Modal decomposition and deep learning-based rainstorm torrential flood disaster-causing element prediction method and system

Abstract

The invention discloses a rainstorm torrential flood disaster-causing element prediction method and system based on modal decomposition and deep learning, and solves the problems that a traditional model is insufficient in non-linear time sequence feature capture, and a physical model depends on complex data and is weak in generalization ability. Comprising the steps of collecting flow, flow velocity and water level data of an upstream site as input, and taking downstream disaster point data as output; preprocessing the data; a frost ice optimization algorithm is adopted to optimize variational mode decomposition parameters; a Fourier transform high and low frequency feature enhanced attention network is constructed, low-frequency and high-frequency components are divided, trend features are extracted through a fluctuation enhancement module, dynamic changes are captured through a multi-path difference calculation unit, self-perception attention is introduced to achieve feature weighted fusion and long-term memory, and a downstream hydrological state prediction result is output. By optimizing a modal decomposition and deep learning cooperation mechanism, the precision, robustness and generalization ability of sudden mountain torrent prediction are significantly improved, and the method is suitable for disaster emergency management in complex scenes of small and medium watersheds.

Description

Technical Field

[0001] The present invention relates to the technical field of flood disaster prediction, and in particular to a method and system for predicting rainstorm and flash flood disaster factors based on modal decomposition and deep learning. Background Art

[0002] Floods are among the most severe natural disasters posing a threat to human life and property. Flash floods in small and medium-sized river basins, characterized by their suddenness, destructive power, and short warning times, pose a significant challenge in flood prevention and disaster reduction. Currently, early warning and forecasting technologies for flash floods in small and medium-sized river basins primarily rely on traditional linear regression models and hydrodynamic models. Due to the linear nature of their assumed relationships, traditional linear regression models struggle to effectively capture the complex and nonlinear dynamics of flash flood data, preventing them from providing accurate predictions. Constructing hydrodynamic models typically requires integrating complex, heterogeneous data from multiple sources, including water regime information, underlying surface properties, and river channel cross-sections. This process is time-consuming and requires a high level of technical expertise, hindering widespread adoption nationwide. Furthermore, hydrodynamic models are computationally expensive and time-consuming to simulate and predict, making them inadequate for the rapid response and real-time decision-making required in emergency response scenarios.

[0003] At present, emergency telemetry alarm station projects have been built in several pilot areas, and telemetry alarm stations have been deployed at key locations in upstream and downstream rivers to achieve real-time, accurate, dynamic and minute-level monitoring of water information. A certain level of flood information database for small and medium-sized river basins has been established.

[0004] Now we need to design a model that uses the established flood information database of small and medium-sized river basins as a driver, which can accurately capture the nonlinear relationship between the upstream water situation information and the water situation information of the downstream disaster-stricken areas 40 minutes, 60 minutes or even several hours later, and realize 40 minutes, 60 minutes or even several hours of early warning and forecast of sudden rainstorms and flash floods in the downstream disaster-stricken areas. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting rainstorm and flash flood disaster factors based on modal decomposition and deep learning, so as to solve the technical problems that traditional models are unable to capture nonlinear time series characteristics, physical models rely on complex data and have weak generalization capabilities.

[0006] To achieve the above objectives, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting rainstorm and flash flood disaster factors based on modal decomposition and deep learning, comprising:

[0008] Collect flow, velocity, and water level data from upstream telemetry stations as input, and flow and water level data from downstream disaster sites as output;

[0009] Preprocess the collected data, including normalization, data set division, and data format conversion;

[0010] The frost optimization algorithm is used to optimize the variational mode decomposition parameters. The decomposition parameters are adjusted through the soft frost search strategy, hard frost puncture strategy and forward greedy mechanism. The input data is adaptively decomposed to obtain subsequences containing different frequency scales.

[0011] Construct a Fourier transform high- and low-frequency feature enhanced attention network, perform Fourier transform on the decomposed subsequence, divide the low-frequency component and high-frequency component according to the preset frequency threshold, process the low-frequency component and high-frequency component respectively, and obtain low-frequency features and high-frequency features;

[0012] The low-frequency and high-frequency features are integrated and adaptively weighted through self-aware attention to output the hydrological status prediction results of the downstream disaster-stricken points.

