Flood intelligent forecasting method and system fusing flood process information physical characteristics

Abstract

The invention discloses an intelligent flood forecasting method and system fusing flood process information physical characteristics, and solves the problems of poor robustness, strong dependence on data quality, weak interpretability and the like of deep learning in flood forecasting modeling. A new data processing mode and an error function are designed by analyzing the main body evolution rule and noise of the flood process, and the quality of model input data is improved; the method comprises the following steps: designing a two-dimensional hidden layer according to hydrometeorological element characteristics, constructing an ensemble attention mechanism screening architecture, and proposing a flood recycling network (FRN) forecasting architecture fused with flood process physical characteristics; an intelligent FRN flood forecasting model fusing flood physical characteristics is constructed, and FRN model visualization application software integrating interpretability and a long forecasting period is developed. The method can improve the accuracy and interpretability of forecasting, and promotes the innovative application of an artificial intelligence method in the field of water science.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological forecasting, and in particular relates to a flood intelligent forecasting method and system integrating physical characteristics of flood process information. Background Art

[0002] Artificial intelligence is a strategic technology leading this round of scientific and technological revolution and industrial transformation, and has a strong spillover effect. With the continuous development of artificial intelligence technology, intelligent flood forecasting has become one of the key measures to improve the level of intelligent flood control in river basins. However, the existing flood forecasting models have certain limitations in dealing with complex flood dynamics, spatiotemporal data characteristics, and prediction accuracy, and lack a flood intelligent forecasting model that integrates new generation artificial intelligence technologies such as deep learning frameworks.

[0003] In order to meet the strategic needs of improving the intelligent level of flood control in river basins, this invention will be guided by the principles of "demand-driven, application-oriented, digital empowerment, and capacity enhancement", focusing on cutting-edge technologies such as deep learning frameworks, strengthening the integration of new-generation artificial intelligence technologies such as deep learning frameworks with water conservancy services, and improving the intelligent level of flood forecasting. Through in-depth research on the spatiotemporal evolution of floods and their influencing factors, designing deep learning models that conform to the dynamic characteristics of floods, and developing high-precision intelligent flood forecasting systems, we will improve the intelligent level of flood control in my country's river basins and provide strong support for the high-quality development of the national economy and river basin security. Summary of the invention

[0004] In order to solve the technical problems existing in the background technology of deep learning in flood forecast modeling, such as poor robustness, strong dependence on data quality and weak interpretability, the present invention aims to provide a flood intelligent forecasting method and system that integrates the physical characteristics of flood process information.

[0005] In order to solve the technical problem, the technical solution of the present invention is:

[0006] A flood intelligent forecasting method integrating physical characteristics of flood process information, the method comprising:

[0007] S1: Using the hydrological and meteorological data in the study basin, the Kalman filter data assimilation method is used to obtain the quantitative relationship between the flood process and the influencing factors;

[0008] S2: Based on the quantitative relationship, the TVF–EMD modal decomposition method is used to explore the evolution law of flood processes;

[0009] S3: Based on the mined evolution law, a step function is constructed to map the flood sequence into a smooth signal;

[0010] S4: Construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take the smoothed signal as input, use the sparse attention mechanism to select key features, and make predictions.

[0011] Furthermore, in step S1, the hydrological and meteorological data include: data of multiple time scales such as hour, day, month and year, and the time scales of the time scale data are consistent.

[0012] Furthermore, in step S1, a Kalman filter robustification method based on approximate Bayesian estimation is used to process the underlying surface, hydrological and meteorological raw data; including:

[0013] S101: Data preprocessing: First, preprocess the collected data, including removing obvious erroneous data and filling missing values;

[0014] S102: Establish state space model: Establish state space model according to the dynamic changes of underlying surface, hydrological and meteorological data; state space model includes state equation and observation equation; state equation describes the evolution of system state over time, and observation equation links system state with observation data; for soil moisture model, state equation is expressed as:

[0015] x k =Fx k-1 +Bu k +w k

[0016] In the formula, x k is the system state at time k, F is the state transition matrix, u k is the control input, B is the control matrix, w k is the process noise, usually assumed to be Gaussian white noise;

[0017] S103: Apply Kalman filtering: input the preprocessed data into the established model, and apply the Kalman filtering method based on approximate Bayesian estimation to perform state estimation;

[0018] S104: Result analysis: Analyze the estimation results after filtering and evaluate the accuracy and robustness of the model.

