Flood simulation real-time correction method considering system non-Gaussian property

Abstract

The invention discloses a flood simulation real-time correction method considering system non-Gaussian property. The method comprises the following steps: constructing a hydrological model for flood simulation in a basin to be simulated; constructing an ensemble Kalman filter coupled with the hydrological model; training a long short-term memory recurrent neural network (LSTM) model based on the hydrological model and an ensemble Kalman filter to process the non-Gaussian property of the hydrological model; and based on the trained LSTM model, carrying out real-time correction on a hydrological model forecast result. The method has the advantages of stable and reliable data source, high calculation efficiency, objective and reasonable result and the like, and the simulation forecast precision of the hydrological model on the flood process is improved.

Description

Technical Field

[0001] The present invention relates to hydrology and deep learning technology, and in particular to a real-time correction method for flood simulation considering system non-Gaussianity. Background Art

[0002] Against the backdrop of global climate change, extreme events are on the rise, and the terrestrial hydrological cycle system is undergoing rapid and profound changes, which are reflected in the non-steady-state characteristics of river basin hydrological processes and frequent floods and waterlogging disasters.

[0003] Flood forecasting is one of the main non-engineering flood prevention and disaster reduction measures. Generally, flood forecasting schemes are built based on historical data. When the actual situation is different from the historical conditions, forecast errors may occur. At this time, it is necessary to update and correct the original forecast results in real time based on the newly obtained observation information. This process of adjusting flood forecasts based on real-time information is called real-time correction of flood forecasts. State variable correction is a widely used theory in real-time correction, that is, the sequential assimilation method is used to gradually absorb the observed river section flow in the forecast process to update the state variables of the hydrological model, so as to achieve real-time correction of flood forecasts.

[0004] The hydrological model involves the simulation of multiple complex physical processes and is a complex high-dimensional nonlinear system in which parameters, simulation outputs, measurement errors, etc. are non-Gaussian. Most of the existing real-time correction methods based on sequential assimilation technology are subject to the high-dimensional nonlinearity and non-Gaussianity of the system to varying degrees, such as the Gaussian distribution assumption of the ensemble Kalman filter, the particle weight degradation of the particle filter in the high-dimensional nonlinear system, etc., which is not conducive to accurate flood forecasting based on the hydrological model. In comparison, the deep learning model is good at extracting complex features in the system and learning non-Gaussian relationships from the data, but it is difficult to handle the non-stationarity of the background error covariance and the results lack interpretability. For example, patent CN117851746A discloses a real-time correction method for time-varying parameters of a hydrological model based on an ensemble Kalman filter. The prior art uses the ensemble Kalman filter technology to update the state of the hydrological model, thereby achieving the purpose of predicting the flood process, but the scheme does not take into account that the non-Gaussianity in the hydrological model may violate the Gaussianity assumption required by the ensemble Kalman filter, which may have an adverse effect on the accuracy of the flood forecast results.

[0005] In view of the above shortcomings, how to combine the existing state variable correction method with the deep learning model to alleviate the constraints of non-Gaussianity in the hydrological model in the real-time correction of flood forecasts and achieve high-precision real-time correction of flood forecasts is exactly the problem that the inventors need to solve. Summary of the invention

[0006] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a real-time correction method for flood simulation taking into account the non-Gaussianity of the system.

[0007] Technical solution: A real-time correction method for flood simulation considering system non-Gaussianity of the present invention comprises the following steps:

[0008] Step 1: Construct a hydrological model M(.) for flood simulation in the basin to be simulated. The hydrological model M(.) has the function of calculating runoff and confluence; obtain the input observed rainfall and evaporation data U, the optimal parameter set and the initial value of the model state

[0009] Step 2: Construct an ensemble Kalman filter coupled with the hydrological model. The specific method is as follows:

[0010] Step 2.1: At time t = 1, the initial value of the model state Apply Gaussian perturbation and repeat N times to obtain the initial value of the model state after perturbation Then set up N hydrological models M that are independent of each other and calculated in parallel in time n (.); then the corresponding input data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow of N groups at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as

