A method and system for predicting runoff data in a dry area based on dataset correction

Abstract

The application provides a drought area runoff data prediction method and system based on dataset correction, and relates to the technical field of hydrology and water resources prediction.The technical points of the application include: obtaining historical runoff data, reference runoff data, auxiliary feature data and error feature information of a target region in a drought area; constructing a hydrological response feature index set of the target region, and dividing the target region into multiple sub-regions based on the index set; according to the sub-regions, the historical runoff data, the reference runoff data and the auxiliary feature data are respectively classified into the sub-regions to which they belong; a statistical correction model is constructed in each sub-region to obtain preliminary corrected runoff data of each sub-region; a residual learning model is constructed in each sub-region to obtain final corrected runoff data of each sub-region; the optimal model combination of each sub-region is determined based on a predetermined evaluation strategy; and the optimal model combination is used for correcting the to-be-predicted runoff data.The application effectively improves the prediction accuracy of the runoff data in the drought area.

Description

Technical Field

[0001] This invention relates to the field of hydrological and water resources prediction technology, and in particular to a method and system for predicting runoff data in arid areas based on dataset correction. Background Technology

[0002] Arid regions have limited total water resources, scarce and unevenly distributed precipitation, and intense evapotranspiration. Runoff processes are influenced by climate change, snowmelt, topography, underlying surface changes, and human activities. Runoff datasets are crucial for conducting water resource assessments, watershed water balance analyses, ecological water demand assessments, agricultural irrigation management, climate change impact studies, and future water security assessments in arid regions. Due to the complex runoff formation mechanisms in arid regions, some areas exhibit characteristics such as zero flow, intermittent runoff, frequent low-flow events, and high-flow events concentrated during snowmelt or precipitation periods. Consequently, different runoff datasets generally exhibit systematic biases when applied to arid regions.

[0003] Existing runoff data correction methods typically employ global bias correction or fixed-parameter quantile mapping. While these methods can reduce historical mean bias to some extent, they still suffer from the following problems: First, the hydrological response varies significantly across different regions of the arid zone. The runoff formation mechanisms differ in mountainous snowmelt areas, oasis agricultural areas, desert areas, and inland river downstream flow interruption areas, making it difficult for a unified correction model to be applied to all regions simultaneously. Second, traditional quantile mapping methods typically assume that the error structure over historical periods is stable, making it difficult to handle non-stationary changes in runoff distribution under the background of climate change. Third, low flow and zero flow are significant phenomena in arid areas, and traditional normal error or mean square error loss functions are difficult to effectively characterize the runoff process with zero expansion and excessive dispersion. Fourth, future reference runoff cannot be directly obtained in future scenario simulations. If the correction model lacks sufficient historical validation and hydrological zoning adaptation, the simulation results of future runoff will have significant uncertainties. Summary of the Invention

[0004] To address the problems of significant systematic bias, strong spatial heterogeneity, insufficient characterization of low-flow and zero-flow processes, and insufficient reliability of future runoff simulation in existing technologies for arid zone runoff datasets, this invention proposes a method and system for predicting arid zone runoff data based on dataset correction.

[0005] According to one aspect of the present invention, a method for predicting runoff data in arid areas based on dataset correction is proposed, the method comprising the following steps: S1. Obtain historical runoff data, reference runoff data, and auxiliary feature data for the target area in the arid zone; perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data; S2. Construct a set of hydrological response characteristic indicators for the target area based on the reference runoff data and the auxiliary feature data; and divide the target area into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. S3. Based on the sub-regions, the historical runoff data, reference runoff data, and auxiliary feature data are respectively assigned to their respective sub-regions to obtain the original partition data, reference partition data, and auxiliary partition data corresponding to each sub-region. S4. Within each sub-region, a statistical correction model is constructed based on the original zoning data, the reference zoning data, and the error characteristic information; the original zoning data is corrected using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. S5. Within each sub-region, a residual learning model is constructed based on the preliminary corrected runoff data and the auxiliary zoning data; the residual learning model is used to learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data; the preliminary corrected runoff data is corrected according to the residual characteristics to obtain the final corrected runoff data for each sub-region. S6. Based on a predetermined evaluation strategy, evaluate the performance of the combination of the statistical correction model and the residual learning model, and determine the optimal model combination for each sub-region based on the evaluation results. S7. Obtain the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine with the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

[0006] Furthermore, the auxiliary feature data mentioned in S1 includes meteorological data, land cover data, and topographic data; the error feature information includes at least two of the following: deviation, relative deviation, absolute deviation, root mean square error, Nash efficiency coefficient, correlation coefficient, and KGE index between the historical runoff data and the reference runoff data. The hydrological response characteristic index set described in S2 includes land cover data, topographic data, multi-year average precipitation, multi-year average temperature, and drought index calculated based on meteorological data; zero-flow frequency and runoff concentration calculated based on reference runoff data. The zero-flow frequency is used to characterize the characteristics of river flow interruption or intermittent runoff in arid areas, which refers to the proportion of time steps in a certain time series where the runoff is lower than a preset zero-flow threshold to the total number of time steps; the runoff concentration is used to characterize the degree of concentration of runoff among months within a year.

[0007] Furthermore, the division of the target region into multiple sub-regions with similar hydrological responses based on the set of hydrological response characteristic indicators in S2 includes: using principal component analysis to reduce the dimensionality of each hydrological response characteristic indicator after standardization; using a Gaussian mixture model to cluster the dimensionality-reduced principal component features, and using the category with the highest probability as the label of the sub-region; wherein the elbow rule is used to determine the optimal number of clusters.

[0008] Furthermore, S3 includes: resampling the historical runoff data, the reference runoff data, and the auxiliary feature data to a unified spatial grid, and performing time alignment to form a spatiotemporally matched partitioned dataset.

[0009] Furthermore, the statistical correction model described in S4 is a non-stationary quantile mapping model. The non-stationary quantile mapping model sets a low flow level threshold and a high flow level threshold according to the error characteristic information, and sets corresponding time-varying scale parameters for the three flow levels of low flow level, medium flow level and high flow level, respectively, to perform quantile mapping correction.

