A multivariable and multi-step runoff prediction method based on dynamic variational mode decomposition

Abstract

The present invention provides a multivariate multi-step runoff prediction method based on dynamic variational mode decomposition, which belongs to the technical field of basin runoff prediction. The method includes collecting a variety of hydrological and meteorological data of the basin under study; using the dynamic variational mode decomposition strategy VMD to decompose the original runoff sequence, performing sliding window slicing processing on the basin underlying surface and meteorological data and the runoff data of the upstream hydrological station, combining the decomposed target hydrological station runoff data subsequence as the first characteristic variable; screening the meteorological factor of the meteorological station data and using the screening result as the second characteristic variable, combining the first characteristic variable and the second characteristic variable together as the input characteristic variable; optimizing the long short-term memory network LSTM parameters by using the sparrow search algorithm SSA to construct a long short-term memory network LSTM, and predicting future runoff. The present invention effectively overcomes the deficiency of bundling decomposition to introduce future forecast information, thereby enhancing the applicability of the decomposition-integrated forecast model in practical engineering.

Description

Technical Field

[0001] The invention belongs to the technical field of basin runoff prediction, and in particular relates to a multivariate multi-step runoff prediction method based on dynamic variational mode decomposition. Background Art

[0002] Basin runoff prediction plays a vital role in water conservancy projects such as flood control, waterlogging prevention, and water resources management. The current research methods for runoff prediction are mainly process-driven hydrological models and data-driven models. In recent years, the signal decomposition algorithm combined with the deep learning model integrated forecasting model is an effective method for hydrological forecasting. However, in practical applications, the signal decomposition in the decomposition-integration model has the problem of "bundling decomposition", which will lead to the leakage of future runoff information and make the model prediction results falsely high. In addition, the traditional LSTM model focuses on studying the relationship between precipitation and runoff at a single hydrological station, and rarely pays attention to the relationship between upstream and downstream runoff. The timeliness and accuracy of hydrological prediction under extreme sudden climate conditions are insufficient.

[0003] In the field of runoff prediction, signal decomposition technology combined with deep learning models, such as the traditional VMD-LSTM hybrid model, has certain advantages in processing complex nonlinear and non-stationary time series data, but it still has many shortcomings in practical applications.

[0004] The main disadvantages of the traditional VMD-LSTM model:

[0005] 1) Future information leakage and bundled decomposition. Traditional VMD often relies on the information of the entire series when decomposing time series. This global decomposition method may introduce future information, leading to the phenomenon of "bundled decomposition", which weakens the model's ability to identify the real runoff pattern and affects the accuracy and reliability of the prediction.

[0006] 2) High sensitivity to parameter selection. The decomposition effect of the VMD algorithm depends on the selection of parameters, especially the number of modes (K) and the bandwidth limit parameter (α). In the traditional VMD-LSTM model, these parameters are usually determined by experience or grid search, which may lead to inaccurate decomposition or long parameter optimization time, lack of systematic optimization methods, and thus affect the training effect and prediction performance of the LSTM model.

[0007] 3) Limitations of static decomposition strategy. Traditional VMD uses a static decomposition method to decompose the entire time series at one time. This method cannot adapt to the dynamic changes and non-stationary characteristics that may exist in the time series, resulting in the decomposition results being difficult to capture the key features of real-time changes, and reducing the model's ability to respond to sudden runoff changes. Summary of the invention

[0008] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multivariable and multi-step runoff prediction method based on dynamic variational mode decomposition, which solves the information leakage and single-input and single-step output problems of traditional VMD-LSTM runoff prediction.

