An intelligent drought disaster risk estimation system
By introducing a dynamic risk sample selection mechanism and optimizing the risk interval loss function, the problem that the drought disaster risk prediction system cannot adapt to the dynamic evolution characteristics and trend features of drought has been solved, and more accurate drought risk prediction and early warning have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN CLIMATE CENT
- Filing Date
- 2026-03-13
- Publication Date
- 2026-05-26
AI Technical Summary
Existing drought disaster risk prediction systems cannot adapt to the dynamic evolution of drought and lack targeted measurement of drought risk boundaries, resulting in the easy omission of key risks and poor risk prediction performance.
A dynamic risk sample selection mechanism is introduced, a risk interval loss function is designed, hyperparameters are optimized, and the dynamic evolution trend characteristics of drought risk are combined to improve sample screening and prediction capabilities.
It improves the reliability and early warning capability of drought risk prediction, reduces the probability of missing critical risks, enhances gradient sensitivity to novel drought samples with low confidence, and improves prediction performance.
Smart Images
Figure CN121836400B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of risk prediction, specifically to an intelligent drought disaster risk prediction system. Background Technology
[0002] Drought disaster risk prediction systems are intelligent decision support systems that integrate multi-source data from meteorology, geography, and hydrology. Through data processing, model analysis, and algorithmic computation, they achieve dynamic monitoring, accurate prediction, and scientific assessment of the probability and risk level of drought disasters. However, general drought disaster risk prediction systems often fail to adapt to the dynamic evolution of drought and lack targeted measurement of drought risk boundaries, leading to the easy omission of key risks. Furthermore, they often fail to focus on the dynamic trend characteristics of drought risks and lack sensitivity to novel drought samples with low confidence levels, resulting in poor risk prediction performance. Summary of the Invention
[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent drought disaster risk prediction system. Addressing the problems of general drought disaster risk prediction systems failing to adapt to the dynamic evolution of drought and lacking targeted measurement of drought risk boundaries, leading to the easy omission of key risks, this solution introduces a dynamic risk sample selection mechanism. This mechanism binds risk assessment status to sample selection, choosing only samples from the drought disaster candidate set that have the highest incremental value for risk assessment, thus improving drought risk capture capabilities. A targeted risk disorder loss is designed to accurately measure the uncertainty of drought risk boundary samples, reducing the probability of missing critical risks. A dynamic threshold is designed to achieve dynamic self-adaptation of sample selection thresholds, thereby improving the reliability of risk prediction. General drought disaster risk prediction systems often fail to focus on the dynamic trend characteristics of drought risks and lack sensitivity to novel drought samples with low confidence levels, leading to poor risk prediction performance. This solution incorporates the dynamic evolution trend of drought risk into the loss function design, measuring not only the error of static risk levels but also quantifying the consistency of trend predictions, thus improving early warning capabilities. It also dynamically adjusts the gradient amplitude based on the drought risk prediction confidence level, avoiding overuse of high-confidence conventional drought samples while enhancing gradient sensitivity to low-confidence novel drought samples. This allows for stable fitting of known risks while effectively exploring unknown risk patterns. Finally, it employs difference-driven search and focus-driven bias optimization of hyperparameters to improve the final drought disaster risk prediction performance.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent drought disaster risk prediction system, including a data acquisition module, a feature dimensionality reduction module, a risk sample selection module, a risk interval loss function design module, a drought risk model design module, a hyperparameter optimization module, and a drought disaster risk prediction module;
[0005] The data acquisition module obtains historical drought risk data and constructs a drought disaster dataset.
[0006] The feature dimensionality reduction module calculates feature importance scores to perform feature dimensionality reduction on the drought disaster dataset, thereby obtaining a drought disaster candidate set.
[0007] The risk sample selection module is based on a drought disaster candidate set and introduces a dynamic risk sample selection mechanism to construct a dynamic training set;
[0008] The risk interval loss function design module defines the trend gain based on the dynamic training set and obtains the risk interval loss function.
[0009] The drought risk model design module obtains a preliminary drought risk model based on the risk interval loss function;
[0010] The hyperparameter optimization module performs hyperparameter optimization on the preliminary drought risk model to obtain the final drought risk model.