[0010] Furthermore, in S4, the correction of the original partition data using the statistical correction model includes: The quantiles of the current runoff data are determined based on the empirical cumulative distribution function of historical runoff data. The quantiles are mapped to quantile mapping values ​​based on the empirical cumulative distribution function of the reference runoff data; Based on the quantile increment preservation relationship, the basic corrected runoff data are obtained from the quantile mapping value; The basic corrected runoff data is scaled according to the time-varying scale parameters corresponding to the runoff data at the current moment to obtain preliminary corrected runoff data.

[0011] Furthermore, the residual learning model in S5 is a recurrent neural network model based on an attention mechanism; the residual feature is a proportional residual, that is, the ratio of the difference between the reference partition data and the preliminary corrected runoff data to the preliminary corrected runoff data.

[0012] Furthermore, the objective function of the attention-based recurrent neural network model described in S5 during the training process... for: ; In the formula, This represents the loss function of a zero-inflated negative binomial distribution; This represents the residual error weighting coefficient; This represents the regularization weight coefficient; Indicates the length of the time series; Indicates the first The parameter set of an LSTM model for each spatial unit; Represents the square of the L2 norm; This represents the target value of the proportional residual. , Indicates reference runoff data, This indicates preliminary corrected runoff data. It is a very small positive number; This represents the predicted value of the proportional residual. ,in This is the weight matrix of the fully connected layer. For bias terms, The context vector output by the attention mechanism.

[0013] Furthermore, the correction in S5 is as follows: multiply the preliminary corrected runoff data by the sum of the predicted proportional residual and 1, and apply a non-negative constraint to the product result to obtain the final corrected runoff data.

[0014] Furthermore, the predetermined evaluation strategy in S6 includes a sliding window verification method, which calculates the performance evaluation index of each model combination in multiple time windows by setting the sliding window length and sliding step size, and determines the optimal model combination based on the average value of the evaluation index of each window.

[0015] According to another aspect of the present invention, a drought-area runoff data prediction system based on dataset correction is proposed. The system is used to implement the aforementioned drought-area runoff data prediction method based on dataset correction; the system includes: The data acquisition module is configured to acquire historical runoff data, reference runoff data, and auxiliary feature data of the target area in the arid zone; and to perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data. The sub-region division module is configured to construct a set of hydrological response characteristic indicators for the target region based on the reference runoff data and the auxiliary feature data; and to divide the target region into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. The data partitioning module is configured to assign historical runoff data, reference runoff data, and auxiliary feature data to their respective sub-regions according to the sub-regions, thereby obtaining the original partitioning data, reference partitioning data, and auxiliary partitioning data corresponding to each sub-region. The statistical correction module is configured to construct a statistical correction model based on the original zoning data, the reference zoning data, and the error characteristic information within each sub-region; and to correct the original zoning data using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. The residual correction module is configured to: construct a residual learning model based on the preliminary corrected runoff data and the auxiliary zoning data within each sub-region; learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data using the residual learning model; and correct the preliminary corrected runoff data according to the residual characteristics to obtain the final corrected runoff data for each sub-region. The model selection module is configured to evaluate the performance of the combination of the statistical correction model and the residual learning model based on a predetermined evaluation strategy, and determine the optimal model combination for each sub-region based on the evaluation results. The runoff prediction module is configured to acquire the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention first performs historical error diagnosis, which can identify systematic biases in runoff scenario datasets in different regions, flow levels, and time periods in arid areas, providing a clear basis for subsequent correction. By dividing the data into sub-regions using multi-year average precipitation, multi-year average temperature, aridity index, snow cover, DEM, glacier cover, zero-flow frequency, and runoff concentration, it can identify regions with similar hydrological responses, avoiding the problem of ignoring spatial heterogeneity in holistic correction models. A partitioned nonstationary quantile mapping model is adopted, and the parameters at the low-flow, medium-flow, and high-flow scales are varied over time to adapt to the nonstationarity of runoff distribution under the background of climate change. Furthermore, a recurrent neural network model based on the attention mechanism (ZINB-Attention-LSTM) is used to learn the proportional residuals, which can handle zero-flow, low-flow, and excessively discrete runoff processes in arid areas, and the attention mechanism is used to identify key time steps and key hydrological driving factors. The optimal model is determined through test period validation and sliding window validation to improve the stability of historical correction results. Finally, the validated model is applied to future scenarios to improve the reliability and physical rationality of future runoff simulations. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1This is a flowchart of a method for predicting runoff data in arid areas based on dataset correction, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a drought area runoff data prediction system based on dataset correction provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] This invention proposes a method for predicting runoff data in arid regions based on dataset correction, such as... Figure 1 As shown, the method includes: S1. Obtain historical runoff data, reference runoff data, and auxiliary feature data for the target area in the arid zone; perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data; S2. Construct a set of hydrological response characteristic indicators for the target area based on the reference runoff data and the auxiliary feature data; and divide the target area into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. S3. Based on the sub-regions, the historical runoff data, reference runoff data, and auxiliary feature data are respectively assigned to their respective sub-regions to obtain the original partition data, reference partition data, and auxiliary partition data corresponding to each sub-region. S4. Within each sub-region, a statistical correction model is constructed based on the original zoning data, the reference zoning data, and the error characteristic information; the original zoning data is corrected using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. S5. Within each sub-region, a residual learning model is constructed based on the preliminary corrected runoff data and the auxiliary zoning data; the residual learning model is used to learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data; the preliminary corrected runoff data is corrected according to the residual characteristics to obtain the final corrected runoff data for each sub-region. S6. Based on a predetermined evaluation strategy, evaluate the performance of the combination of the statistical correction model and the residual learning model, and determine the optimal model combination for each sub-region based on the evaluation results. S7. Obtain the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine with the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

[0021] First, in S1, historical runoff data, reference runoff data, and auxiliary feature data of the target area in the arid zone are acquired; based on the reference runoff data, error diagnosis is performed on the historical runoff data to obtain error feature information of the historical runoff data; the auxiliary feature data includes meteorological data, land cover data, and topographic data.

