Deep Learning-Based Drought Prediction Method and System for the Yellow River Basin

Abstract

The present application relates to the technical field of drought prediction, and particularly to a method and system for predicting drought in the Yellow River Basin based on deep learning. The method includes: obtaining adjacent meteorological stations of a hydrological station to be measured, and obtaining a fused meteorological parameter sequence of the hydrological station to be measured based on the meteorological parameter sequences of the adjacent meteorological stations; inputting the hydrological parameter sequence and the fused meteorological parameter sequence of the hydrological station to be measured into the drought prediction network of the hydrological station to be measured to predict the future drought index; obtaining the future drought index of the adjacent hydrological stations corresponding to the hydrological station to be measured; calculating the target drought index of the hydrological station to be measured based on the correlation between the hydrological station to be measured and the adjacent hydrological stations, the future drought indices of the hydrological station to be measured and the adjacent hydrological stations, and the determination coefficient of the drought prediction network, and then predicting the drought level of the hydrological station to be measured. Through the technical solution of the present application, the drought level of each hydrological station can be accurately predicted.

Description

Technical Field

[0001] The present application generally relates to the technical field of drought prediction, and more particularly to a method and system for predicting drought in the Yellow River Basin based on deep learning. Background Art

[0002] Drought refers to a state of long-term lack of water in meteorology, climate, hydrology and ecosystems, which usually leads to a serious lack of water on the land, affecting agricultural, ecological and socio-economic development. How to accurately predict the drought level in the Yellow River Basin is an urgent problem to be solved. Summary of the invention

[0003] In order to solve the above-mentioned technical problems of the present application, the present application provides a deep learning-based drought prediction method and system for the Yellow River Basin to accurately predict the drought level of each hydrological station in the Yellow River Basin.

[0004] In a first aspect, the present application provides a deep learning-based drought prediction method for the Yellow River Basin, which is used to predict the drought level of each hydrological station in the Yellow River Basin. The prediction method comprises: obtaining adjacent meteorological stations of the hydrological station to be tested, and obtaining a fusion sequence of meteorological parameters of the hydrological station to be tested based on the meteorological parameter sequence of the adjacent meteorological station within a preset time period, wherein the preset time period includes the current moment and a set number of moments before the current moment; collecting the hydrological parameter sequence of the hydrological station to be tested within the preset time period, and inputting the hydrological parameter sequence and the meteorological parameter fusion sequence into a drought prediction network corresponding to the hydrological station to be tested to predict the future drought index of the hydrological station to be tested. number, the future drought index is the drought index at the next adjacent moment to the current moment; the adjacent hydrological station of the hydrological station to be tested is obtained, and the future drought index of the adjacent hydrological station is predicted according to the drought prediction network corresponding to the adjacent hydrological station; the correlation between the hydrological station to be tested and the adjacent hydrological station is calculated, and the target drought index of the hydrological station to be tested is calculated based on the correlation, the future drought indexes of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network, wherein the determination coefficient is used to characterize the prediction ability of the drought prediction network; the drought level of the hydrological station to be tested at the next adjacent moment is determined based on the target drought index.

[0005] In one embodiment, the method of obtaining the adjacent meteorological stations of the hydrological station to be measured includes: taking all meteorological stations within the neighborhood of the Yellow River Basin as candidate meteorological stations; calculating the distance value between a candidate meteorological station and each hydrological station in the Yellow River Basin, taking the candidate meteorological station as the adjacent meteorological station of the hydrological station corresponding to the minimum value of the distance value, traversing all candidate meteorological stations to obtain the adjacent meteorological stations of each hydrological station, and the hydrological station to be measured is any one of all the hydrological stations in the Yellow River Basin.

[0006] In one embodiment, a hydrological site corresponds to a drought prediction network, and the drought prediction network includes a first time series module, a second time series module and a regression module; the first time series module is used to perform time series feature extraction on the meteorological parameter fusion sequence to obtain meteorological parameter features; the second time series module is used to perform feature extraction on the hydrological parameter sequence to obtain hydrological parameter features; the hydrological parameter features and the meteorological parameter features are input into the regression module to output the future drought index of the hydrological site corresponding to the drought prediction network.

[0007] In one embodiment, the training method of the drought prediction network corresponding to the hydrological station to be tested includes: at any historical moment, collecting the hydrological parameter sequence and the meteorological parameter fusion sequence of the hydrological station to be tested as a group of training samples, and collecting the drought index of the hydrological station to be tested at the historical moment as a sample label; after using the particle swarm algorithm to determine the hyperparameters of the drought prediction network, the training samples are input into the drought prediction network to output the prediction results, and the hyperparameters include learning rate, Dropout ratio and gradient clipping; based on the prediction results and the sample labels, the mean square error loss function value is calculated; the drought prediction network is back-propagated according to the hyperparameters and the mean square error loss function value, the training parameters of the drought prediction network are updated, and one training is completed; iteratively collecting training samples to train the drought prediction network, and the hyperparameters and training parameters of the drought prediction network are continuously updated until the mean square error loss function value is less than the preset loss value, and the drought prediction network that has been trained is obtained.