[0019] Furthermore, the Kalman filtering step includes two stages: prediction and update:

[0020] The formula for the prediction phase is:

[0021]

[0022] P k|k-1 =FP k-1|k-1 F T +Q

[0023] In the formula, is the state prediction estimate at time k, P k|k-1 is the covariance matrix of the forecast error, Q is the covariance matrix of the process noise;

[0024] The formula for the update phase is:

[0025] K k =P k|k-1 H T (HP k|k-1 H T +R) -1

[0026]

[0027] P k =(IK k H)P k|k-1

[0028] In the formula, K k is the Kalman gain, z k is the observation at time k, H is the observation matrix, and R is the covariance matrix of the observation noise.

[0029] Further, the step S2 specifically includes:

[0030] Using Hilbert transform to calculate the instantaneous amplitude A(t) and instantaneous frequency of the signal sequence X(t) The calculation formula is:

[0031]

[0032] In the formula, is the Hilbert transform of the signal sequence X(t);

[0033] Determine the local maximum value A({t max {}}()) and the local minimum A({t min {}}());

[0034] For A({t min {}}()) and A({t max {}}()) are interpolated respectively, and the obtained curves are β1(t) and β2(t), and then the instantaneous mean a1(t) and instantaneous envelope a2(t) are calculated. The calculation formula is:

[0035]

[0036] right and Interpolate respectively to get η1(t) and η2(t), and then calculate and The calculation formula is:

[0037]

[0038] Calculate the local cutoff frequency The calculation formula is:

[0039]

[0040] Rearrange To solve the intermittent problem;

[0041] Defining Signals The extreme points of h(t) are used as nodes, and the signal sequence is approximated by B-spline interpolation. The approximation result is m(t), which represents the local mean curve.

[0042] Calculate the cutoff criterion θ(t). If θ(t)≤ξ, then determine that X(t) is an IMF. Otherwise, let x1(t)=x(t)-m(t) and repeat the above steps. The calculation formula of the cutoff criterion θ(t) is:

[0043]

[0044] In the formula, B Loughlin (t) is the Loughlin instantaneous bandwidth of the two component signals; is the weighted average instantaneous frequency of a single component;

[0045] After the above steps, the value of the original signal sequence X(t) is all modal components X i (t) and.

[0046] Further, the step S3 specifically includes:

[0047] S301: Setting flood step value;

[0048] S302: Conversion sequence, the formula is as follows:

[0049]

[0050] In the formula, Q j is the flood step value, To round down, Q o is the flood observation value, and J is the step value.

[0051] Further, in step S4, constructing a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements refers to constructing a time dimension hidden layer and a space dimension hidden layer respectively to enhance the model's feature extraction capability for information of different dimensions; using a sparse attention mechanism to select key features and perform prediction, specifically including:

[0052] S401: Constructing a temporal attention mechanism:

[0053] First, use RNN to process time series data and obtain the hidden state h at each time step. t :

[0054] h t =σ(W hh h t-1 +W xh x t +b h )

[0055] In the formula, h t is the hidden state at time step t, h t-1 is the hidden state at time step t-1, x t is the input feature at time step t, W hh is the weight matrix of the hidden state, W xh is the weight matrix input to the hidden state, b h is the bias of the hidden layer, σ is the activation function;

[0056] Calculate the temporal attention weight α t :

[0057]

[0058] In the formula, e t =v T tanh(Wh t +b), W, v and b are the parameters to be learned;

[0059] Weight the hidden states at each time step to get the output of the temporal attention:

[0060]

[0061] S402: Building a spatial attention mechanism:

[0062] Each position i has a feature representation s at each time point. ti ;

[0063] At each time point t, the observations x at all locations t1 ,x t2 ...x tn Concatenate to form an input vector xt ;

[0064] Use this combined vector as the input of the RNN to obtain the overall hidden state h t ;

[0065] x t =Concat(x t1 ,x t2 ,...,x tn )

[0066] h t =RNN(x t ,h t-1 )