[0009] In order to achieve the above objectives, the technical solution adopted by the present invention is: a multivariate multi-step runoff prediction method based on dynamic variational mode decomposition, comprising the following steps:

[0010] S1. Collect various hydrological and meteorological data of the study basin;

[0011] S2, using the dynamic variational mode decomposition strategy VMD to decompose the original runoff sequence of the target hydrological station, and determine the parameters of the dynamic variational mode decomposition strategy VMD;

[0012] S3, based on the hydrological and meteorological data, the basin underlying surface and meteorological data and the runoff data of the upstream hydrological station of the target hydrological station are processed by sliding window slicing, and the subsequence of the runoff data of the target hydrological station obtained by VMD decomposition is combined into the first characteristic variable;

[0013] S4, screening meteorological factors of the meteorological station data, selecting meteorological factors whose cumulative contribution is greater than a preset threshold as the second characteristic variable, combining the first characteristic variable with the second characteristic variable and dividing them into a training set, a test set and a validation set;

[0014] S5. Optimize the parameters of the long short-term memory network LSTM by using the sparrow search algorithm SSA to construct the long short-term memory network LSTM;

[0015] S6. Use the training set to train the long short-term memory network LSTM, and use the test validation and validation sets for testing and validation respectively;

[0016] S7. Use the tested and verified long short-term memory network LSTM to predict future runoff.

[0017] Beneficial effects of the present invention: The present invention proposes a multivariate and multi-step runoff prediction method based on a dynamic variational mode decomposition strategy VMD and a long short-term memory network LSTM. The model continuously adjusts relevant parameters of the decomposition method according to new information, and simultaneously considers the runoff relationship between upstream and downstream hydrological stations, as well as precipitation, temperature radiation and basin underlying surface information, to achieve multi-forecast period runoff prediction, which can effectively overcome the deficiency of bundled decomposition in introducing future forecast information, thereby enhancing the applicability of the decomposition-integrated forecast model in actual engineering.

[0018] Furthermore, the S2 is specifically:

[0019] Define the modal decomposition objective: represent the signal of the original runoff sequence of the target hydrological station as the superposition of several intrinsic mode functions;

[0020] Construct a variational problem: Estimate the bandwidth of each sub-signal through square regularization and Gaussian smoothing, and constrain the sum of the bandwidths of each sub-signal to be the smallest, and obtain the variational model of the bandwidth of each sub-signal;

[0021] Introducing Lagrange multipliers and optimization constraints: Based on the variational model, the bandwidth and center frequency of the mode are used as constraints, and the quadratic penalty factor and Lagrange operator are introduced to transform the constructed constrained variational problem into an unconstrained variational problem;

[0022] Iterative optimization process: The Lagrangian function is iteratively optimized by the alternating direction multiplier method, and the intrinsic mode function, center frequency and Lagrangian operator are updated alternately until the preset convergence conditions are met;

[0023] Extracting intrinsic mode functions: After the iterative optimization process is completed, each intrinsic mode function is obtained to complete the decomposition of the original runoff sequence of the target hydrological station;

[0024] Determine parameters: The Bayesian optimization algorithm is used to determine the parameters in the dynamic variational mode decomposition strategy VMD.

[0025] The beneficial effect of the above further scheme is that the present invention shows significant advantages in runoff prediction. By introducing the dynamic variational mode decomposition strategy VMD, the model can decompose the runoff sequence in real time under the sliding window mechanism, effectively avoiding the common future information leakage and "bundled decomposition" problems in the traditional VMD method.

[0026] Furthermore, the S5 is specifically:

[0027] Initialization phase: determine the structural parameters of the long short-term memory network LSTM, set the parameters of the sparrow search algorithm SSA, and initialize the sparrow population, where each sparrow individual represents a set of parameters of the long short-term memory network LSTM;

[0028] Define the fitness function: The prediction error of the long short-term memory network LSTM on the validation set is used as the fitness function evaluation criterion;

[0029] Optimize the sparrow search algorithm SSA: including the discoverer stage, follower stage, alert stage and update stage:

[0030] Discoverer stage: According to the positioning behavior of the discoverer in the current population, guide the search direction, adjust the LSTM parameter set of the LSTM network, and generate a new LSTM parameter combination of the LSTM network;

[0031] Follower stage: simulate the follower's following behavior to the discoverer and optimize the parameters of the long short-term memory network LSTM;

[0032] Alert stage: monitor the search process;

[0033] Update phase: According to the fitness function evaluation results, the positions and parameters of each individual in the sparrow population are updated to approach the global optimal solution;

[0034] Parameter optimization and model training: In each iteration, the sparrow search algorithm SSA is used to generate new LSTM parameter combinations, train and evaluate the LSTM, and select the optimal parameter set according to the fitness function until the preset termination condition is met;

[0035] Model output: Select the LSTM parameter set with the best fitness function value to build the LSTM network.