[0013] Furthermore, 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 training set, validation set and test set in a ratio of 7:1:2;

[0016] Convert the data into a time series window format, with the input being the historical data of the past 32 time steps and the output being the forecast results of the current 60 time steps;

[0017] The input shape is: number of samples, time steps, number of features; the output shape is: number of samples, 60, 2.

[0018] Furthermore, the calculation formula of the soft cream search strategy is as follows:

[0019]

[0020] in, is the new position of the updated particle, ij represents the jth particle in the i-th frost agent; is the jth particle of the best frost agent in the frost population R; parameter r1 is a random number in the range [-1, 1], which controls the direction of particle movement together with cos and changes with the number of iterations; cosθ represents the cosine value of the angle between the particle movement direction and the best particle, which is used to control the direction of particle movement; β is an environmental factor that 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, a random number in the range [0, 1], which is used to control the distance between the centers of two frost particles; b UijRefers to exploratory information, which represents the search path information of the jth particle in the i-th population; b Lij Refers to learning information, indicating the learning path information of the jth particle in the i-th population; r2 represents a random number in the range [0,1], which is used to introduce randomness into the algorithm to affect the behavior or decision-making of particles; E represents the environmental factor, which is a parameter used to simulate the influence of the external environment; 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, which is a random number in the range [0,1] and is used to control the distance between the centers of two frost particles; t is the current iteration number; T is the maximum number of iterations of the algorithm; the mathematical model of β is a step function; [·] indicates rounding; the default value of w is 5, which is used to control the number of segments of the step function;

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

[0022]

[0023] in, represents the new position of the particle after the update, is the position of the jth particle of the best frost agent in the frost population R; F normr The normalized value of the current agent fitness value, which means the probability of the i-th ice agent being selected; r3 is a random number in the interval [-1,1];

[0024] The forward greedy mechanism determines whether to replace an individual by comparing the updated fitness value of the individual with the value before the update, and replaces the solutions of the two individuals at the same time.

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

[0026] The data acquisition module is used to collect flow, flow velocity, and water level data from upstream sites and construct a hydrological time series dataset;

[0027] The preprocessing module is used to preprocess the collected data, including normalization, data set division and data format conversion;

[0028] The signal decomposition module is used to decompose the preprocessed data and optimize the variational mode decomposition (VMD) parameters using the frost optimization algorithm to achieve the optimal decomposition for different input features.

[0029] The prediction module predicts sudden rainstorms and flash floods in the disaster-stricken areas by constructing a Fourier transform high and low frequency feature enhanced attention network.

[0030] Furthermore, the structure of the Fourier transform high- and low-frequency feature enhanced attention network includes:

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

[0032] The low-frequency processing module performs deep feature extraction on the low-frequency signal through the fluctuation feature enhancement module FEM;

[0033] The high-frequency processing module processes high-frequency signals through the multi-path difference feature extraction module DEM to capture dynamic changes and sudden features in time series data;

[0034] Self-aware attention is used to weightedly fuse low-frequency and high-frequency features.

[0035] Furthermore, the fluctuation feature enhancement module performs Avgpool1d average pooling with a convolution kernel size of 3 and FluctuationEnhancer fluctuation separation enhancement unit processing on the input signal 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 original trend of the fundamental frequency part;

[0036] The Fluctuation Enhancer subtracts the average pooled signal from the input signal, obtaining the fluctuating portion. This portion is then enhanced using the BoostBlock feature. Boost Block is a feature enhancement module that includes depthwise separable convolution and channel expansion-contraction design, and introduces the GELU activation function to increase nonlinearity.

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

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

[0039] Furthermore, the central difference convolution is performed by re-parameterizing the standard convolution kernel W conv ∈R Cout×Cin×K Construct a center-neighborhood comparison mechanism to modify the weight of the convolution kernel center position; where W conv is the 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 comparison mechanism is constructed to perform weight correction on the center position of the convolution kernel;

[0040] The channel differential convolution introduces a cross-channel weight coupling mechanism to establish a difference relationship between channels. The operation of exchanging channel weights offsets the convolution kernel weights to form a difference response between channels to capture the relationship pattern of the input signal on different channels.