[0067] Then, to obtain the feature representation s for each position t , applying a fully connected layer from the global hidden state h t Extract and transform information from

[0068] s ti =W i h t +b i

[0069] Where W i and b i is a learning parameter specific to each position, used to obtain the value from the overall hidden state h t Extract specific features of the location;

[0070] Calculate the spatial attention weight β i :

[0071]

[0072] Where, d i =u T tanh(Vs ti +c), V, u and c are the parameters to be learned;

[0073] Weight the feature vectors at each spatial position to obtain the output of spatial attention:

[0074]

[0075] S403: Feature Fusion:

[0076] The results of temporal and spatial attention are fused to form a comprehensive feature representation:

[0077] f t =Concat(h t ',s t ')

[0078] S404: Use sparse attention mechanism to select the most critical features and make predictions:

[0079] Calculate sparse attention weights:

[0080] γ k =Sparsemax(g k )

[0081]

[0082] In the formula, g k =z T tanh(Xf t +e), z, X and e are learning parameters, Δ K-1 represents a K-dimensional probability simplex;

[0083] S405: Use sparse attention weights to weight the merged features and make final predictions:

[0084]

[0085] In the formula, γ k is the sparse attention weight, f t is the comprehensive eigenvector, W y , b y are the parameters of the output layer.

[0086] A flood intelligent forecasting system integrating physical characteristics of flood process information, the system comprising:

[0087] Data assimilation and quantitative relationship extraction module: used to use the hydrological and meteorological data in the study basin and adopt the Kalman filter data assimilation method to obtain the quantitative relationship between the flood process and the influencing factors;

[0088] Dynamic evolution law mining module: used to mine the evolution law of flood process based on the quantitative relationship by using TVF-EMD modal decomposition method;

[0089] Signal smoothing and conversion module: used to construct a step function based on the mined evolution law and map the flood sequence into a smooth signal;

[0090] Feature selection and flood prediction module: used to construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take smooth signals as input, use sparse attention mechanism to select key features, and make predictions.

[0091] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the above-mentioned intelligent flood forecasting methods integrating physical characteristics of flood process information is implemented.

[0092] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any one of the above-mentioned intelligent flood forecasting methods integrating physical characteristics of flood process information.

[0093] Compared with the prior art, the advantages of the present invention are:

[0094] The present invention discloses a flood intelligent forecasting method and system integrating the physical characteristics of flood process information, which solves the problems of poor robustness, strong dependence on data quality and weak interpretability of deep learning in flood forecast modeling. By analyzing the main evolution law and noise of the flood process, a new data processing method and error function are designed to improve the quality of model input data; a two-dimensional hidden layer is designed according to the characteristics of hydrological and meteorological elements, a collective attention mechanism screening architecture is constructed, and a flood recurrent network (FRN) forecasting architecture integrating the physical characteristics of the flood process is proposed; an FRN flood intelligent forecasting model integrating the physical characteristics of the flood is constructed, and an FRN model visualization application software with both interpretability and a longer forecast period is developed. The present invention can improve the accuracy and interpretability of the forecast and promote the innovative application of artificial intelligence methods in the field of water science. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is the structural flow chart of the FRN flood intelligent forecasting model;

[0096] Figure 2 This is a schematic diagram of Kalman filtering;

[0097] Figure 3 It is a schematic diagram of step function change;

[0098] Figure 4 It is a schematic diagram of the focal band interval;

[0099] Figure 5 It is the hidden layer structure change diagram;

[0100] Figure 6 Diagram of the integrated attention architecture. DETAILED DESCRIPTION

[0101] The specific implementation mode of the present invention is described below in conjunction with embodiments:

[0102] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0103] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0104] Embodiment 1:

[0105] The present invention provides a flood intelligent forecasting method named FRN. In order to achieve the above object, the present invention provides the following technical solutions:

[0106] Step 1: Determine the study basin and collect hydrological and meteorological data in the study basin. The hydrological and meteorological data can be data of various time scales such as hours, days, months, and years, but the data time scale should be kept unified;

[0107] Step 2: Use Kalman filter data assimilation method to deeply explore the quantitative relationship between flood process and its influencing factors;

[0108] Step 3: Use TVF–EMD modal decomposition technology to explore the evolution law of flood process

[0109] Step 4: Design a step function to map the flood series into a smooth signal to provide high-quality data for subsequent research.

