Flood forecasting method based on multi-modal intelligent algorithm

Abstract

The invention discloses a flood forecasting method based on a multi-modal intelligent algorithm, which belongs to the technical field of hydrological forecasting, and is technically characterized by comprising the following steps: S1, collecting historical rainfall and flow data of a drainage basin, processing data missing values, and forming continuous data through a small section of runoff; s2, using a TVFEMD decomposition method to decompose the flood data into a plurality of relatively simple intrinsic mode functions; s3, calculating an entropy value according to the sample entropy, and aggregating and reconstructing IMFs in the same entropy value interval; s4, dividing the data set into a training set and a test set, and extracting data features by using CNN to obtain a multi-dimensional matrix; s5, inputting the feature matrix extracted by the CNN into the BiLSTM model, and optimizing hyper-parameters of the BiLSTM model by using MWOA; and S6, predicting each component subjected to TVFEMD-SE decomposition and aggregation, and superposing predicted values of each component to obtain a flood predicted value, thereby effectively solving the problems of flood sequence multi-scale feature extraction and noise robustness, and remarkably improving the model convergence speed and the prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological prediction, and in particular to a flood forecasting method based on a multimodal intelligent algorithm. Background Art

[0002] Hydrological forecasting technology is a core component of water resources management and disaster prevention and mitigation. Traditional methods are mostly based on physical models (such as hydrodynamic and hydrological models) and statistical models. Their accuracy is limited by simplifying assumptions in parameterization processes and the complex nonlinear characteristics of data. In recent years, while machine learning methods (such as support vector machines (SVMs) and random forests (RFs)) have made some progress in runoff forecasting, they still have limitations in modeling coupled spatiotemporal features and capturing long-term dependencies. Deep learning technology, with its powerful nonlinear mapping capabilities, provides a new approach for extracting high-dimensional features from hydrological time series data. Convolutional neural networks (CNNs) can effectively extract spatial distribution characteristics of rainfall and runoff, while long short-term memory networks (LSTMs) excel at characterizing the temporal dynamics of flood processes.

[0003] However, a single model cannot take into account the synergistic effects of spatiotemporal characteristics, and hydrological data are often affected by noise and non-stationarity, which limits the generalization ability of the model. To this end, signal decomposition techniques (such as empirical mode decomposition (EMD) and variational mode decomposition (VMD)) have been introduced into the field of hydrology. By decomposing non-stationary sequences into multiple intrinsic mode functions, the data complexity is effectively reduced. However, traditional decomposition methods have problems such as modal aliasing and endpoint effects, which limit their application in hydrological forecasting. In addition, neural network hyperparameter optimization is the key to improving model performance. Traditional methods (such as grid search and random search) are inefficient and prone to falling into local optimality. Although optimization algorithms based on swarm intelligence (such as particle swarm optimization and whale optimization algorithm) can find global optimality, they still have shortcomings in convergence speed and stability.

[0004] Therefore, the present invention proposes a hybrid deep learning flood forecasting method based on time-varying filter empirical mode decomposition (TVFEMD) and sample entropy signal reconstruction (SE). By integrating multimodal signal decomposition, adaptive feature reconstruction and a hybrid deep learning framework, combined with an efficient optimization strategy, a high-precision and robust hydrological forecasting method system is constructed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, an embodiment of the present invention aims to provide a flood forecasting method based on a multimodal intelligent algorithm to solve the problems in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A flood forecasting method based on a multimodal intelligent algorithm comprises the following steps: S1. Collect historical precipitation and flow data of the basin, process missing values, and connect different flood data into continuous data through a small section of runoff; S2, using TVFEMD decomposition method to decompose flood data into multiple relatively simple intrinsic mode functions (IMFs); S3, calculate the entropy value based on the sample entropy, aggregate and reconstruct the IMFs in the same entropy value range; S4. Divide the dataset into training and test sets, and use CNN to extract data features to obtain a multidimensional matrix; S5. Input the feature matrix extracted by CNN into the BiLSTM model and use MWOA to optimize the hyperparameters of the BiLSTM model. S6. Predict each component of TVFEMD-SE decomposition and aggregation separately, and superimpose the predicted values ​​of each component to obtain the flood forecast value.