[0022] According to embodiments of the present invention, historical runoff data includes any one or more of the following: climate model-driven runoff data, hydrological model-simulated runoff data, land surface model runoff data, reanalysis runoff data, or remote sensing inversion runoff data. Specifically, climate model-driven runoff data refers to runoff simulation data obtained by driving a hydrological model with meteorological variables such as precipitation and temperature output from a global or regional climate model; hydrological model-simulated runoff data refers to runoff sequences simulated by a hydrological model based on watershed parameters and meteorological inputs; land surface model runoff data refers to surface runoff or subsurface runoff data simulated by a land surface process model; reanalysis runoff data refers to gridded runoff products formed by fusing observational data and numerical model results through data assimilation methods; and remote sensing inversion runoff data refers to runoff data indirectly estimated using remote sensing information such as satellite altimetry, remotely sensed water area, and changes in surface water storage. A historical runoff dataset can be represented as: In the formula, Indicates the length of the time series.

[0023] Reference runoff data refers to quality-controlled measured runoff data or gridded reference runoff data, which can be derived from national hydrological yearbooks, watershed hydrological station observation data, data released by watershed management agencies, or publicly available scientific data platforms. As the benchmark data for diagnosing and correcting historical runoff data biases, reference runoff data can be represented as: .

[0024] Auxiliary feature data includes meteorological data, land cover data, and topographic data. Land cover data includes snow cover data and glacier cover data; snow cover data represents the proportion of snow cover area within each spatial cell to the total area of ​​that cell, ranging from 0 to 1; glacier cover data represents the proportion of glacier cover area within each spatial cell to the total area of ​​that cell, also ranging from 0 to 1. Snow cover data can be derived from MODIS satellite remote sensing snow cover products or reanalysis snow depth / snow water equivalent data conversion results, while glacier cover data can be derived from glacier cataloging data or remote sensing interpretation data. Topographic data is digital elevation model data, i.e., DEM data, in meters, used to characterize the surface elevation information of each spatial cell, and is not simply geographic coordinate data.

[0025] Historical error diagnosis of historical runoff datasets based on reference runoff data includes calculating at least two of the following error characteristics: deviation, relative deviation, absolute deviation, root mean square error, Nash efficiency coefficient, correlation coefficient, and Kling-Gupta Efficiency (KGE) index between historical runoff data and reference runoff data. The KGE index is a hydrological model evaluation index used to comprehensively measure the degree of fit between simulated runoff and reference runoff in three dimensions: correlation, variability, and mean deviation. A KGE value closer to 1 indicates a better simulation effect. Absolute deviation includes the absolute deviation between historical runoff data and reference runoff data during high-flow, low-flow, and zero-flow phases. The division of high-flow, low-flow, and zero-flow phases is based on the 90th quantile, 10th quantile, and zero-flow threshold of the reference runoff data, respectively.

[0026] Among them, deviation The calculation formula is: ; In the formula, This represents the simulated runoff value of the j-th historical runoff dataset at time step t; This represents the reference runoff value at time step t. A value greater than 0 indicates that the simulated runoff is generally too high. <0 indicates that the simulated runoff is generally low.

[0027] relative deviation The calculation formula is: ; In the formula, This represents a very small positive number, used to avoid computational instability caused by a denominator of zero; its value is usually 10. -6 .

[0028] Root mean square error The calculation formula is: Nash efficiency coefficient The calculation formula is: ; In the formula, , representing the average value of the reference runoff; Correlation coefficient The calculation formula is: ; In the formula, , representing the average value of the simulated runoff.

[0029] KGE index The calculation formula is: ; in, This represents the ratio of the simulated mean runoff value to the reference mean runoff value. ; Indicates the standard deviation of simulated runoff values Standard deviation of reference runoff The ratio, , , .

[0030] Through the above error diagnosis, the deviation direction, deviation magnitude, correlation, distribution differences, and high- and low-flow simulation defects of different runoff datasets within the target area are determined, and historical deviation characteristics are obtained. These error characteristics are used to: provide a basis for index selection in S2 for the division of hydrological response units; provide a basis for setting low-flow and high-flow thresholds in S4; and provide error structure priors for input feature selection and loss function selection in S5 for the residual learning model.

[0031] Then, in S2, a set of hydrological response characteristic indicators for the target area is constructed based on the reference runoff data and the auxiliary feature data; and the target area is divided into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. The set of hydrological response characteristic indicators includes land cover data, topographic data, multi-year average precipitation, multi-year average temperature, and aridity index calculated based on meteorological data, and zero-flow frequency and runoff concentration calculated based on reference runoff data.

[0032] According to an embodiment of the present invention, the multi-year average precipitation and multi-year average temperature in the hydrological response characteristic index set are calculated from the meteorological data obtained by S1 through multi-year averaging. The aridity index is calculated as follows: In the formula, Indicates the first Aridity index of each spatial unit, This represents the multi-year average potential evapotranspiration. This represents the multi-year average precipitation; the multi-year average potential evapotranspiration can be obtained directly from the product dataset or calculated based on meteorological data. A spatial unit refers to the basic analytical unit obtained by dividing the target area according to a uniform spatial resolution, such as a 0.1°×0.1° or 0.25°×0.25° regular grid; each spatial unit has a unique spatial location and corresponds to a set of runoff, meteorological, topographic, and land cover data.

[0033] Zero-flow frequency (ZIF) characterizes the flow interruption or intermittent runoff characteristics of rivers in arid regions. It refers to the proportion of time steps within a given time series where the runoff volume is below a preset ZIF threshold. A higher ZIF indicates stronger intermittent runoff and more significant flow interruption in that spatial unit. The preset ZIF threshold is used to define this characteristic. For determining an approximate no-flow state, a value of 0.001 m³ / s is preferred; however, when the spatial unit scale is large or the runoff level is high, it can be adjusted to 0.01 m³ / s according to the hydrological characteristics of the study area. Zero current frequency of each spatial unit The calculation formula is as follows: ; In the formula, Indicates the first spatial units Reference runoff value at time; indicator function for: .

[0034] Runoff concentration is used to characterize the degree of concentration of runoff among months within a year. Runoff concentration is high when runoff is mainly concentrated in a few months; it is low when runoff is relatively evenly distributed throughout the year. This index can be used to identify areas with concentrated snowmelt runoff, areas with concentrated precipitation runoff, and areas with relatively stable runoff throughout the year. Runoff concentration in a spatial unit The following formula is used for calculation: ; In the formula, Indicates the first The average number of spatial units over many years Monthly reference runoff.