[0110] Step 5: Design a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, use a sparse attention mechanism to select the most critical features, and make predictions.

[0111] The data assimilation refers to the use of a Kalman filter robustification method based on approximate Bayesian estimation to process raw data such as underlying surface, hydrology, and meteorology. This method extracts accurate state estimates from complex nonlinear systems while effectively suppressing the influence of noise and outliers. The Kalman filter data assimilation method includes: data preprocessing, including removing erroneous data and filling missing values; establishing a state space model, including state equations and observation equations; applying Kalman filtering for state estimation, including two stages of prediction and update; and result analysis to evaluate the accuracy and robustness of the model.

[0112] The specific process is as follows:

[0113] 1) Data preprocessing: First, the collected data is preprocessed, including removing obvious erroneous data and filling missing values. The purpose of this step is to ensure the quality of the input data and lay a good foundation for subsequent filtering processing.

[0114] 2) Establish a state space model: Establish an appropriate state space model based on the dynamic changes of the underlying surface, hydrological and meteorological data. The state space model usually consists of two parts: the state equation and the observation equation. The state equation describes the evolution of the system state over time, while the observation equation links the system state to the observed data. For the soil moisture model, the state equation can be expressed as:

[0115] x k =Fx k-1 +Bu k +w k

[0116] In the formula, x k is the system state at time k, F is the state transition matrix, u k is the control input, B is the control matrix, w k is the process noise, usually assumed to be Gaussian white noise.

[0117] 3) Apply Kalman filtering: The preprocessed data is input into the established model, and the Kalman filtering method based on approximate Bayesian estimation is applied for state estimation.

[0118] The basic steps of Kalman filtering include two stages: prediction and update.

[0119] The formula for the prediction phase is:

[0120]

[0121] P k|k-1 =FP k-1|k-1 F T +Q

[0122] In the formula, is the state prediction estimate at time k, P k|k-1 is the covariance matrix of the forecast errors, and Q is the covariance matrix of the process noise.

[0123] The formula for the update phase is:

[0124] K k =P k|k-1 H T (HP k|k-1 H T +R) -1

[0125]

[0126] P k =(IK k H)P k|k-1

[0127] In the formula, K k is the Kalman gain, z k is the observation at time k, H is the observation matrix, and R is the covariance matrix of the observation noise.

[0128] 4) Result analysis: Analyze the estimation results after filtering and evaluate the accuracy and robustness of the model.

[0129] The TVF-EMD is an improvement on EMD, which effectively solves the mode mixing problem existing in EMD. The TVF-EMD modal decomposition technology includes: using Hilbert transform to calculate the instantaneous amplitude and instantaneous frequency of the signal sequence; determining the local maximum and local minimum of the instantaneous amplitude; interpolating the local maximum and minimum, calculating the instantaneous mean and instantaneous envelope; calculating the local cutoff frequency, rearranging to solve the intermittent problem; defining the signal, using spline interpolation to approximate the signal sequence; calculating the cutoff criterion, and determining the modal component.

[0130] The specific steps are as follows:

[0131] ① Use Hilbert transform to calculate the instantaneous amplitude A(t) and instantaneous frequency of the signal sequence X(t) The calculation formula is:

[0132]

[0133] In the formula, is the Hilbert transform of the signal sequence X(t).

[0134] ② Determine the local maximum value A({t max {}}()) and the local minimum A({t min {}}());

[0135] ③ For A({t min {}}()) and A({t max {}}()) are interpolated respectively, and the obtained curves are β1(t) and β2(t), and then the instantaneous mean a1(t) and instantaneous envelope a2(t) are calculated. The calculation formula is:

[0136]

[0137] ④Yes and Interpolate respectively to get η1(t) and η2(t), and then calculate and The calculation formula is:

[0138]

[0139] ⑤Calculate the local cutoff frequency The calculation formula is:

[0140]

[0141] ⑥Rearrange To solve the intermittent problem;

[0142] ⑦Define signal The extreme points of h(t) are used as nodes, and the signal sequence is approximated by B-spline interpolation. The approximation result is m(t), which represents the local mean curve.