[0011] The drought disaster risk prediction module performs risk prediction on real-time drought risk data based on a drought risk model.
[0012] Furthermore, the data acquisition module acquires historical drought risk data, including meteorological data, hydrological data, soil data, and other data; preprocesses the acquired data, including missing value imputation, outlier handling, and data normalization, and labels the drought risk level; thus obtaining a drought disaster dataset.
[0013] Furthermore, the feature dimensionality reduction module specifically includes:
[0014] Calculate the feature importance score, and quantify the correlation between features and risk by the ratio of covariance to standard deviation;
[0015] Feature dimensionality reduction is achieved through weighted summation to obtain a candidate set of drought disasters.
[0016] Furthermore, the risk sample selection module specifically includes:
[0017] Initialization: For each sample in the drought disaster candidate set, calculate the initial risk feature vector; train a multilayer perceptron using the drought disaster dataset to obtain the basic model, and obtain the initial risk prediction probability based on the initial risk feature vector; generate an initial risk state for each sample, and divide the drought disaster candidate set into a dynamic training set and a pool of candidate samples;
[0018] To design the risk entropy loss, for each sample in the pool of candidate samples, calculate the feature centroid with each risk category to obtain the risk entropy loss of each candidate sample.
[0019] The design of risk sample selection reward involves calculating the corresponding reward value for the candidate sample, updating the dynamic threshold, and selecting samples. After K iterations, the final dynamic training set is finally selected from the drought disaster candidate set.
[0020] Furthermore, the risk interval loss function design module specifically includes:
[0021] Define the trend gain quantity to characterize the consistency and incremental value of single-sample risk trend prediction;
[0022] Define a risk interval loss function and dynamically adjust the gradient magnitude through the interaction term between the gain and the confidence level;
[0023] The risk interval loss function is standardized by standardizing the gain, resulting in a standardized risk interval loss function.
[0024] Furthermore, the drought risk model design module specifically includes:
[0025] Model architecture design: A four-layer architecture is adopted, consisting of feature extraction layer, dynamic risk assessment layer, loss function optimization layer, and output layer. The dynamic risk sample selection mechanism of the risk sample selection module is embedded into the dynamic risk assessment layer. At the same time, the standardized risk interval loss function of the risk interval loss function design module is used as the core optimization criterion of the loss function optimization layer to achieve drought risk prediction.
[0026] Model training process: The dynamic training set after iterative screening by the risk sample selection module is divided into training set, validation set, and test set; the dynamic risk sample selection mechanism of the risk sample selection module is embedded throughout the training process, the risk status of the samples is re-evaluated after each round of training, high-value samples are dynamically added to the training set, and the model gradient is continuously optimized based on the standardized risk interval loss function; the model performance is verified by cross-validation; and the preliminary drought risk model is output.
[0027] Furthermore, the hyperparameter optimization module specifically includes:
[0028] Hyperparameter initialization: Initialize candidate solutions for hyperparameters of the drought risk model;
[0029] Clustering: Fuzzy mean clustering is used to cluster the candidate hyperparameter solutions into K clusters. The average objective function value of each cluster on the dynamic training set is calculated, and the cluster with the highest average objective function value is selected.
[0030] Candidate solution update involves performing two partial modification operations simultaneously on each candidate solution within the optimal cluster; then, a crossover operation is used to fuse these two partial modification vectors with the original candidate solution to generate a new hyperparameter candidate solution.
[0031] Convergence judgment: When the fitness value of the hyperparameters converges or the maximum number of iterations is reached, the optimization stops; if the objective function value of the optimal solution meets the target, the optimal hyperparameters are loaded into the preliminary drought risk model, the model parameters are iteratively updated, and the final drought risk model is obtained; otherwise, the iteration starts again.
[0032] Furthermore, the drought disaster risk prediction module acquires real-time drought risk data, and after preprocessing and dimensionality reduction, it is input into the drought risk model after hyperparameter optimization to obtain the drought disaster risk level, which serves as the drought disaster risk prediction result.