[0011] Step 2.2: At time t>1, calculate the predicted state quantities of all N hydrological models M(.) And simulated flow ys n,t ;

[0012] Step 2.3: At time t>1, calculate the analytical state quantities of all N hydrological models M(.) at time t

[0013] Step 2.4: The optimal analytical value of the state variable at time t>1 Set as the ensemble average of the analyzed state quantities, the optimal simulated flow of the model at each moment Set to the ensemble mean of the simulated flow;

[0014] Step 3: Based on the hydrological model M(.) and the ensemble Kalman filter, a long short-term memory recursive neural network LSTM model is trained to transform the flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the LSTM model simulates the traffic ysl t and state quantity Xl t As the output value, the training objective is set to minimize the loss function between the LSTM analysis traffic and the observed traffic until the training is completed;

[0015] Step 4: Use the LSTM model trained in step 3 to perform real-time correction on the forecast results of the hydrological model M(.). The specific method is as follows:

[0016] Step 4.1: When using the hydrological model M(.) for flood forecasting, at the forecast start time tf, repeat steps 2.2 to 2.4 to obtain the corresponding optimal analysis value Optimal simulation flow and flow observation value y tf , and then input it into the trained LSTM model to obtain the collective forecast flow at the forecast start time tf and collection state variables

[0017] Step 4.2: Set the forecast period to LT hours and The ensemble average of is the optimal initial state of the hydrological model at the beginning of the forecast, which is substituted into the hydrological model for flood forecasting.

[0018] Furthermore, the detailed steps of step 1 are as follows:

[0019] Step 1.1, obtain flood data and evaporation data of the hydrological station at the outlet of the basin, interpolate the flood data into a flood data set with a time step of 1 hour by interpolation, evenly distribute the evaporation of the observation time step in time, and interpolate it into an evaporation data set with a time step of 1 hour;

[0020] Step 1.2, obtain the coordinate position and rainfall data of the rain gauge station in the basin, evenly distribute the rainfall of the observation time step in time, interpolate it into a rainfall data set with a time step of 1 hour, and then use the Thiessen polygon method to divide the basin into k sub-basins with the rain gauge station as the center, k is the number of rain gauge stations in the basin, the rainfall of each sub-basin is set to the rainfall data set of the corresponding rain gauge station in the basin, and the evaporation of each sub-basin is set to the evaporation data set of the hydrological station at the outlet of the basin;

[0021] Step 1.3, select a hydrological model M(.) with runoff generation and runoff calculation functions, the number of sub-basins calculated in the hydrological model M(.) is k, the state quantity set representing soil moisture in runoff generation calculation is marked as X1, the state quantity set representing river water storage in runoff calculation is marked as X2, and the model state quantity set is recorded as X = [X1, X2];

[0022] Step 1.4: At the first calculation time t = 1, the initial value of X1 is recorded as X1ini, and the value of each element in X1ini is set to half of the maximum soil moisture content; the initial value of X2 is recorded as X2ini, and the value of each element in X2ini is set to the observed flow of the outlet hydrological station at t = 1 divided by the number of sub-basins, and the two are recorded together as the initial value of the model state quantity

[0023] Step 1.5: Take the observed rainfall and evaporation as the model input data, denoted as U, set the insensitive parameters in the hydrological model M(.) to the recommended values, and use the shuffled complex evolution algorithm SCE-UA to automatically calibrate the parameters of the hydrological model M(.) with the objective function of maximizing the Nash efficiency coefficient between the observed flow and the simulated flow to obtain the optimal parameter set.