[0007] As a further solution of the present invention, the input of the flood data in step 2 is Perform Hilbert transform to obtain analytical signal, and record the result , (1) (2) In the formula, and are respectively expressed as the instantaneous amplitude and instantaneous phase of the flood discharge series; Determine the local maximum and local minimum of the instantaneous amplitude, which are expressed as and Then, interpolate the maximum point set and the minimum point set to get the curve and , and Calculated according to formula (3) and (4) respectively;

[0008] (3) (4) right and Perform interpolation and find and , and then calculate the instantaneous frequency component according to equations (5) and (6): and ; (5) (6) The cutoff frequency is calculated according to formula (7); (7) By calculating , calculate the signal according to formula (8) , The extreme points of the time-varying filter are used to construct the nodes, and the B-spline interpolation is used to After further filtering, the approximate result is ; (8) When the stop frequency meets hour, That is IMF; when the stop frequency does not meet the conditions, that is , , then repeat the above steps, The calculation formula is as follows: (9) In the formula, is the instantaneous bandwidth of the two-component signal, is the weighted average of the instantaneous frequency of the single-component signal, the threshold Set to 0.1.

[0009] As a further solution of the present invention, the step of reconstructing the sample entropy signal into the SE aggregation reconstructed signal includes: Calculate the sample entropy, the calculation formula is as follows: (10) In the formula, m is the reconstruction dimension, r is the similarity tolerance, N is the length of flood data, is the probability that two sequences match m points under the similarity tolerance r.

[0010] As a further embodiment of the present invention, the step of extracting high-level features from the reconstructed signal by the CNN includes: Establish a convolution layer, set the appropriate convolution kernel size, extract the grid matrix as the feature map matrix and input it into the CNN model. In the convolution layer, the spatial features of the basin meteorological information are extracted by reading the matrix. c A filter is used to identify the input feature map matrix. According to formula (11), the matrix is ​​extracted and the weight matrix and bias of the filter are calculated. The filter is used to scan the feature map matrix. Formula (12) is used to convert D c The filters are convolved with the input matrix area they cover, and the final output is D c The feature matrix after convolution is F conv ; (11) (12) In the formula, A feature graph representing the input; Represents the feature graph output after convolution calculation; express The height of the characteristic figure; express The width of the feature pattern; express The height of the characteristic figure; express The width of the feature pattern; Indicates the number of convolution kernels; Represents the convolution kernel output result, which represents the calculated value at the height h and width w of the d-th convolution kernel matrix; Represents the weight value in the d-th convolution kernel filter, Represents the deviation value corresponding to the d-th convolution kernel filter, and n represents the convolution step size of the convolution kernel; Secondly, since the input matrix dimension is large, in order to reduce the optimization difficulty and the number of parameters, the convolution matrix is ​​downsampled in the pooling layer, which is also called pooling. The process is shown in formula (13). The maximum value method is used for pooling calculation, which plays the role of dimensionality reduction and principal component extraction for the input matrix. (13) (14) In the formula, is the pooling step size; Indicates the output result of pooling calculation; Represents the pooling layer feature matrix; The features output by the pooling layer are expanded into a one-dimensional matrix through the flatten layer, and the fully connected layer performs linear transformation to output the matrix Y that is consistent with the number of BiLSTM cells. out , this matrix Y out As the input information of the BiLSTM network, the BiLSTM cell layer is used to learn the time change process of the input data from front to back and from back to front, and the nonlinear relationship between meteorological data and hydrological data is established.

[0011] As a further solution of the present invention, the steps of adaptively optimizing the hyperparameters of the BiLSTM model using the MWOA optimization algorithm include: Set the two parameters of MWOA population size N and maximum number of iterations T, and generate the initial population in the search space according to Circle mapping. The Circle mapping calculation formula is as follows: (15) In the formula, mod(a, b) means mapping the value of a to [0, b); Calculate the fitness value of each individual in the population and record the optimal position X(t); The adaptive weights of the fusion beta distribution and the inverse incomplete Γ function are updated according to formula (16), and the parameter ω is updated according to formula (23) (24). 、 、 The specific formula is as follows: (16) (17) (18) In the formula: = 0.9, = 0.4; is the inverse of the incomplete Γ function Matlab calls the function; λ (λ⩾0) is a random variable; represents a random number that follows a Beta distribution, b1=1, b2=2; Inertia weight adjustment factor, used to control inertia weight The degree of deviation; and is a random number between [0, 1]; is a parameter that decreases from 2 to 0 as the number of iterations increases, defined as , is the current iteration number, is the maximum number of iterations of the algorithm; Compare the size of |A|, compare the randomly generated p value with 0.5, and select the corresponding position update formula. If p<0.5 and |A|<1, update the current individual position according to formula (19); if p<0.5 and |A|⩾1, randomly search for prey according to formula (20); if p⩾0.5, update the current position according to formula (21). The specific formula is as follows: (19) (20) (twenty one) In the formula, is the current location, is the current optimal position vector, represents the distance between the whale and its prey, b is the constant defining the spiral equation, and in this paper, b=1. is a random number between [−1, 1], is a random number between [0, 1], , represents the position vector of a randomly selected whale from the group; For the saved solutions, the pinhole imaging reverse learning is performed dimension by dimension according to formula (22). After the pinhole imaging reverse learning, the solution of a certain dimension is combined with the values ​​of other dimensions to form a new solution. Then, the new solution is evaluated according to the fitness of the objective function. If the quality of the solution is better than the previous solution, the updated result of the pinhole imaging reverse learning of the dimension is retained; otherwise, the information of the solution before the pinhole imaging reverse learning is retained. In this way, all dimensions are updated;