[0036] The beneficial effect of the above further scheme is: the present invention adopts the Bayesian optimization algorithm and the sparrow search algorithm SSA to systematically optimize the key parameters of the dynamic variational mode decomposition strategy VMD and the long short-term memory network LSTM, which not only improves the efficiency and accuracy of parameter selection, reduces the risk of inaccurate decomposition and long parameter optimization time, but also enhances the overall stability and generalization ability of the model.

[0037] Furthermore, the expression of the long short-term memory network LSTM is as follows: ; ; ; ; ; ; ;in, Indicates the runoff prediction result output at the current moment, represents the weight matrix of the output layer, represents the output layer bias term, Represents the hidden state of the long short-term memory network LSTM at the current moment, represents the output vector of the output gate, Indicates the current cell state. represents the hyperbolic tangent activation function, represents the matrix element-wise product, represents the Sigmoid activation function, represents the weight matrix of the output gate, Represents the hidden state of the long short-term memory network LSTM at the previous moment, It represents the input feature vector received by the long short-term memory network LSTM at the current moment. The input feature vector includes the basin average NDVI, basin average precipitation, basin average temperature, basin average air pressure, basin average specific humidity, basin average long-wave radiation, basin average short-wave radiation, meteorological station precipitation, meteorological station relative humidity, meteorological station actual evapotranspiration, daily runoff data of four upstream hydrological stations, and the dynamic variational mode decomposition strategy VMD to decompose the target hydrological station runoff data subsequence. represents the output gate bias term, represents the output vector of the forget gate, Indicates the cell state at the previous moment, represents the input gate output vector, represents the new candidate cell state, represents the candidate cell state weight matrix, represents the candidate value bias term, represents the input gate weight matrix, represents the input layer bias term, represents the forget gate weight matrix, Represents the bias term of the forget gate.

[0038] The beneficial effects of the above further scheme are: the long short-term memory network LSTM integrates multivariate inputs, including various hydrological and meteorological data such as precipitation, temperature, air pressure, radiation, and normalized vegetation index, which comprehensively reflects the multiple factors affecting runoff and significantly improves the ability to capture runoff changes under complex environmental conditions. In addition, the long short-term memory network LSTM has multi-forecasting capabilities and can simultaneously provide runoff forecast results for multiple days in the future to meet the needs of different application scenarios, enhancing its practicality and flexibility. In summary, the present invention has obvious technical advantages over traditional methods in terms of prediction accuracy, parameter optimization efficiency, multivariate fusion and multi-forecast capabilities, showing broad application prospects in actual hydrological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0040] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0041] Example

[0042] like Figure 1 As shown, the present invention provides a multivariate multi-step runoff prediction method based on dynamic variational mode decomposition, wherein the multivariate is reflected in the input data having multiple runoff data, including the previous runoff data of the target hydrological station and the runoff data of the upstream hydrological station, the basin underlying surface information, the basin average meteorological conditions, and the meteorological station data; the multi-step is reflected in the forecast period of 3 days, and the runoff results of 1-3 days can be output. The implementation method of the present invention is as follows:

[0043] S1. Collect various hydrological and meteorological data of the study basin;

[0044] In this example, hydrological and meteorological data such as precipitation, temperature, air pressure, radiation and runoff in the study basin and daily data of vegetation information of normalized difference vegetation index are collected. The data includes reanalysis data and site measured data, and the average precipitation of the basin is obtained by processing. Temperature , air pressure Specific humidity , long wave radiation , short wave radiation and Normalized Difference Vegetation Index (NDVI), and precipitation measured at 19 meteorological stations ,humidity , evaporation The data and daily runoff data of the target hydrological station and its upstream four hydrological stations. The specific data and spatial resolution are shown in Table 1:

[0045] Table 1

[0046]

[0047] S2. The original runoff sequence of the target hydrological station is decomposed using the dynamic variational mode decomposition strategy VMD, and the parameters of the dynamic variational mode decomposition strategy VMD are determined, which are specifically:

[0048] Define the modal decomposition objective: represent the signal of the original runoff sequence of the target hydrological station as the superposition of several intrinsic mode functions;

[0049] Construct a variational problem: Estimate the bandwidth of each sub-signal through square regularization and Gaussian smoothing, and constrain the sum of the bandwidths of each sub-signal to be the smallest, and obtain the variational model of the bandwidth of each sub-signal;

[0050] Introducing Lagrange multipliers and optimization constraints: Based on the variational model, the bandwidth and center frequency of the mode are used as constraints, and the quadratic penalty factor and Lagrange operator are introduced to transform the constructed constrained variational problem into an unconstrained variational problem;

[0051] Iterative optimization process: The Lagrangian function is iteratively optimized by the alternating direction multiplier method, and the intrinsic mode function, center frequency and Lagrangian operator are updated alternately until the preset convergence conditions are met;

[0052] Extracting intrinsic mode functions: After the iterative optimization process is completed, each intrinsic mode function is obtained to complete the decomposition of the original runoff sequence of the target hydrological station;

[0053] Determine parameters: The Bayesian optimization algorithm is used to determine the parameters in the dynamic variational mode decomposition strategy VMD.

[0054] In this embodiment, the dynamic variational mode decomposition strategy VMD is used to decompose the original runoff sequence of the target hydrological station. Subsequence ,in, Indicates subsequences, K subsequences as different feature variables. The Bayesian optimization algorithm is used to determine the dynamic variational mode decomposition strategy VMD and Parameters. The specific steps of VMD decomposition are as follows:

[0055] Define the modal decomposition objective: transform the input signal Expressed as several eigenmode functions The superposition of, that is: ;in, represents the original runoff data of the target hydrological station, represents the mode number, k Indicates k The number of modes.

[0056] Constructing a variational problem: estimating each sub-signal through square regularization and Gaussian smoothing The bandwidth of each sub-signal is constrained to be the smallest, and the variational model of the bandwidth of each sub-signal can be expressed as: ;in, Indicates k The center frequency of the mode, represents the Dirichlet function, Indicates time, represents the imaginary unit, Represents a convolution operation.

[0057] Introducing Lagrange multipliers and optimization constraints: Taking the bandwidth and center frequency of the mode as constraints, introducing the quadratic penalty factor and Lagrange operator, the constrained variational problem is transformed into an unconstrained variational problem. The augmented Lagrange operator is: ;in, represents the Lagrangian operator, represents the Lagrange multiplier, represents the quadratic penalty factor.

[0058] Iterative optimization process: Iteratively optimize the Lagrangian function through the alternating direction method of multipliers (ADMM), and alternately update the intrinsic mode function , center frequency and Lagrangian operator , looking for the "saddle point" of the extended Lagrangian expression. Iterate the above steps until the preset convergence conditions are met.

[0059] Extracting intrinsic mode functions: After the optimization process is completed, each intrinsic mode function is obtained ,These modal functions have clear center frequencies and limited bandwidths, and can effectively reflect the different frequency components of the signal.

[0060] S3, based on the hydrological and meteorological data, the basin underlying surface and meteorological data and the runoff data of the upstream hydrological station of the target hydrological station are processed by sliding window slicing, and the subsequence of the runoff data of the target hydrological station obtained by VMD decomposition is combined into the first characteristic variable;

[0061] In this embodiment, the basin underlying surface, meteorological data, and runoff data of the hydrological station upstream of the target hydrological station are processed by sliding window slicing. The CMFD precipitation, temperature, radiation, air pressure, specific humidity and other climate factor raster data and the precipitation, evaporation, humidity and other data of multiple meteorological stations are processed by sliding window slicing and taken as the first characteristic variable.