[0033] The beneficial effects achieved by the present invention using the above solution are as follows:
[0034] (1) In view of the fact that general drought disaster risk prediction systems cannot adapt to the dynamic evolution characteristics of drought and lack targeted measurement of drought risk boundaries, which leads to the easy omission of key risks, this solution introduces a dynamic risk sample selection mechanism, which binds the risk assessment status with the sample selection, selects only the samples with the highest incremental value for risk assessment in the drought disaster candidate set, and improves the ability to capture drought risks; it designs a risk disorder loss to accurately measure the uncertainty of drought risk boundary samples and reduce the probability of missing critical risks; it designs a dynamic threshold to realize the dynamic self-adaptation of sample screening threshold; and thus improves the reliability of risk prediction.
[0035] (2) In view of the problem that general drought disaster risk prediction systems do not focus on the dynamic trend characteristics of drought risk and are not sensitive enough to new drought samples with low confidence, resulting in poor risk prediction effect, this scheme incorporates the dynamic evolution trend of drought risk into the loss function design. This not only measures the error of static risk level, but also quantifies the consistency of trend prediction and improves the early warning capability. Combined with the dynamic adjustment of gradient magnitude of drought risk prediction confidence, it avoids the overuse of high-confidence conventional drought samples, while enhancing the gradient sensitivity to low-confidence new drought samples. It can effectively explore unknown risk patterns while stably fitting known risks. The hyperparameters are optimized by difference-driven search and focus-driven bias modification to improve the final drought disaster risk prediction effect. Attached Figure Description
[0036] Figure 1 A schematic diagram of an intelligent drought disaster risk prediction system provided by the present invention;
[0037] Figure 2 A schematic diagram of the risk model for the drought risk model design module.
[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0040] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0041] Example 1, see Figure 1 The present invention provides an intelligent drought disaster risk prediction system, including a data acquisition module, a feature dimensionality reduction module, a risk sample selection module, a risk interval loss function design module, a drought risk model design module, a hyperparameter optimization module, and a drought disaster risk prediction module.
[0042] The data acquisition module obtains historical drought risk data, constructs a drought disaster dataset, and sends the data to the feature dimensionality reduction module.
[0043] The feature dimensionality reduction module calculates feature importance scores to reduce the dimensionality of the drought disaster dataset; and sends the data to the risk sample selection module to obtain a drought disaster candidate set.
[0044] The risk sample selection module is based on the drought disaster candidate set, introduces a dynamic risk sample selection mechanism, constructs a dynamic training set, and sends the data to the risk interval loss function design module.
[0045] The risk interval loss function design module defines a trend gain based on a dynamic training set to obtain the risk interval loss function; and sends the data to the drought risk model design module.
[0046] The drought risk model design module obtains a preliminary drought risk model based on the risk interval loss function and sends the data to the hyperparameter optimization module.
[0047] The hyperparameter optimization module optimizes the hyperparameters of the preliminary drought risk model to obtain the final drought risk model; and sends the data to the drought disaster risk prediction module.
[0048] The drought disaster risk prediction module performs risk prediction on real-time drought risk data based on a drought risk model.
[0049] Example 2, see Figure 1 This embodiment is based on the above embodiment. The data acquisition module acquires historical drought risk data, including meteorological data, hydrological data, soil data, and other data. The meteorological data includes daily precipitation, evaporation, average temperature, relative humidity, and wind speed. The hydrological data includes surface runoff and groundwater level. The soil data includes soil type, soil moisture, organic matter content, and porosity. The other data includes cultivated land area, irrigated area, and population density. The acquired data is preprocessed, including missing value imputation (linear interpolation / K-nearest neighbor imputation), outlier handling (3σ principle), and data normalization (min-max normalization). The drought risk level is labeled, including normal, mild, moderate, and severe. A drought disaster dataset is obtained.
[0050] Example 3, see Figure 1 This embodiment, based on the above embodiment, uses a feature dimensionality reduction module to screen the core features with the greatest impact on drought risk for drought disaster datasets through feature importance scoring, thereby improving the accuracy of risk prediction; specifically including:
[0051] The feature importance score is calculated by using the ratio of covariance to standard deviation to quantify the correlation between the feature and risk, expressed as: ;in, It is the importance score of feature f; It is the covariance between feature f and risk label y; and These are the standard deviations of feature f and risk label y, respectively.