[0024] Furthermore, in step 2, in the process of constructing an ensemble Kalman filter coupled with the hydrological model,

[0025] For step 2.1, at time t = 1, first initialize the model state Apply a Gaussian perturbation ω with a mean of 0 and a covariance of Q n , for N hydrological models M that are independent and calculated in parallel in time n (.), let n = 1, 2, ..., N respectively, and repeat the following process: input the model at this moment into the data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow rate of the nth group at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as

[0026] For step 2.2, let n = 1, 2, ..., N and repeat the process one by one until all N hydrological model simulation processes are completed at time t>1, and we get

[0027] ω n That is, the Gaussian noise with a mean of 0 and a covariance of Q added to the model state in step 2.1, and Q is set by expert experience;

[0028] In step 2.3, analyze the state quantity The calculation formula is as follows:

[0029]

[0030]

[0031] In the formula, y tis the observed flow at time t, ys n,t is the flow simulated by the hydrological model at time t, K t is the Kalman gain matrix at time t, v is Gaussian noise with mean 0 and covariance R, and R is set by expert experience;

[0032] Optimal analysis value in step 2.4 And the optimal simulation flow The calculation formula is as follows:

[0033]

[0034] Starting from t=2, repeat steps 2.2 to 2.4 moment by moment until all moments are calculated.

[0035] Furthermore, the model training process in step 3 is as follows:

[0036] Step 3.1: Construct a long short-term memory recursive neural network LSTM model, and use the Adam algorithm as the back propagation optimization algorithm of the model. LSTM consists of multiple neuron layers, and each layer contains multiple neurons. The neurons pass through the gating unit, that is, the input gate i t 、Forget gate t and output gate o t To memorize and learn, and prevent gradient explosion, the calculation formula is as follows:

[0037] i t =σ(W ix x t +W ih h t-1 +b i )

[0038] f t =σ(W fx x t +W fh h t-1 +b f )

[0039] o t =σ(W ox x t +W oh h t-1 +b o )

[0040]

[0041] h t =φ(c t )⊙o t

[0042] In the formula, c t , and h t are the memory state, candidate memory state and hidden state at time t, σ and φ are the changes of sigmoid function and tanh function respectively, W ix , W ih , W fx , W fh , W ox , W oh , are input gate i t 、Forget gate t , output gate o t and candidate memory states The matrix weights of the input and output calculations, b i 、b f 、b o , are input gate i t 、Forget gate t , output gate o t and candidate memory states The bias term of , ⊙ represents the element-wise multiplication of each vector;

[0043] Step 3.2: Randomly select 80% of all the calculation moments as training set samples, and the remaining 20% ​​as validation set samples;

[0044] Step 3.3: The flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the LSTM model simulates the traffic ysl t and state quantity Xl t As the output value, the training goal is set to minimize the loss function between the LSTM analysis flow and the observed flow. The loss function is set to mean squared error loss. Based on the performance of the validation set, the particle swarm algorithm is used to automatically optimize the hyperparameters (learning rate, number of hidden layers, number of neurons, batch size, number of iterations), determine the LSTM model hyperparameters, and complete the LSTM training.

[0045] Furthermore, the collective forecast flow at time tf in step 4.1 is and collection state variables The calculation formula is as follows:

[0046]

[0047] Among them, ysl tfand Xl tf They are respectively the simulated flow and state quantity at time tf output by the LSTM model, is the analytical state quantity at time tf, They are the optimal simulated flow and optimal analysis value of the Kalman filter at time tf, ys n,tf is the hydrological model M n (.)Simulated traffic at time tf.

[0048] Beneficial effects: The present invention introduces deep learning technology into the real-time correction process of flood forecast based on ensemble Kalman filter, and solves the problem of conflict between non-Gaussianity in hydrological model and Gaussianity assumption of ensemble Kalman filter on the basis of ensuring objectivity and reliability of calculation results. In addition, the method mainly uses observed rainfall, observed evaporation and observed flow, and the data source is stable and reliable. The functional relationship between variables in the method is clear, which is conducive to computer automation execution of real-time correction of flood forecast, and is conducive to the use of ensemble Kalman filter to perform real-time correction of flood forecast in hydrological model with non-Gaussianity, and is conducive to the in-depth development of digital hydrological forecast research. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the process of the present invention.

[0050] Figure 2 The sub-basin distribution map of the Tunxi Basin in the embodiment

[0051] Figure 3 This is the LSTM model training flow chart in the embodiment.

