Rainstorm flood disaster-causing element prediction method and system based on modal decomposition and deep learning

Abstract

The application discloses a rainstorm flood disaster-causing factor prediction method and system based on modal decomposition and deep learning, solves the problems of insufficient capture of nonlinear time sequence characteristics, dependence on complex data and weak generalization ability of traditional models, and comprises the following steps: collecting flow, flow rate and water level data of an upstream station as input and collecting data of a downstream disaster point as output; preprocessing the data; optimizing variational modal decomposition parameters by using frost ice optimization algorithm; constructing a Fourier transform high-low frequency feature enhanced attention network, dividing low frequency and high frequency components, extracting trend characteristics through a fluctuation enhancement module, capturing dynamic changes through a multi-path difference calculation unit, introducing self-perception attention to realize feature weighted fusion and long-term memory, and outputting downstream hydrological state prediction results. Through the optimization of the modal decomposition and deep learning collaborative mechanism, the precision, robustness and generalization ability of the sudden mountain flood prediction are significantly improved, and the application is suitable for disaster emergency management in complex scenes of small and medium-sized river basins.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flood disaster prediction, in particular to a rainstorm and mountain flood disaster-causing element prediction method and system based on modal decomposition and deep learning. BACKGROUND

[0002] Flood disaster is one of the natural disasters that most seriously threatens human life and property safety. The sudden mountain flood in small and medium-sized river basins is a key and difficult point in the field of flood control and disaster reduction because of its suddenness, destructive power and short warning time. At present, the warning and forecasting technology for mountain flood in small and medium-sized river basins mainly relies on traditional linear regression model and hydrodynamic model. The traditional linear regression model is difficult to effectively capture the complex and nonlinear dynamic characteristics in the sudden mountain flood data due to the linear characteristics of the assumed relationship, and cannot provide accurate prediction. The construction of the hydrodynamic model usually needs to integrate complex data such as water regime information, underlying surface properties and river section data, which not only consumes time in the model construction process but also has a high professional technical threshold, making it difficult to achieve popular application in the whole country. In addition, the hydrodynamic model has high calculation cost and long time required in the simulation and prediction process, which is difficult to meet the demand for rapid response and real-time decision-making in emergency scenarios.

[0003] At present, emergency telemetry warning stations have been built in several pilot areas, and telemetry warning stations have been deployed at key places in the upstream and downstream river channels to realize real-time, accurate, dynamic and minute-level monitoring of water regime information. A certain degree of small and medium-sized river basin flood information library has been established.

[0004] It is necessary to design a model that can accurately capture the nonlinear relationship between upstream water regime information and downstream disaster area water regime information 40 minutes, 60 minutes or even several hours later, and realize the 40-minute, 60-minute or even several-hour warning and forecasting of the sudden rainstorm and mountain flood in the downstream disaster area, using the established small and medium-sized river basin flood information library as the driving. SUMMARY

[0005] The present application aims to provide a rainstorm and mountain flood disaster-causing element prediction method and system based on modal decomposition and deep learning to solve the technical problems of insufficient capture of nonlinear time series characteristics by traditional models, dependence of physical models on complex data and weak generalization ability.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a rainstorm and mountain flood disaster-causing element prediction method based on modal decomposition and deep learning, comprising:

[0008] Collecting the flow, flow rate and water level data of the upstream telemetry station as input and the flow and water level data of the downstream disaster point as output;

[0009] The collected data is preprocessed, including normalization, dataset division and data format conversion;

[0010] The frost-ice optimization algorithm is used to optimize the variational modal decomposition parameters, the decomposition parameters are adjusted through the soft frost search strategy, the hard frost puncture strategy and the forward greedy mechanism, the input data is adaptively modal decomposed, and sub-sequences containing different frequency scales are obtained;

[0011] The Fourier transform high-low frequency feature enhanced attention network is constructed, the sub-sequences after decomposition are subjected to Fourier transform, the low frequency components and the high frequency components are divided according to a preset frequency threshold, the low frequency components and the high frequency components are processed respectively, and the low frequency features and the high frequency features are obtained;

[0012] The low frequency features and the high frequency features are fused, self-perception attention is used for adaptive weighting, and the hydrological state prediction result of the downstream disaster point is output.

[0013] Further, the specific process of preprocessing the collected data is as follows:

[0014] The input feature normalization interval is [-1, 1], and the output feature normalization interval is [0, 1];

[0015] The dataset is divided into a training set, a validation set and a test set according to a ratio of 7:1:2;

[0016] The data is converted into a time series window form, the input is the historical data of the past 32 time steps, and the output is the prediction result of the current 60 time steps;

[0017] The input shape is: sample number, time step length, feature number; and the output shape is: sample number, 60, 2.