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

[0042] The weighted 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 adding the output features of each path, a 1D convolutional layer is input to integrate the above features.

[0045] Furthermore, the self-aware attention module includes four submodules: 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 a 7×1 convolution to capture the key spatial positions.

[0047] After global average pooling, the channel attention part generates channel attention through two layers of 1×1 convolution in a bottleneck structure, modeling the dependency between channels while reducing the number of parameters to 1 / 8 of the original number of channels, reducing the risk of overfitting;

[0048] The scaled dot product attention transposes the input F and projects it into Query, Key, and Value, calculates the self-attention matrix and performs weighted aggregation to output global context features;

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

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

[0051] (1) The present invention combines a frosting optimization algorithm that can automatically adjust parameters to achieve the optimal decomposition scheme based on the characteristics of different input features. The purpose of the variational mode decomposition module is to deeply explore the potential features in the water regime data of small and medium-sized watersheds, especially for water regime data with complex time-varying characteristics, to ensure that the decomposed signal has high usability. This processing not only provides more accurate input data for the subsequent deep learning module, but also lays a solid foundation for the subsequent model to capture the nonlinear spatiotemporal relationship in the water regime data.

[0052] (2) Aiming at the characteristics of time series signals of sudden mountain torrents in small and medium-sized watersheds, the present invention proposes a Fourier transform high- and low-frequency feature enhanced attention network. This model is a deep learning network based on Fourier transform and multi-scale feature extraction, and its design integrates deep processing of time domain and frequency domain information. The core innovation of the model lies in the use of Fourier transform for frequency separation (low frequency and high frequency) and the innovative modules multi-path difference calculation unit (DEM), fluctuation enhancement module (FEM), and self-perception attention (SPRA) for high and low frequency signals respectively, which effectively capture and enhance the main trends and detailed features in time series signals. This multi-level processing method is particularly suitable for the suddenness, nonlinearity and variability in time series data, such as the prediction of sudden mountain torrents in small and medium-sized watersheds.

[0053] (3) The present invention organically combines optimized variational mode decomposition with deep learning: through the seamless integration of optimization algorithm (RIME), data preprocessing (VMD) and efficient intelligent prediction algorithm process, it can better capture the characteristics of flood time series processes in small and medium-sized watersheds, so that the new algorithm has good adaptability to multi-dimensional and multi-type emergency telemetry station perception data, and is applicable to different spatial scales and complex watershed underlying surfaces such as small and medium-sized watersheds, mountain torrents, sub-basins / sub-catchment areas, etc., which enhances the robustness of the model to noise and abnormal data, improves the overall performance of mountain torrent disaster factor prediction, and has important practical application value for mountain torrent disaster emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0055] Figure 1 This is the decomposition diagram after RIME optimization of VMD in the present invention;

[0056] Figure 2 It is an architectural diagram of the deep learning model of the present invention;

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

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

[0059] Figure 5 It is a workflow diagram of the SPRA self-perception attention of the present invention. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0061] The method for predicting factors causing rainstorm and flash flood disasters based on modal decomposition and deep learning includes the following steps:

[0062] Step 1: Construct the dataset: collect the flow, velocity, and water level data of 12 telemetry stations upstream as input, and the flow and water level data of the downstream disaster site as output;

[0063] Specifically, this dataset is used to model and predict hydrological time series. Its input variable X and output variable Y represent the hydrological observation data of the upstream telemetry station and the hydrological response data of the downstream disaster site, respectively. The data preprocessing and construction process is as follows:

[0064] Input features (X): Contains observation data from 12 upstream telemetry alarm stations. The characteristics of each station include: discharge (Q), velocity (V), and water level (H).

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

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

[0067] 2-1. Normalization: The input feature (X) is normalized to the range [-1, 1] using MinMaxScaler to enhance model stability.