[0052] Feature dimensionality reduction, achieved through weighted summation, reduces redundant information and is represented as: ;in, It is the dimensionality reduction feature of the i-th drought disaster data; It is the weight of the k-th feature, given by Obtained by performing max-min normalization; is the feature value of the k-th dimension of the i-th drought disaster data; m is the total number of feature dimensions; the dimension is reduced to 1 / 3 of the original dimension; thus, a drought disaster candidate set is obtained.
[0053] Example 4, see Figure 1This embodiment, based on the above embodiment, introduces a dynamic risk sample selection mechanism in the risk sample selection module to avoid the sample selection strategy failing to adapt to the dynamic changes in drought risk (the evolution of drought from mild to severe). This mechanism binds the risk assessment status to sample selection, selecting only the samples from the drought disaster candidate set that have the highest incremental value for risk assessment, thereby improving the drought risk capture capability. Specifically, this includes:
[0054] Initialization: For each sample in the drought disaster candidate set, calculate the initial risk feature vector (the feature values after dimensionality reduction); train a multilayer perceptron using the drought disaster dataset to obtain the base model, and obtain the initial risk prediction probability based on the initial risk feature vector; generate an initial risk state for each sample to ensure that the sample selection decision is based on the current evaluation level of the model; divide the drought disaster candidate set into a dynamic training set (initially empty) and a candidate sample pool (initially the entire drought disaster candidate set); the risk state table is represented as follows: ;in, It represents the risk state of the i-th sample at step t. It is the risk feature vector of the i-th sample at step t; It is the model's probability prediction of the risk of the sample; the sample corresponds to drought disaster data;
[0055] The uncertainty of drought risk is mainly reflected in the risk boundary of the candidate samples. By designing a risk disorder loss, we measure the risk uncertainty of the candidate samples. The higher the entropy loss, the closer the sample is to the risk boundary, and the higher the sample value. This is expressed as: ;in, These are candidate samples The risk entropy loss; C is the number of risk categories; It is the feature vector of the candidate sample; is the feature centroid of the j-th risk class, and the feature mean of historical samples (dimensionality reduction needs to be synchronized); for each sample in the pool of candidate samples, calculate the feature centroid with each risk class to obtain the risk entropy loss of each candidate sample;
[0056] The design incorporates a risk sample selection reward system, using a reward and penalty mechanism to ensure that only the most valuable samples for risk assessment are selected, expressed as follows: ; ;in, It is the reward value for sample selection; It is a dynamic threshold that increases as the accuracy of the model's risk assessment improves. It is the average risk disorder loss of the candidate sample pool. is the real-time risk assessment accuracy of the model at step t, β is the threshold calibration factor, with a value of 0.2~0.5; calculate the corresponding reward value for the candidate sample, update the dynamic threshold; and perform sample selection, encouraging the selection of high-value samples and penalizing the selection of low-value samples; after K iterations, finally select the final dynamic training set (containing only the samples with the highest incremental value for risk assessment) from the drought disaster candidate set.
[0057] By performing the above operations, this solution addresses the problem that general drought disaster risk prediction systems cannot adapt to the dynamic evolution of drought and lack targeted measurement of drought risk boundaries, leading to the easy omission of key risks. It introduces a dynamic risk sample selection mechanism, binding risk assessment status with sample selection, choosing only samples from the drought disaster candidate set that have the highest incremental value for risk assessment, thus improving the ability to capture drought risks. A targeted risk disorder loss is designed to accurately measure the uncertainty of drought risk boundary samples, reducing the probability of missing critical risks. A dynamic threshold is designed to achieve dynamic self-adaptation of sample selection thresholds, thereby improving the reliability of risk prediction.