[0008] In one embodiment, collecting the drought index of the hydrological station to be tested at the historical moment as a sample label includes: calculating a standardized runoff index according to a hydrological parameter sequence of the hydrological station to be tested at the historical moment; calculating a standardized precipitation index according to a fusion sequence of meteorological parameters of the hydrological station to be tested at the historical moment; and calculating the drought index based on the standardized runoff index and the standardized precipitation index, wherein the drought index satisfies the relationship:

[0009] Wherein, SPI is the standardized precipitation index of the hydrological station to be measured, SRI is the standardized runoff index of the hydrological station to be measured, and Copula (SPI, SRI) represents the calculation of the Copula function value between SPI and SRI. represents the inverse function of the standard normal distribution, and G is the drought index of the hydrological station to be tested.

[0010] In one embodiment, the method of determining the hyperparameters of the drought prediction network using a particle swarm algorithm includes: deploying a plurality of particles that can move freely in a hyperparameter space, wherein a spatial position in the hyperparameter space corresponds to a hyperparameter combination of a learning rate, a dropout ratio, and a gradient clipping; calculating an evaluation value of a particle at any spatial position based on the training sample, wherein the evaluation value is a mean square error loss function value obtained by inputting the training sample into the drought prediction network after setting the hyperparameters of the drought prediction network to the values ​​corresponding to the spatial position; for any particle, obtaining the spatial position of the minimum evaluation value of the particle in the historical iteration process as the individual optimal point of the particle, and taking the spatial position of the minimum evaluation value among all individual optimal points as the global optimal point; calculating the spatial position of the particle in the next iteration based on the individual optimal point, the global optimal point, and the spatial position of the particle in the current iteration, wherein the spatial position of the particle in the next iteration satisfies the relationship:

[0011]

[0012] in, is the spatial position of the ith particle in the next iteration, is the spatial position of the ith particle in the current iteration, V i k+1 is the moving speed of the ith particle in the next iteration, V i k is the moving speed of the ith particle in the current iteration, is the individual optimal point of the i-th particle in the current iteration, is the global optimal point of the current iteration, ω is the inertia weight, c 1 and c 2 are the individual learning factor and the group learning factor, r 1 and r 2 are individual random numbers and group random numbers in the interval [0,1] respectively; the spatial positions of all particles are updated iteratively, and in response to the number of iterations being equal to the preset number of iterations, the iteration is stopped and the global optimal point of the last iteration is used as the hyperparameter of the drought prediction network.

[0013] In one embodiment, the inertia weight in the current iteration satisfies the relationship:

[0014] Among them, ω max is the maximum inertia weight, ω min is the minimum inertia weight, n is the number of iterations of the current iteration, n max is the preset number of iterations, ω is the inertia weight in the current iteration; the individual learning factor in the current iteration satisfies the relationship:

[0015] The group learning factor in the current iteration satisfies the relationship:

[0016] Among them, c max is the maximum individual learning factor, c min is the minimum individual learning factor, n is the number of iterations of the current iteration, n max is the preset number of iterations, c 1 is the individual learning factor in the current iteration, c 2 is the group learning factor in the current iteration.

[0017] In one embodiment, calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network includes: calculating the determination coefficient of the drought prediction network corresponding to the hydrological station to be tested and each adjacent hydrological station; the target drought index of the hydrological station to be tested satisfies the relationship:

[0018] Where N(τ) is the number of all adjacent hydrological stations, ρ fτ is the correlation between the hydrological station f to be tested and the τth adjacent hydrological station, G f and G τ are the future drought index of the tested hydrological station f and the τth adjacent hydrological station, respectively, f and p τ are the determination coefficients of the drought prediction network corresponding to the tested hydrological station f and the τth adjacent hydrological station, ∑ρ represents the sum of the correlations ρ between the tested hydrological station f and all adjacent hydrological stations, ∑p represents the sum of the determination coefficients p of the drought prediction network corresponding to all adjacent hydrological stations, G * is the target drought index of the hydrological station to be tested.

[0019] In one embodiment, calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network also includes: for the hydrological station to be tested or any adjacent hydrological station, obtaining a meteorological parameter sequence of all adjacent meteorological stations within a preset time period; at any time within the preset time period, calculating the variance of the meteorological parameters of all adjacent meteorological stations, and arranging all variances in the order of time to obtain a variance sequence; performing a stationarity test on the variance sequence to obtain an unsteady probability, and updating the determination coefficient of the corresponding drought prediction network based on the unsteady probability, and the updated determination coefficient satisfies the relationship:

[0020] Among them, p τ is the determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station, W τ is the probability of instability of the τth adjacent hydrological station, The updated determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station; the target drought index of the hydrological station to be tested is calculated based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the updated determination coefficient of each drought prediction network.