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

[0069] 2-2. Dataset division: training set (70%), validation set (10%), and test set (20%) to ensure that the model is fully learned during the training phase and generalized on the test set.

[0070] 2-3. In order 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 of each sample consists of 32 consecutive time steps, that is, the model can observe the historical data of the past 32 time steps (each time step is 5 minutes) of the twelve upstream stations for prediction.

[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 steps, number of features) = (N, 32, 36);

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

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

[0077] The target output is: Y 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 sites affects future hydrological conditions downstream.

[0079] Step 3: Optimize the variational mode decomposition (VMD): Figure 1 As shown, the frost ice optimization algorithm (RIEM) is used to optimize the variational mode decomposition parameters, and the input data is adaptively decomposed to obtain subsequences containing different frequency scales;

[0080] The frost optimization algorithm adjusts the decomposition parameters through a soft frost search strategy, a hard frost puncture strategy, and a forward greedy mechanism to achieve efficient signal decomposition. The algorithm process includes:

[0081] 3-1. Soft Frost Search Strategy: Simulating the movement of soft frost particles in frost ice, a step-by-step search and development method is proposed for the exploration algorithm, allowing the algorithm to switch between large-scale exploration and small-scale development, thereby achieving efficient and high-precision optimization:

[0082]

[0083] in, is the new position of the updated particle, ij represents the jth particle in the i-th 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 [-1, 1] that controls the particle's motion direction together with cos and changes with the number of iterations; β is an environmental factor that 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, a random number in the range [0, 1] that controls the distance between the centers of two frost particles; t is the current iteration number; T is the maximum number of iterations of the algorithm; the mathematical model of β is a step function; [·] indicates rounding; the default value of w is 5 and is used to control the number of segments in the step function.

[0084] 3-2. Hard frost puncture strategy: enables particles in the algorithm to exchange with each other, improving the algorithm's convergence speed and ability to escape local optimality:

[0085]

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

[0087] 3-3. Forward Greedy Mechanism: This is used for population updates in metaheuristic optimization algorithms. This mechanism compares the updated fitness value of an individual with the value before the update to decide whether to replace the individual and simultaneously replaces the solutions of both individuals.

[0088] Step 4: Build a deep learning model;

[0089] The optimally decomposed data then enters the deep learning model. In view of the characteristics of the time series signals of sudden mountain torrents in small and medium-sized basins, the Fourier transform high and low frequency features are used to enhance the attention network. This model is a deep learning network based on Fourier transform and multi-scale feature extraction. The design integrates deep processing of time domain and frequency domain information. The model effectively captures and enhances the main trends and detailed features in the time series signal through frequency separation (low frequency and high frequency) and multi-scale convolutional networks. This multi-level processing method is particularly suitable for the suddenness, nonlinearity and variability in time series data, such as the prediction of sudden mountain torrents in small and medium-sized basins. Figure 2 Specifically, the structure of the model includes:

[0090] 4-1. Frequency Split Layer: This model performs frequency analysis on time series data through Fourier transform (FFT), fully mining the frequency domain characteristics in the data.

[0091] Frequency-domain processing methods offer several advantages for flash flood time-series forecasting: First, frequency thresholding: By setting threshold_freq to distinguish between low-frequency and high-frequency components, the model can more flexibly adjust its reliance on low-frequency and high-frequency information based on frequency characteristics, accurately capturing critical flash flood moments. Second, low-frequency-high-frequency separation: Smoothing low-frequency components reduces the impact of noise and improves the model's ability to learn trend data. Detail extraction of high-frequency components enhances the model's sensitivity to subtle fluctuations in sudden floods, localized precipitation, and flow changes, significantly improving short-term warning capabilities.

[0092] 4-2. Low-frequency processing module (low_processor): Use the fluctuation feature enhancement module to perform deep feature extraction on the low-frequency signal.