[0018] Further, the calculation formula of the soft frost search strategy is as follows:

[0019]

[0020] wherein, is the new position of the updated particle, and ij represents the jth particle in the ith frost agent; is the jth particle of the best frost agent in the frost population R; the parameter r1 is a random number in the range of [-1, 1], which controls the motion direction of the particle together with cos, and changes with the number of iterations, and cosθ represents the cosine value of the included angle between the motion direction of the particle and the best particle, which is used to control the direction of the particle motion; β is an environmental factor, which changes with the number of iterations to simulate the influence of the external environment, and is used to ensure the convergence of the algorithm; h is the adhesion degree, which is a random number in the range of [0, 1], and is used to control the distance between the centers of two frost particles; b Uijdenotes exploratory information, representing the search path information of the jth particle in the ith group; b Lij denotes learning information, representing the learning path information of the jth particle in the ith group; r2 represents a random number in the range of [0, 1], which is used to introduce randomness in the algorithm and affect the behavior or decision of the particle; E represents an environmental factor, which is a parameter used to simulate the influence of the external environment; β is an environmental factor, which changes with the number of iterations to simulate the influence of the external environment, and is used to ensure the convergence of the algorithm; h is the adhesion degree, which is a random number in the range of [0, 1], used to control the distance between the centers of the two frost particles; t is the current number of iterations; T is the maximum number of iterations of the algorithm; the mathematical model of β is a step function; [·] represents rounding; the default value of w is 5, which is used to control the number of segments of the step function;

[0021] The hard frost penetration strategy enables particles in the algorithm to exchange with each other, improving the convergence speed and the ability to jump out of the local optimum of the algorithm, and the calculation formula is:

[0022]

[0023] wherein, denotes the new position of the updated particle, is the position of the jth particle of the best frost agent in the frost population R; F normr denotes the normalized value of the current agent fitness value, which means the probability of the ith ice agent being selected; r3 is a random number in the interval [-1, 1];

[0024] The forward greedy mechanism compares the updated fitness value of the individual with the value before updating to determine whether to replace the individual, and simultaneously replaces the solutions of the two individuals.

[0025] In a second aspect, the present application provides a rainstorm flood disaster-causing factor prediction system based on modal decomposition and deep learning, comprising:

[0026] A data acquisition module is configured to acquire flow, flow rate and water level data of an upstream station, and construct a hydrological time series dataset.

[0027] A preprocessing module is configured to preprocess the collected data, including normalization processing, dataset division and data format conversion.

[0028] A signal decomposition module is configured to decompose the preprocessed data, and optimize the variational modal decomposition VMD parameters through the frost-ice optimization algorithm, so as to perform optimal decomposition for different input features.

[0029] A prediction module is configured to predict the sudden rainstorm flood in the affected area through the construction of a Fourier transform high-low frequency feature enhanced attention network.

[0030] Further, the structure of the Fourier transform high-low frequency feature enhancement attention network comprises:

[0031] a frequency separation layer, which performs frequency analysis on the time series data through Fourier transform, and sets a threshold frequency to distinguish low frequency components and high frequency components;

[0032] a low frequency processing module, which performs deep feature extraction on the low frequency signal through a fluctuation feature enhancement module FEM;

[0033] a high frequency processing module, which processes the high frequency signal through a multi-path difference feature extraction module DEM to capture dynamic changes and burst features in the time series data;

[0034] self-aware attention, which is used to weight and fuse low frequency features and high frequency features.

[0035] Further, the fluctuation feature enhancement module processes the input signal through an Avgpool1d average pooling with a convolution kernel size of 3 and a FluctuationEnhancer fluctuation separation enhancement unit to obtain a fundamental frequency part and an enhanced fluctuation part; the fundamental frequency part is processed using a convolution network with a convolution kernel size of 1 to retain the original trend of the fundamental frequency part;

[0036] The FluctuationEnhancer fluctuation separation enhancement unit obtains the fluctuation part by subtracting the Avgpool1d average pooled signal from the input signal, and then the fluctuation part is processed through BoostBlock feature enhancement; wherein Boost Block is a feature enhancement module containing deep separable convolution and channel expansion-shrinkage design, and GELU activation function is introduced to increase nonlinearity;

[0037] The fundamental frequency part and the output result of the FluctuationEnhancer fluctuation separation enhancement unit are spliced in the channel dimension to obtain the final output of the fluctuation feature enhancement module.

[0038] Further, the multi-path difference feature extraction module comprises a multi-path fusion of center difference convolution, channel difference convolution, time gradient difference convolution, weight difference convolution and standard convolution.