[0035] To avoid the holistic correction model ignoring the spatial heterogeneity of arid regions, this embodiment divides the target area into multiple sub-regions with similar hydrological responses based on a set of hydrological response characteristic indicators. These sub-regions are named hydrological response units. Specifically, this includes: S21, standardizing each indicator in the set of hydrological response characteristic indicators to obtain a standardized hydrological response characteristic matrix; the standardization formula is: ; In the formula, This represents the standardized hydrological response index; Indicates the first The spatial unit of the first One hydrological response indicator; and They represent the first The mean and standard deviation of each hydrological response index are used to form a standardized hydrological response feature vector. , Indicates the number of indicators.

[0036] S22. Principal component analysis is used to reduce the dimensionality of the standardized hydrological response indicators, resulting in the principal component feature matrix. First, the standardized hydrological response feature matrix... Calculate the covariance matrix : ; In the formula, This indicates the number of spatial units, i.e., the number of samples. This represents the degree-of-freedom correction term for the unbiased estimate.

[0037] Then, eigenvalue decomposition is performed: ; Indicates the first One eigenvalue; Indicates the first 1 feature vector; select the top 10 feature vectors whose cumulative contribution rate meets the preset threshold. Principal components: ; In the formula, Indicates the preceding Cumulative variance contribution rate of each principal component; This indicates the preset cumulative contribution rate threshold. For example, the values ​​can be 0.85 or 0.90.

[0038] The principal component features after dimensionality reduction are: ; Indicates the first The dimensionality-reduced principal component eigenvectors of each spatial unit; Indicates from the previous The projection matrix consists of eigenvectors.

[0039] S23. Use a Gaussian mixture model to perform cluster analysis on the principal component feature matrix, and take the category with the highest probability as the label of the sub-region, that is, obtain the label of the hydrological response unit.

[0040] The probability density of the Gaussian mixture model is: ; In the formula, Indicates sample The probability density function value; This represents the total number of Gaussian components, i.e., the number of hydrological response unit categories; Indicates the first The mixing weights of the Gaussian components satisfy the following conditions: ; Indicates a multivariate Gaussian distribution; Indicates the first A vector of mean values ​​of Gaussian components; Indicates the first The covariance matrix of Gaussian components.

[0041] Before clustering, the elbow rule is used to determine the number of candidate clusters; the elbow rule is applied using the following within-cluster sum of squares... Determine the number of candidate clusters: ; In the formula, To represent the square of the Euclidean norm (L2 norm), This represents the standardized hydrological response feature vector of the i-th spatial unit. Indicates the first One cluster, Indicates the first The cluster centers and the mean vector of the aforementioned Gaussian components are the same value. (Selection) The positions where the decreasing trend of cluster number slows down significantly with increasing cluster number are selected as candidate cluster numbers.

[0042] The label of hydrological response unit i is obtained as : ; ; In the formula, h is the summation variable, representing the index number of the Gaussian component, h = 1, 2, ... ; Indicates the first The spatial unit belongs to the first The posterior probability of each Gaussian component; This indicates that the category with the highest probability is selected.

[0043] This step obtains similar regions of hydrological response, providing a spatial unit basis for subsequent regional runoff correction modeling.

[0044] Then, in S3, historical runoff data, reference runoff data, and auxiliary feature data are respectively assigned to their respective sub-regions according to the sub-regions, resulting in original partition data, reference partition data, and auxiliary partition data for each sub-region. This includes: resampling the historical runoff data, the reference runoff data, and the auxiliary feature data to a unified spatial grid and performing time alignment to form a spatiotemporally matched partition dataset.

[0045] According to an embodiment of the present invention, each spatial unit i within the target area is traversed, based on the label of its hydrological response unit. It is assigned to the corresponding partition z; subsequently, the historical runoff data, reference runoff data, and auxiliary feature data corresponding to spatial unit i are synchronously assigned to the partition data set. Therefore, each hydrological response unit forms an independent dataset, and the subsequent non-stationary quantile mapping model and residual learning model are trained and applied separately within each partition.

[0046] Historical runoff data, reference runoff data, and auxiliary feature data are resampled to a unified spatial grid and time-aligned, including unifying the time scale and time range. Continuous variables such as temperature and topography data can be resampled using bilinear interpolation; data with conservation properties, such as precipitation and runoff data, can be resampled using conservative remapping methods; categorical or proportional cover data, including snow cover data, glacier cover data, and hydrological response unit labels, can be resampled using area-weighted methods; thus forming a dataset. as follows: ; In the formula, Represents historical runoff data; Indicates reference runoff data; Indicates precipitation; Indicates temperature; Represents digital elevation topographic data; Indicates the percentage of snow cover; Indicates the percentage of glacier coverage; This indicates the label for the hydrological response unit.

[0047] Furthermore, the historical period can be divided into a training period, a validation period, and a testing period. The training period is used to fit the distribution relationship and train the model, the validation period is used to optimize the model parameters, and the testing period is used to independently evaluate the model's generalization ability.

[0048] Then, in S4, within each sub-region, a statistical correction model is constructed based on the original partition data and the reference partition data; the original partition data is corrected using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. The statistical correction model is a non-stationary quantile mapping model. This model sets low-flow-level and high-flow-level thresholds based on the error characteristic information, and sets corresponding time-varying scale parameters for the low-flow-level, medium-flow-level, and high-flow-level flows respectively to perform quantile mapping correction.

[0049] According to an embodiment of the present invention, in each hydrological response unit Within this invention, a non-stationary quantile mapping model is constructed. The non-stationary quantile mapping model is a statistical model used for correcting runoff data bias. Its basic idea is to establish a quantile correspondence between historical simulated runoff distributions and reference runoff distributions, and to introduce time-varying scale parameters based on traditional quantile mapping to reflect the non-stationary changes in the runoff distribution error structure under the background of climate change. In this invention, the model sets time-varying scale parameters according to three flow levels: low flow, medium flow, and high flow, thereby improving the correction adaptability for low-flow, zero-flow, and high-flow processes in arid regions.