[0093] Specifically, the separated low-frequency part mainly reflects the trend characteristics of the time series signal of small and medium-sized watersheds compared with the high-frequency part, and the signal change rate is smaller than that of the high-frequency part. Therefore, a "fluctuation feature enhancement module" abbreviated as FEM is proposed for the low-frequency part. Through multi-scale fluctuation decomposition and fluctuation enhancement mechanism, the subtle dynamic characteristics and trend characteristics in its stable characteristics are better captured. Its structure is as follows Figure 3 The fluctuation feature enhancement module performs Avgpool1d average pooling with a convolution kernel size of 3 and Fluctuation Enhancer fluctuation separation enhancement unit processing on the input signal 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 original fundamental frequency part trend:

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

[0095] Among them, AvgPool1D 3×1 is the global average pooling; GELU is the activation function; Conv1D 1×1 It is a convolutional network with a convolution kernel size of 1.

[0096] The Fluctuation Enhancer fluctuation separation enhancement unit obtains the fluctuation part by subtracting the signal after the average pooling of Avgpool1d from the input signal, and then the fluctuation part is enhanced by the BoostBlock feature.

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

[0098] Among them, Boost Block is a feature enhancement module that includes depth-wise separable convolution and channel expansion-contraction design, and introduces GELU activation function to increase nonlinearity.

[0099] The fundamental frequency part and the output of the Fluctuation Enhancer fluctuation separation and enhancement unit are concatenated 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] Among them, represents splicing in the Concat channel dimension.

[0102] 4-3. High-frequency processing module (high_processor): Use the multi-path difference feature extraction module to process high-frequency signals and capture dynamic changes and sudden features in time series data.

[0103] Specifically, the differential convolution technique is used to extract high-frequency features of sudden new mountain torrent time series data, and a multi-path differential feature extraction module (DEM) is proposed. Its core structure is as follows: Figure 4 As shown, the first step in the multi-dimensional differential convolution fusion framework is the central differential convolution, which reparameterizes the standard convolution kernel W conv ∈r Cout×Cin×KA center-neighborhood comparison mechanism is constructed to modify the weight of the center position of the convolution kernel. This change-sensitive weight adjustment can capture rapidly changing positions. The second differential convolution is a channel differential convolution. By introducing a cross-channel weight coupling mechanism to establish a difference relationship between channels, this operation of exchanging channel weights can offset the convolution kernel weights, forming a differential response between channels to capture the relationship pattern of the input signal on different channels. The third differential convolution is a time gradient differential 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 differential convolution is a weight differential convolution. By keeping the first weight of the convolution kernel unchanged and inverting the last weight value, a symmetric differential structure is formed to enhance the difference in local mutation features. The fifth standard convolution retains the original convolution kernel to maintain the baseline feature extraction capability. The outputs w1, w2, w3, w4, and w5 of the four differential convolutions and a 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]-θ∑ k≠c W[:,:,k];

[0105] W chd =W-θ·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 size of the convolution kernel, which is 3 in this embodiment; c is the index of the center position of the convolution kernel, which enhances the model's sensitivity to local feature mutations; shift means moving one position to the right 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 intensity of information interaction between channels, which is 1 in this embodiment.

[0111] 4-4. Self-Perceived Attention (SPRA): The SPRA mechanism of this model further improves the adaptability to watershed hydrological data.

[0112] Specifically, after the feature extraction of the time series data of small and medium-sized river basins by the special convolution module, it enters the self-perception attention module SPRA, which shows significant advantages in enhancing feature representation ability, computational efficiency and adaptability through the multimodal attention coordination mechanism. Its structure is as follows Figure 5 As shown, it contains four sub-modules: spatial attention, channel attention, scaled dot product attention, and fusion attention.

[0113] Spatial Attention: Calculate the mean and maximum values ​​along the channel dimension, concatenate them, and generate a spatial attention map S through a 7×1 convolution (Padding=3) to capture key spatial locations:

[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] Among them, MeanPool is the global average pooling; MaxPool is the global maximum pooling; Concat is splicing in the channel dimension.

[0119] After global average pooling, the channel attention part generates channel attention C through two layers of 1×1 convolution in a bottleneck structure (dimensionality reduction rate R=8). This modeling of the inter-channel dependency reduces the number of parameters to 1 / 8 of the original number of channels, 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] Among them, AdaptiveAvgPool1D is adaptive global average pooling; AdaptiveMaxPool1D is adaptive global maximum pooling; W1 is a shrinking fully connected network; W2 is a restoring fully connected network.