[0021] In the second aspect of the present application, a deep learning-based drought prediction system for the Yellow River Basin is provided, comprising a processor and a memory, wherein the memory stores computer program instructions. When the computer program instructions are executed by the processor, a deep learning-based drought prediction method for the Yellow River Basin described in the first aspect of the present application is implemented.

[0022] The technical solution of this application has the following beneficial technical effects:

[0023] Through the technical solution provided in the present application, the adjacent meteorological stations of the hydrological station to be measured are obtained, and the meteorological parameter fusion sequence of the hydrological station to be measured is obtained based on the meteorological parameter sequence of the adjacent meteorological stations, and the hydrological parameter sequence and the meteorological parameter fusion sequence of the hydrological station to be measured are input into the drought prediction network corresponding to the hydrological station to be measured to predict the future drought index of the hydrological station to be measured; further, the adjacent hydrological stations of the hydrological station to be measured and the future drought index of each adjacent hydrological station are obtained, and the future drought indexes of the hydrological station to be measured and each adjacent hydrological station are fused based on the correlation between the hydrological station to be measured and each adjacent hydrological station and the determination coefficient of each drought prediction network to obtain the target drought index of the hydrological station to be measured, and the drought level of the hydrological station to be measured is accurately predicted based on the target drought index, thereby improving the accuracy of drought prediction in the Yellow River Basin. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become easy to understand. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0025] Figure 1 is a flowchart of a method for predicting drought in the Yellow River Basin based on deep learning according to an embodiment of the present application;

[0026] Figure 2 It is a block diagram of the Yellow River Basin drought prediction system based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0028] It should be understood that when the terms "first", "second", etc. are used in the claims, specification and drawings of the present application, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the specification and claims of the present application indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0029] According to the first aspect of the present application, the present application provides a deep learning-based drought prediction method for the Yellow River Basin, which is used to predict the drought level of each hydrological station in the Yellow River Basin. A plurality of hydrological stations are deployed along the Yellow River Basin, and the hydrological stations can collect hydrological parameters at any time at their locations, including water level, flow rate, sand content, and water temperature. A plurality of meteorological stations are also deployed around the Yellow River Basin, and the meteorological stations are used to collect meteorological data at any time at their locations, including temperature, evaporation, sunshine hours, and precipitation.

[0030] Understandably, the hydrological stations are deployed along the Yellow River Basin, while the meteorological stations are deployed at any location where meteorological data needs to be monitored, that is, the location of the hydrological stations is related to the Yellow River Basin, while the location of the meteorological stations is not related to the Yellow River Basin.

[0031] Figure 1 : is a flowchart of a method for predicting drought in the Yellow River Basin based on deep learning according to an embodiment of the present application. Figure 1 As shown, the Yellow River Basin drought prediction method 100 based on deep learning includes steps S101 to S105, which are described in detail below.

[0032] S101, obtaining the adjacent meteorological sites of the hydrological site to be measured, and obtaining a fusion sequence of meteorological parameters of the hydrological site to be measured based on the meteorological parameter sequence of the adjacent meteorological sites within a preset time period, wherein the preset time period includes the current moment and a set number of moments before the current moment.

[0033] In one embodiment, the hydrological station to be tested is any one of all the hydrological stations in the Yellow River Basin; since the hydrological station to be tested can only collect hydrological parameters at its location, in order to accurately predict the drought level of the hydrological station to be tested, it is also necessary to obtain the meteorological parameters of the location where the hydrological station to be tested is located. Since no meteorological station is set up at the hydrological station to be tested, the meteorological parameter fusion sequence of the hydrological station to be tested is obtained based on the meteorological parameter sequence of the adjacent meteorological station corresponding to the hydrological station to be tested.

[0034] Specifically, the method of obtaining the adjacent meteorological stations of the hydrological station to be tested includes: taking all meteorological stations within the neighborhood of the Yellow River Basin as candidate meteorological stations; calculating the distance value between a candidate meteorological station and each hydrological station in the Yellow River Basin, taking the candidate meteorological station as the adjacent meteorological station of the hydrological station with the minimum value of the distance value, traversing all candidate meteorological stations to obtain the adjacent meteorological stations of each hydrological station, and the hydrological station to be tested is any one of all the hydrological stations in the Yellow River Basin.

[0035] The Yellow River Basin neighborhood range is the area where the straight-line distance to the Yellow River Basin is less than the preset distance; the preset distance is 15 km. Due to the large number of meteorological stations and their dense distribution, it can be ensured that one hydrological station corresponds to at least one adjacent meteorological station during the implementation process.

