Flood forecasting method considering flood peak propagation time

Through the improved RFDTW algorithm and BiLSTM-IBP model combined with multiple improved whale optimization algorithm, the problem of neglected correlation between upstream and downstream stations is solved, the accuracy and efficiency of flood forecasting are improved, and more accurate capture of flood propagation laws is achieved. and forecast results.

CN119025890APending Publication Date: 2024-11-26NAT ELECTRIC POWER INVESTMENT GRP YELLOW RIVER UPSTREAM HYDROPOWER DEV CO LTD +2
View PDF 11 Cites 0 Cited by

Patent Information

Application Number
CN202411118466.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing flood forecasting methods tend to ignore the correlation between observation elements between upstream and downstream stations, resulting in low forecast accuracy.

Method used

The improved RFDTW algorithm optimizes the dynamic bending distance measurement method to match the flood process between upstream and downstream related stations in a similar manner, extract the flood propagation law as the input of the forecast model, and extract it from the forward and reverse aspects through the BiLSTM-IBP forecast model. Time sequence features, combined with multiple improved whale optimization algorithms to optimize model parameters.

Benefits of technology

The efficiency and accuracy of the flood forecast model are improved, and the flood propagation laws can be captured more accurately and the accuracy of forecast results can be improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025890A_ABST
    Figure CN119025890A_ABST
Patent Text Reader

Abstract

The invention discloses a flood forecasting method considering flood peak propagation time, and the method specifically comprises the steps: collecting and preprocessing data, and dividing the data into a training set and a test set; a correlation analysis method and a flood propagation time mining method are adopted to optimize forecast factors; constructing a BiLSTMIBP forecasting model, extracting time sequence characteristics between forecasting factors from the forward direction and the reverse direction, and taking the output of a hidden layer of a BiLSTM network model as the input of an IBP neural network; the parameters of the BiLSTMIBP forecasting model are optimized; and forecasting the test set data by adopting the forecasting model and the optimal forecasting model parameters, and outputting a result. According to the method, the parameters of the BiLSTM-IBP forecasting model are optimized by constructing the whale optimization algorithm with multiple improvements, so that the model parameters are quickly optimized, and the efficiency and precision of the model are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of hydrological time series prediction, and in particular relates to a flood forecasting method taking into account flood peak propagation time. Background Art

[0002] Hydrological data has the characteristics of large data volume, fast update speed, and complex types, and there are complex relationships between seasons, landforms, and hydrological laws. Although physical models can achieve good prediction results in the process of hydrological forecasting, the establishment of the model needs to consider the physical mechanism and variables of the basin, such as soil moisture, etc. Therefore, it is difficult to set the parameter rate in the process of establishing the physical model. Due to its own limitations, machine learning models are difficult to extract the spatiotemporal characteristics of flood processes and reasonably explain the operation of the model, resulting in unsatisfactory results in hydrological forecasting. Based on this, the hidden laws of flood peak propagation time are mined from historical hydrological data, and it is used as a forecasting factor to construct a hydrological time series prediction model based on machine learning. The swarm intelligence optimization algorithm is used to optimize the parameters of the model, improve the performance of the hydrological time series prediction model based on machine learning, and provide data and knowledge support for basin flood warning, scheduling, and flood prevention and disaster reduction, which has important research significance and practical value. Summary of the invention

[0003] The purpose of the present invention is to provide a flood forecasting method taking into account the flood peak propagation time, so as to solve the problem that the correlation between observation elements between upstream and downstream stations is easily ignored in the existing flood forecasting method, resulting in low forecasting accuracy.

[0004] The technical solution adopted by the present invention is a flood forecasting method considering the flood peak propagation time, which is specifically implemented according to the following steps:

[0005] S1, data collection and preprocessing, and dividing the preprocessed hydrological data set into training set and test set;

[0006] S2. Prediction factor optimization: The prediction factors are optimized by using correlation analysis method and flood propagation time mining method;

[0007] S3. Build the BiLSTM_IBP prediction model, extract the temporal characteristics between prediction factors from both the forward and reverse aspects, and use the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network;

[0008] S4. Optimizing the parameters of the BiLSTM_IBP prediction model;

[0009] S5. Use the constructed BiLSTM-IBP prediction model and the optimal prediction model parameters to predict the test set data and obtain the output result of the prediction model.

[0010] The present invention is also characterized in that:

[0011] In step S1, specifically:

[0012] S11. Collect hydrological data from target hydrological stations and related rainfall stations to form a hydrological data set; the hydrological data include water level time series data, flow time series data, rainfall time series data, hydrological station spatial location information, and rainfall station spatial location information;

[0013] S12. Preprocess the hydrological data set, and divide the preprocessed hydrological data set into a training set and a test set, with the ratio of the training set to the test set being 7:3.