[0012] (twenty two) In the formula, and They represent the optimal solution of the j-th dimension and the optimal solution of the j-th dimension reverse solution, n is the adjustment factor, , h and h' are the height of the flame in the pinhole imaging and the height of the flame projection on the receiving screen, respectively. and are the upper and lower limits of the solution of the j-th dimension respectively; Determine whether the algorithm has reached the maximum number of iterations. If so, stop the calculation and output the optimal position and fitness value; otherwise, repeat the execution until the global optimal solution and optimal fitness are obtained.

[0013] As a further solution of the present invention, the step of constructing the BiLSTM model includes: BiLSTM cell layer construction, the bidirectional long short-term memory network (BiLSTM) is composed of two LSTMs, forward and backward. The specific steps and formulas are as follows: Forget Gate: The forget gate is the key calculation unit that determines how much long-term information C to retain. Its mathematical essence is to zero the memory cell that was input in the previous time step. Multiply by a ratio between 0 and 1 to indicate the information retention ratio, thereby filtering out some old information. The calculation formula is as follows: (twenty three) In the formula, is the output of the forget gate, and are the weight and bias of the forget gate, is the hidden state at the previous moment, is the input at the current moment, It is the sigmoid activation function, and the output range is [0, 1]; Input Gate: The input gate determines what new information will be written into the memory cell. Its output consists of two parts: one is the input that determines the new memory information, and the other is the update of the old memory information. (twenty four) (25) in is the output of the input gate, is a candidate memory unit, which determines which new information will be stored in the memory unit, W and b are the corresponding weight and bias of the input gate, and tanh is the activation function; Update memory unit: (26) In the formula, is the state of the memory cell at the previous moment, is the current state of the memory unit, and They are the outputs of the forget gate and the input gate, which control the updating and forgetting of memory; Output Gate: The output gate determines the hidden state at the next moment , that is, the output of LSTM. The output gate decides how much information to output based on the current memory cell state and input information; (27) (28) In the formula, is the output of the output gate, is the current state of the memory unit, is the hidden state at the current moment, W and b are the corresponding weight and bias of the output gate respectively, and tanh is the activation function; The features extracted by CNN are input into the constructed BiLSTM model for time series modeling, completing the establishment of the CNN-BiLSTM hybrid model.

[0014] As a further solution of the present invention, the missing values ​​of the processed data are interpolated to unify the time intervals to 1 hour.

[0015] In summary, the embodiments of the present invention have the following beneficial effects compared with the prior art: (1) The present invention takes into account the noise interference, modal aliasing and redundant component problems in flood data, and uses time-varying filtered empirical mode decomposition (TVFEMD) and sample entropy signal reconstruction (SE) technology to convert the original non-stationary signal into input data with low noise, high regularity and clear physical meaning, so that the subsequent model can focus on the essential laws of flood evolution rather than being disturbed by noise or redundant fluctuations, thereby significantly improving the prediction accuracy and reliability.

[0016] (2) This paper considers the limitations of single prediction models and uses CNN feature extraction to compensate for BiLSTM's lack of sensitivity to local detail features. BiLSTM prediction also compensates for CNN's weak ability to model long-term dependencies. The synergistic effect of the two can simultaneously capture the local details and global evolution trends of flood data, thereby more accurately predicting complex hydrological processes.

[0017] (3) The present invention takes into account the problem that traditional optimization algorithms are prone to falling into local optimal solutions, and uses an improved whale optimization algorithm to improve the optimization efficiency of the optimization algorithm and increase the stability of the algorithm, thereby compensating for the defects of traditional optimization algorithms such as being prone to falling into local optimal solutions and having slow convergence speed.

[0018] In order to more clearly illustrate the structural features and effects of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the flow chart of the TVFEMD-SE-CNN-MWOA-BiLSTM model.

[0020] Figure 2 This is the CNN structure diagram.

[0021] Figure 3 This is the MWOA flow chart.