[0036] After determining the adjacent meteorological stations of the hydrological station to be measured, the meteorological parameter fusion sequence of the hydrological station to be measured can be obtained based on the meteorological parameter sequence of the adjacent meteorological station within a preset time period. Specifically, obtaining the meteorological parameter fusion sequence of the hydrological station to be measured based on the meteorological parameter sequence of the adjacent meteorological station within a preset time period includes: for an adjacent meteorological station, collecting the meteorological parameter sequence within the preset time period; calculating the average sequence of the meteorological parameter sequences of all adjacent meteorological stations to obtain the meteorological parameter fusion sequence of the hydrological station to be measured.

[0037] Among them, the meteorological parameter fusion sequence of the hydrological station to be measured can reflect the meteorological data of the location of the hydrological station to be measured; the preset time period includes the current moment and a set number of moments before the current moment. In this embodiment, the set number is 19, that is, the preset time period includes a total of 20 moments including the current moment.

[0038] S102, collecting the hydrological parameter sequence of the hydrological station to be tested within the preset time period, and inputting the hydrological parameter sequence and the meteorological parameter fusion sequence into the drought prediction network corresponding to the hydrological station to be tested to predict the future drought index of the hydrological station to be tested, wherein the future drought index is the drought index at the next adjacent moment to the current moment.

[0039] In one embodiment, a hydrological site corresponds to a drought prediction network, and the drought prediction network includes a first time series module, a second time series module and a regression module. The first time series module is used to perform time series feature extraction on the meteorological parameter fusion sequence to obtain meteorological parameter features; the second time series module is used to perform feature extraction on the hydrological parameter sequence to obtain hydrological parameter features; the hydrological parameter features and the meteorological parameter features are input into the regression module to output the future drought index of the hydrological site corresponding to the drought prediction network.

[0040] Among them, the first timing module and the second timing module both adopt recurrent neural networks such as LSTM or GRU; and the regression module adopts a fully connected neural network.

[0041] In one embodiment, in order to ensure that the drought prediction network can accurately output the future drought index of the corresponding hydrological station, all drought prediction networks need to be trained. This application takes the drought prediction network corresponding to the hydrological station to be tested as an example to describe in detail the training method of the drought prediction network.

[0042] Specifically, the training method of the drought prediction network corresponding to the hydrological station to be tested includes: at any historical moment, collecting the hydrological parameter sequence and the meteorological parameter fusion sequence of the hydrological station to be tested as a group of training samples, and collecting the drought index of the hydrological station to be tested at the historical moment as a sample label; after using the particle swarm algorithm to determine the hyperparameters of the drought prediction network, the training samples are input into the drought prediction network to output the prediction results, and the hyperparameters include learning rate, Dropout ratio and gradient clipping; based on the prediction results and the sample labels, the mean square error loss function value is calculated; the drought prediction network is back-propagated according to the hyperparameters and the mean square error loss function value, the training parameters of the drought prediction network are updated, and one training is completed; iteratively collecting training samples to train the drought prediction network, and the hyperparameters and training parameters of the drought prediction network are continuously updated until the mean square error loss function value is less than the preset loss value, and the drought prediction network that has been trained is obtained.

[0043] Among them, the hyperparameters of the drought prediction network are not updated during the back propagation process, and their values ​​need to be preset before the training process, while the training parameters of the drought prediction network can be updated during the back propagation process. The learning rate, dropout ratio and gradient clipping are common hyperparameters in the neural network training process, which are common knowledge of those skilled in the art and will not be repeated here.

[0044] In one embodiment, collecting the drought index of the hydrological station to be tested at the historical moment as a sample label includes: calculating a standardized runoff index according to a hydrological parameter sequence of the hydrological station to be tested at the historical moment; calculating a standardized precipitation index according to a fusion sequence of meteorological parameters of the hydrological station to be tested at the historical moment; and calculating the drought index based on the standardized runoff index and the standardized precipitation index, wherein the drought index satisfies the relationship:

[0045] Wherein, SPI is the standardized precipitation index of the hydrological station to be measured, SRI is the standardized runoff index of the hydrological station to be measured, and Copula (SPI, SRI) represents the calculation of the Copula function value between SPI and SRI. represents the inverse function of the standard normal distribution, and G is the drought index of the hydrological station to be tested.

[0046] Among them, the Standardized Runoff Index (SRI) can reflect the hydrological drought conditions of hydrological stations; the Standardized Precipitation Index (SPI) can characterize the degree of meteorological drought on different time scales through precipitation; both the Standardized Runoff Index and the Standardized Precipitation Index follow the Meteorological Drought Classification (GB / T20481-2017) formulated by the National Technical Committee for Climate and Climate Change Standardization, see Table 1 for details.

