CN-122024411-A - Drought and flood emergency grading early warning method based on convolutional neural network

Abstract

The invention relates to a drought and flood emergency grading early warning method based on a convolutional neural network, which comprises the steps of grading definition and label construction of the drought and flood emergency based on objective clustering; the method comprises the steps of constructing a multisource climate driving feature set based on nonlinear dynamic optimization, constructing a GDFAA-CNN prediction model with time sequence perception and cost sensitivity characteristics, and applying the multiscale rolling prediction to medium and long periods of drought and waterlogging emergency. The invention utilizes the unique one-dimensional convolution and pooling layer structure of the convolution neural network, can automatically extract time sequence accumulation and mutation characteristics from high-dimensional and complex input climate data, adopts a data-driven model training mode, only needs to input current climate index data in practical application, outputs a prediction result of 1-3 months in the future in second-level time, greatly shortens the prediction time, avoids the uncertainty accumulation caused by the parameter calibration of a physical model, and strives for precious advance for the establishment of emergency plans by the flood disaster defense department.

Inventors

• DU JUNKAI
• LIU HAIYING
• Lv Xianglin
• QIU YAQIN
• CHENG XIAOMAN
• CHEN XIN
• TAN XINGYAN
• JIA LING
• WANG DONGDONG
• HAO CHUNFENG

Assignees

• 中国水利水电科学研究院
• 水利部信息中心(水利部水文水资源监测预报中心)

Dates

Publication Date

20260512

Application Date

20260130

Claims (1)