[0062] S4, screening meteorological factors of the meteorological station data, selecting meteorological factors whose cumulative contribution is greater than a preset threshold as the second characteristic variable, combining the first characteristic variable with the second characteristic variable and dividing them into a training set, a test set and a validation set;

[0063] In this example, the LightGBM lightweight gradient boosting machine is used to screen meteorological factors of meteorological station data, and meteorological factors with a cumulative contribution greater than 70% are selected as the second characteristic variable. The sequence data, the basin underlying surface after sliding window slicing processing, meteorological data, and the runoff data of the upstream hydrological station of the target hydrological station are combined with the second characteristic variable as the input characteristic variable. The input characteristic variables are divided into training set, validation set and test set, which are used for training and parameter adjustment of long short-term memory network LSTM, hyperparameter adjustment and evaluation of model performance, respectively.

[0064] S5. The long short-term memory network LSTM is constructed by optimizing the parameters of the long short-term memory network LSTM using the sparrow search algorithm SSA, which is specifically as follows:

[0065] Initialization phase: determine the structural parameters of the long short-term memory network LSTM, set the parameters of the sparrow search algorithm SSA, and initialize the sparrow population, where each sparrow individual represents a set of parameters of the long short-term memory network LSTM;

[0066] Define the fitness function: The prediction error of the long short-term memory network LSTM on the validation set is used as the fitness function evaluation criterion;

[0067] Optimize the sparrow search algorithm SSA: including the discoverer stage, follower stage, alert stage and update stage:

[0068] Discoverer stage: According to the positioning behavior of the discoverer in the current population, guide the search direction, adjust the LSTM parameter set of the LSTM network, and generate a new LSTM parameter combination of the LSTM network;

[0069] Follower stage: simulate the follower's following behavior to the discoverer and optimize the parameters of the long short-term memory network LSTM;

[0070] Alert stage: monitor the search process;

[0071] Update phase: According to the fitness function evaluation results, the positions and parameters of each individual in the sparrow population are updated to approach the global optimal solution;

[0072] Parameter optimization and model training: In each iteration, the sparrow search algorithm SSA is used to generate new LSTM parameter combinations, train and evaluate the LSTM, and select the optimal parameter set according to the fitness function until the preset termination condition is met;

[0073] Model output: Select the LSTM parameter set with the best fitness function value to build the LSTM network.

[0074] In this embodiment, the time step of the long short-term memory network LSTM is set to 30 days, and the sparrow search algorithm SSA is used to optimize the hyperparameters of the long short-term memory network LSTM, such as the number of hidden layers, dropout method, and learning rate. The process of using the sparrow search algorithm SSA to optimize the parameters is as follows:

[0075] Initialization phase: determine the structural parameters of the long short-term memory network LSTM, including the number of layers, the number of neurons in each layer, etc.; set the relevant parameters of the sparrow search algorithm SSA, such as population size, maximum number of iterations, the ratio of discoverers to followers, etc.; randomly initialize the sparrow population, and each sparrow individual represents a set of parameters of the long short-term memory network LSTM (including weights, biases and other hyperparameters).

[0076] Fitness function design: Define the fitness function and use the prediction error (such as mean square error, cross entropy, etc.) of the long short-term memory network LSTM on the validation set as the fitness evaluation standard. The smaller the fitness function, the higher the individual fitness.

[0077] The optimization process of the sparrow search algorithm: discoverer stage: according to the positioning behavior of the discoverer in the current population, guide the search direction, adjust the LSTM parameter set of the long short-term memory network, and generate a new parameter combination; follower stage: simulate the follower's following behavior of the discoverer, further explore the parameter space, and optimize the LSTM parameters of the long short-term memory network; alert stage: monitor potential threats in the search process, avoid the algorithm from falling into local optimality, and enhance the diversity of the population through appropriate disturbance mechanisms; update stage: according to the fitness function evaluation results, update the position and parameters of each individual in the sparrow population, and gradually approach the global optimal solution.