[0050] The time-varying scale parameters corresponding to low, medium, and high flow rates are named low-flow-scale parameter, medium-flow-scale parameter, and high-flow-scale parameter, respectively, and their linear changes with time are as follows: ; In the formula, , and These represent the low-flow, medium-flow, and high-flow scale parameters of the z-th hydrological response unit at time step t, respectively. Indicates normalized time. The value ranges from 0 to 1. Indicates the start time of the time series. Indicates the end time of the time series; , , These represent the intercept terms for low-flow, medium-flow, and high-flow scale parameters, respectively. , , These represent the trend terms of the parameters at the low-flow, medium-flow, and high-flow scales over time, respectively. These six parameters together constitute the parameter vector to be optimized for the z-th hydrological response unit: .

[0051] The original zoning data were corrected using the aforementioned non-stationary quantile mapping model to obtain preliminary corrected runoff data for each sub-region, i.e., nsQDM correction. The original zoning data per hydrological response unit... Historical runoff data. Specifically, this includes: S41, and the empirical cumulative distribution function based on historical runoff data. Determine the quantiles corresponding to the runoff data at the current time. : ; In the formula, Let represent the historical runoff data of the z-th hydrological response unit at time step t. The empirical cumulative distribution function is denoted as . It is a distribution function constructed based on all historical runoff data samples of the z-th hydrological response unit during the training period, after arranging the runoff observations in ascending order and using empirical cumulative probability (i.e., the proportion of samples less than or equal to a certain runoff value to the total number of samples); for runoff values ​​that do not fall on the sample points, the corresponding cumulative probability is calculated using a linear interpolation method.

[0052] S42. Obtain quantile mapping values ​​by mapping the quantiles according to the empirical cumulative distribution function of the reference runoff data. : ; In the formula, The inverse function of the empirical cumulative distribution function of the reference runoff data for the z-th hydrological response unit is expressed at the quantile. The runoff value at that location.

[0053] S43. Based on the quantile increment preservation relationship, the baseline corrected runoff data are obtained from the quantile mapping values. : ; In the formula, Indicates the first The inverse function of the empirical cumulative distribution function of historical runoff data for each hydrological response unit at the quantile The formula is used to maintain consistency between the simulated runoff change ratio and the historical distribution mapping.

[0054] S44. Scale the baseline corrected runoff data according to the time-varying scale parameters corresponding to the current runoff data to obtain preliminary corrected runoff data. : ; In the formula, Indicates based on current runoff data The time-varying scale parameters selected for the flow level; ; The low flow threshold (i.e., low flow level threshold) of the z-th hydrological response unit is preferably taken as the 10th percentile of the historical runoff data during the training period. This represents the high-flow threshold (i.e., high-discharge level threshold) for the z-th hydrological response unit, preferably taken as the 90th percentile of historical runoff data during the training period. The low-flow and high-flow thresholds can be adjusted based on the error characteristic information obtained in S1.

[0055] Among them, six parameters to be optimized from the low-flow-scale parameters, medium-flow-scale parameters, and high-flow-scale parameters were obtained through particle swarm optimization algorithm - PSO parameter optimization.

[0056] Specifically, the particle position is represented as : ; The particle velocity update formula is: ; The particle position update formula is: ; in, Indicates the first The first hydrological response unit The particle in the first The parameter vector to be optimized at the next iteration is the set of parameters consisting of low-flow-scale parameters, medium-flow-scale parameters, and high-flow-scale parameters; k represents the iteration number. Indicates particle velocity; Indicates inertia weight; and These represent individual learning factors and group learning factors, respectively. and Represents a random number between 0 and 1; This represents the historical optimal position of the p-th particle in the k-th iteration; This represents the globally optimal position of the particle swarm in the k-th iteration; The iteration step size represents the time step corresponding to one iteration in the particle swarm optimization algorithm, and its value is 1 (unit iteration step).

[0057] The objective function is: ; in, , , , They represent the first KGE index, Nash efficiency coefficient, high-flow Nash efficiency coefficient and relative deviation of each hydrological response unit during the validation period; This represents the mean penalty term. In the formula, Indicates the first The mean of corrected runoff during the verification period of each hydrological response unit. This represents the mean of the reference runoff data. It is a very small positive number, used to avoid the denominator being zero; Indicates the variable rate penalty term. , Indicates the first The standard deviation of corrected runoff during the validation period of each hydrological response unit This represents the standard deviation of the reference runoff data; This represents the Kolmogorov-Smirnov statistic. , and Let these represent the empirical cumulative distribution functions of the corrected runoff and the reference runoff data, respectively. x Indicates the points where the runoff variable takes values; to These are the weighting coefficients for each evaluation indicator.

[0058] As an embodiment of the present invention, the above-mentioned weighting coefficient can take the following values: =0.25, =0.20, =0.15, =0.15, =0.10, =0.10, =0.05. The specific values ​​of each weight coefficient can be adjusted using grid search or Bayesian optimization methods based on the model performance during the validation period, so as to make the objective function... It has reached its maximum value.

[0059] The optimization objective of PSO is: ; This represents the optimal parameter vector. PSO searches for the optimal parameter combination to achieve the objective function during the validation period by continuously updating particle velocity and position.

[0060] Then, in S5, within each sub-region, a residual learning model is constructed based on the preliminary corrected runoff data and the auxiliary zoning data; the residual learning model is used to learn the residual features of the preliminary corrected runoff data relative to the reference zoning data; the preliminary corrected runoff data is corrected according to the residual features to obtain the final corrected runoff data for each sub-region; wherein, the residual learning model is a recurrent neural network model based on an attention mechanism, named the ZINB-Attention-LSTM model; the residual feature is a proportional residual, that is, the ratio of the difference between the reference zoning data and the preliminary corrected runoff data to the preliminary corrected runoff data.

[0061] According to an embodiment of the present invention, since nsQDM correction mainly corrects distribution bias, but low flow, zero flow, peak flow and time-dependent residuals may still exist, a partitioned adaptive residual learning model is further constructed.