[0126] The scaled dot product attention transposes the input F and projects it into Query, Key, and Value, calculates the self-attention matrix and performs weighted aggregation to output 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] Among them, TokenEmbedding refers to the embedding layer, PositionEmbedding is the position encoding, H is the number of attention heads; D is the mapped feature dimension; FFN is the feedforward neural network; Projection is the linear layer

[0133] The fusion attention concatenates the original input X with the global attention feature D, and generates the final attention output P through group convolution (Groups = C) and Sigmoid activation function, realizing the adaptive weighted fusion of local details and global context to avoid 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] Among them, Concat refers to splicing in the channel dimension; SiLU is the activation function; GroupConv1D 7×1 It is grouped convolution.

[0137] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting factors of rainstorm and flash flood disasters based on modal decomposition and deep learning, characterized by: include: Collect flow, velocity, and water level data from upstream telemetry stations as input, and flow and water level data from downstream disaster sites as output; Preprocess the collected data, including normalization, data set division, and data format conversion; The frost optimization algorithm is used to optimize the variational mode decomposition parameters. The decomposition parameters are adjusted through the soft frost search strategy, hard frost puncture strategy and forward greedy mechanism. The input data is adaptively decomposed to obtain subsequences containing different frequency scales. Construct a Fourier transform high- and low-frequency feature enhanced attention network, perform Fourier transform on the decomposed subsequence, divide the low-frequency component and high-frequency component according to the preset frequency threshold, process the low-frequency component and high-frequency component respectively, and obtain low-frequency features and high-frequency features; The low-frequency and high-frequency features are integrated and adaptively weighted through self-aware attention to output the hydrological status prediction results of the downstream disaster-stricken points.

2. The method for predicting rainstorm and flash flood disaster factors based on modal decomposition and deep learning according to claim 1 is characterized in that: The specific process of preprocessing the collected data is as follows: The input feature normalization interval is [-1, 1], and the output feature normalization interval is [0, 1]; The dataset is divided into training set, validation set and test set in a ratio of 7:1:2; Convert the data into a time series window format, with the input being the historical data of the past 32 time steps and the output being the forecast results of the current 60 time steps; The input shape is: number of samples, time steps, number of features; the output shape is: number of samples, 60, 2.

3. The method for predicting rainstorm and flash flood disaster factors based on modal decomposition and deep learning according to claim 1 is characterized in that: The calculation formula of the soft cream search strategy is as follows: in, is the new position of the updated particle, ij represents the jth particle in the i-th frost agent; is the jth particle of the best frost agent in the frost population R; parameter r1 is a random number in the range [-1, 1], which controls the direction of particle movement together with cos and changes with the number of iterations; cosθ represents the cosine value of the angle between the particle movement direction and the best particle, which is used to control the direction of particle movement; β is an environmental factor that 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, a random number in the range [0, 1], which is used to control the distance between the centers of two frost particles; b Uij Refers to exploratory information, which represents the search path information of the jth particle in the i-th population; b Lij Refers to learning information, indicating the learning path information of the jth particle in the i-th population; r2 represents a random number in the range [0,1], which is used to introduce randomness into the algorithm and affect the behavior or decision-making of the particle; E represents the environmental factor, which is a parameter used to simulate the influence of the external environment; t is the current iteration number; T is the maximum iteration number of the algorithm; the mathematical model of β is a step function; [·] indicates rounding; the default value of w is 5, which is used to control the number of segments of the step function; The frost penetration strategy enables the particles in the algorithm to exchange with each other, improving the convergence speed of the algorithm and the ability to escape the local optimum. The calculation formula is: in, represents the new position of the particle after the update, is the position of the jth particle of the best frost agent in the frost population R; F normr The normalized value of the current agent fitness value, which means the probability of the i-th ice agent being selected; r3 is a random number in the interval [-1,1]; The forward greedy mechanism determines whether to replace an individual by comparing the updated fitness value of the individual with the value before the update, and replaces the solutions of the two individuals at the same time.