[0014] In step S2, specifically:

[0015] S21. Extract the upstream and downstream hydrological stations and rainfall stations that have an impact on the current forecast section and their impact on the current forecast section based on the spatial relationship distance, geographic attribute distance and time series correlation distance between the rainfall station, hydrological station and the forecast section;

[0016] S22. Use the improved RFDTW algorithm to optimize the dynamic bending distance measurement method to perform similarity matching on the flood processes between the upstream and downstream related hydrological stations extracted in step S21, build a flood propagation law analysis model, use the least squares method for curve fitting, and mine the time propagation relationship of the hydrological flood processes related to the upstream and downstream stations. Use the mined time series propagation relationship to extract the time series data of the related stations as the input of the BiLSTM_IBP forecasting model.

[0017] In step S22, specifically:

[0018] Step S22.1, using the sliding window technology to extract the characteristic points of the peak and trough of the flow process of the two upstream and downstream hydrological stations, and using the improved RFDTW dynamic time warping algorithm to match the peak flow and trough flow characteristics of the two upstream and downstream stations, to obtain the time difference between the characteristic points in the water flow process of the two time series, and then to obtain the propagation time of the water flow between the two upstream and downstream stations;

[0019] RFDTW uses the improved LB_FR lower bound function to train the training set sequence X=(x1,x2,…,x i ,…,x n ) and Y=(y1,y2,…,y,…,y n ) to filter and obtain a candidate training set with better similarity;

[0020] The upper bound value U in the LB_FR lower bound function i Using formula (1) to calculate, the lower limit value L i Calculated using formula (2);

[0021]

[0022] In the formula, and is the sequence point x i Respectively with point x i-1 and x i+1 The first derivative of ;

[0023] Furthermore, RFDTW uses a weighted dynamic time warping method to calculate the distance of the candidate training set; wherein the weighted dynamic time warping mathematical model is shown in equations (3) to (5);

[0024] WDTW(1,1)=dis W (x1,y1) (3);

[0025]

[0026] dis W (x i ,y i )=||w |i-j| (x i -y i )|| p (5);

[0027] Where dis represents the distance between two sequence points; WDTW and DTW represent the weighted dynamic time warping and dynamic time warping measurement methods; || p represents the Lp distance metric, w |i-j Represents sequence point x i ,y i The weight of the distance measure between them;

[0028] Let l = ij, l represents the distance between the corresponding points of two time series, and the weight w l The calculation formula is shown in formula (6):

[0029]

[0030] Among them, w max Yes l The maximum value, n mid is the coordinate of the middle point of the time series length n, and g is a constant parameter;

[0031] Step S22.2, based on the peak flow and propagation time obtained in step S22.1, the least squares method is used to perform curve fitting, and the time series data of the associated sites are extracted based on the mined time series propagation relationship as the input of the forecast model.

[0032] In step S3, specifically:

[0033] Step S31, construct a BiLSTM network model, obtain the temporal features of the associated site input data extracted in step S22 through the three-gate structure of the forward and reverse LSTM respectively; and use the two regularization structures of BatchNormalization and Dropout to avoid network overfitting, so as to obtain the prediction results, as shown in formula (7) and formula (8):

[0034]

[0035] is the hidden state vector of BiLSTM; are the hidden state vectors of the forward and reverse LSTM respectively; h t It is the forecast output result after the fully connected layer;

[0036] Step S32, constructing a BP neural network module improved by Bayesian regularization, adding a penalty term to the Bayesian regularization algorithm, adjusting the training performance function of the BP neural network, and finally obtaining the mathematical model of the IBP neural network, as shown in equations (9) and (10):

[0037] E=k1E W +k2E D (9);

[0038]

[0039] E w is the mean square error of all weights of the IBP neural network; E D is the mean square error of the IBP neural network output, ω r′ is the weight of the IBP neural network, r′=1, 2, 3…N, N is a positive integer; k1 and k2 are regularization coefficients;

[0040] Step S33, using the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network to build a BiLSTM-IBP prediction model based on time series.