[0022] Figure 4 This is the BiLSTM structure diagram.

[0023] Figure 5 This is the LSTM structure diagram.

[0024] Figure 6 This is the CNN-BiLSTM structure diagram. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0027] A hybrid deep learning flood forecasting method based on time-varying filter empirical mode decomposition (TVFEMD) and sample entropy signal reconstruction (SE) is proposed. Figure 1 As shown, it specifically includes the following steps: Collect historical precipitation and flow data of the basin, interpolate missing values ​​of flood data, unify the time interval to 1 hour, select typical flood data and connect them into a continuous sequence through a small section of runoff, and set the model forecast period; The TVFEMD decomposition method is used to extract and decompose flood data to generate multiple relatively simple intrinsic mode functions (IMFs) and extract multi-scale characteristic signals; The entropy of each intrinsic mode function (IMFs) is calculated based on the sample entropy, and then the IMFs with different entropy values ​​are reconstructed, and the IMFs in the same entropy value range are aggregated into different components (IMF's); The decomposed and reconstructed dataset is divided into a training set and a test set, which are then fed into a CNN to extract features. The generated multidimensional matrix is ​​then fed into the BiLSTM model. A unified initialization strategy and a nonlinear improvement strategy were adopted to train the BiLSTM model using the training set. MWOA was used to optimize the three hyperparameters of BiLSTM: the number of hidden units, the number of iterations, and the learning rate, to obtain the optimal hyperparameters of the model. The test set data is decomposed and reconstructed by TVFEMD-SE. The component functions IMF's are input into the trained CNN-MWOA-BiLSTM model for prediction. The prediction results of each component are superimposed to obtain the final flood forecast result.

[0028] In this embodiment, the steps of collecting and processing historical precipitation and flow data of the watershed and setting a forecast period include: Historical precipitation and flow data for the basin were collected. Considering computational convenience and the continuity of flood runoff rise and fall, missing flood data were interpolated, using spline interpolation to round off the time interval to 1 hour. To eliminate interference from rainfall events between flood events, a short runoff segment was inserted between each flood event, linking all flood events into a continuous flood event. The model's flood forecast horizon was set to n hours. Rainfall and evaporation data from rain gauges and hydrological stations for the previous n hours, as well as flow data from hydrological control stations for the previous n hours, served as the model's input features. The cross-sectional flow at the hydrological control station at the current time (t) served as the model's output variable.

[0029] In this embodiment, the step of extracting and decomposing flood data using the TVFEMD decomposition method includes: 1. Input flood data Perform Hilbert transform to obtain analytical signal, and record the result . (1)

[0030] (2) In the formula, and They are respectively expressed as the instantaneous amplitude and instantaneous phase of the flood flow series.

[0031] 2. Determine the local maximum and local minimum of the instantaneous amplitude, expressed as and Then, interpolate the maximum point set and the minimum point set to get the curve and , and Calculate according to formula (3) and (4) respectively.

[0032] (3) (4) 3. Yes and Perform interpolation and find and , and then calculate the instantaneous frequency component according to equations (5) and (6): and .

[0033] (5) (6) 4. Calculate the cutoff frequency according to formula (7).

[0034] (7) 5. Through calculated , calculate the signal according to formula (8) . The extreme points of the time-varying filter are used to construct the nodes, and the B-spline interpolation is used to After further filtering, the approximate result is .

[0035] (8) 6. When the stop frequency meets hour, That is IMF; when the stop frequency does not meet the conditions, that is , , then repeat steps 1 to 5. The calculation formula is as follows:

[0036] (9) In the formula, is the instantaneous bandwidth of the two-component signal, It is the weighted average of the instantaneous frequency of the single component signal. Set to 0.1.

[0037] In this embodiment, the step of reconstructing the SE aggregated reconstructed signal using the sample entropy signal includes: Calculate the sample entropy, the calculation formula is as follows: (10) In the formula, m is the reconstruction dimension, r is the similarity tolerance, N is the length of flood data, is the probability that two sequences match m points under the similarity tolerance r.

[0038] In this embodiment, the step of extracting high-level features from the reconstructed signal using CNN includes: The steps of CNN feature extraction are as follows Figure 2 As shown in Figure 1, first, a convolution layer is established, the appropriate convolution kernel size is set, and the grid matrix is ​​extracted as the feature map matrix to input into the CNN model. The spatial features of the basin meteorological information are extracted by reading the matrix in the convolution layer. Then, D c A filter is used to identify the input feature map matrix. According to formula (11), the matrix is ​​extracted and the weight matrix and bias of the filter are calculated. The filter is used to scan the feature map matrix, starting from the upper left corner and moving from left to right and from top to bottom until the lower right corner of the matrix is ​​scanned. Using formula (12), D c The filters are convolved with the input matrix area they cover, and the final output is D c The feature matrix after convolution is F conv .