[0078] Parameter optimization and model training: In each iteration, the sparrow search algorithm SSA is used to generate new long short-term memory network LSTM parameter combinations, and the long short-term memory network LSTM is trained and evaluated; the optimal parameter set is selected according to the fitness function, and it is continuously iterated until the preset termination condition is met (such as reaching the maximum number of iterations or the fitness function converges).

[0079] Final model output: Select the LSTM parameter set with the best fitness function value, build the final LSTM deep learning model, and achieve efficient prediction and analysis of the target task.

[0080] In this embodiment, the hyperparameters obtained by optimizing the sparrow search algorithm SSA are used to establish a long short-term memory network LSTM, and a multi-forecast period scheme is set to train the model.

[0081] In this embodiment, the storage unit It is the core component of the LSTM network. At each time step, the input data and the output data of the previous moment (in this solution, the output data of the previous moment refers to the prediction result of the LSTM network on the runoff of the target hydrological station at the previous moment, that is, the hidden layer output at the previous moment) will enter the LSTM network and be used to calculate the values ​​of the gate unit and memory unit. At the current moment t, the LSTM inputs are: current network input value (include t The basin average NDVI, basin average precipitation, basin average temperature, basin average air pressure, basin average specific humidity, basin average long-wave radiation, basin average short-wave radiation, meteorological station precipitation, meteorological station relative humidity, meteorological station actual evapotranspiration, daily runoff of four upstream hydrological stations, and VMD dynamic decomposition to obtain the target hydrological station runoff data subsequence) 30 days before the moment, the output value of the long short-term memory network LSTM at the previous moment (Long Short-Term Memory Network LSTM t -1 moment to the target hydrological station runoff prediction results), the previous moment memory unit state ; The outputs of the long short-term memory network LSTM are: The output value of the long short-term memory network LSTM at the current moment (Long Short-Term Memory Network LSTM t The prediction result of the runoff of the target hydrological station in the next 1-3 days at the moment), the memory unit status at the current moment and the hidden layer output The calculation process is as follows:

[0082] Forget Gate :Calculate the state variables of the memory unit at the previous moment The degree of information discarded is calculated according to the Sigmoid function as a value between 0 and 1, where 1 means all information is retained and 0 means all information is discarded: .

[0083] Input Gate : Select the current moment input value and the memory unit state variable at the previous moment How much to add to the memory cell state variable middle: .

[0084] Candidate Gate : Calculate the current candidate unit state:

[0085] The current memory unit consists of two parts: one part is selected by the forget gate to remember the information after forgetting; the other part is filtered by the input gate to update the current information: .

[0086] Output Gate : Calculate the current information output level: ; .

[0087] Hidden layer output After being passed to the output layer of the long short-term memory network LSTM, it is calculated t Final output of the moment model : .

[0088] in, Indicates the runoff prediction result output at the current moment, represents the weight matrix of the output layer, represents the output layer bias term, Represents the hidden state of the long short-term memory network LSTM at the current moment, represents the output vector of the output gate, Indicates the current cell state. represents the hyperbolic tangent activation function, represents the matrix element-wise product, represents the Sigmoid activation function, represents the weight matrix of the output gate, Represents the hidden state of the long short-term memory network LSTM at the previous moment, It represents the input feature vector received by the long short-term memory network LSTM at the current moment. The input feature vector includes the basin average NDVI, basin average precipitation, basin average temperature, basin average air pressure, basin average specific humidity, basin average long-wave radiation, basin average short-wave radiation, meteorological station precipitation, meteorological station relative humidity, meteorological station actual evapotranspiration, daily runoff data of four upstream hydrological stations, and the dynamic variational mode decomposition strategy VMD to decompose the target hydrological station runoff data subsequence. represents the output gate bias term, represents the output vector of the forget gate, Indicates the cell state at the previous moment, represents the input gate output vector, represents the new candidate cell state, represents the candidate cell state weight matrix, represents the candidate value bias term, represents the input gate weight matrix, represents the input layer bias term, represents the forget gate weight matrix, Represents the bias term of the forget gate.

[0089] S6. Use the training set to train the long short-term memory network LSTM, and use the test validation set and validation set to test and validate respectively. The forecast period is set to 3 days, and the output results include the runoff of 1-3 days.