[0041] In step S4, specifically:

[0042] Step S41, initialize the number of iterations epoch, batch size batch_size, number of hidden layer nodes, and IBP regularization coefficient of the BiLSTM-IBP prediction model, and randomly generate a population using the chaotic initialization strategy of formula (11);

[0043] z(t m +1)=1-m(cos(n cos -1 z(t m))) 2 , z(t m )∈[-1,1] (11);

[0044] In the formula, z(t m )、z(t m +1) is at t m ,t m +1 moment chaos map state, m and n are control variables;

[0045] Step S42, introduce the salp group leader mechanism to improve the whale optimization algorithm WOA, and the improved population update formula is shown in formula (12):

[0046]

[0047] In the formula, x 1,s represents the first salp position in the s-dimensional search space, F s Indicates the location of food, ub s and lb s Respectively represent the upper and lower bounds of the s-th dimensional search space;

[0048] c2 and c3 represent two independent random numbers that obey the distribution, c1 is the control factor in the algorithm, and the calculation formula is shown in formula (13):

[0049]

[0050] In the formula, t represents the current number of iterations, T represents the maximum number of iterations, and e is a natural constant;

[0051] Step S43, using a new nonlinear decreasing strategy to calculate the convergence factor a, so as to better update the search range A and disturbance factor C of the new population; the calculation formula of a is shown in formula (14):

[0052]

[0053] The calculation formulas of A and C are shown in formula (15) and formula (16):

[0054] A = 2 × a × ra (15);

[0055] C = 2 × r (16);

[0056] r is a random number between 0 and 1;

[0057] Step S44, fitness and position update: update the new position and fitness of each individual according to the search range A, and use the nonlinear parameter c1 of SSA to optimize the attack hunting and predation process of WOA; the position update formula is shown in equations (17)-(19):

[0058] X(t+1)=c1X * (t)-A·|C·X * (t)-X(t)| (17)

[0059] X(t+1)=D t ·e bt ·cos(2πd)+c1X * (t) (18)

[0060] X(t+1)=X rand (t)-A·|C·X rand (t)-X(t)| (19)

[0061] Where, X * (t) is the current optimal position, X(t) is the current position of the whale D t represents the distance between the prey and the whale, b is the logarithmic spiral coefficient, X rand (t) is a random position;

[0062] Step S45, using the above steps to obtain the optimal prediction model parameters, including the number of iterations epoch, batch size batch_size, the number of hidden layer nodes, and the IBP regularization coefficient, and determine whether the current number of iterations t reaches the maximum number of loops T and the loop end condition; if the maximum number of loops T is reached, save the current model and its parameters, and jump to step S5; otherwise, reinitialize a set of parameters, repeat the steps in the above S4 until the optimal model parameters are obtained, and assign the optimal solution to the BiLSTM-IBP prediction model and save it.

[0063] The beneficial effects of the present invention are as follows: the present invention adopts an improved RFDTW optimized dynamic bending distance measurement method to perform similarity matching on flood processes between upstream and downstream related measuring stations, extracts the propagation law between upstream and downstream floods as the input of the prediction model, and then uses the BiLSTM neural network to extract the memory features of the time series context, and uses the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network to construct a BiLSTM-IBP forecasting model; and optimizes the parameters of the BiLSTM-IBP forecasting model by constructing a multiple improved whale optimization algorithm (MIWOA), thereby quickly optimizing the model parameters and improving the efficiency and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of a flood forecasting method considering flood peak propagation time of the present invention;

[0065] Figure 2It is a flow chart of the improved RFDTW algorithm optimization dynamic bending distance measurement method of the present invention;

[0066] Figure 3 It is a structural diagram of the BiLSTM-IBP model of the present invention;

[0067] Figure 4 It is a schematic diagram of the multi-improved whale optimization algorithm MIWOA of the present invention;

[0068] Figure 5 It is the accuracy test chart of various models. DETAILED DESCRIPTION

[0069] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Example 1

[0071] The flood forecasting method considering the flood peak propagation time of the present invention is specifically implemented according to the following steps:

[0072] S1, data collection and preprocessing, and dividing the preprocessed hydrological data set into training set and test set;

[0073] S2. Prediction factor optimization: The prediction factors are optimized by using correlation analysis method and flood propagation time mining method;

[0074] S3. Build the BiLSTM_IBP prediction model, extract the temporal characteristics between prediction factors from both the forward and reverse aspects, and use the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network;

[0075] S4. Optimizing the parameters of the BiLSTM_IBP prediction model;

[0076] S5. Use the constructed BiLSTM-IBP prediction model and the optimal prediction model parameters to predict the test set data and obtain the output result of the prediction model.