[0058] Example 5, see Figure 1 This embodiment is based on the above embodiment. The risk interval loss function design module uses the dynamic training set output by the risk sample selection module as the core data input and designs the risk interval loss function based on trend loss; specifically, it includes:
[0059] Define the trend gain, which characterizes the consistency and incremental value of single-sample risk trend prediction, and measures the accuracy of the model's prediction of the drought risk evolution trend of the sample. It is expressed as: ;in, It is the initial trend gain of the i-th sample; It is the rate of change over time of the predicted risk level of the sample; It is the rate of change over time of the sample's true risk level; It is a smoothing term, take 10. -6 ;
[0060] Define a risk interval loss function, and dynamically adjust the gradient magnitude through the interaction term between the gain and the confidence level: when the risk prediction confidence is high, reduce the gradient update strength to avoid over-utilization; when the risk prediction confidence is low, enhance gradient sensitivity to promote the exploration of new risk patterns, expressed as: ; ;in, It is the initial risk interval depletion of the sample; It is the gradient calibration factor, with a value ranging from 0.1 to 0.3; It is the probability of drought risk prediction. The higher the probability of drought risk occurring in the model output for the sample, the stronger the confidence of the model in predicting the risk level of the sample. Conversely, the lower the probability, the weaker the confidence. It is the basic risk trend loss of the sample;
[0061] The risk interval loss function is standardized, and the gain is standardized to avoid the difficulty in adjusting the calibration factor due to differences in the gain distribution at different training stages. This is expressed as: ;in, It is the standardized trend gain. and These are the mean and standard deviation of the trend gain in the current complete dynamic training set, respectively; output the standardized risk interval loss function; replace the trend gain with the standardized trend gain to obtain the standardized risk interval loss function.
[0062] Example 6, see Figure 1 and Figure 2 This embodiment, based on the above embodiments, constructs a risk model using a drought risk model design module, outputs the risk level and impact range, and deeply integrates the dynamic mechanism and output results of the risk sample selection module to achieve accurate prediction of drought risk. Specifically, it includes:
[0063] Model architecture design: A four-layer architecture is adopted, consisting of a feature extraction layer, a dynamic risk assessment layer, a loss function optimization layer, and an output layer. The dynamic risk sample selection mechanism of the risk sample selection module is embedded into the dynamic risk assessment layer. Simultaneously, the standardized risk interval loss function of the risk interval loss function design module is used as the core optimization criterion for the loss function optimization layer, achieving accurate prediction of drought risk. The model employs a hybrid CNN+LSTM architecture, where CNN is used to extract spatiotemporal features, and LSTM is used to capture the dynamic evolution of drought risk. The output is the drought risk probability, which is converted into a risk level, expressed as: ;in, It represents the drought risk level of the i-th sample; It represents the probability of the i-th drought risk occurring;
[0064] Model training process: The dynamic training set after iterative screening by the risk sample selection module is divided into training set, validation set, and test set in a 7:2:1 ratio; the dynamic risk sample selection mechanism of the risk sample selection module is embedded throughout the training process, and the risk status of the samples is re-evaluated after each round of training, and high-value samples are dynamically added to the training set. At the same time, the model gradient is continuously optimized based on the standardized risk interval loss function; the model performance is verified by cross-validation, and the risk status and entropy loss characteristics of the samples are retained during the verification process to ensure the model's ability to capture the dynamic evolution of drought and its generalization ability; the preliminary drought risk model is output.
[0065] Example 7, see Figure 1This embodiment is based on the above embodiment. The hyperparameter optimization module, adapted to the high-dimensional heterogeneity of drought data, optimizes the hyperparameters of the preliminary drought risk model optimized by the risk sample selection module and dynamic training mechanism through partial modification-crossover-selection, focusing on the optimal risk model parameter cluster and accelerating convergence to the globally optimal hyperparameters. The objective function of hyperparameter optimization is the risk prediction accuracy of the model on the dynamic training set; specifically including:
[0066] Hyperparameter initialization: Initialize candidate solutions for hyperparameters of the drought risk model (learning rate, batch size, regularization coefficient, number of network layers, number of neurons).
[0067] Clustering: Fuzzy mean clustering is used to cluster the candidate hyperparameter solutions into K clusters. The average objective function value (risk prediction accuracy) of each cluster on the dynamic training set is calculated, and the cluster with the highest average objective function value (the region where the optimal solution is located) is selected.