[0143] ⑧ Calculate the cut-off criterion θ(t). If θ(t)≤ξ, X(t) is determined to be an IMF. Otherwise, let x1(t)=x(t)-m(t). Repeat ① to ⑧. The calculation formula of the cut-off criterion θ(t) is:

[0144]

[0145]

[0146] In the formula, B Loughlin (t) is the Loughlin instantaneous bandwidth of the two component signals; is the weighted average instantaneous frequency of the individual components.

[0147] After the above steps, the value of the original signal sequence X(t) is all modal components X i (t) and.

[0148] The step function is to transform the fluctuating flood sequence process into several stable steps (such as Figure 4 As shown in the figure, the input factors are concentrated on a few limited targets to avoid the impact of flood process fluctuations. The processing method is as follows: set the flood step value; transform the sequence and map the flood observation value to the step value;

[0149] 1) Set the flood step value (assuming it is 100, that is, every 100m 3 / s flood is one level).

[0150] 2) Conversion sequence, the formula is as follows:

[0151]

[0152] In the formula, Q j is the flood step value, To round down, Q o is the flood observation value, and J is the step value.

[0153] The structure conforming to the two-dimensional physical characteristics of flood space and time refers to respectively constructing a "time dimension hidden layer" and a "space dimension hidden layer" to enhance the model's ability to extract features of information of different dimensions.

[0154] The temporal attention mechanism is as follows

[0155] ① First, use RNN to process time series data and obtain the hidden state h of each time step t :

[0156] h t =σ(W hh h t-1 +W xh x t +b h )

[0157] In the formula, h t is the hidden state at time step t, h t-1 is the hidden state at time step t-1, x t is the input feature at time step t, W hh is the weight matrix of the hidden state, W xh is the weight matrix input to the hidden state, b h is the bias of the hidden layer and σ is the activation function.

[0158] ② Calculate the time attention weight α t :

[0159]

[0160] In the formula, e t =v T tanh(Wh t +b), W, v and b are the parameters to be learned.

[0161] ③ Weight the hidden state of the time step to get the output of temporal attention:

[0162]

[0163] 2) Spatial Attention Mechanism

[0164] Each position i has a feature representation s at each time point. ti .

[0165] ①At each time point t, the observation values ​​x at all locations t1 ,x t2 ...x tnConcatenate to form an input vector x t .

[0166] ②Use this combined vector as the input of RNN to obtain the overall hidden state h t .

[0167] x t =Concat(x t1 ,x t2 ,...,x tn )

[0168] h t =RNN(x t ,h t-1 )

[0169] Then, to obtain the feature representation s for each position t , applying a fully connected layer from the global hidden state h t Extract and transform information.

[0170] s ti =W i h t +b i

[0171] Where W i and b i is a learning parameter specific to each position, used to obtain the value from the overall hidden state h t Extract specific features of the location.

[0172] ③Calculate the spatial attention weight β i :

[0173]

[0174] Where, d i =u T tanh(Vs ti +c), V, u and c are the parameters to be learned.

[0175] ④Weight the feature vectors of each spatial position to obtain the output of spatial attention:

[0176]

[0177] 3) Feature Fusion

[0178] The results of temporal and spatial attention are fused to form a comprehensive feature representation:

[0179] f t =Concat(h t ',s t ')

[0180] 4) Sparse Attention Mechanism

[0181] Use sparse attention mechanism to select the most critical features and make predictions.

[0182] Calculate sparse attention weights:

[0183] γ k =Sparsemax(g k )

[0184]

[0185] In the formula, g k =z T tanh(Xf t +e), z, X and e are learning parameters, Δ K-1 represents a K-dimensional probability simplex.

[0186] 5) Prediction output

[0187] Use sparse attention weights to weight the merged features and make the final prediction:

[0188]

[0189] In the formula, γ k is the sparse attention weight, f t is the comprehensive eigenvector, W y , b y are the parameters of the output layer.