[0039] (11) (12) In the formula, A feature graph representing the input; Represents the output after convolution calculation Feature graphics; express The height of the characteristic figure; express The width of the feature pattern; express The height of the characteristic figure; express The width of the feature pattern; Indicates the number of convolution kernels; Represents the convolution kernel output result, which represents the calculated value at the height h and width w of the d-th convolution kernel matrix; Represents the weight value in the d-th convolution kernel filter, It represents the deviation value corresponding to the d-th convolution kernel filter, and n represents the convolution step size of the convolution kernel.

[0040] Secondly, since the input matrix dimension is large, in order to reduce the optimization difficulty and the number of parameters, the convolution matrix is ​​downsampled in the pooling layer, which is also called pooling. The process is shown in formula (13). The maximum value method is used for pooling calculation, which plays the role of dimensionality reduction and principal component extraction for the input matrix.

[0041] (13) (14) In the formula, is the pooling step size; Indicates the output result of pooling calculation; Represents the pooling layer feature matrix.

[0042] Finally, the features output by the pooling layer are expanded into a one-dimensional matrix through the flatten layer, and then linearly transformed through the fully connected layer to output the matrix Y that is consistent with the number of BiLSTM cells. out , this matrix Y out As the input information of the BiLSTM network, the BiLSTM cell layer is then used to learn the time change process of the input data from front to back and from back to front, and the nonlinear relationship between meteorological data and hydrological data is established.

[0043] In this embodiment, the step of adaptively optimizing the hyperparameters of the BiLSTM model using the MWOA optimization algorithm includes: 1. Set the two parameters of MWOA population size N and maximum number of iterations T, and generate the initial population in the search space according to Circle mapping. The Circle mapping calculation formula is as follows: (15) In the formula, mod(a, b) means mapping the value of a to [0, b).

[0044] 2. Calculate the fitness value of each individual in the population and record the optimal position X(t).

[0045] 3. Integrate the adaptive weights of the Beta distribution and the inverse incomplete Γ function, update the parameter ω according to formula (16), and update according to formulas (23) and (24) 、 、 The specific formula is as follows: (16) (17) (18) In the formula, = 0.9, = 0.4; is the inverse of the incomplete Γ function Matlab calls the function; λ (λ⩾0) is a random variable; represents a random number that follows a Beta distribution, b1=1, b2=2; Inertia weight adjustment factor, used to control inertia weight The degree of deviation enables it to better balance the algorithm's global search and local development capabilities; and is a random number between [0, 1]; is a parameter that decreases from 2 to 0 as the number of iterations increases, defined as , is the current iteration number, is the maximum number of iterations of the algorithm.

[0046] 4. Compare the size of |A|, compare the randomly generated p value with 0.5, and select the corresponding position update formula. If p < 0.5 and |A| < 1, update the current individual position according to formula (19); if p < 0.5 and |A| ⩾ 1, randomly search for prey according to formula (20); if p ⩾ 0.5, update the current position according to formula (21). The specific formula is as follows: (19) (20) (twenty one) In the formula, is the current location, is the current optimal position vector, represents the distance between the whale and its prey, b is the constant defining the spiral equation, and in this paper, b=1. is a random number between [−1, 1], is a random number between [0, 1], , represents the position vector of a randomly selected whale from the population.

[0047] 5. For the solution saved in step 4, perform pinhole imaging reverse learning dimension by dimension according to equation (22). After pinhole imaging reverse learning, the solution of a certain dimension is combined with the values ​​of other dimensions to form a new solution. The new solution is then evaluated according to the fitness of the objective function. If the quality of the solution is better than the previous solution, the updated result of pinhole imaging reverse learning is retained for that dimension; otherwise, the information of the solution before pinhole imaging reverse learning is retained. In this way, all dimensions are updated.

[0048] (twenty two) In the formula, and They represent the optimal solution of the j-th dimension and the optimal solution of the j-th dimension reverse solution, n is the adjustment factor, , h and h' are the height of the flame in the pinhole imaging and the height of the flame projection on the receiving screen, respectively. and are the upper and lower limits of the solution of the j-th dimension respectively.

[0049] 6. Determine whether the algorithm has reached the maximum number of iterations. If so, stop the calculation and output the optimal position and fitness value; otherwise, repeat steps 3 to 5 until the global optimal solution and optimal fitness are obtained.