[0052] Figure 4 This is the flood forecast result of Tunxi Basin "TX-20150809082010041008" in the example. Figure 4 a is a real-time correction method using only ensemble Kalman filtering; Figure 4 b is the real-time correction effect using the solution of the present invention;

[0053] Figure 5 This is the flood forecast result of Tunxi River Basin “TX-20110609082017062308” in the example. Figure 5 a is a real-time correction method using only ensemble Kalman filtering; Figure 5 b is the real-time correction effect using the solution of the present invention. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0055] like Figure 1As shown, the real-time correction method for flood simulation considering the non-Gaussianity of the system of the present invention comprises the following steps:

[0056] Step 1: Construct a hydrological model M(.) for flood simulation in the basin to be simulated. The hydrological model M(.) has the function of calculating runoff and confluence; obtain the input observed rainfall and evaporation data U, the optimal parameter set and the initial value of the model state

[0057] Step 2: Construct an ensemble Kalman filter coupled with the hydrological model. The specific method is as follows:

[0058] Step 2.1: At time t = 1, the initial value of the model state Apply Gaussian perturbation and repeat N times to obtain the initial value of the model state after perturbation Then set up N hydrological models M that are independent of each other and calculated in parallel in time n (.); then the corresponding input data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow of N groups at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as

[0059] Step 2.2: At time t>1, calculate the predicted state quantities of all N hydrological models M(.) And simulated flow ys n,t ;

[0060] Step 2.3: At time t>1, calculate the analytical state quantities of all N hydrological models M(.) at time t

[0061] Step 2.4: The optimal analytical value of the state variable at time t>1 Set as the ensemble average of the analyzed state quantities, the optimal simulated flow of the model at each moment Set to the ensemble mean of the simulated flow;

[0062] Step 3: Based on the hydrological model M(.) and the ensemble Kalman filter, a long short-term memory recursive neural network LSTM model is trained to transform the flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the LSTM model simulates the traffic ysl t and state quantity Xl tAs the output value, the training objective is set to minimize the loss function between the LSTM analysis traffic and the observed traffic until the training is completed;

[0063] Step 4: Use the LSTM model trained in step 3 to perform real-time correction on the forecast results of the hydrological model M(.). The specific method is as follows:

[0064] Step 4.1: When using the hydrological model M(.) for flood forecasting, at the forecast start time tf, repeat steps 2.2 to 2.4 to obtain the corresponding optimal analysis value Optimal simulation flow and flow observation value y tf , and then input it into the trained LSTM model to obtain the collective forecast flow at the forecast start time tf and collection state variables

[0065] Step 4.2: Set the forecast period to LT hours and The ensemble average of is the optimal initial state of the hydrological model at the beginning of the forecast, which is substituted into the hydrological model for flood forecasting.

[0066] Example

[0067] This embodiment performs a real-time calibration demonstration on the Tunxi hydrological station at the outlet of the Tunxi Basin and the Xin'anjiang hydrological model. The specific method is as follows:

[0068] Step 1: Construct a hydrological model for flood simulation in the basin to be simulated:

[0069] Step 1.1, obtain the flood data and evaporation data of the Tunxi hydrological station at the outlet of the Tunxi River Basin, interpolate the flood data into a flood data set with a time step of 1 hour by interpolation, evenly distribute the evaporation of the observation time step in time, and interpolate it into an evaporation data set with a time step of 1 hour;

[0070] Step 1.2: Obtain the coordinates and rainfall data of 9 rain gauges in the Tunxi Basin, evenly distribute the rainfall of the observation time step in time, interpolate it into a rainfall dataset with a time step of 1 hour, and use the Thiessen polygon method to divide the basin into 9 sub-basins with the rain gauge as the center (see Figure 2 ), the rainfall of each sub-basin is set to the rainfall dataset of the corresponding rain gauge station in the basin, and the evaporation of each sub-basin is set to the evaporation dataset of the hydrological station at the outlet of the basin;