[0190] Embodiment 2:

[0191] This embodiment is applied to the first embodiment, and takes the middle reaches of the Yangtze River Basin in China as the research object. Due to the complex terrain and uneven rainfall distribution, the region is often threatened by floods. The present invention provides a flood intelligent forecasting model named FRN, such as Figure 1 As shown, it mainly includes the following steps:

[0192] Step 1: Identify the study area and collect data

[0193] A specific section of the middle reaches of the Yangtze River was assumed as the study basin, and hourly hydrological and meteorological data of the basin were collected from 2010 to 2020, including rainfall, flow, water level, and temperature.

[0194] Step 2: Data preprocessing and Kalman filter data assimilation

[0195] In the data preprocessing stage, abnormal rainfall data exceeding 3% were found and removed, and missing water level data were filled. Figure 2As shown, the Kalman filter data assimilation method is used to process the data, where the state transfer matrix F is set to 0.9 and the control input u k is 0.05, and the process noise covariance matrix Q is 0.01.

[0196] Step 3: TVF-EMD modal decomposition technique

[0197] Through TVF-EMD modal decomposition, the team identified three main modes in the flood process, of which the first mode is closely related to changes in rainfall, the second mode is related to upstream water, and the third mode reflects the hydrological process within the basin.

[0198] Step 4: Design the step function

[0199] like Figure 3 As shown in Figure 1, a step function is designed to map the flood flow observation value to the step value. For example, the flood step value is set to 100m 3 / s, round the observed flow value down to the nearest 100m 3 / multiples of s.

[0200] Step 5: Construct a 2D hidden layer and sparse attention mechanism

[0201] like Figure 5 and 6 As shown in the figure, a two-dimensional hidden layer structure is constructed, which includes a time dimension hidden layer and a space dimension hidden layer. The time dimension hidden layer considers the flow changes in the past 24 hours, while the space dimension hidden layer considers the hydrological characteristics of different sub-regions in the basin. The sparse attention mechanism performs weighted merging of features by calculating the key feature weights.

[0202] The FRN model was trained using data from 2010 to 2019. The model parameters included a temporal attention weight matrix W of 3x3 and a spatial attention weight matrix V of 5x5. After training, the model was used to predict floods on data from 2020. The model prediction results showed that for a flood event in July 2020, the FRN model accurately predicted the flood peak 12 hours in advance, with a peak flow of 8000m 3 / s, and the error with the actual observed value is within 5%.

[0203] This example demonstrates the application of the FRN model in flood forecasting in the middle reaches of the Yangtze River. Through the assumed data and parameters, it is proved that the model can effectively improve the accuracy and interpretability of flood forecasting, providing strong technical support for flood control in the basin.

[0204] Embodiment 3:

[0205] The present invention provides a flood intelligent forecasting system integrating information and physical characteristics of flood process. The system can be used to implement the above-mentioned flood intelligent forecasting method integrating information and physical characteristics of flood process, specifically, including:

[0206] Data assimilation and quantitative relationship extraction module: used to use the hydrological and meteorological data in the study basin and adopt the Kalman filter data assimilation method to obtain the quantitative relationship between the flood process and the influencing factors;

[0207] Dynamic evolution law mining module: used to mine the evolution law of flood process based on the quantitative relationship by using TVF-EMD modal decomposition method;

[0208] Signal smoothing and conversion module: used to construct a step function based on the mined evolution law and map the flood sequence into a smooth signal;

[0209] Feature selection and flood prediction module: used to construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take smooth signals as input, use sparse attention mechanism to select key features, and make predictions.

[0210] Embodiment 4:

[0211] This embodiment provides a terminal device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a flood intelligent forecasting method that integrates the physical characteristics of flood process information, including the following steps:

[0212] S1: Using the hydrological and meteorological data in the study basin, the Kalman filter data assimilation method is used to obtain the quantitative relationship between the flood process and the influencing factors;

[0213] S2: Based on the quantitative relationship, the TVF–EMD modal decomposition method is used to explore the evolution law of flood processes;

[0214] S3: Based on the mined evolution law, a step function is constructed to map the flood sequence into a smooth signal;

[0215] S4: Construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take the smoothed signal as input, use the sparse attention mechanism to select key features, and make predictions.

[0216] Embodiment 5:

[0217] This embodiment provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understandable that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory.