4. A rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning, used to implement the method described in any one of claims 1 to 3, characterized in that: include: The data acquisition module is used to collect flow, flow velocity, and water level data from upstream sites and construct a hydrological time series dataset; The preprocessing module is used to preprocess the collected data, including normalization, data set division and data format conversion; The signal decomposition module is used to decompose the preprocessed data and optimize the variational mode decomposition (VMD) parameters using the frost optimization algorithm to achieve the optimal decomposition for different input features. The prediction module predicts sudden rainstorms and flash floods in the disaster-stricken areas by constructing a Fourier transform high and low frequency feature enhanced attention network.

5. The rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning according to claim 4 is characterized in that: The structure of the Fourier transform high and low frequency feature enhanced attention network includes: The frequency separation layer performs frequency analysis on the time series data through Fourier transform and sets the threshold frequency to distinguish low-frequency components from high-frequency components; The low-frequency processing module performs deep feature extraction on the low-frequency signal through the fluctuation feature enhancement module FEM; The high-frequency processing module processes high-frequency signals through the multi-path difference feature extraction module DEM to capture dynamic changes and sudden features in time series data; Self-aware attention is used to weightedly fuse low-frequency and high-frequency features.

6. The rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning according to claim 5 is characterized in that: The fluctuation feature enhancement module performs Avgpool1d average pooling with a convolution kernel size of 3 and FluctuationEnhancer fluctuation separation enhancement unit processing on the input signal 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 original fundamental frequency part trend; The Fluctuation Enhancer fluctuation separation enhancement unit obtains the fluctuation part by subtracting the signal after the average pooling of Avgpool1d from the input signal, and then enhances the fluctuation part through the BoostBlock feature; The Boost Block is a feature enhancement module that includes depthwise separable convolution and channel expansion-contraction design, and introduces the GELU activation function to increase nonlinearity; The fundamental frequency part and the output result of the Fluctuation Enhancer fluctuation separation and enhancement unit are feature spliced ​​in the channel dimension to obtain the final output of the fluctuation feature enhancement module.

7. The rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning according to claim 5 is characterized in that: The multi-path difference feature extraction module includes multi-path fusion of center difference convolution, channel difference convolution, time gradient difference convolution, weight difference convolution and standard convolution.

8. The rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning according to claim 7 is characterized in that: The central difference convolution is performed by reparameterizing the standard convolution kernel W conv ∈R Cout×Cin×K Construct a center-neighborhood comparison mechanism to modify the weight of the convolution kernel center position; where W conv is the three-dimensional convolution kernel weight matrix; Cout is the number of output channels; Cin is the number of input channels; K is the convolution kernel spatial dimension; The channel differential convolution introduces a cross-channel weight coupling mechanism to establish a difference relationship between channels. The operation of exchanging channel weights offsets the convolution kernel weights to form a difference response between channels to capture the relationship pattern of the input signal on different channels. The time gradient differential convolution is used to calculate the first-order or second-order time gradient of time series data; The weighted differential convolution forms a symmetric differential 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 convolutional layer is input to integrate the above features.

9. The rainstorm and flash flood disaster factor prediction system based on modal decomposition and deep learning according to claim 5 is characterized in that: The self-aware attention module includes four submodules: spatial attention, channel attention, scaled dot product attention, and fusion attention; The spatial attention calculates the mean and maximum values ​​along the channel dimension, concatenates them, and generates a spatial attention map through a 7×1 convolution to capture the key spatial positions. After global average pooling, the channel attention part generates channel attention through two layers of 1×1 convolution in a bottleneck structure, modeling the dependency between channels while reducing the number of parameters to 1 / 8 of the original number of channels, reducing the risk of overfitting; The scaled dot product attention transposes the input F and projects it into query, key, and value, calculates the self-attention matrix and performs weighted aggregation to output the global context feature; The fused attention concatenates the original input with the global attention feature, generates the final attention output through grouped convolution and Sigmoid activation function, and realizes the adaptive weighted fusion of local details and global context to avoid information loss.

Images