[0071] Step 1.3, the hydrological model Xin'anjiang model is recorded as M(.), the number of calculated sub-basins is 9, the tension water content W is used as the state quantity representing soil moisture, the state quantity set X1 = [W1, W2, ..., W9], the outflow P at the outlet of each sub-basin is used as the state quantity representing the water storage capacity of the river channel, the state quantity set is marked as X2 = [P1, P2, ..., P9], and the model state quantity set is recorded as X = [X1, X2];

[0072] Step 1.4: At the first calculation time t = 1, the initial value of X1 is recorded as X1ini, and the value of each element in X1ini is set to half of the maximum soil moisture content, that is, 80 mm; the initial value of X2 is recorded as X2ini, and the value of each element in X2ini is set to the observed flow rate of the outlet hydrological station at t = 1 divided by the number of sub-basins, that is, 1.67m 3 / s, and the two together are recorded as the initial value of the model state quantity

[0073] Step 1.5: Use observed rainfall and evaporation as model input data, denoted as U.

[0074] According to the Xin'anjiang model manual, the insensitive parameters of the Xin'anjiang model are set as follows: upper layer tension water capacity WU = 20 mm, lower layer tension water capacity WL = 80 mm, basin average tension water capacity WM = 160 mm, deep evapotranspiration conversion coefficient C = 0.17, tension water storage capacity curve power B = 0.3, free water storage capacity curve power EX = 1.5, soil midflow recession coefficient CI = 0.6, Muskingum method calculation parameter XE = 0.4.

[0075] The sensitive parameters of the Xin'anjiang model include the evapotranspiration conversion coefficient K, the free water storage capacity SM, the free water outflow coefficient KI to the subsoil flow, the free water outflow coefficient KG to the groundwater runoff, the groundwater recession coefficient CG, the river network recession coefficient CS and the hysteresis L. The objective function is to maximize the Nash efficiency coefficient between the observed flow and the simulated flow. The shuffled complex evolution algorithm (abbreviated as SCE-UA algorithm) is used to automatically calibrate the sensitive parameters of the Xin'anjiang model, and the optimal parameter set of the sensitive parameters of the Xin'anjiang model is obtained.

[0076]

[0077] Step 2: Construct an ensemble Kalman filter coupled with the hydrological model:

[0078] Step 2.1, at t = 1, apply a Gaussian perturbation with a mean of 0 and a covariance of 5.0 to the initial value X1ini of the model state X1 of the Xinanjiang model, and apply a Gaussian perturbation with a mean of 0 and a covariance of 10.0 to the initial value X2ini of the model state X2, repeat 50 times, and obtain the initial value of the model state after the perturbation Set up 50 independent and time-parallel Xinanjiang models M n (.), let n = 1, 2, ..., 50 respectively, and repeat the following process:

[0079] The model at this moment is input into the data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow rate of the nth group at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as

[0080] Step 2.2: At time t (t>1), the predicted state quantity is calculated by the hydrological model: And simulated flow ys n,t , repeat set by set (i.e., let n = 1, 2, ..., 50) until all 50 hydrological model simulation processes are calculated:

[0081]

[0082] In the formula, ω n To add Gaussian noise with mean 0 and covariance Q to the model state quantity, all non-diagonal elements of Q are 0. For X1, all diagonal elements in Q are 5.0, and for X2, all diagonal elements in Q are 10.0.

[0083] Step 2.3, at time t (t>1), repeat each set (i.e., let n = 1, 2, ..., 50), and calculate the state quantity by the following formula

[0084]

[0085]

[0086] In the formula, y t is the observed flow at time t, ys n,t is the simulated flow of the hydrological model at time t, K t is the Kalman gain matrix at time t, v is Gaussian noise with mean 0 and covariance R. Since there is only one outlet in the Tunxi Basin, the matrix y t 、ys n,t , v and R are both 1 row and 1 column, and R is set to 10.0;

[0087] Step 2.4: Optimal analytical value of state variables at time t (t>1) Set to analyze the state quantity to get the ensemble average value, the optimal simulated flow of the model at each moment Set to the ensemble average of the simulated flow, the expression is as follows:

[0088]

[0089] Step 2.5, starting from t=2, repeat steps 2.2 to 2.4 moment by moment until all moments are calculated.