[0218] The processor may load and execute one or more instructions stored in a computer-readable storage medium to implement the corresponding steps of the above-mentioned embodiment of a flood intelligent forecasting method integrating physical characteristics of flood process information; the processor may load and execute the following steps of one or more instructions in the computer-readable storage medium:

[0219] S1: Using the hydrological and meteorological data in the study basin, the Kalman filter data assimilation method is used to obtain the quantitative relationship between the flood process and the influencing factors;

[0220] S2: Based on the quantitative relationship, the TVF–EMD modal decomposition method is used to explore the evolution law of flood processes;

[0221] S3: Based on the mined evolution law, a step function is constructed to map the flood sequence into a smooth signal;

[0222] S4: Construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take the smoothed signal as input, use the sparse attention mechanism to select key features, and make predictions.

[0223] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

[0224] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A flood intelligent forecasting method integrating physical characteristics of flood process information, characterized in that: The method comprises: S1: Using the hydrological and meteorological data in the study basin, the Kalman filter data assimilation method is used to obtain the quantitative relationship between the flood process and the influencing factors; S2: Based on the quantitative relationship, the TVF–EMD modal decomposition method is used to explore the evolution law of flood processes; S3: Based on the mined evolution law, a step function is constructed to map the flood sequence into a smooth signal; S4: Construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take the smoothed signal as input, use the sparse attention mechanism to select key features, and make predictions.

2. The method for intelligent flood forecasting integrating physical characteristics of flood process information according to claim 1 is characterized in that: In step S1, the hydrological and meteorological data include: data of multiple time scales such as hour, day, month and year, and the time scales of the time scale data are consistent.

3. The method for intelligent flood forecasting integrating physical characteristics of flood process information according to claim 1 is characterized in that: In the step S1, a Kalman filter robustification method based on approximate Bayesian estimation is used to process the underlying surface, hydrological and meteorological raw data; including: S101: Data preprocessing: First, preprocess the collected data, including removing obvious erroneous data and filling missing values; S102: Establish state space model: Establish state space model according to the dynamic changes of underlying surface, hydrological and meteorological data; state space model includes state equation and observation equation; state equation describes the evolution of system state over time, and observation equation links system state with observation data; for soil moisture model, state equation is expressed as: x k =Fx k-1 +This k +w k In the formula, x k is the system state at time k, F is the state transition matrix, u k is the control input, B is the control matrix, w k is the process noise, usually assumed to be Gaussian white noise; S103: Apply Kalman filtering: input the preprocessed data into the established model, and apply the Kalman filtering method based on approximate Bayesian estimation to perform state estimation; S104: Result analysis: Analyze the estimation results after filtering and evaluate the accuracy and robustness of the model.

4. The method for intelligent flood forecasting integrating physical characteristics of flood process information according to claim 1 is characterized in that: The steps of Kalman filtering include two stages: prediction and update: The formula for the prediction phase is: P k|k-1 =FP k-1|k-1 F T +Q In the formula, is the state prediction estimate at time k, P k|k-1 is the covariance matrix of the forecast error, Q is the covariance matrix of the process noise; The formula for the update phase is: K k =P k|k-1 H T (HP k|k-1 H T +R) -1 P k =(I-K k H)P k|k-1 In the formula, K k is the Kalman gain, z k is the observation at time k, H is the observation matrix, and R is the covariance matrix of the observation noise.

5. The method for intelligent flood forecasting integrating information and physical characteristics of flood process according to claim 1 is characterized in that: The step S2 specifically includes: Using Hilbert transform to calculate the instantaneous amplitude A(t) and instantaneous frequency of the signal sequence X(t) The calculation formula is: In the formula, is the Hilbert transform of the signal sequence X(t); Determine the local maximum value A({t max {}}()) and the local minimum A({t min {}}()); For A({t min {}}()) and A({t max {}}()) are interpolated respectively, and the obtained curves are β1(t) and β2(t), and then the instantaneous mean a1(t) and instantaneous envelope a2(t) are calculated. The calculation formula is: right and Interpolate respectively to get η1(t) and η2(t), and then calculate and The calculation formula is: Calculate the local cutoff frequency The calculation formula is: Rearrange To solve the intermittent problem; Defining Signals The extreme points of h(t) are used as nodes, and the signal sequence is approximated by B-spline interpolation. The approximation result is m(t), which represents the local mean curve. Calculate the cutoff criterion θ(t). If θ(t)≤ξ, then determine that X(t) is an IMF. Otherwise, let x1(t)=x(t)-m(t) and repeat the above steps. The calculation formula of the cutoff criterion θ(t) is: In the formula, B Loughlin (t) is the Loughlin instantaneous bandwidth of the two component signals; is the weighted average instantaneous frequency of a single component; After the above steps, the value of the original signal sequence X(t) is all modal components X i (t) and.