[0050] In this embodiment, the step of constructing the BiLSTM model includes: BiLSTM cell layer construction, such as Figure 4 As shown in the figure, the bidirectional long short-term memory network (BiLSTM) is composed of two LSTMs, forward and backward. The structure of LSTM is as follows: Figure 5 The specific steps and formulas are as follows:

[0051] (1) Forget Gate The forget gate is the key calculation unit that determines how much long-term information C to retain. Its mathematical essence is to zero the memory cell that was input in the previous time step. Multiply by a ratio between 0 and 1 to indicate the information retention ratio, thereby filtering out some old information. The calculation formula is as follows: (twenty three) In the formula, is the output of the forget gate, and are the weight and bias of the forget gate, is the hidden state at the previous moment, is the input at the current moment, It is a sigmoid activation function with an output range of [0, 1].

[0052] (2) Input gate The input gate determines what new information will be written into the memory cell. Its output consists of two parts: one is the input that determines the new memory information, and the other is the update of the old memory information.

[0053] (twenty four) (25) In the formula, is the output of the input gate, is a candidate memory cell, which determines which new information will be stored in the memory cell. W and b are the corresponding weight and bias of the input gate, respectively, and tanh is the activation function.

[0054] (8) Update memory unit (26) In the formula, is the state of the memory cell at the previous moment, is the current state of the memory unit, and They are the outputs of the forget gate and the input gate, which control the updating and forgetting of memory.

[0055] (4) Output gate The output gate determines the hidden state at the next moment , which is the output of LSTM. The output gate determines how much information to output based on the current state of the memory cell and the input information.

[0056] (27) (28) In the formula, is the output of the output gate, is the current state of the memory unit, is the hidden state at the current moment. W and b are the corresponding weight and bias of the output gate, respectively, and tanh is the activation function.

[0057] The features extracted by CNN are input into the constructed BiLSTM model for time series modeling, completing the establishment of the CNN-BiLSTM hybrid model.

[0058] According to the above steps, a hybrid deep learning flood forecasting method based on time-varying filter empirical mode decomposition (TVFEMD) and sample entropy signal reconstruction (SE) is constructed.

[0059] The TVFEMD module is used to adaptively decompose non-stationary and nonlinear flood time series data and extract multi-scale features; the SE module is used to screen key components and aggregate signals in the same entropy value range; the TVFEMD-SE hybrid module is used for data pre-processing; The CNN module is used to extract spatial features from the reconstructed signal, convert the original time series data into a feature matrix, and enhance the expression of local features. The BiLSTM module is used to receive the feature sequence output by the CNN and capture the long-term dependencies and bidirectional dynamic features in the flood time series data. The CNN-BiLSTM hybrid module is used for feature extraction and model prediction. The MOWA module intelligently optimizes the model's hyperparameters, improving the adaptability of BiLSTM's key configurations to flood data, thereby significantly improving the model's prediction accuracy and generalization capabilities. The five modules are combined in sequence for decomposition, reconstruction, feature extraction, optimization, and prediction, respectively, to establish a hybrid deep learning flood forecasting method based on (TVFEMD) and sample entropy signal reconstruction (SE).

[0060] In summary, the present invention aims to solve the problems of strong noise interference, difficult feature extraction and low efficiency of model parameter optimization in non-stationary time series data in basin flood forecasting. The method uses time-varying filter empirical mode decomposition (TVFEMD) technology and sample entropy signal reconstruction (SE) to construct an adaptive noise reduction mechanism, combines convolutional neural network (CNN) to extract local spatial features in the reconstructed signal, and introduces bidirectional long short-term memory network (BiLSTM) to capture the full-cycle temporal dependency of flood evolution in a bidirectional manner. The improved whale optimization algorithm (MWOA) is further nested to intelligently optimize the model hyperparameters. With the goal of improving the accuracy of flood forecasting, a hybrid deep learning flood forecasting method based on (TVFEMD) and sample entropy signal reconstruction (SE) is proposed. The present invention integrates the technical advantages of time-varying filter decomposition, sample entropy reconstruction, convolution feature extraction and intelligent optimization algorithm, with the core goal of optimizing flood forecasting accuracy and improving model generalization capabilities. It constructs a collaborative computing framework of signal noise reduction-spatial feature extraction-hyperparameter adaptive optimization-time series dynamic modeling, and realizes noise suppression, spatiotemporal feature extraction and dynamic optimization of model hyperparameters of non-stationary hydrological sequences. It can provide flood process simulation with high time series fitting degree and strong generalization performance for basin flood control scheduling.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flood forecasting method based on a multimodal intelligent algorithm, characterized in that: The following steps are involved: S1. Collect historical precipitation and flow data of the basin, process missing values, and connect different flood data into continuous data through a small section of runoff; S2, using TVFEMD decomposition method to decompose flood data into multiple intrinsic mode functions IMFs; S3, calculate the entropy value based on the sample entropy, aggregate and reconstruct the IMFs in the same entropy value range; S4. Divide the data set into a training set and a test set, and use CNN to extract data features to obtain a multidimensional matrix; S5. Input the feature matrix extracted by CNN into the BiLSTM model and use MWOA to optimize the hyperparameters of the BiLSTM model. S6. Predict each component of TVFEMD-SE decomposition and aggregation separately, and superimpose the predicted values ​​of each component to obtain the flood forecast value.