[0090] Step 3: Train the long short-term memory recursive neural network (LSTM) model based on the hydrological model and the ensemble Kalman filter. The training process is shown in Figure 3 :

[0091] Step 3.1: Construct a long short-term memory recursive neural network model. The Adam algorithm is used as the back propagation optimization algorithm of the model. The LSTM model consists of multiple neuron layers, and each layer contains multiple neurons. The neurons pass through the gating unit, that is, the input gate i t 、Forget gate t and output gate o t To memorize and learn, and prevent gradient explosion, the calculation formula is as follows:

[0092] i t =σ(W ix x t +W ih h t-1 +b i )

[0093] f t =σ(W fx x t +W fh h t-1 +b f )

[0094] o t =σ(W ox x t +W oh h t-1 +b o )

[0095]

[0096] h t =φ(c t )⊙o t

[0097] In the formula, c t , and h tare the memory state, candidate memory state and hidden state at time t, σ and φ are the changes of sigmoid function and tanh function respectively, W ix , W ih , W fx , W fh , W ox , W oh , are input gate i t 、Forget gate t , output gate o t and candidate memory states The matrix weights of the input and output calculations, b i 、b f 、b o , are input gate i t 、Forget gate t , output gate o t and candidate memory states The bias term of , ⊙ represents the element-wise multiplication of each vector;

[0098] Step 3.2: Randomly select 80% of all the calculation moments as training set samples, and the remaining 20% ​​as validation set samples;

[0099] Step 3.3: The flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the LSTM model simulates the traffic ysl t and state quantity Xl t As the output value, the training goal is set to minimize the loss function between the LSTM analysis flow and the observed flow, and the loss function is set to mean squared error loss (Mean Squared Error Loss). Based on the performance of the validation set, the particle swarm algorithm is used to automatically optimize the hyperparameters (learning rate, number of hidden layers, number of neurons, batch size, number of iterations), determine the LSTM model hyperparameters, and complete the LSTM training. The final hyperparameter values ​​used are as follows: learning rate is 0.002, the number of hidden layers is 2, the number of neurons is 100, the batch size is 64, and the number of iterations is 300;

[0100] Step 4: Perform real-time correction on the hydrological model forecast results based on the trained LSTM model:

[0101] Step 4.1: When using the hydrological model for flood forecasting, at the start of the forecast, denoted as time tf, repeat steps 2.2 to 2.4 to obtain and tf, input it into the trained LSTM model, then the ensemble forecast flow at time tf and collection state variables Calculated by the following formula:

[0102]

[0103]

[0104] Among them, ysl tf and Xl tf They are respectively the simulated flow and state quantity at time tf output by the LSTM model, is the analytical state quantity at time tf, They are the optimal simulated flow and optimal analysis value of the Kalman filter at time tf, ys n,tf is the hydrological model M n (.)Simulated flow at time tf;

[0105] Step 4.2: The forecast period is set to 6 hours. The ensemble average of is the optimal initial state of the hydrological model at the start of the forecast. It is substituted into the hydrological model for flood forecasting. The forecast results of two floods in the Tunxi Basin with a forecast period of 6 hours are shown in Figure 4 and Figure 5 .

[0106] In the flood forecast No. "TX-2015080908" for the Tunxi River Basin in this embodiment, compared with the observed value, the relative error of the flood peak of the present invention is 20%, which is 31% less than when no real-time correction is performed, and is 7% less than when only the ensemble Kalman filter is used for real-time correction; in the flood forecast No. "TX-2017062308" for the Tunxi River Basin, compared with the observed value, the relative error of the flood peak of the present invention is 8%, which is 25% less than when no real-time correction is performed, and is 10% less than when only the ensemble Kalman filter is used for real-time correction, which proves that the present invention is more beneficial to improving the accuracy of flood forecasts than using only the ensemble Kalman filter for real-time correction of flood forecasts.