[0068] Candidate solution update: for each candidate solution in the optimal cluster Simultaneously, two partial modification operations are performed: differential evolution partial modification, which guides the search through the difference of hyperparameter candidate solutions to avoid local optima, as expressed as: Focusing on traction bias modification and optimal hyperparameter clusters to reduce invalid searches, this can be represented as: Then, through a crossover operation, these two biased vectors are fused with the original candidate solutions (fusing according to the binomial crossover rule) to generate new hyperparameter candidate solutions; where, It is the partial modification vector of the o-th candidate solution for hyperparameters; and These are two randomly selected candidate solutions for hyperparameters, where g is the number of iterations and F is the scaling factor, ranging from 0.5 to 1.0. It is the partial change vector driven by clustering; It is the optimal solution in the optimal hyperparameter cluster;
[0069] Convergence is determined when the fitness values of the hyperparameters converge (the improvement is less than 10). -3 If the maximum number of iterations (20~50) is reached, the optimization stops; if the objective function value of the optimal solution is higher than 0.95, the optimal hyperparameters are loaded into the preliminary drought risk model to complete the model parameter iterative update and obtain the final drought risk model; otherwise, the iteration is restarted.
[0070] By performing the above operations, this solution addresses the problem that general drought disaster risk prediction systems often fail to focus on the dynamic trend characteristics of drought risks and lack sensitivity to novel drought samples with low confidence levels, leading to poor risk prediction performance. Instead, it incorporates the dynamic evolution trend of drought risk into the loss function design. This not only measures the error of static risk levels but also quantifies the consistency of trend predictions, improving early warning capabilities. Furthermore, it dynamically adjusts the gradient magnitude based on the drought risk prediction confidence level, avoiding overuse of high-confidence conventional drought samples while enhancing gradient sensitivity to novel drought samples with low confidence levels. This allows for stable fitting of known risks while effectively exploring unknown risk patterns. Finally, it performs difference-driven search and focus-driven bias optimization on hyperparameters to improve the final drought disaster risk prediction performance.
[0071] Example 8, see Figure 1 This embodiment is based on the above embodiment. The drought disaster risk prediction module obtains real-time drought risk data, and after preprocessing and dimensionality reduction, it is input into the drought risk model after hyperparameter optimization to obtain the drought disaster risk level, which is used as the drought disaster risk prediction result.
[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0073] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. An intelligent drought disaster risk prediction system, characterized in that: The system includes a data acquisition module, a feature dimensionality reduction module, a risk sample selection module, a risk interval loss function design module, a drought risk model design module, a hyperparameter optimization module, and a drought disaster risk prediction module. The data acquisition module obtains historical drought risk data and constructs a drought disaster dataset. The feature dimensionality reduction module calculates feature importance scores to perform feature dimensionality reduction on the drought disaster dataset; Obtain a candidate set of drought disasters; The risk sample selection module is based on a drought disaster candidate set and introduces a dynamic risk sample selection mechanism to construct a dynamic training set; The risk interval loss function design module defines the trend gain based on the dynamic training set and obtains the risk interval loss function. The drought risk model design module obtains a preliminary drought risk model based on the risk interval loss function; The hyperparameter optimization module performs hyperparameter optimization on the preliminary drought risk model to obtain the final drought risk model. The drought disaster risk prediction module predicts the risk of real-time drought risk based on the drought risk model; The data acquisition module obtains historical drought risk data, including meteorological data, hydrological data, soil data, and other data. The meteorological data includes daily precipitation, evaporation, average temperature, relative humidity, and wind speed. The hydrological data includes surface runoff and groundwater level. The soil data includes soil type, soil moisture, organic matter content, and porosity. The other data includes cultivated land area, irrigated area, and population density. The acquired data is preprocessed, including missing value imputation, outlier handling, and data normalization, and drought risk levels are labeled. Obtain drought disaster dataset; The risk sample selection module specifically includes: Initialization: For each sample in the drought disaster candidate set, calculate the initial risk feature vector; train a multilayer perceptron using the drought disaster dataset to obtain the base model, and obtain the initial risk prediction probability based on the initial risk feature vector; generate an initial risk state for each sample; divide the drought disaster candidate set into a dynamic training set and a candidate sample pool; the risk state is represented as: ;in, It represents the risk state of the i-th sample at step t. It is the risk feature vector of the i-th sample at step t; It is the probability of the model predicting the risk of