2. The flood forecasting method based on multimodal intelligent algorithm according to claim 1 is characterized in that: The input of S2 is flood data Perform Hilbert transform to obtain the analytical signal, and the result is recorded as ; ;(1) ；(2) and are respectively expressed as the instantaneous amplitude and instantaneous phase of the flood discharge series; Determine the local maximum and local minimum of the instantaneous amplitude, which are expressed as and Then, interpolate the maximum point set and the minimum point set to get the curve and , and Calculated according to formula (3) and (4) respectively; ；(3) ； (4) right and Perform interpolation and find and , and then calculate the instantaneous frequency component according to equations (5) and (6): and ; ；(5) ； (6) The cutoff frequency is calculated according to formula (7); ； (7) By calculating , calculate the signal according to formula (8) , The extreme points of the time-varying filter are used to construct the nodes, and the B-spline interpolation is used to After further filtering, the approximate result is ; ； (8) When the stop frequency meets hour, That is IMF; when the stop frequency does not meet the conditions, that is , , then repeat the above steps, The calculation formula is as follows: ； (9) In the formula, is the instantaneous bandwidth of the two-component signal, is the weighted average of the instantaneous frequency of the single-component signal, the threshold Set to 0.

1.

3. The flood forecasting method based on multimodal intelligent algorithm according to claim 2 is characterized in that: The step of reconstructing the sample entropy signal into an SE aggregated reconstructed signal comprises: Calculate the sample entropy, the calculation formula is as follows: ；(10) In the formula, m is the reconstruction dimension, r is the similarity tolerance, N is the length of flood data, is the probability that two sequences match m points under the similarity tolerance r.

4. The flood forecasting method based on multimodal intelligent algorithm according to claim 3 is characterized in that: The steps of extracting high-level features from the reconstructed signal by CNN include: Establish a convolution layer, set the convolution kernel size, extract the grid matrix as the feature map matrix and input it into the CNN model. In the convolution layer, read the matrix to extract the spatial features of the basin meteorological information. Use D c A filter is used to identify the input feature map matrix. According to formula (11), the matrix is ​​extracted and the weight matrix and bias of the filter are calculated. The filter is used to scan the feature map matrix. Formula (12) is used to convert D c The filters are convolved with the input matrix area they cover, and the final output is D c The feature matrix after convolution is F conv ; ；（11） ；（12） In the formula, A feature graph representing the input; Represents the feature graph output after convolution calculation; express The height of the characteristic figure; express The width of the feature pattern; express The height of the characteristic figure; express The width of the feature pattern; Indicates the number of convolution kernels; Represents the convolution kernel output result, which represents the calculated value at the height h and width w of the d-th convolution kernel matrix; Represents the weight value in the d-th convolution kernel filter, Represents the deviation value corresponding to the d-th convolution kernel filter, and n represents the convolution step size of the convolution kernel; Secondly, in the pooling layer, the convolution matrix is ​​downsampled, also known as pooling. The process is shown in formula (13). The maximum value method is used for pooling calculation, which plays the role of dimensionality reduction and principal component extraction for the input matrix. ；（13） ；（14） In the formula, is the pooling step size; Indicates the output result of pooling calculation; Represents the pooling layer feature matrix; The features output by the pooling layer are expanded into a one-dimensional matrix through the flatten layer, and the fully connected layer performs linear transformation to output the matrix Y that is consistent with the number of BiLSTM cells. out , this matrix Y out As the input information of the BiLSTM network, the BiLSTM cell layer is used to learn the time change process of the input data from front to back and from back to front, and the nonlinear relationship between meteorological data and hydrological data is established.