[0090] In this example, the performance of the deep learning runoff prediction model on the test set data is evaluated, and the Nash coefficient (NSE), correlation coefficient (R), Kling-Gupta efficiency (KGE), root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the performance of the model in different forecast periods. The expression of each statistical evaluation index is as follows: ; ; ; ;in, and Represent the measured and simulated flow series respectively ( ), and represent the standard deviation of measured and simulated flow rates, and represent the average flow rates of the measured and simulated values, respectively. ), Indicates the number of measured values.

[0091] S7. Use the tested and verified long short-term memory network LSTM to predict future runoff.

[0092] In this embodiment, the present invention can better capture the seasonal fluctuation and peak characteristics of runoff, can capture most of the extreme hydrological events (such as flood peaks) and accurately characterize the amplitude and occurrence time of the peak.

[0093] In summary, the present invention aims to solve the following key technical problems based on the dynamic model VMD-LSTM:

[0094] 1) Avoid future information leakage and bundled decomposition. The dynamic variational mode decomposition strategy is adopted, that is, a dynamic variational mode decomposition strategy VMD decomposition is performed every 30 data points, and then a unit is slid for the next decomposition. This sliding window method ensures that each decomposition is based only on current and past information, avoiding the leakage of future information.

[0095] 2) Optimize parameter selection and reduce parameter sensitivity. The dynamic variational mode decomposition strategy VMD decomposition parameters are determined by the Bayesian optimization algorithm, and the optimal parameter combination is systematically searched, which reduces the subjective judgment and traversal time of parameter selection and improves the stability and consistency of the decomposition process.

[0096] 3) Dynamic decomposition strategy to adapt to non-stationary characteristics. The dynamic variational mode decomposition strategy VMD can capture the dynamic changes and non-stationary characteristics in the time series in real time through sliding window decomposition, ensuring that the decomposition results can reflect the latest status of the data in a timely manner and enhance the model's ability to respond to real-time changes.

[0097] 4) Efficiency of parameter optimization of LSTM network. The sparrow search algorithm SSA is used to optimize the parameters of LSTM network. As a novel swarm intelligence optimization algorithm, the sparrow search algorithm SSA has strong global search capabilities and fast convergence characteristics. It can find a near-optimal parameter combination in a short time, improve the model training efficiency and ensure that the LSTM network has good generalization ability.

[0098] 5) Multi-variable input and multi-forecasting capability. The LSTM network uses multi-variable input, not only considering precipitation information, but also integrating various hydrological and meteorological data such as temperature, air pressure, radiation, and vegetation index, to fully reflect the various factors affecting runoff and enhance the LSTM network's ability to capture runoff changes under complex environmental conditions. In addition, the LSTM network can achieve multi-forecasting, that is, predicting runoff conditions for multiple days in the future at the same time, meeting the needs of short-term and medium-term forecasts in practical applications, and enhancing the practicality and flexibility of the model.