[0047] Table 1 Meteorological drought classification

[0048] Drought Index ≤-2.0 -2.0～-1.5 -1.5～-1.0 -1.0～-0.5 ≥-0.5 Drought level Extreme drought Severe drought Moderate Drought Mild drought No drought

[0049] In one embodiment, in order to accelerate the training of the drought prediction network, a particle swarm algorithm is used to determine the value of the hyperparameters in the drought prediction network during each training. Specifically, the use of the particle swarm algorithm to determine the hyperparameters of the drought prediction network includes: deploying a plurality of particles that can move freely in the hyperparameter space, wherein a spatial position in the hyperparameter space corresponds to a hyperparameter combination of a learning rate, a Dropout ratio, and a gradient clipping; calculating the evaluation value of the particle at any spatial position based on the training sample, wherein the evaluation value is the mean square error loss function value obtained by inputting the training sample into the drought prediction network after the hyperparameter of the drought prediction network is set to the numerical value corresponding to the spatial position; for any particle, obtaining the spatial position of the minimum evaluation value of the particle in the historical iteration process as the individual optimal point of the particle, and taking the spatial position of the minimum evaluation value among all individual optimal points as the global optimal point; calculating the spatial position of the particle in the next iteration based on the individual optimal point, the global optimal point, and the spatial position of the particle in the current iteration, wherein the spatial position of the particle in the next iteration satisfies the relationship:

[0050]

[0051] in, is the spatial position of the ith particle in the next iteration, X i k is the spatial position of the ith particle in the current iteration, V i k+1 is the moving speed of the ith particle in the next iteration, V i k is the moving speed of the ith particle in the current iteration, is the individual optimal point of the i-th particle in the current iteration, is the global optimal point of the current iteration, ω is the inertia weight, c 1 and c 2 are the individual learning factor and the group learning factor, r 1 and r 2 are individual random numbers and group random numbers in the interval [0,1] respectively; the spatial positions of all particles are iteratively updated, and in response to the number of iterations being equal to the preset number of iterations, the iteration is stopped and the global optimal point of the last iteration is used as the hyperparameter of the drought prediction network.

[0052] Among them, the number of particles in the hyperparameter space can be set in advance. The number of particles in this embodiment is 10. The evaluation value can reflect the accuracy of the output result of the drought prediction network under a hyperparameter combination. The smaller the evaluation value, the greater the accuracy of the output result of the drought prediction network, and the better the corresponding hyperparameter combination.

[0053] In one iteration, one particle corresponds to an individual optimal point, and all particles correspond to a global optimal point. Multiple iterations are performed continuously, and all particles search in the hyperparameter space until the number of iterations is equal to the preset number of iterations, and the optimal hyperparameter combination is obtained. Under the optimal hyperparameter combination, the drought prediction network is back-propagated to update the training parameters of the drought prediction network and complete one training.

[0054] In one embodiment, in the process of determining the hyperparameters of the drought prediction network using the particle swarm algorithm, in order to ensure that particles can search for the optimal hyperparameter combination globally in the hyperparameter space, it is necessary to adaptively adjust the inertia weight, individual learning factor and group learning factor.

[0055] Specifically, the inertia weight in the current iteration satisfies the relationship:

[0056] Among them, ω max is the maximum inertia weight, ω min is the minimum inertia weight, n is the number of iterations of the current iteration, n max is the preset number of iterations, ω is the inertia weight in the current iteration; the individual learning factor in the current iteration satisfies the relationship:

[0057] Among them, c max is the maximum individual learning factor, c min is the minimum individual learning factor, n is the number of iterations of the current iteration, n max is the preset number of iterations, c 1 is the individual learning factor in the current iteration; the group learning factor in the current iteration satisfies the relationship:

[0058] Among them, c max is the maximum individual learning factor, c min is the minimum individual learning factor, n is the number of iterations of the current iteration, n max is the preset number of iterations, c 2 is the group learning factor in the current iteration.

[0059] Among them, the tanh function is introduced into the inertia weight, which can control the inertia weight ω at ω in the early stage of iteration. max , ensuring a strong global search capability; as the number of iterations increases, ω decreases nonlinearly, which has stronger flexibility than the linear inertia weight. In the later stage of iteration, ω is close to but not equal to ω min , ensuring a certain local search capability, which helps to find the global optimal solution. At the same time, in the initial iteration, the individual learning factor c 1 Take a larger value, the group learning factor c 2 Take a smaller value to make the particles search in a larger range; in the later iteration, the group learning factor c 2 Take a larger value, individual learning factor c 1 Taking a smaller value will make the particles move closer to the global optimal value.

[0060] In this way, during a training process of the drought prediction network, the particle swarm algorithm is used to determine the values ​​of the hyperparameters in the drought prediction network, and the back propagation is used to update the values ​​of the training parameters in the drought prediction network, so as to accelerate the training speed of the drought prediction network and improve the accuracy of the drought prediction network; the trained drought prediction network can accurately predict the future drought index of the hydrological station, and one hydrological station corresponds to one drought prediction network, so as to ensure the accuracy of the future drought index of each hydrological station.

[0061] S103, obtaining adjacent hydrological sites of the hydrological site to be measured, and predicting the future drought index of the adjacent hydrological sites according to the drought prediction network corresponding to the adjacent hydrological sites.