6. The method for intelligent flood forecasting integrating physical characteristics of flood process information according to claim 1 is characterized in that: The step S3 specifically includes: S301: Setting flood step value; S302: Conversion sequence, the formula is as follows: In the formula, Q j is the flood step value, To round down, Q o is the flood observation value, and J is the step value.

7. The method for intelligent flood forecasting integrating physical characteristics of flood process information according to claim 1 is characterized in that: In the step S4, the construction of a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements refers to respectively constructing a time dimension hidden layer and a space dimension hidden layer to enhance the model's ability to extract features of information of different dimensions; The sparse attention mechanism is used to select key features and make predictions, including: S401: Constructing a temporal attention mechanism: First, use RNN to process time series data and obtain the hidden state h at each time step. t : h t =σ(W hh h t-1 +W xh x t +b h ) In the formula, h t is the hidden state at time step t, h t-1 is the hidden state at time step t-1, x t is the input feature at time step t, W hh is the weight matrix of the hidden state, W xh is the weight matrix input to the hidden state, b h is the bias of the hidden layer, σ is the activation function; Calculate the temporal attention weight α t : In the formula, e t =v T tanh(Wh t +b), W, v and b are the parameters to be learned; Weight the hidden states at each time step to get the output of the temporal attention: S402: Building a spatial attention mechanism: Each position i has a feature representation s at each time point. ti ; At each time point t, the observations x at all locations t1 ,x t2 ...x tn Concatenate to form an input vector x t ; Use this combined vector as the input of the RNN to obtain the overall hidden state h t ; x t =Concat(x t1 ,x t2 ,...,x tn ) h t =RNN(x t ,h t-1 ) Then, to obtain the feature representation s for each position t , applying a fully connected layer from the global hidden state h t Extract and transform information from s ti =W i h t +b i Where W i and b i is a learning parameter specific to each position, used to obtain the value from the overall hidden state h t Extract specific features of the location; Calculate the spatial attention weight β i : Where, d i =u T tanh(Vs ti +c), V, u and c are the parameters to be learned; Weight the feature vectors at each spatial position to obtain the output of spatial attention: S403: Feature Fusion: The results of temporal and spatial attention are fused to form a comprehensive feature representation: f t =Concat(h t ',s t ') S404: Use sparse attention mechanism to select the most critical features and make predictions: Calculate sparse attention weights: c k =Sparsemax(g k ) In the formula, g k =z T tanh(Xf t +e), z, X and e are learning parameters, Δ K-1 represents a K-dimensional probability simplex; S405: Use sparse attention weights to weight the merged features and make final predictions: In the formula, γ k is the sparse attention weight, f t is the comprehensive eigenvector, W y , b y are the parameters of the output layer.

8. A flood intelligent forecasting system integrating the physical characteristics of flood process information, characterized in that: The system comprises: Data assimilation and quantitative relationship extraction module: used to use the hydrological and meteorological data in the study basin and adopt the Kalman filter data assimilation method to obtain the quantitative relationship between the flood process and the influencing factors; Dynamic evolution law mining module: used to mine the evolution law of flood process based on the quantitative relationship by using TVF-EMD modal decomposition method; Signal smoothing and conversion module: used to construct a step function based on the mined evolution law and map the flood sequence into a smooth signal; Feature selection and flood prediction module: used to construct a two-dimensional hidden layer that conforms to the spatiotemporal physical characteristics of hydrological and meteorological elements, take smooth signals as input, use sparse attention mechanism to select key features, and make predictions.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, an intelligent flood forecasting method integrating physical characteristics of flood process information as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a flood intelligent forecasting method that integrates physical characteristics of flood process information as described in any one of claims 1 to 7.

Images