[0107] In summary, the present invention combines the update results of the ensemble Kalman filter with the LSTM deep learning model, fully utilizing the advantages of the deep learning model to further process the non-Gaussianity that may exist in the hydrological model, thereby having a favorable impact on the accuracy of flood forecasting results.

Claims

1. A real-time correction method for flood simulation considering system non-Gaussianity, characterized in that: The following steps are involved: Step 1: Construct a hydrological model M(.) for flood simulation in the basin to be simulated. The hydrological model M(.) has the function of calculating runoff and confluence; obtain the input observed rainfall and evaporation data U, the optimal parameter set and the initial value of the model state Step 2: Construct an ensemble Kalman filter coupled with the hydrological model M(.), the specific method is as follows; Step 2.1: At time t = 1, the initial value of the model state Apply Gaussian perturbation and repeat N times to obtain the initial value of the model state after perturbation Then set up N hydrological models M that are independent of each other and calculated in parallel in time n (.); then the corresponding input data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow of N groups at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as Step 2.2: At time t>1, calculate the predicted state quantities of all N hydrological models M(.) And simulated flow ys n,t ; Step 2.3: At time t>1, calculate the analytical state quantities of all N hydrological models M(.) at time t Step 2.4: The optimal analytical value of the state variable at time t>1 Set as the ensemble average of the analyzed state quantities, the optimal simulated flow of the model at each moment Set to the ensemble mean of the simulated flow; Step 3: Based on the hydrological model M(.) and the ensemble Kalman filter, a long short-term memory recursive neural network LSTM model is trained to transform the flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the simulated traffic ysl of the LSTM model is t and state quantity Xl t As the output value, the training goal is to minimize the loss function between the LSTM analysis flow and the observed flow until the training is completed; Step 4: Use the LSTM model trained in step 3 to perform real-time correction on the forecast results of the hydrological model M(.). The specific method is as follows: Step 4.1: When using the hydrological model M(.) for flood forecasting, at the forecast start time tf, repeat steps 2.2 to 2.4 to obtain the corresponding optimal analysis value Optimal simulation flow and flow observation value y tf , and then input it into the trained LSTM model to obtain the collective forecast flow at the forecast start time tf and collection state variables Step 4.2: Set the forecast period to LT hours and The ensemble average of is the optimal initial state of the hydrological model at the beginning of the forecast, which is substituted into the hydrological model for flood forecasting.

2. The real-time correction method for flood forecasting considering system non-Gaussianity according to claim 1 is characterized in that: The detailed steps of step 1 are as follows: Step 1.1, obtain flood data and evaporation data of the hydrological station at the outlet of the basin, interpolate the flood data into a flood data set with a time step of 1 hour by interpolation, evenly distribute the evaporation of the observation time step in time, and interpolate it into an evaporation data set with a time step of 1 hour; Step 1.2, obtain the coordinate position and rainfall data of the rain gauge station in the basin, evenly distribute the rainfall of the observation time step in time, interpolate it into a rainfall data set with a time step of 1 hour, and then use the Thiessen polygon method to divide the basin into k sub-basins with the rain gauge station as the center, k is the number of rain gauge stations in the basin, the rainfall of each sub-basin is set to the rainfall data set of the corresponding rain gauge station in the basin, and the evaporation of each sub-basin is set to the evaporation data set of the hydrological station at the outlet of the basin; Step 1.3, select a hydrological model M(.) with runoff generation and runoff calculation functions, the number of sub-basins calculated in the hydrological model M(.) is k, the state quantity set representing soil moisture in runoff generation calculation is marked as X1, the state quantity set representing river water storage in runoff calculation is marked as X2, and the model state quantity set is recorded as X = [X1, X2]; Step 1.4: At the first calculation time t = 1, the initial value of X1 is recorded as X1ini, and the value of each element in X1ini is set to half of the maximum soil moisture content; the initial value of X2 is recorded as X2ini, and the value of each element in X2ini is set to the observed flow of the outlet hydrological station at t = 1 divided by the number of sub-basins, and the two are recorded together as the initial value of the model state quantity Step 1.5: Take the observed rainfall and evaporation as the model input data, denoted as U, set the insensitive parameters in the hydrological model M(.) to the recommended values, and use the shuffled complex evolution algorithm SCE-UA to automatically calibrate the parameters of the hydrological model M(.) to obtain the optimal parameter set.