[0062] In one embodiment, after obtaining the future drought index of the hydrological station to be measured, since the meteorological parameter fusion sequence of the hydrological station to be measured is based on the fusion result of the adjacent meteorological stations, in order to ensure the accuracy of the future drought index of the hydrological station to be measured, the adjacent hydrological stations of the hydrological station to be measured are further obtained, and the adjacent hydrological stations are all hydrological stations in the adjacent area of ​​the hydrological station to be measured, and the adjacent area of ​​the hydrological station to be measured is a circular area with the hydrological station to be measured as the center point and a radius of a preset radius.

[0063] The future drought index of each adjacent hydrological station is obtained in the same way. It can be understood that at the current moment, the future drought index of the hydrological station to be tested is obtained based on the drought prediction network of the hydrological station to be tested, and the future drought index of each adjacent hydrological station is predicted based on the drought prediction network corresponding to each adjacent hydrological station.

[0064] S104, calculating the correlation between the hydrological station to be tested and the adjacent hydrological station, and calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought indexes of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network, wherein the determination coefficient is used to characterize the prediction ability of the drought prediction network.

[0065] In one embodiment, for an adjacent hydrological station, the correlation between the hydrological station to be tested and the adjacent hydrological station is calculated based on the drought index of the hydrological station to be tested and the adjacent hydrological station in historical time, and the correlation is the Pearson correlation coefficient.

[0066] After obtaining the correlation between the hydrological station to be tested and each adjacent hydrological station, calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network includes: calculating the determination coefficient of the drought prediction network corresponding to the hydrological station to be tested and each adjacent hydrological station; the target drought index of the hydrological station to be tested satisfies the relationship:

[0067] Where N(τ) is the number of all adjacent hydrological stations, ρ fτ is the correlation between the hydrological station f to be tested and the τth adjacent hydrological station, G f and G τ are the future drought index of the tested hydrological station f and the τth adjacent hydrological station, respectively, f and p τ are the determination coefficients of the drought prediction network corresponding to the tested hydrological station f and the τth adjacent hydrological station, ∑ρ represents the sum of the correlations ρ between the tested hydrological station f and all adjacent hydrological stations, ∑p represents the sum of the determination coefficients p of the drought prediction network corresponding to all adjacent hydrological stations, G * is the target drought index of the hydrological station to be tested.

[0068] The determination coefficient of the drought prediction network is common knowledge to those skilled in the art and will not be elaborated here.

[0069] In another embodiment, calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network also includes: for the hydrological station to be tested or any adjacent hydrological station, obtaining a meteorological parameter sequence of all adjacent meteorological stations within a preset time period; at any time within the preset time period, calculating the variance of the meteorological parameters of all adjacent meteorological stations, and arranging all variances in the order of time to obtain a variance sequence; performing a stationarity test on the variance sequence to obtain an unsteady probability, and updating the determination coefficient of the corresponding drought prediction network based on the unsteady probability, and the updated determination coefficient satisfies the relationship:

[0070] Among them, p τ is the determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station, W τ is the probability of instability of the τth adjacent hydrological station, The updated determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station; the target drought index of the hydrological station to be tested is calculated based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the updated determination coefficient of each drought prediction network.

[0071] The stationarity test includes ADF test, PP test or KPSS test, which is not limited in this application.

[0072] It can be understood that the determination coefficient is used to characterize the prediction ability of the drought prediction network. The larger the determination coefficient, the more accurate the output result of the drought prediction network; the smaller the probability of instability of a hydrological station, the more accurately the fusion sequence of meteorological parameters of the hydrological station can reflect the meteorological parameters at the hydrological station, and the more accurate the output result of the drought prediction network is. Therefore, the updated determination coefficient measures the accuracy of the output results of the drought prediction network from two aspects: the prediction ability of the drought prediction network and the accuracy of the meteorological parameters of the hydrological station; the accuracy of the target drought index of the hydrological station to be tested can be improved based on the updated determination coefficient.

[0073] In this way, the target drought index of the hydrological station to be tested is accurately calculated based on the hydrological station to be tested and its adjacent hydrological stations, thereby improving the accuracy of drought prediction for the hydrological station to be tested.

[0074] S105: Determine the drought level of the hydrological station to be measured at the next adjacent moment based on the target drought index.

[0075] In one embodiment, referring to Table 1, the target drought index is mapped to a corresponding drought grade according to the meteorological drought grades in Table 1, so as to accurately predict the drought grade of the hydrological station to be tested at the next adjacent moment.