Claims

1. A multivariate multi-step runoff prediction method based on dynamic variational mode decomposition, characterized in that: The following steps are involved: S1. Collect various hydrological and meteorological data of the study basin; S2, using the dynamic variational mode decomposition strategy VMD to decompose the original runoff sequence of the target hydrological station, and determine the parameters of the dynamic variational mode decomposition strategy VMD; S3, based on the hydrological and meteorological data, the basin underlying surface and meteorological data and the runoff data of the upstream hydrological station of the target hydrological station are processed by sliding window slicing, and the subsequence of the runoff data of the target hydrological station obtained by VMD decomposition is combined into the first characteristic variable; S4, screening meteorological factors of the meteorological station data, selecting meteorological factors whose cumulative contribution is greater than a preset threshold as the second characteristic variable, combining the first characteristic variable with the second characteristic variable and dividing them into a training set, a test set and a validation set; S5. Optimize the parameters of the long short-term memory network LSTM by using the sparrow search algorithm SSA to construct the long short-term memory network LSTM; S6. Use the training set to train the long short-term memory network LSTM, and use the test validation and validation sets for testing and validation respectively; S7. Use the tested and verified long short-term memory network LSTM to predict future runoff; The S2 is specifically: Define the modal decomposition objective: represent the signal of the original runoff sequence of the target hydrological station as the superposition of several intrinsic mode functions; Construct a variational problem: Estimate the bandwidth of each sub-signal through square regularization and Gaussian smoothing, and constrain the sum of the bandwidths of each sub-signal to be the smallest, and obtain the variational model of the bandwidth of each sub-signal; Introducing Lagrange multipliers and optimization constraints: Based on the variational model, the bandwidth and center frequency of the mode are used as constraints, and the quadratic penalty factor and Lagrange operator are introduced to transform the constructed constrained variational problem into an unconstrained variational problem; Iterative optimization process: The Lagrangian function is iteratively optimized by the alternating direction multiplier method, and the intrinsic mode function, center frequency and Lagrangian operator are updated alternately until the preset convergence conditions are met; Extracting intrinsic mode functions: After the iterative optimization process is completed, each intrinsic mode function is obtained to complete the decomposition of the original runoff sequence of the target hydrological station; Determine parameters: The Bayesian optimization algorithm is used to determine the parameters in the dynamic variational mode decomposition strategy VMD.

2. The multivariate multi-step runoff prediction method based on dynamic variational mode decomposition according to claim 1 is characterized in that: The S5 is specifically: Initialization phase: determine the structural parameters of the long short-term memory network LSTM, set the parameters of the sparrow search algorithm SSA, and initialize the sparrow population, where each sparrow individual represents a set of parameters of the long short-term memory network LSTM; Define the fitness function: The prediction error of the long short-term memory network LSTM on the validation set is used as the fitness function evaluation criterion; Optimize the sparrow search algorithm SSA: including the discoverer stage, follower stage, alert stage and update stage: Discoverer stage: According to the positioning behavior of the discoverer in the current population, guide the search direction, adjust the LSTM parameter set of the LSTM network, and generate a new LSTM parameter combination of the LSTM network; Follower stage: simulate the follower's following behavior to the discoverer and optimize the parameters of the long short-term memory network LSTM; Alert stage: monitor the search process; Update phase: According to the fitness function evaluation results, the positions and parameters of each individual in the sparrow population are updated to approach the global optimal solution; Parameter optimization and model training: In each iteration, the sparrow search algorithm SSA is used to generate new LSTM parameter combinations, train and evaluate the LSTM, and select the optimal parameter set according to the fitness function until the preset termination condition is met; Model output: Select the LSTM parameter set with the best fitness function value to build the LSTM network.

3. The multivariable multi-step runoff prediction method based on dynamic variational mode decomposition according to claim 2 is characterized in that: The expression of the long short-term memory network LSTM is as follows: in, Indicates the runoff prediction result output at the current moment, represents the weight matrix of the output layer, represents the output layer bias term, Represents the hidden state of the long short-term memory network LSTM at the current moment, represents the output vector of the output gate, Indicates the current cell state. represents the hyperbolic tangent activation function, represents the matrix element-wise product, represents the Sigmoid activation function, represents the weight matrix of the output gate, Represents the hidden state of the long short-term memory network LSTM at the previous moment, It represents the input feature vector received by the long short-term memory network LSTM at the current moment. The input feature vector includes the basin average NDVI, basin average precipitation, basin average temperature, basin average air pressure, basin average specific humidity, basin average long-wave radiation, basin average short-wave radiation, meteorological station precipitation, meteorological station relative humidity, meteorological station actual evapotranspiration, daily runoff data of four upstream hydrological stations, and the dynamic variational mode decomposition strategy VMD to decompose the target hydrological station runoff data subsequence. represents the output gate bias term, represents the output vector of the forget gate, Indicates the cell state at the previous moment, represents the input gate output vector, represents the new candidate cell state, represents the candidate cell state weight matrix, represents the candidate value bias term, represents the input gate weight matrix, represents the input layer bias term, represents the forget gate weight matrix, Represents the bias term of the forget gate.

Images