[0039] Further, the center difference convolution is constructed by reparameterizing the standard convolution kernel W conv ∈R Cout×Cin×K A center-neighborhood contrast mechanism is constructed to correct the weight of the center position of the convolution kernel; wherein W conv is a three-dimensional convolution kernel weight matrix; Cout is the number of output channels; Cin is the number of input channels; K is the spatial dimension of the convolution kernel, and a center-neighborhood contrast mechanism is constructed to correct the weight of the center position of the convolution kernel;

[0040] The channel differential convolution establishes the difference relationship between channels by introducing a cross-channel weight coupling mechanism, and the operation of exchanging channel weights offsets the convolution kernel weights, forms the difference response between channels to capture the relationship mode of the input signal on different channels;

[0041] The time gradient differential convolution is used to calculate the first or second order time gradient of the time series data;

[0042] The weight differential convolution forms a symmetric differential structure by keeping the first weight of the convolution kernel unchanged and inverting the last weight value;

[0043] The standard convolution retains the original convolution kernel and maintains the baseline feature extraction capability;

[0044] After the features of each path are added, a 1D convolution layer is input to integrate the above features.

[0045] Further, the self-perception attention module includes four sub-modules of spatial attention, channel attention, scaled dot product attention and fusion attention;

[0046] The spatial attention calculates the mean and maximum values along the channel dimension, concatenates them, and generates a spatial attention map through 7x1 convolution to capture key spatial positions;

[0047] The channel attention part generates channel attention through two 1x1 convolutions of the bottleneck structure after global average pooling, models the inter-channel dependency while reducing the parameter quantity to 1 / 8 of the original channel number, and reduces the risk of overfitting;

[0048] The scaled dot product attention projects the input F into Query, Key and Value after transposition, calculates the self-attention matrix and weighted aggregation, and outputs the global context features;

[0049] The fusion attention concatenates the original input and the global attention features, generates the final attention output through grouped convolution and Sigmoid activation function, realizes adaptive weighted fusion of local details and global context, and avoids information loss.

[0050] Based on the above technical solutions, the embodiments of the present application can at least produce the following technical effects:

[0051] (1) The present application combines the frost optimization algorithm, which can automatically adjust the parameters according to the characteristics of different input features to realize the optimal decomposition scheme. The purpose of the variational modal decomposition module is to deeply mine the potential characteristics of the water regime data in the middle and small basins, especially for the water regime data with complex time-varying characteristics, to ensure that the signals obtained after decomposition have high usability. This processing not only provides more accurate input data for the subsequent deep learning module, but also lays a foundation for the subsequent model to capture the nonlinear spatiotemporal relationship in the water regime data.

[0052] (2) The present application proposes a Fourier transform high-low frequency feature enhanced attention network for the characteristics of the sudden mountain flood time series signal in the middle and small basins. The model is a deep learning network based on Fourier transform and multi-scale feature extraction, which integrates the deep processing of time domain and frequency domain information. The core innovation of the model lies in the frequency separation (low frequency and high frequency) by Fourier transform and the innovative modules of multi-path difference calculation unit (DEM), fluctuation enhancement module (FEM) and self-perception attention (SPRA) for high and low frequency signals, which effectively capture and enhance the main trend and detail features in the time series signal. This multi-level processing method is especially suitable for suddenness, nonlinearity and variability in time series data, such as sudden mountain flood prediction in middle and small basins.

[0053] (3) The present application combines the optimized variational modal decomposition and deep learning: through the seamless integration of optimization algorithm (RIME), data preprocessing (VMD) and efficient intelligent prediction algorithm process, the characteristics of the flood time series process in the middle and small basins are better captured, so that the new algorithm has good adaptability to multi-element and multi-type emergency telemetry station sensing data, is applicable to different spatial scales and complex basin underlying surfaces in middle and small basins, mountain flood gully, sub-basin / sub-converging area, etc., enhances the robustness of the model to noise and abnormal data, improves the overall performance of mountain flood disaster element prediction, and has important practical application value for mountain flood disaster emergency. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings shown.

[0055] Figure 1 is the decomposition graph of the RIME optimized VMD of the present application;

[0056] Figure 2 is the architecture diagram of the deep learning model of the present application;

[0057] Figure 3 is a structural diagram of the fluctuation feature enhancement module of the present application;

[0058] Figure 4 is a structural diagram of the multi-path difference feature extraction module of the present application;

[0059] Figure 5 is a workflow diagram of the SPRA self-perceived attention of the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application. In addition, the technical solutions of various embodiments can be combined with each other, but it must be based on the fact that those skilled in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope of the present application.