the sample; Design risk disorder loss is expressed as: ;in, These are candidate samples The risk entropy loss; C is the number of risk categories; It is the feature vector of the candidate sample; is the feature centroid of the j-th risk class, and the feature mean of the historical samples; for each sample in the candidate sample pool, calculate the feature centroid with each risk class to obtain the risk entropy loss of each candidate sample; The design incorporates a risk sample selection reward system, using a reward and penalty mechanism to ensure that only the most valuable samples for risk assessment are selected, expressed as follows: ; ;in, It is the reward value for sample selection; It is a dynamic threshold. It is the average risk disorder loss of the candidate sample pool. is the real-time risk assessment accuracy of the model at step t, and β is the threshold calibration factor; calculate the corresponding reward value for the candidate sample, update the dynamic threshold; and perform sample selection, encouraging the selection of high-value samples and penalizing the selection of low-value samples; after K iterations, the final dynamic training set is finally selected from the drought disaster candidate set; The risk interval loss function design module includes: Define the trend gain, which characterizes the consistency and incremental value of single-sample risk trend prediction, and measures the accuracy of the model's prediction of the drought risk evolution trend of the sample. It is expressed as: ;in, It is the initial trend gain of the i-th sample; It is the rate of change over time of the predicted risk level of the sample; It is the rate of change over time of the sample's true risk level; It is a smoothing term; Define a risk interval loss function, dynamically adjusting the gradient magnitude through the interaction term between the gain and confidence level, expressed as: ; ;in, It is the initial risk interval depletion of the sample; It is the gradient calibration factor; It is the probability of drought risk prediction; It is the basic risk trend loss of the sample.
2. The intelligent drought disaster risk prediction system according to claim 1, characterized in that: The risk interval loss function design module also includes: The risk interval loss function is standardized by standardizing the gain, resulting in a standardized risk interval loss function.
3. The intelligent drought disaster risk prediction system according to claim 1, characterized in that: The drought risk model design module specifically includes: Model architecture design: A four-layer architecture is adopted, consisting of feature extraction layer, dynamic risk assessment layer, loss function optimization layer, and output layer. The dynamic risk sample selection mechanism of the risk sample selection module is embedded into the dynamic risk assessment layer. At the same time, the standardized risk interval loss function of the risk interval loss function design module is used as the core optimization criterion of the loss function optimization layer to achieve drought risk prediction. Model training process: The dynamic training set after iterative screening by the risk sample selection module is divided into training set, validation set, and test set; the dynamic risk sample selection mechanism of the risk sample selection module is embedded throughout the training process, the risk status of the samples is re-evaluated after each round of training, high-value samples are dynamically added to the training set, and the model gradient is continuously optimized based on the standardized risk interval loss function; the model performance is verified by cross-validation; and the preliminary drought risk model is output.
4. The intelligent drought disaster risk prediction system according to claim 1, characterized in that: The hyperparameter optimization module specifically includes: Hyperparameter initialization: Initialize candidate solutions for hyperparameters of the drought risk model; Clustering: Fuzzy mean clustering is used to cluster the candidate hyperparameter solutions into K clusters. The average objective function value of each cluster on the dynamic training set is calculated, and the cluster with the highest average objective function value is selected. Candidate solution update involves performing two partial modification operations simultaneously on each candidate solution within the optimal cluster; then, a crossover operation is used to fuse these two partial modification vectors with the original candidate solution to generate a new hyperparameter candidate solution. Convergence judgment: When the fitness value of the hyperparameters converges or the maximum number of iterations is reached, the optimization stops; if the objective function value of the optimal solution meets the target, the optimal hyperparameters are loaded into the preliminary drought risk model, the model parameters are iteratively updated, and the final drought risk model is obtained; otherwise, the iteration starts again.
5. The intelligent drought disaster risk prediction system according to claim 1, characterized in that: The feature dimensionality reduction module specifically includes: Calculate the feature importance score, and quantify the correlation between features and risk by the ratio of covariance to standard deviation; Feature dimensionality reduction is achieved through weighted summation to obtain a candidate set of drought disasters.
6. The intelligent drought disaster risk prediction system according to claim 1, characterized in that: The drought disaster risk prediction module acquires real-time drought risk data, and after preprocessing and dimensionality reduction, it is input into the drought risk model after hyperparameter optimization to obtain the drought disaster risk level, which serves as the drought disaster risk prediction result.