3. The real-time correction method for flood forecasting considering system non-Gaussianity according to claim 1 is characterized in that: In step 2, the process of constructing an ensemble Kalman filter coupled with the hydrological model is as follows: For step 2.1, at time t = 1, first initialize the model state Apply a Gaussian perturbation ω with a mean of 0 and a covariance of Q n , for N hydrological models M that are independent and calculated in parallel in time n (.), let n = 1, 2, ..., N respectively, and repeat the following process: input the model at this moment into the data U t＝1 and the optimal parameter set and the initial value of the model state after disturbance Input hydrological model M n (.), and obtain the simulated flow rate of the nth group at the first moment and the initial X1 and X2 values ​​at the second moment, recorded as For step 2.2, let n = 1, 2, ..., N and repeat the process one by one until all N hydrological model simulation processes are completed at time t>1, and we get In step 2.3, analyze the state quantity The calculation formula is as follows: In the formula, K t is the Kalman gain matrix at time t, v is the Gaussian noise with mean 0 and covariance R; Optimal analysis value in step 2.4 And the optimal simulation flow The calculation formula is as follows: Starting from t=2, repeat steps 2.2 to 2.4 moment by moment until all moments are calculated.

4. The real-time correction method for flood forecasting considering system non-Gaussianity according to claim 1 is characterized in that: The model training process in step 3 is as follows: Step 3.1: Construct a long short-term memory recursive neural network LSTM model, and use the Adam algorithm as the back propagation optimization algorithm of the model. LSTM consists of multiple neuron layers, and each layer contains multiple neurons. The neurons pass through the gating unit, that is, the input gate i t 、Forget gate t and output gate o t To memorize and learn, and prevent gradient explosion, the calculation formula is as follows: i t =σ(W ix x t +W ih h t-1 +b i ) f t =σ(W fx x t +W fh h t-1 +b f ) o t =σ(W ox x t +W oh h t-1 +b o ) h t =φ(c t )⊙o t In the formula, c t , and h t are the memory state, candidate memory state and hidden state at time t, σ and φ are the changes of sigmoid function and tanh function respectively, W ix , W ih , W fx , W fh , W ox , W oh , are input gate i t 、Forget gate t , output gate o t and candidate memory states The matrix weights of the input and output calculations, b i , b f , b o , are input gate i t 、Forget gate t , output gate o t and candidate memory states The bias term of , ⊙ represents the element-wise multiplication of each vector; Step 3.2: Randomly select 80% of all the calculation moments as training set samples, and the remaining 20% ​​as validation set samples; Step 3.3: The flow observation value y t , optimal simulation flow and the optimal analysis value As the input data for training the LSTM model, the simulated traffic ysl of the LSTM model t and state quantity Xl t As the output value, the training objective is set to minimize the loss function between the LSTM analysis flow and the observed flow. The loss function is set to mean square error loss. The particle swarm algorithm is used to automatically optimize the hyperparameters, determine the LSTM model hyperparameters, and complete the LSTM training.

5. The real-time correction method for flood forecasting considering system non-Gaussianity according to claim 1 is characterized in that: The ensemble forecast flow at time tf in step 4.1 and collection state variables The calculation formula is as follows: Among them, ysl tf and Xl tf They are respectively the simulated flow and state quantity at time tf output by the LSTM model, is the analytical state quantity at time tf, and They are the optimal simulated flow and optimal analysis value of the Kalman filter at time tf, ys n,tf is the hydrological model M n (.)Simulated traffic at time tf.

Images