[0061] The rainstorm flood disaster-causing factor prediction method based on modal decomposition and deep learning includes the following steps:

[0062] Step 1, constructing a data set: collecting the flow, flow rate and water level data of 12 upstream telemetry sites as input, and the flow and water level data of downstream disaster points as output;

[0063] Specifically, the data set is used for modeling and predicting hydrological time series, and the input variable X and the output variable Y represent the hydrological observation data of the upstream telemetry site and the hydrological response data of the downstream disaster point, respectively. The data preprocessing and construction process is as follows:

[0064] Input features (X): include the observation data of 12 upstream telemetry alarm sites, and the features of each site include: flow (Discharge, Q), flow rate (Velocity, V), and water level (Water Level, H).

[0065] Output variable (Y): represents the hydrological state of the downstream disaster point, including: flow (Q) and water level (H) as two target variables.

[0066] Step 2, data preprocessing: preprocessing the collected data, including normalization processing, data set division and data format conversion;

[0067] 2-1, normalization processing: the input features (X) are normalized to the [-1, 1] interval by using MinMaxScaler to enhance the stability of the model;

[0068] The output variable (Y) is normalized to [0, 1] using MinMaxScaler to ensure consistent numerical ranges and improve model convergence.

[0069] 2-2, Data set division: training set (70%), validation set (10%), test set (20%) to ensure that the model learns sufficiently during the training phase and generalizes well on the test set.

[0070] 2-3, To adapt to the input format of deep learning models such as LSTM and GRU, the data is converted into a time series window format:

[0071] Time window length (look_back = 32): The input data for each sample consists of 32 consecutive time steps, i.e., the model can observe the historical data of the upstream twelve stations for 32 time steps (5 minutes each) to make predictions.

[0072] Prediction step (T = 60): The goal is to predict the downstream flow and water level data 5 hours after the current time step.

[0073] Data format:

[0074] Input shape: (number of samples, time step length, number of features) = (N, 32, 36);

[0075] Output shape: (number of samples, 60, 2);

[0076] For example, for sample t, the input data is: t = [X t-31 , X t-30 ,..., X t ];

[0077] The target output is: t = [Y t+1 , Y t+2 ,..., t +60 ].

[0078] The purpose of this is to allow the model to learn how historical information from upstream stations affects future hydrological conditions downstream.

[0079] Step 3, optimize the variational modal decomposition (VMD): as shown in Figure 1 , use the Riemannian Ice and Manifold (RIEM) optimization algorithm to optimize the variational modal decomposition parameters, and perform adaptive modal decomposition on the input data to obtain sub-sequences containing different frequency scales;

[0080] The Riemannian Ice and Manifold (RIEM) optimization algorithm adjusts the decomposition parameters through a soft frost search strategy, hard frost puncture strategy, and forward greedy mechanism to achieve efficient decomposition of signals. The algorithm process includes:

[0081] 3-1, soft frost search strategy: simulate the movement of soft frost particles in frost ice, and propose a step-by-step search and development method for exploring algorithms, allowing algorithms to switch between large-scale exploration and small-scale development, so as to realize efficient and high-precision optimization:

[0082]

[0083] wherein, is the new position of the updated particle, and ij represents the jth particle in the ith frost agent; is the jth particle of the best frost agent in the frost population R; The parameter r1 is a random number in the range of [-1, 1], which controls the direction of particle movement together with cos, and changes with the number of iterations; β is the environmental factor, which changes with the number of iterations to simulate the influence of the external environment, and is used to ensure the convergence of the algorithm; h is the adhesion degree, which is a random number in the range of [0, 1], used to control the distance between the centers of two frost particles; t is the current iteration number; T is the maximum iteration number of the algorithm; The mathematical model of β is a step function; [·] represents rounding; The default value of w is 5, which is used to control the number of segments of the step function.

[0084] 3-2, hard frost piercing strategy: make particles in the algorithm exchange with each other, improve the convergence speed and the ability to jump out of local optimum of the algorithm:

[0085]

[0086] wherein, represents the new position of the updated particle, and is the position of the jth particle of the best frost agent in the frost population R; F normr represents the normalized value of the current agent fitness value, which means the probability of the ith ice agent being selected; r3 is a random number in the interval [-1, 1].

[0087] 3-3, forward greedy mechanism: used for population update in meta-heuristic optimization algorithm. This mechanism compares the updated fitness value of the individual with the value before updating to decide whether to replace the individual, and at the same time replaces the solutions of the two individuals.