[0062] The ZINB-Attention-LSTM model incorporates a long short-term memory network and an attention mechanism. The input features of the partitioned adaptive residual learning model include pre-corrected runoff data, precipitation, temperature, DEM, glacier cover, snow cover, runoff lag terms, and month and season. The feature matrix is ​​represented as follows: : ; In the formula, Indicates the lag order; This indicates preliminary corrected runoff data; Indicates precipitation; Indicates temperature; Indicates the average elevation of the zone; This indicates the average glacier coverage percentage for each region; Indicates the average snow cover percentage of the zone; Indicates the month; Indicates the season; The runoff lag term refers to the preliminary corrected runoff data from one or more previous time steps, used to reflect the temporal continuity of the runoff process and the impact of historical conditions on the present.

[0063] Long Short-Term Memory (LSTM) network outputs hidden state vectors : ; In the formula, Indicates time step The hidden state vector. The cell state vector representing time step t; and These represent the hidden state and cell state of the previous time step, respectively; This represents a unit in the Long Short-Term Memory (LSTM) network.

[0064] The attention weights in the attention mechanism are : ; In the formula, This represents the attention score at time step t. The query vector representing the attention mechanism. The weight matrix representing the attention mechanism. The bias term representing the attention mechanism. It is the hyperbolic tangent activation function.

[0065] The objective function, i.e., the comprehensive loss function, of the attention-based recurrent neural network model during training is... for: ; In the formula, This represents the loss function of a zero-inflated negative binomial distribution; This represents the residual error weighting coefficient, used to control the intensity of supervised learning, and can be set to 0.4; This represents the regularization weight coefficient, used to control model complexity, and can be set to 0.6; Indicates the length of the time series; Indicates the first The parameter set of an LSTM model for each spatial unit, including weight matrix, bias term, etc. This represents the squared L2 norm, used for weight decay. It can prevent overfitting and improve generalization ability; This represents the target value of the proportional residual. , Indicates reference runoff, This indicates preliminary corrected runoff data; This represents the predicted value of the proportional residual. , This represents the weight matrix of the fully connected layer; Indicates the bias term; This represents the context vector obtained by weighted summation of the hidden states at each time step. .

[0066] The zero-inflated negative binomial distribution loss function (ZINB loss function) is as follows: ; In the formula, the zero-inflation negative binomial probability This is used to maximize the fitting ability of the residual distribution, and is particularly suitable for data with a high proportion of zero values, asymmetric distribution, and excessively discrete data. Let be a random variable, representing the time step. The residual distribution; The zero inflation probability represents the probability that the residual is a structurally zero value, i.e., a non-random zero flow or an error-free state. Indicates non-zero distributed weights; This represents the probability that the negative binomial distribution is at zero. This represents the negative binomial probability of non-zero residuals, used to characterize the distribution of continuous, non-negative, but overly discrete residuals. ; Indicates time step Proportional residual target value ; represents the mean parameter of the negative binomial distribution, and represents the expected level of the residuals; This represents the discrete parameter of the negative binomial distribution, used to control the magnitude of the variance. The smaller the value, the stronger the dispersion. This is a gamma function.

[0067] Furthermore, the optimal LSTM hyperparameters can be searched using the Particle Swarm Optimization (PSO) algorithm.

[0068] Specifically, the hyperparameter vector is : ; in, These represent the number of hidden layer neurons, the number of LSTM layers, the random inactivation rate, the learning rate, the batch size, the number of training rounds, the lag order, the residual error weight coefficient, and the regularization weight coefficient, respectively.

[0069] The optimization objective is: ; In the formula, Indicates the peak deviation rate; , , , , These are the weighting coefficients for each indicator. As an embodiment of the present invention, the weighting coefficients in the above LSTM hyperparameter optimization objective function can take the following values: =0.30, =0.25, =0.15, =0.15, =0.15. The above weighting coefficients reflect the emphasis placed on high flow fitting accuracy, overall efficiency, and flood peak deviation control in the overall performance of the model. The specific values ​​can be fine-tuned based on the evaluation results during the validation period.

[0070] Optimal LSTM Hyperparameters .

[0071] The preliminary corrected runoff data is then corrected based on the proportional residuals. This involves multiplying the preliminary corrected runoff data by the sum of the predicted proportional residuals and 1, and applying a non-negativity constraint to the product. This yields the final corrected runoff data for each sub-region. : ; In the formula, This represents a non-negative constraint function to avoid negative runoff.

[0072] Then, in S6, based on a predetermined evaluation strategy, the performance of the combination of the statistical correction model and the residual learning model is evaluated, and the optimal model combination for each sub-region is determined based on the evaluation results. The predetermined evaluation strategy includes a sliding window validation method, which calculates the performance evaluation index of each model combination in multiple time windows by setting the sliding window length and sliding step size, and determines the optimal model combination based on the mean of the evaluation indexes of each window.

[0073] According to an embodiment of the present invention, the final corrected runoff data of the historical period are verified based on the training period, validation period, testing period and sliding window validation method to determine the optimal model combination for each hydrological response unit.

[0074] Specifically, let the length of the sliding window be... The sliding step size is Then the first The time intervals corresponding to each sliding window are: ; , Indicates the first The start time of each sliding window; Indicates the window end time.

[0075] Evaluation metrics are calculated within each sliding window, and the optimal model combination is determined based on the mean and stability of the evaluation metrics for each window. Evaluation metrics include: Nash efficiency coefficient NSE90 under high flow conditions and peak flow deviation rate. RMSE improvement rate and KGE improvement rate. The calculation formulas are shown below.

[0076] ; In the formula, This represents the Nash efficiency coefficient under high flow conditions; Indicates the corrected runoff; Indicates reference runoff data; This represents the set of high-flow events, i.e., the reference runoff ≥ 90th percentile; This represents the average value of a subset of high-flow events; Represents a very small positive number, which can take the value 10. -6 To prevent the denominator from being 0.

[0077] ; In the formula, Indicates peak deviation rate Indicates the maximum corrected runoff; Indicates the maximum reference runoff, in percentages.

[0078] ; In the formula, Indicates the RMSE improvement rate; This represents the original simulation error; This indicates the error after correction.

[0079] ; In the formula, Indicates the KGE improvement rate; Indicates KGE after correction; This represents the original KGE.