[0076] The above introduces the technical principle and implementation details of the Yellow River Basin drought prediction method based on deep learning of the present application through specific embodiments. Through the technical solution provided by the present application, the adjacent meteorological stations of the hydrological station to be tested are obtained, and the meteorological parameter fusion sequence of the hydrological station to be tested is obtained based on the meteorological parameter sequence of the adjacent meteorological station, and the hydrological parameter sequence and the meteorological parameter fusion sequence of the hydrological station to be tested are input into the drought prediction network corresponding to the hydrological station to be tested to predict the future drought index of the hydrological station to be tested; further, the adjacent hydrological stations of the hydrological station to be tested and the future drought index of each adjacent hydrological station are obtained, and the future drought index of the hydrological station to be tested and each adjacent hydrological station are fused according to the correlation between the hydrological station to be tested and each adjacent hydrological station and the determination coefficient of each drought prediction network, and the target drought index of the hydrological station to be tested is obtained, and the drought level of the hydrological station to be tested is accurately predicted according to the target drought index, so as to improve the accuracy of drought prediction in the Yellow River Basin.

[0077] According to the second aspect of the present application, the present application also provides a Yellow River Basin drought prediction system based on deep learning. Figure 2 is a block diagram of a Yellow River Basin drought prediction system based on deep learning according to an embodiment of the present application. Figure 2 As shown, the system 50 includes a processor and a memory, the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a method for predicting drought in the Yellow River Basin based on deep learning according to the first aspect of the present application is implemented. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, whose settings and functions are known in the art, so they are not described here.

[0078] In the present application, the aforementioned memory may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, system, or device. For example, a computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in this application may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0079] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A method for predicting drought in the Yellow River Basin based on deep learning, characterized in that: Used to predict the drought level of each hydrological station in the Yellow River Basin, the prediction method includes: Acquire the adjacent meteorological sites of the hydrological site to be measured, and acquire the meteorological parameter fusion sequence of the hydrological site to be measured based on the meteorological parameter sequence of the adjacent meteorological sites within a preset time period, wherein the preset time period includes the current moment and a set number of moments before the current moment; Collecting a hydrological parameter sequence of the hydrological station to be tested within the preset time period, inputting the hydrological parameter sequence and the meteorological parameter fusion sequence into a drought prediction network corresponding to the hydrological station to be tested to predict a future drought index of the hydrological station to be tested, wherein the future drought index is a drought index at a next adjacent moment to the current moment; Acquire the adjacent hydrological stations of the hydrological station to be measured, and predict the future drought index of the adjacent hydrological stations according to the drought prediction network corresponding to the adjacent hydrological stations; Calculating the correlation between the hydrological station to be tested and the adjacent hydrological station, and calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological station, and the determination coefficient of each drought prediction network, wherein the determination coefficient is used to characterize the prediction ability of the drought prediction network; The drought level of the hydrological station to be measured at the next adjacent moment is determined based on the target drought index.

2. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 1, characterized in that: The method of obtaining the adjacent meteorological stations of the hydrological station to be measured includes: All meteorological stations within the Yellow River Basin neighborhood are taken as candidate meteorological stations; Calculate the distance value between a candidate meteorological station and each hydrological station in the Yellow River Basin, take the candidate meteorological station as the adjacent meteorological station corresponding to the minimum value of the distance value, traverse all candidate meteorological stations, and obtain the adjacent meteorological station of each hydrological station. The hydrological station to be tested is any one of all the hydrological stations in the Yellow River Basin.

3. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 1, characterized in that: One hydrological site corresponds to one drought prediction network, and the drought prediction network includes a first time series module, a second time series module and a regression module; The first time series module is used to extract time series features from the meteorological parameter fusion sequence to obtain meteorological parameter features; The second time series module is used to extract features from the hydrological parameter sequence to obtain hydrological parameter features; The hydrological parameter characteristics and the meteorological parameter characteristics are input into the regression module to output the future drought index of the corresponding hydrological station of the drought prediction network.

4. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 3, characterized in that: The training method of the drought prediction network corresponding to the hydrological station to be tested includes: At any historical moment, the hydrological parameter sequence and the meteorological parameter fusion sequence of the hydrological station to be tested are collected as a set of training samples, and the drought index of the hydrological station to be tested at the historical moment is collected as a sample label; After determining the hyperparameters of the drought prediction network by using a particle swarm algorithm, the training samples are input into the drought prediction network to output prediction results, wherein the hyperparameters include a learning rate, a dropout ratio, and a gradient clipping; Calculate the mean square error loss function value based on the prediction result and the sample label; Back-propagating the drought prediction network according to the hyperparameters and the mean square error loss function value, updating the training parameters of the drought prediction network, and completing one training; Iteratively collect training samples to train the drought prediction network, continuously update the hyperparameters and training parameters of the drought prediction network, and stop when the mean square error loss function value is less than a preset loss value, thereby obtaining a trained drought prediction network.

5. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 4, characterized in that: Collecting the drought index of the hydrological station to be tested at the historical moment as a sample label includes: Calculate the standardized runoff index based on the hydrological parameter sequence of the hydrological station to be measured at the historical moment; Calculate the standardized precipitation index based on the fusion sequence of meteorological parameters of the hydrological station to be measured at the historical moment; The drought index is calculated based on the standardized runoff index and the standardized precipitation index, and the drought index satisfies the relationship: Wherein, SPI is the standardized precipitation index of the hydrological station to be measured, SRI is the standardized runoff index of the hydrological station to be measured, and Copula (SPI, SRI) represents the calculation of the Copula function value between SPI and SRI. represents the inverse function of the standard normal distribution, and G is the drought index of the hydrological station to be tested.

6. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 4, characterized in that: The method of using a particle swarm algorithm to determine the hyper parameters of the drought prediction network includes: Deploy a plurality of particles that can move freely in a hyperparameter space, wherein a spatial position in the hyperparameter space corresponds to a hyperparameter combination of a learning rate, a dropout ratio, and a gradient clipping; Calculating an evaluation value of a particle at any spatial position according to the training sample, wherein the evaluation value is a mean square error loss function value obtained by inputting the training sample into the drought prediction network after setting a hyperparameter of the drought prediction network to a value corresponding to the spatial position; For any particle, the spatial position of the minimum evaluation value of the particle in the historical iteration process is obtained as the individual optimal point of the particle, and the spatial position of the minimum evaluation value among all individual optimal points is taken as the global optimal point; The spatial position of the particle in the next iteration is calculated based on the individual optimal point, the global optimal point and the spatial position of the particle in the current iteration, and the spatial position of the particle in the next iteration satisfies the relationship: in, is the spatial position of the ith particle in the next iteration, is the spatial position of the ith particle in the current iteration, is the moving speed of the i-th particle in the next iteration, is the moving speed of the ith particle in the current iteration, is the individual optimal point of the i-th particle in the current iteration, is the global optimal point of the current iteration, ω is the inertia weight, c1 and c2 are the individual learning factor and the group learning factor, r1 and r2 are the individual random number and the group random number in the interval [0,1] respectively; The spatial positions of all particles are updated iteratively. In response to the number of iterations being equal to a preset number of iterations, the iteration is stopped and the global optimal point of the last iteration is used as the hyperparameter of the drought prediction network.

7. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 6, characterized in that: The inertia weight in the current iteration satisfies the relationship: Among them, ω max is the maximum inertia weight, ω min is the minimum inertia weight, n is the number of iterations of the current iteration, n max is the preset number of iterations, ω is the inertia weight in the current iteration; The individual learning factor in the current iteration satisfies the relationship: The group learning factor in the current iteration satisfies the relationship: Among them, c max is the maximum individual learning factor, c min is the minimum individual learning factor, n is the number of iterations of the current iteration, n max is the preset number of iterations, c1 is the individual learning factor in the current iteration, and c2 is the group learning factor in the current iteration.

8. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 1, characterized in that: Calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological stations, and the determination coefficient of each drought prediction network includes: Calculate the determination coefficient of the drought prediction network corresponding to the hydrological station to be tested and each adjacent hydrological station; The target drought index of the hydrological station to be tested satisfies the relationship: Where N(τ) is the number of all adjacent hydrological stations, ρ fτ is the correlation between the hydrological station f to be tested and the τth adjacent hydrological station, G f and G τ are the future drought index of the tested hydrological station f and the τth adjacent hydrological station, respectively, f and p τ are the determination coefficients of the drought prediction network corresponding to the tested hydrological station f and the τth adjacent hydrological station, ∑ρ represents the sum of the correlations ρ between the tested hydrological station f and all adjacent hydrological stations, ∑p represents the sum of the determination coefficients p of the drought prediction network corresponding to all adjacent hydrological stations, G * is the target drought index of the hydrological station to be tested.

9. The method for predicting drought in the Yellow River Basin based on deep learning according to claim 8, characterized in that: Calculating the target drought index of the hydrological station to be tested based on the correlation, the future drought index of the hydrological station to be tested and the adjacent hydrological stations, and the determination coefficient of each drought prediction network further includes: For the hydrological station to be measured or any adjacent hydrological station, obtaining a meteorological parameter sequence of all adjacent meteorological stations within a preset time period; At any time within the preset time period, the variance of the meteorological parameters of all adjacent meteorological stations is calculated, and all variances are arranged in order of time to obtain a variance sequence; The variance sequence is subjected to a stationarity test to obtain the probability of instability, and the determination coefficient of the corresponding drought prediction network is updated based on the instability probability. The updated determination coefficient satisfies the relationship: Among them, p τ is the determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station, W τ is the probability of instability of the τth adjacent hydrological station, is the updated determination coefficient of the drought prediction network corresponding to the τth adjacent hydrological station; The target drought index of the hydrological station to be measured is calculated based on the correlation, the future drought indexes of the hydrological station to be measured and the adjacent hydrological stations, and the updated determination coefficient of each drought prediction network.

10. A Yellow River Basin drought prediction system based on deep learning, characterized in that: It comprises a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a method for predicting drought in the Yellow River Basin based on deep learning according to any one of claims 1 to 9 is implemented.

Images