[0088] Step 4, construct a deep learning model;

[0089] The optimal decomposed data then enters the deep learning model. According to the characteristics of the time series signal of sudden mountain torrents in small and medium-sized river basins, a Fourier transform high-low frequency feature enhanced attention network is used. This model is a deep learning network based on Fourier transform and multi-scale feature extraction, which integrates time domain and frequency domain information in depth processing. The model effectively captures and enhances the main trend and detail features in the time series signal through frequency separation (low frequency and high frequency) and multi-scale convolution network. This multi-level processing method is especially suitable for suddenness, nonlinearity and variability in time series data, such as sudden mountain flood prediction in small and medium-sized river basins. Figure 2 As shown in the following figure, specifically, the structure of the model includes:

[0090] 4-1, Frequency separation layer (Freq Split Layer): The model uses Fourier transform (FFT) to analyze the frequency of time series data, fully exploiting the frequency domain features in the data.

[0091] The frequency domain processing method has the following advantages in dealing with mountain torrent time series prediction: first, frequency threshold division: by setting threshold_freq to distinguish low frequency and high frequency components, the model can more flexibly adjust the dependence on low frequency and high frequency information according to the frequency characteristics, and accurately capture the key moment of mountain flood occurrence. The second is the separation of low frequency and high frequency: the smoothing processing of low frequency part can reduce the influence of noise and improve the learning ability of the model to trend data. The detail extraction of high frequency part enhances the sensitivity of the model to sudden floods, local precipitation and the subtle fluctuations of flow changes, greatly improving the short-term warning ability.

[0092] 4-2, Low frequency processing module (low_processor): a fluctuation feature enhancement module is used to extract deep features from low frequency signals.

[0093] Specifically, the separated low frequency part mainly reflects the trend characteristics of time series signals in small and medium-sized river basins, and the change rate is smaller than that of high frequency part. Therefore, a "fluctuation feature enhancement module" is proposed for low frequency part, which uses multi-scale fluctuation decomposition and fluctuation enhancement mechanism to better capture the subtle dynamic features and trend features in the smooth characteristics. The structure is as shown in the following figure: Figure 3 The fluctuation feature enhancement module processes the input signal using an average pool 1d average pooling with a convolution kernel size of 3 and a fluctuation enhancer fluctuation separation enhancement unit to obtain the fundamental frequency part and the enhanced fluctuation part. The fundamental frequency part is processed using a convolution network with a convolution kernel size of 1 to retain the trend of the original fundamental frequency part:

[0094] Y F = GELU (Conv1D 1×1AvgPool1D 3×1 (X)));

[0095] wherein AvgPool1D 3×1 is the global average pooling; GELU is the activation function; Conv1D 1×1 is the convolution network with the kernel size of 1.

[0096] Fluctuation Enhancer is used to obtain the fluctuation part by subtracting the signal after the Avgpool1d average pooling from the input signal, and then the fluctuation part is enhanced by BoostBlock,

[0097] Y E = BoostBlock(X-AvgPool1D 3×1 (X))+X;

[0098] wherein Boost Block is a feature enhancement module containing depth separable convolution and channel expansion-shrinkage design, and GELU activation function is introduced to increase nonlinearity.

[0099] The output results of the fundamental frequency part and the Fluctuation Enhancer are spliced in the channel dimension to obtain the final output of the fluctuation feature enhancement module:

[0100] Y FEM = Concat([Y F ,Y E ], dim = 1);

[0101] wherein Concat represents channel dimension splicing.

[0102] 4-3, high frequency processing module (high_processor): a multi-path differential feature extraction module is used to process high frequency signals to capture dynamic changes and burst features in time series data.

[0103] Specifically, the differential convolution technology is used for the extraction of high frequency features of burst new flood time series data, and a multi-path differential feature extraction module (Multi-path Differential Feature Extraction Model, DEM for short) is proposed, the core structure of which is shown in Figure 4 As shown in the figure, the multi-dimensional differential convolution fusion framework first uses central difference convolution. This operator reparameterizes the standard convolution kernel W conv ∈r Cout×Cin×KA center-neighbor contrast mechanism is constructed to correct the weight of the center position of the convolution kernel. This change-sensitive weight adjustment can capture rapidly changing positions. The second difference convolution is a channel difference convolution. A cross-channel weight coupling mechanism is introduced to establish the difference relationship between channels. This exchange of channel weights can offset the convolution kernel weights and form a difference response between channels to capture the relationship patterns of the input signal in different channels. The third difference convolution is a time gradient difference convolution used to calculate the first or second order time gradient of time series data, which helps to extract the instantaneous rate of change of time series data. The fourth difference convolution is a weight difference convolution. A symmetric difference structure is formed by keeping the first weight of the convolution kernel unchanged and inverting the last weight value to enhance the difference of local abrupt features. The fifth standard convolution preserves the original convolution kernel to maintain the baseline feature extraction capability. The w1, w2, w3, w4 and w5 outputs of the four difference convolutions and one standard convolution are added to obtain the final w, and b1, b2, b3, b4 and b5 are added to obtain the final b. A 1D convolution layer is input to integrate the above features:

[0104] W cd = W[:,:,c] - theta * sum k≠c W[:,:,k];

[0105] W chd = W - theta * shift(W);

[0106] W td = W[:,:,[-0.5,0,0.5]];

[0107] W wd [:,:,0] = W[:,:,0], W wd [:,:,2] = -W[:,:,2];

[0108] W SC = W conv ;

[0109] Y MTDC = CELU(Conv1D(x,W cd +W chd +W td +W wd +W sc ,B cd +B chd +B td +B wd +B sc ));

[0110] where k is the convolution kernel size, which is 3 in this embodiment; c is the convolution kernel center position index, which enhances the sensitivity of the operation enhancement model to local feature mutations; shift represents a one-bit right circular shift in the variable dimension of W (equivalent to the previous variable of W affecting the current variable); θ is the degree coefficient, which dynamically adjusts the information interaction intensity between channels, which is 1 in this embodiment.

[0111] 4-4, self-perception attention (SPRA): the SPRA mechanism of this model further improves the adaptive ability to the basin hydrological data.

[0112] Specifically, after the small and medium-sized basin flood time series data is extracted by the special convolution module described above, it enters the self-perception attention module SPRA. This module, through the multi-modal attention collaborative mechanism, shows significant advantages in enhancing feature representation ability, computational efficiency and adaptability. Its structure is shown in Figure 5 , which includes spatial attention, channel attention, scaled dot-product attention and fusion attention four sub-modules.

[0113] Spatial attention (Spatial Attention) calculates the mean and maximum along the channel dimension, splices them, and then generates a spatial attention map S through 7x1 convolution (Padding=3) to capture key spatial positions:

[0114] X avg =MeanPool(X,dim=1)∈R B×1×L ;

[0115] X max =MaxPool(X,dim=1)∈R B×1×L ;

[0116] X cat =Concat([X avg ,X max ],dim=1)∈R B×1×L ;

[0117] S attn =Sigmoid((Conv1D k＝7 (X cat ));

[0118] Where MeanPool is global average pooling; MaxPool is global maximum pooling; Concat concatenates in the channel dimension.

[0119] The channel attention part generates channel attention C through two layers of 1x1 convolution of the bottleneck structure (dimension reduction rate R=8) after global average pooling, which models the inter-channel dependency while reducing the parameter quantity to 1 / 8 of the original channel number, reducing the risk of overfitting:

[0120] X avg = AdaptiveAvgPool1D(X) ∈ R B×C×1 ;

[0121] X max = AdaptiveMaxPool1D(X) ∈ R B×C×1 ;

[0122] A avg = W2(ReLU(W1X avg )) ∈ R B×C ;

[0123] A max = W2(ReLU(W1X max )) ∈ R B×C ;

[0124] C attn = Sigmoid((A avg + A max )) ∈ R B×C×1 ;

[0125] where AdaptiveAvgPool1D is adaptive global average pooling; AdaptiveMaxPool1D is adaptive global max pooling; W1 is a shrinkage fully connected network; W2 is a restoration fully connected network.

[0126] Scaled dot-product attention takes the input F transposed and projected into Query, Key, Value, computes the self-attention matrix and weighted aggregation, outputs the global context feature:

[0127] X emb = TokenEmbedding((C attn ⊙ X + S attn ⊙ X + X) T ) + PositionEmbedding(L) ∈ R B×L×D ;

[0128] Q = X emb W Q , K = X emb W K , V = X emb W V ;

[0129]

[0130] D attn = FFN(LayerNorm(X emb + Attention)) ∈ R B×L×D ;

[0131] D attn =Projection(D attn )∈R B×L×C ;

[0132] Where TokenEmbedding refers to the embedding layer, PositionEmbedding is the position encoding, H is the number of attention heads, D is the dimension of the mapped features, FFN is the feed-forward neural network, Projection is the linear layer

[0133] The fusion attention splices the original input X with the global attention feature D, generates the final attention output P through grouped convolution (Groups=C) and Sigmoid activation function, realizes adaptive weighted fusion of local details and global context, and avoids information loss:

[0134] X concat =Concat([X,D attn ],dim=1)∈R B×2C×L ;

[0135] P=SiLU((GroupConv1D 7×1 (X concat )))∈R B×C×L ;

[0136] Where Concat refers to splicing in the channel dimension, SiLU is the activation function, and GroupConv1D 7×1 is the grouped convolution.