[0080] Average performance of Nash efficiency coefficient for each sliding window Average performance of peak deviation rate RMSE improvement rate and average performance KGE improvement rate average performance for: ; ; ; ; In the formula, This indicates the total number of sliding windows.

[0081] Based on the evaluation results during the testing period, the Pareto multi-objective optimization method was used to determine the optimal model combination for each hydrological response unit.

[0082] Multi-objective evaluation vector The definition is as follows: ; In the formula, Indicates the first The first hydrological response unit A multi-objective evaluation vector is formed by the average performance of the candidate runoff correction models during the testing period's sliding window. Used to characterize the model's fitting ability under high-flow conditions; Used to reflect the degree of systematic bias of the model towards the flood peak process; Used to measure the ability of a model to reduce error; Used to measure the improvement in the overall fitting performance of the model.

[0083] The objective function is to maximize the average performance of the high-flow Nash efficiency coefficient, the average performance of the RMSE improvement rate, and the average performance of the KGE improvement rate, while minimizing the average performance of the absolute value of the peak deviation rate, so as to achieve synergistic optimization among the model's fitting ability, overall error control ability, and system deviation control ability in high-flow processes. ; ; In the formula, and Indicates the first Two candidate correction models for each hydrological response unit; and Representing the model respectively and model In the Performance values ​​on each evaluation metric; This indicates that all evaluation indicators are applicable. This indicates that at least one evaluation indicator exists; the symbol " "Indicates that both conditions must be met simultaneously." express Its performance on all evaluation metrics is no less than that of the model. And the model Its performance is significantly better than the model at least on one evaluation metric. When both of the above conditions are met, the model is called a valid model. Pareto Dominance Model Therefore, the model This is not a preferred model. Specifically, for the flood peak deviation rate index, it is transformed into a minimization problem by taking its absolute value, thus maintaining an optimization direction consistent with other maximization indices.

[0084] ; In the formula, Indicates the first The Pareto optimal solution set corresponding to each hydrological response unit is the set of models for which no other candidate model is superior to the model in all evaluation indicators and is strictly superior to the model in at least one indicator.

[0085] ; In the formula, Indicates the first Optimal runoff correction model for each hydrological response unit; Let represent the ideal performance vector, where each component is the optimal value of the corresponding evaluation index in the candidate model, used to construct the ideal reference point; the above formula selects the model with the best overall performance from the Pareto optimal solution set as the final correction model by calculating the Euclidean distance between the candidate model and the ideal point.

[0086] Finally, in S7, the runoff data to be predicted and the auxiliary feature data to be predicted for the target area are obtained; the runoff data to be predicted is corrected by combining the optimal model combination with the auxiliary feature data to be predicted, and the corrected prediction data for the target area is generated and output.

[0087] According to an embodiment of the present invention, the optimal model combination is applied to predict future scenario runoff data to generate a future corrected predicted runoff dataset.

[0088] Specifically, the runoff data to be predicted is represented as follows: ; In the formula, This represents the quantile mapping value of the runoff data to be predicted; This represents the trained correction model; This represents the runoff data to be predicted; This represents the optimal parameters of the trained correction model.

[0089] Predicted proportional residuals ; This represents the trained ZINB-Attention-LSTM model; Represents the set of optimal parameters for the model; The input features of the runoff data to be predicted include ,precipitation Temperature ,terrain glacier coverage rate Snow cover ratio Month, season, and lagged runoff.

[0090] The corrected predicted runoff data are as follows: ; And output the set ; , Indicates the total number of hydrological response units; , This indicates the predicted length of the future time step.

[0091] This invention also proposes a drought-area runoff data prediction system based on dataset correction. This system is used to implement the drought-area runoff data prediction method based on dataset correction described in the above embodiments. Figure 2 As shown, the system includes: The data acquisition module 210 is configured to acquire historical runoff data, reference runoff data, and auxiliary feature data of a target area in an arid region; and to perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data. The sub-region division module 220 is configured to construct a set of hydrological response characteristic indicators for the target region based on the reference runoff data and the auxiliary characteristic data; and to divide the target region into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. The data partitioning module 230 is configured to assign historical runoff data, reference runoff data and auxiliary feature data to their respective sub-regions according to the sub-regions, so as to obtain the original partitioning data, reference partitioning data and auxiliary partitioning data corresponding to each sub-region. The statistical correction module 240 is configured to construct a statistical correction model based on the original zoning data, the reference zoning data, and the error characteristic information within each sub-region; and to use the statistical correction model to correct the original zoning data to obtain preliminary corrected runoff data for each sub-region. The residual correction module 250 is configured to construct a residual learning model based on the preliminary corrected runoff data and the auxiliary zoning data in each sub-region; use the residual learning model to learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data; and correct the preliminary corrected runoff data according to the residual characteristics to obtain the final corrected runoff data for each sub-region. The model selection module 260 is configured to evaluate the performance of the combination of the statistical correction model and the residual learning model based on a predetermined evaluation strategy, and determine the optimal model combination for each sub-region based on the evaluation results. The runoff prediction module 270 is configured to acquire the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

[0092] It should be noted that the function of the arid zone runoff data prediction system based on dataset correction described in the embodiments of the present invention can be described by the aforementioned arid zone runoff data prediction method based on dataset correction. Therefore, for the parts not described in detail in the system embodiments, please refer to the above method embodiments, and they will not be repeated here.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting runoff data in arid regions based on dataset correction, characterized in that, Includes the following steps: S1. Obtain historical runoff data, reference runoff data, and auxiliary feature data for the target area in the arid zone; perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data; S2. Construct a set of hydrological response characteristic indicators for the target area based on the reference runoff data and the auxiliary characteristic data; The target area is then divided into multiple sub-regions with similar hydrological responses based on the set of hydrological response characteristic indicators. S3. Based on the sub-regions, the historical runoff data, reference runoff data, and auxiliary feature data are respectively assigned to their respective sub-regions to obtain the original partition data, reference partition data, and auxiliary partition data corresponding to each sub-region. S4. Within each sub-region, a statistical correction model is constructed based on the original zoning data, the reference zoning data, and the error characteristic information; the original zoning data is corrected using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. S5. Within each sub-region, a residual learning model is constructed based on the preliminary corrected runoff data and the auxiliary zoning data; the residual learning model is used to learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data; The preliminary corrected runoff data are corrected based on the residual characteristics to obtain the final corrected runoff data for each sub-region; S6. Based on a predetermined evaluation strategy, evaluate the performance of the combination of the statistical correction model and the residual learning model, and determine the optimal model combination for each sub-region based on the evaluation results. S7. Obtain the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine with the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

