting in N trained machine learning models that serve as land water storage forecast models. The joint forecasting module enables drought forecasting and early warning based on this terrestrial water storage forecasting model.

Images