[0137] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting disaster-causing factors of rainstorm and flood based on modal decomposition and deep learning, characterized in that, The method comprises the following steps: Collecting flow, flow rate and water level data of upstream telemetry sites as input and flow and water level data of downstream disaster sites as output; Preprocessing the collected data, including normalization processing, data set division and data format conversion; Optimizing the variational mode decomposition parameters by using the frost and ice optimization algorithm, adjusting the decomposition parameters through the soft frost search strategy, the hard frost puncture strategy and the forward greedy mechanism, and adaptively decomposing the input data to obtain sub-sequences containing different frequency scales; Building a Fourier transform high and low frequency feature enhanced attention network, performing Fourier transform on the decomposed sub-sequences, dividing the low frequency components and the high frequency components according to a preset frequency threshold, processing the low frequency components and the high frequency components respectively, and obtaining low frequency features and high frequency features; Fusing the low frequency features and the high frequency features, adaptively weighting through self-perception attention, and outputting the hydrological state prediction result of the downstream disaster site; The self-perception attention comprises spatial attention, channel attention, scaled dot-product attention and fusion attention; The spatial attention calculates the mean and the maximum along the channel dimension, generates a spatial attention map through 7x1 convolution after splicing, and captures the key spatial position of the fused features; The channel attention generates channel attention through two layers of 1x1 convolution of the bottleneck structure after global average pooling, models the inter-channel dependency of the fused features, reduces the parameter quantity to 1 / 8 of the original channel number, and reduces the risk of overfitting; The scaled dot-product attention projects the input F into query, key and value after transposition, calculates a self-attention matrix and weighted aggregation, and outputs the global context features of the aggregated fused features; The fusion attention splices the original input and the global attention features, generates the final attention output through grouped convolution and Sigmoid activation function, realizes adaptive weighting fusion of local details and global context, and avoids information loss. 2.The method of claim 1, wherein, The specific process of preprocessing the collected data is as follows: The input feature normalization interval is [ The output feature normalization interval is [0, 1]; Divide the data set into a training set, a validation set and a test set according to the ratio of 7:1:2; Convert the data into a time series window form, input the past 32 time steps of historical data, and output the prediction result of the current 60 time steps; The input shape is: sample number, time step length, feature number; the output shape is: sample number, 60, 2. 3.The method of claim 1, wherein, The calculation formula of the soft frost search strategy is as follows: ； ； ； wherein, is the new position of the updated particle, The hard frost puncture strategy enables particles in the algorithm to exchange with each other, improves the convergence speed and the ability to jump out of local optimum of the algorithm, and the calculation formula is: denotes the i-th particle in the j-th snow agent; i denotes the i-th particle in the j-th snow agent; j denotes the i-th particle in the j-th snow agent; is the i-th particle in the best snow agent of the snow population; R is the i-th particle in the best snow agent of the snow population; j is the i-th particle in the best snow agent of the snow population; r 1 is a random number in the range of [-1, 1] that controls the direction of particle motion together with cos, and changes with the number of iterations, denotes the cosine value of the angle between the direction of particle motion and the best particle, which is used to control the direction of particle motion; β is an environmental factor that changes with the number of iterations to simulate the influence of the external environment, which is used to ensure the convergence of the algorithm; h is the adhesion degree, which is a random number in the range of [0, 1] that is used to control the distance between the centers of two snow particles; refers to exploratory information, which represents the search path information of the i-th particle in the j-th group; j refers to exploratory information, which represents the search path information of the i-th particle in the j-th group; i refers to exploratory information, which represents the search path information of the i-th particle in the j-th group; refers to learning information, which represents the learning path information of the i-th particle in the j-th group; j refers to learning information, which represents the learning path information of the i-th particle in the j-th group; i refers to learning information, which represents the learning path information of the i-th particle in the j-th group; denotes a random number in the range of [0, 1] that is used to introduce randomness in the algorithm and affect the behavior or decision of the particle; denotes an environmental factor, which is a parameter used to simulate the influence of the external environment; t is the current number of iterations; T is the maximum number of iterations of the algorithm; β The mathematical model of is a step function; [·] represents rounding; w The default value of is 5, which is used to control the number of segments of the step function; The forward greedy mechanism compares the updated fitness value of an individual with the value before updating to determine whether to replace the individual, and simultaneously replaces the solutions of the two individuals. ； in, This indicates the new position of the particle after the update. It is a frost-body population R The best cream agent in China j The position of each particle; This represents the normalized value of the current agent's fitness value, meaning that the first... i The probability of an ice agent being selected; r 3 is a value in the interval [ Random numbers within the range [1,1]; The method comprises the following steps:

4. A rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning, for implementing the method of any one of claims 1-3, characterized in that, A data acquisition module is configured to acquire flow, flow rate and water level data of upstream sites and to construct a hydrological time series data set; A preprocessing module is configured to preprocess the acquired data, including normalization processing, data set division and data format conversion; A signal decomposition module is configured to perform signal decomposition on the preprocessed data, and to optimize variational mode decomposition (VMD) parameters by using a frost and ice optimization algorithm, so as to perform optimal decomposition on different input features. ​ The prediction module predicts the sudden rainstorm and mountain flood in the disaster area by constructing a Fourier transform high-low frequency feature enhanced attention network. 5.The rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning according to claim 4, characterized in that, The Fourier transform high-low frequency feature enhanced attention network comprises the following structure: A frequency separation layer performs frequency analysis on the time series data through Fourier transform, and sets a threshold frequency to distinguish low frequency components and high frequency components; A low frequency processing module performs deep feature extraction on the low frequency signal through a fluctuation feature enhancement module FEM; A high frequency processing module processes the high frequency signal through a multi-path difference feature extraction module DEM to capture dynamic changes and sudden features in the time series data; A self-aware attention module is used to weight and fuse low frequency features and high frequency features. 6.The rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning according to claim 5, characterized in that, The fluctuation feature enhancement module processes the input signal through an Avgpool1d average pooling with a convolution kernel size of 3 and a Fluctuation Enhancer fluctuation separation enhancement unit to obtain a fundamental frequency part and an enhanced fluctuation part; the fundamental frequency part is processed by a convolution network with a convolution kernel size of 1 to retain the trend of the original fundamental frequency part; The Fluctuation Enhancer fluctuation separation enhancement unit obtains the fluctuation part by subtracting the signal after Avgpool1d average pooling from the input signal, and then the fluctuation part is processed through Boost Block feature enhancement; The Boost Block is a feature enhancement module including a depth separable convolution and a channel expansion-shrinkage design, and a GELU activation function is introduced to increase nonlinearity; The fundamental frequency part and the output result of the Fluctuation Enhancer fluctuation separation enhancement unit are spliced in the channel dimension to obtain the final output of the fluctuation feature enhancement module. 7.The rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning according to claim 5, characterized in that, The multi-path difference feature extraction module includes a center difference convolution, a channel difference convolution, a time gradient difference convolution, a weight difference convolution and a standard convolution. 8.The rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning according to claim 7, characterized in that, The center difference convolution is realized by reparameterizing a standard convolution kernel A center-neighbor contrast mechanism is constructed to correct the weight of the center position of the convolution kernel, wherein, is a three-dimensional convolution kernel weight matrix; is the number of output channels; is the number of input channels; is the spatial dimension of the convolution kernel; The channel difference convolution establishes the difference relationship between channels by introducing a cross-channel weight coupling mechanism, and the operation of exchanging channel weights offsets the convolution kernel weight, forms the difference response between channels to capture the relationship mode of the input signal in different channels; The time gradient difference convolution is used to calculate the first or second time gradient of the time series data; The weight difference convolution forms a symmetric difference structure by keeping the first weight of the convolution kernel unchanged and inverting the last weight value; The standard convolution retains the original convolution kernel and maintains the baseline feature extraction capability; After adding the output features of each path, a 1d convolution layer is input to integrate and obtain high frequency processing features. 9.The rainstorm flood disaster-causing element prediction system based on modal decomposition and deep learning of claim 5, wherein, The self-aware attention module includes four sub-modules: spatial attention, channel attention, scaled dot-product attention and fusion attention; The spatial attention calculates the mean and maximum along the channel dimension, splices them, and generates a spatial attention map through a 7x1 convolution to capture the key spatial position of the fused features; The channel attention part generates channel attention through two 1x1 convolutions in a bottleneck structure after global average pooling, models the inter-channel dependency of the fused features, reduces the parameter quantity to 1 / 8 of the original channel number, and reduces the risk of overfitting; The scaled dot-product attention projects the input F transposed into query, key, value, calculates a self-attention matrix, and performs weighted aggregation, and outputs a global context feature after fusion and aggregation of the features; The fusion attention splices the original input and the global attention feature, generates the final attention output through grouped convolution and a sigmoid activation function, realizes adaptive weighted fusion of local details and global context, and avoids information loss.

Images