2. The method according to claim 1, characterized in that, The auxiliary feature data mentioned in S1 includes meteorological data, land cover data, and topographic data; the error feature information includes at least two of the following: deviation, relative deviation, absolute deviation, root mean square error, Nash efficiency coefficient, correlation coefficient, and KGE index between the historical runoff data and the reference runoff data. The set of hydrological response characteristic indicators described in S2 includes land cover data, topographic data, and multi-year average precipitation, multi-year average temperature, and drought index calculated based on meteorological data. The zero-flow frequency and runoff concentration are calculated based on reference runoff data. The zero-flow frequency is used to characterize the characteristics of intermittent or interrupted runoff in arid areas, and refers to the proportion of time steps in a certain time series where the runoff volume is lower than a preset zero-flow threshold to the total number of time steps. The runoff concentration is used to characterize the degree of concentration of runoff among months within a year.

3. The method according to claim 2, characterized in that, The method described in S2 for dividing the target region into multiple sub-regions with similar hydrological responses based on the set of hydrological response characteristic indicators includes: using principal component analysis to reduce the dimensionality of each hydrological response characteristic indicator after standardization; using a Gaussian mixture model to cluster the dimensionality-reduced principal component features, and using the category with the highest probability as the label of the sub-region; wherein, the elbow rule is used to determine the optimal number of clusters.

4. The method according to claim 1, characterized in that, S3 includes: resampling the historical runoff data, the reference runoff data, and the auxiliary feature data to a unified spatial grid and performing time alignment to form a spatiotemporally matched partitioned dataset.

5. The method according to claim 1, characterized in that, The statistical correction model described in S4 is a non-stationary quantile mapping model. The non-stationary quantile mapping model sets low flow level thresholds and high flow level thresholds based on the error characteristic information, and sets corresponding time-varying scale parameters for the three flow levels of low flow level, medium flow level and high flow level, respectively, to perform quantile mapping correction.

6. The method according to claim 5, characterized in that, S4 uses the statistical correction model to correct the original partition data, including: The quantiles of the current runoff data are determined based on the empirical cumulative distribution function of historical runoff data. The quantiles are mapped to quantile mapping values ​​based on the empirical cumulative distribution function of the reference runoff data; Based on the quantile increment preservation relationship, the basic corrected runoff data are obtained from the quantile mapping value; The basic corrected runoff data is scaled according to the time-varying scale parameters corresponding to the runoff data at the current moment to obtain preliminary corrected runoff data.

7. The method according to claim 1, characterized in that, The residual learning model described in S5 is a recurrent neural network model based on an attention mechanism; the residual feature is a proportional residual, which is the ratio of the difference between the reference partition data and the preliminary corrected runoff data to the preliminary corrected runoff data.

8. The method according to claim 7, characterized in that, The objective function of the attention-based recurrent neural network model described in S5 during the training process for: ； In the formula, This represents the loss function of a zero-inflated negative binomial distribution; This represents the residual error weighting coefficient; This represents the regularization weight coefficient; Indicates the length of the time series; Indicates the first The parameter set of an LSTM model for each spatial unit; Represents the square of the L2 norm; This represents the target value of the proportional residual. , Indicates reference runoff data, This indicates preliminary corrected runoff data. It is a very small positive number; This represents the predicted value of the proportional residual. ,in This is the weight matrix of the fully connected layer. For bias terms, The context vector output by the attention mechanism.

9. The method according to claim 8, characterized in that, The correction described in S5 is as follows: multiply the preliminary corrected runoff data by the sum of the predicted proportional residual and 1, and apply a non-negative constraint to the product result to obtain the final corrected runoff data.

10. The method according to claim 1, characterized in that, The predetermined evaluation strategy described in S6 includes a sliding window verification method. The sliding window verification method calculates the performance evaluation index of each model combination in multiple time windows by setting the sliding window length and sliding step size, and determines the optimal model combination based on the average value of the evaluation index of each window.

11. A system for predicting runoff data in arid regions based on dataset correction, characterized in that, The system is used to implement the arid zone runoff data prediction method based on dataset correction according to any one of claims 1-10; the system includes: The data acquisition module is configured to acquire historical runoff data, reference runoff data, and auxiliary feature data of the target area in the arid zone; and to perform error diagnosis on the historical runoff data based on the reference runoff data to obtain error feature information of the historical runoff data. The sub-region division module is configured to construct a set of hydrological response characteristic indicators for the target region based on the reference runoff data and the auxiliary feature data; and to divide the target region into multiple sub-regions with similar hydrological responses according to the set of hydrological response characteristic indicators. The data partitioning module is configured to assign historical runoff data, reference runoff data, and auxiliary feature data to their respective sub-regions according to the sub-regions, thereby obtaining the original partitioning data, reference partitioning data, and auxiliary partitioning data corresponding to each sub-region. The statistical correction module is configured to construct a statistical correction model based on the original zoning data, the reference zoning data, and the error characteristic information within each sub-region; and to correct the original zoning data using the statistical correction model to obtain preliminary corrected runoff data for each sub-region. The residual correction module is configured to: construct a residual learning model based on the preliminary corrected runoff data and the auxiliary zoning data within each sub-region; learn the residual characteristics of the preliminary corrected runoff data relative to the reference zoning data using the residual learning model; and correct the preliminary corrected runoff data according to the residual characteristics to obtain the final corrected runoff data for each sub-region. The model selection module is configured to evaluate the performance of the combination of the statistical correction model and the residual learning model based on a predetermined evaluation strategy, and determine the optimal model combination for each sub-region based on the evaluation results. The runoff prediction module is configured to acquire the runoff data to be predicted and the auxiliary feature data to be predicted for the target area; use the optimal model combination to combine the auxiliary feature data to be predicted to perform correction processing on the runoff data to be predicted, and generate and output the corrected prediction data for the target area.

Images