1. 1. A drought and flood emergency grading early warning method based on a convolutional neural network is characterized by comprising the following steps of: step 1, hierarchical definition and label construction of drought and waterlogging emergency events based on objective clustering, which comprises the following sub-steps: Step 11, identifying drought and flood emergency historical events, namely selecting a standardized precipitation and evapotranspiration index SPEI as a drought index, defining a flood event by adopting a relative threshold method, and integrating uncertainty analysis and former research, wherein a drought and flood emergency event is judged to be a drought and flood emergency event if the time interval between the end of the drought event and the beginning of the flood event is less than 10 days, and the drought and flood emergency event judging condition DFAA event is as follows: Wherein SPEI t is drought index of T days, T is transition days, T ds is drought starting time, T de is drought ending time, T fs is flood starting time, P t is accumulated precipitation amount of T days, and q th is relative precipitation threshold; And (12) evaluating the event intensity of a drought and flood mechanism, namely comprehensively evaluating the intensity of the drought and flood emergency event by adopting an entropy weight method, wherein the intensity of the drought and flood emergency event is determined by three core indexes, namely the intensity of drought and flood, and the alternation time between the drought and the flood, and the intensity calculation formula based on the entropy weight method is as follows: Wherein w j is the weight of the jth index, P i,j is the specific gravity of the ith drought and flood emergency on the jth index; The information entropy value of the j index is DFAA i , the comprehensive score of the ith drought and flood emergency event and the intensity of the drought and flood emergency event are obtained, m is the index number, and n is the event number; Is a standardized index; Step 13, event classification and label construction based on event intensity, namely converting the identified discrete day scale event into a continuous month scale intensity sequence, carrying out classification construction on the event intensity by adopting a K-means++ unsupervised clustering algorithm, and carrying out evaluation by adopting an elbow method when determining the optimal clustering quantity, wherein the elbow valve evaluation mode is that curve characteristics of the square sum WSS changing along with K values are analyzed by calculating the square sum in clusters under different clustering quantities K, and when obvious elbow inflection points appear on the curve, the K value corresponding to the position is the optimal clustering quantity, and the intra-group square error and SSE under different clustering quantities K are expressed as follows: Wherein C j is the j-th cluster, x is the sample point in the cluster, i.e., the historical DFAAI i value, and μ j is the centroid of the cluster; Drawing a relation curve of k and SSE (k), and selecting a curve curvature maximum point, namely an elbow point, wherein a corresponding k value is used as the optimal clustering number; Based on the optimal clustering number, the K-means++ algorithm is utilized to cluster the historical DFAAI i sequences, and the selection probability P (x) of the (r+1) th centroid is: Wherein: representing the distance of the sample x from the nearest selected centroid, iteratively updating the centroids until convergence, classifying each time step t as a class Forming a tag set ; For the sample Square of the distance to the nearest selected centroid; The method comprises the steps of collecting all samples to be clustered; step 2, constructing a multisource climate driving feature set based on nonlinear dynamic optimization, namely dynamically constructing a high signal-to-noise ratio space-time feature matrix containing the optimal physical precursor signals for different foreseeable periods of 1 month, 2 months and 3 months, and taking the space-time feature matrix as model input, wherein the method comprises the following substeps: sub-step 21, acquisition and preprocessing of Multi-Source climate index, collecting N remote-related climate indexes, and recording as a set Normalizing all indexes to eliminate dimension difference, and adopting a Min-Max normalization formula: Wherein X i,t is the normalized value of the ith climate index at the time t, f i is the historical value of the ith climate index, and f i,t is the historical value of the ith climate index at the time t; Sub-step 22, non-linear characteristics based on MIC-mRMR are preferably for different foresight periods Respectively constructing optimal feature subsets, and screening by adopting a maximum mutual information coefficient MIC and a maximum correlation minimum redundancy mRMR algorithm; Step 221, calculating the maximum correlation, namely calculating the MIC value between each climate index X i and the target predictive period drought and fast rotation label Y tau, and recording as D (X i , Y tau), wherein the MIC can capture any functional relation between the climate index X i and the target predictive period drought and fast rotation label Y tau, and the formula is approximate to: Wherein MIC (X, Y) is the maximum mutual information coefficient between variables X and Y, X represents a certain climate index, Y represents drought and flood emergency labels in the target forestation period, n x ,n y is grid number, and I (X, Y) is variable And (3) with Mutual information under the current grid division, B is the upper limit of the total number of grid divisions (usually taking the sample size A function of (d), and (d) a function of (d) a number of (c) a. Step 222, minimum redundancy calculation, namely calculating MIC values between the feature X i to be selected and the feature X s in the selected feature set S, and marking the MIC values as R (X i ,X s ) for measuring repeated information among the features; step 223, incremental search strategy, defining an evaluation function phi, each time selecting a feature addition set S that maximizes phi from the remaining features: wherein D (X i , Y tau) is the association degree between the feature to be selected and the forecast object, S is the selected feature set, X i is the feature to be selected, and X s is a feature in the selected set S. Repeating the above steps until K optimal features (such as K=20) are selected, to form an optimal feature subset for the foresight period τ 。 Sub-step 23, constructing a space-time sliding window feature matrix based on the preferred feature subset Constructing a time sliding window matrix of the input GDFAA-CNN, setting the sliding window length as L (i.e. predicting the future by using the information of the past L months), and constructing an input matrix for the current time t : Wherein x i,t is the characteristic value of the Kth characteristic at the time t. The matrix has the specification of The evolution sequence of the characteristics in the time dimension is reserved; Step 3, constructing GDFAA-CNN prediction models with time sequence perception and cost sensitivity characteristics, wherein the method comprises the following substeps: step 31, constructing a time sequence convolution layer, extracting the double characteristics of accumulation and mutation, namely constructing a one-dimensional convolution neural network aiming at the characteristic that the drought and waterlogging sharp turning process comprises long-term weather background and short-term weather mutation, and setting a convolution kernel in an input matrix along the time dimension The upper sliding, the shallow convolution kernel is used for capturing the high-frequency oscillation signal of the short-term climate factor, the deep convolution kernel is used for capturing the low-frequency evolution trend of the long-term climate factor, and the activation value of the jth feature map of the ith layer is obtained The method comprises the following steps: wherein D is the time span of the convolution kernel, x is the input feature, delta is the time delay index; weighting the learned 'time sequence evolution mode'; is a bias term; Step 32, a time sequence maximum pooling strategy is adopted to screen key physical time, namely considering that drought and waterlogging are always triggered by strong signals at specific time instead of full-period average action, adopting the maximum pooling strategy, realizing dimension reduction while retaining the strongest activation signals, automatically screening the key physical time with the greatest contribution to future drought and waterlogging on a time axis, and filtering noise interference in non-key time: Wherein: Omega is a pooling window; Output for the convolution layer; In the substep 33, embedding a cost sensitive loss function to solve the problem of scarcity of extreme samples, namely aiming at the problem that drought and waterlogging are suddenly turned to a typical low probability extreme event, a conventional loss function is easily submerged by a large number of non-suddenly turned samples, and forcing the model to pay attention to the extreme samples which are difficult to predict by defining a weighted cross entropy loss function and giving higher penalty weight to the scarce 'suddenly turned event' samples to solve the problem of 'missing report' caused by sample unbalance: Wherein L (theta) is a weighted cross entropy loss function value, theta is a parameter set to be optimized of the model, and y i is the first True class label of each sample, X i is the first The real class labels of the samples, P is the conditional probability distribution of model prediction, M is the total number of the samples, C is the class number, a c is the class weight coefficient, and the class weight coefficient of the heavy drought and water fast turning class is set for the rare heavy drought and water fast turning class C rare Setting for common sample ; Step 34, integration and robustness enhancement based on Dropout, namely, introducing a random inactivation Dropout mechanism into a full-connection layer for simulating uncertainty in climate prediction and preventing overfitting, randomly disconnecting part of neurons in the training process, and enhancing generalization capability and robustness of the model in the face of unseen extreme climate situations; and 4, the medium-long-term multi-scale rolling forecast application of the drought and waterlogging emergency event comprises the following sub-steps: a substep 41, real-time data acquisition and input construction, namely acquiring the latest N items of climate index data from a meteorological data interface at the current moment T now , and constructing a current input feature matrix X now through normalization and sliding window processing according to the substep 21 and the substep 22; Step 42, multiple prediction period probability deduction, namely inputting X now into a GDFAA-CNN prediction model set of trained 1 month, 2 months and 3 months prediction periods respectively, performing forward propagation calculation on the model, and outputting corresponding probability distribution vectors ; Step 43, judging the grading result, and determining the drought and flood level of T now +tau at the future time according to the principle of maximum posterior probability : Output of month tau occurs And the accurate grading early warning of the middle and long periods is realized at drought and waterlogging emergency.