5. The flood forecasting method based on multimodal intelligent algorithm according to claim 4 is characterized in that: The steps of adaptively optimizing the hyperparameters of the BiLSTM model using the MWOA optimization algorithm include: Set the two parameters of MWOA population size N and maximum number of iterations T, and generate the initial population in the search space according to Circle mapping. The Circle mapping calculation formula is as follows: ；(15) In the formula, mod(a, b) means mapping the value of a to [0, b); Calculate the fitness value of each individual in the population and record the optimal position X(t); The adaptive weights of the fusion beta distribution and the inverse incomplete Γ function are updated according to formula (16), and the parameter ω is updated at the same time. 、 、 Parameters, the specific formula is as follows: ；(16) ；(17) ；(18) In the formula, =0.9, =0.4; is the inverse of the incomplete Γ function Matlab calls the function; λ (λ⩾0) is a random variable; represents a random number that follows the Beta distribution, b1=1, b2=2; Inertia weight adjustment factor, used to control inertia weight The degree of deviation; and is a random number between [0, 1]; is a parameter that decreases from 2 to 0 as the number of iterations increases, defined as , is the current iteration number, is the maximum number of iterations of the algorithm; Compare the size of |A|, compare the randomly generated p value with 0.5, and select the corresponding position update formula. If p<0.5 and |A|<1, update the current individual position according to formula (19); if p<0.5 and |A|⩾1, randomly search for prey according to formula (20); if p⩾0.5, update the current position according to formula (21). The specific formula is as follows: ；(19) ；(20) ；(21) In the formula, is the current location, is the current optimal position vector, represents the distance between the whale and its prey, b is the constant defining the spiral equation, and in this paper, b=1. is a random number between [−1, 1], is a random number between [0, 1], , represents the position vector of a randomly selected whale from the group; For the saved solutions, the pinhole imaging reverse learning is performed dimension by dimension according to formula (22). After the pinhole imaging reverse learning, the solution of a certain dimension is combined with the values ​​of other dimensions to form a new solution. Then, the new solution is evaluated according to the fitness of the objective function. If the quality of the solution is better than the previous solution, the updated result of the pinhole imaging reverse learning is retained in this dimension; otherwise, the information of the solution before the pinhole imaging reverse learning is retained. In this way, all dimensions are updated. ；(22) In the formula, and They represent the optimal solution of the j-th dimension and the optimal solution of the j-th dimension reverse solution, n is the adjustment factor, , h and h' are the height of the flame in the pinhole imaging and the height of the flame projection on the receiving screen, respectively. and are the upper and lower limits of the solution of the j-th dimension respectively; Determine whether the algorithm has reached the maximum number of iterations. If so, stop the calculation and output the optimal position and fitness value; otherwise, repeat the execution until the global optimal solution and optimal fitness are obtained.

6. The flood forecasting method based on multimodal intelligent algorithm according to claim 5 is characterized in that: The steps for building the BiLSTM model include: BiLSTM cell layer construction, the bidirectional long short-term memory network (BiLSTM) is composed of two LSTMs, forward and backward. The specific steps and formulas are as follows: Forget Gate: The forget gate is the key calculation unit that determines how much long-term information C to retain. Its mathematical essence is to zero the memory cell that was input in the previous time step. Multiply by a ratio between 0 and 1 to indicate the information retention ratio, thereby filtering out some old information. The calculation formula is as follows: ；（23） In the formula, is the output of the forget gate, and are the weight and bias of the forget gate, is the hidden state at the previous moment, is the input at the current moment, It is the sigmoid activation function, and the output range is [0, 1]; Input Gate: The input gate determines what new information will be written into the memory cell. Its output consists of two parts: one is the input that determines the new memory information, and the other is the update of the old memory information. ；（24） ；（25） In the formula, is the output of the input gate, is a candidate memory unit, which determines which new information will be stored in the memory unit, W and b are the corresponding weight and bias of the input gate, and tanh is the activation function; Update memory unit: ；（26） In the formula, is the state of the memory cell at the previous moment, is the current state of the memory unit, and They are the outputs of the forget gate and the input gate, which control the updating and forgetting of memory; Output Gate: The output gate determines the hidden state at the next moment , that is, the output of LSTM. The output gate decides how much information to output based on the current memory cell state and input information; ；（27) ；(28) In the formula, is the output of the output gate, is the current state of the memory unit, is the hidden state at the current moment, W and b are the corresponding weight and bias of the output gate respectively, and tanh is the activation function; The features extracted by CNN are input into the constructed BiLSTM model for time series modeling, completing the establishment of the CNN-BiLSTM hybrid model.

7. The flood forecasting method based on multimodal intelligent algorithm according to claim 6 is characterized in that: The missing values ​​of the processed data are interpolated to unify the time intervals to 1 hour.

Images