Description

Drought and flood emergency grading early warning method based on convolutional neural network Technical Field The invention relates to a drought and waterlogging emergency grading early warning method based on a convolutional neural network, which is a hydrologic analysis method, is a method for disaster prevention and reduction, and is a method for predicting and warning disastrous weather. Background Drought and water logging is a typical extreme hydrologic event, which refers to the phenomenon that the regional hydrologic state is rapidly changed from drought to water logging (or vice versa) within a certain time interval. At present, a direct prediction model specially aiming at the medium-long drought and waterlogging emergency is not established. Existing technical means are not usually modeled for drought and waterlogging emergency events per se, but rely on long-term weather forecast products (such as output results of global weather patterns GCMs or numerical weather forecast patterns) issued by the weather department. Technicians indirectly deduce or identify the drought and waterlogging emergency process possibly happening in the future by analyzing the change trend of weather elements such as future rainfall, air temperature and the like in the forecast products and combining with a hydrologic model or an empirical formula. The indirect prediction mode has the obvious defects of limited medium-long term prediction precision and low reference value in practical application, and the prediction precision of the existing weather prediction products can be remarkably attenuated along with the extension of the prediction period on the medium-long term prediction of the extension period, month and season scale. Because drought and waterlogging are extremely abnormal events, the drought and waterlogging rapid transit system is extremely sensitive to the threshold value of meteorological conditions, high uncertainty commonly exists in medium-long-term weather forecast, and the accuracy requirements of actual disaster prevention and reduction are difficult to meet. The modeling method relies on weather forecast products to combine with physical process models to carry out deduction, and relates to complex atmospheric-ocean-land multi-turn layer coupling simulation. The calculation process is extremely tedious and takes a long time, and the requirement of quick emergency response cannot be met. In addition, a large number of physical parameters involved in the physical model are rated to have larger uncertainty, and the uncertainty can be accumulated step by step in the deduction process, so that the reliability of a final drought and waterlogging tight-turning prediction result is further reduced. How to find a special prediction model, and accurately early warning of drought and waterlogging on the premise of reducing calculated amount as much as possible, is a problem to be solved. Disclosure of Invention In order to overcome the problems in the prior art, the invention provides a drought and flood emergency grading early warning method based on a convolutional neural network. The method realizes medium-long-term, high-precision and automatic grading direct prediction of drought and waterlogging emergency by constructing an intelligent special drought and waterlogging emergency prediction model and using a convolutional neural network to drive by using multi-source climate signals. The invention aims to realize a drought and flood emergency grading early warning method based on a convolutional neural network, which comprises the following steps: step 1, hierarchical definition and label construction of drought and waterlogging emergency events based on objective clustering, which comprises the following sub-steps: Step 11, identifying drought and flood emergency historical events, namely selecting a standardized precipitation and evapotranspiration index SPEI as a drought index, defining a flood event by adopting a relative threshold method, and integrating uncertainty analysis and former research, wherein a drought and flood emergency event is judged to be a drought and flood emergency event if the time interval between the end of the drought event and the beginning of the flood event is less than 10 days, and the drought and flood emergency event judging condition DFAA event is as follows: Wherein SPEI t is drought index of T days, T is transition days, T ds is drought starting time, T de is drought ending time, T fs is flood starting time, P t is accumulated precipitation amount of T days, and q th is relative precipitation threshold; And (12) evaluating the event intensity of a drought and flood mechanism, namely comprehensively evaluating the intensity of the drought and flood emergency event by adopting an entropy weight method, wherein the intensity of the drought and flood emergency event is determined by three core indexes, namely the intensity of drought and flood, and the alternation time be