[0077] Example 2

[0078] The flood forecasting method considering the flood peak propagation time of the present invention is specifically implemented according to the following steps:

[0079] S1. Data collection and preprocessing; specifically:

[0080] S11. Collect hydrological data from target hydrological stations and related rainfall stations to form a hydrological data set;

[0081] Hydrological data include water level time series data, flow time series data, rainfall time series data, hydrological station spatial location information, rainfall station spatial location information;

[0082] S12, preprocessing the hydrological data set, dividing the preprocessed hydrological data set into a training set and a test set, and the ratio of the training set to the test set is 7:3;

[0083] The preprocessing process specifically includes: data quality control including missing value detection and repair, inconsistent data detection and repair; and normalization of the data set after quality control; then dividing the preprocessed hydrological data set into a training set and a test set;

[0084] S2. Prediction factor optimization: The prediction factors of the forecast model are optimized by using the correlation analysis method and the flood propagation time mining method; specifically:

[0085] S21. Extract the upstream and downstream hydrological (water level) stations and rainfall stations that have an impact on the current forecast section and their impact on the current forecast section based on the spatial relationship distance, geographical attribute distance and time series correlation distance between the rainfall station, hydrological station and the forecast section;

[0086] S22, using the improved RFDTW algorithm to optimize the dynamic bending distance measurement method to perform similarity matching on the flood processes between the upstream and downstream related hydrological (water level) stations extracted in step S21, constructing a flood propagation law analysis model, using the least squares method for curve fitting, mining the time propagation relationship of the upstream and downstream stations related hydrological (water level) flood processes, and extracting the time series data of the associated stations based on the mined time series propagation relationship as the input of the subsequent BiLSTM_IBP forecast model; specifically:

[0087] Step S22.1, using the sliding window technology to extract the characteristic points of the peak and trough of the flow process of the two upstream and downstream hydrological (water level) stations, and using the improved RFDTW dynamic time warping algorithm to match the peak flow and trough flow characteristics of the two upstream and downstream stations, to obtain the time difference between the characteristic points in the water flow process of the two time series, and then to obtain the propagation time of the water flow between the two upstream and downstream stations;

[0088] Furthermore, RFDTW uses the improved LB_FR lower bound function to calculate the training set sequence X = (x1, x2, ..., x i ,…,x n ) and Y=(y1,y2,…,y,…,y n ) to filter and obtain a candidate training set with better similarity;

[0089] The upper bound value U in the LB_FR lower bound function i Using formula (1) to calculate, the lower limit value L i Calculated using formula (2);

[0090]

[0091] In the formula, and is the sequence point x i Respectively with point x i-1 and x i+1 The first-order derivative of , from which we can see that for any x i Meet U i ≥x i ≥L i , which meets the conditions of the lower bound function.

[0092] Furthermore, RFDTW uses a weighted dynamic time warping method to calculate the distance of the candidate training set; wherein the weighted dynamic time warping mathematical model is shown in equations (3) to (5);

[0093] WDTW(1,1)=dis W (x1,y1) (3);

[0094]

[0095] dis W (x i ,y i )=||w |i-j| (x i -y i )|| p (5);

[0096] Where dis represents the distance between two sequence points; WDTW and DTW represent the weighted dynamic time warping and dynamic time warping measurement methods; || p represents the Lp distance metric, p=1 and p=2 represent the Manhattan distance and Euclidean distance respectively; w |i-j| Represents sequence point x i ,y i The weight of the distance measure between them;

[0097] Let l = ij, l represents the distance between the corresponding points of two time series, and the weight w l The calculation formula is shown in formula (6):

[0098]

[0099] Among them, w max Yes l The maximum value, usually w max =1,n mid is the coordinate of the middle point of the time series length n, g is a constant parameter, its size determines w l The telescopic range;

[0100] Step S22.2, based on the peak flow and propagation time obtained in step S22.1, the least square method is used to perform curve fitting, and the time propagation relationship of the hydrological (water level) flood process related to the upstream and downstream stations is mined, and the time series data of the related stations are extracted based on the mined time series propagation relationship as the input of the subsequent forecast model;

[0101] S3. Construct the BiLSTM_IBP forecast model, extract the time series characteristics between the forecast input factors (hydrological data) from both the forward and reverse aspects, and use the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network to improve the prediction accuracy and generalization ability of the model; specifically:

[0102] Step S31, construct a bidirectional long short-term memory network BiLSTM module, and obtain the temporal characteristics of the associated site input data extracted in step S22 through the three-gate structure of the forward and reverse LSTMs; and use the two regularization structures of BatchNormalization and Dropout to avoid network overfitting and reduce errors, so as to obtain the prediction results, as shown in formulas (7) and (8):

[0103]

[0104] is the hidden state vector of BiLSTM; are the hidden state vectors of the forward and reverse LSTM respectively; h t It is the forecast output result after the fully connected layer.

[0105] Step S32, constructing a BP neural network module improved by Bayesian regularization, adding a penalty term to the Bayesian regularization algorithm, adjusting the training performance function of the BP neural network, and finally obtaining the mathematical model of the IBP neural network, as shown in equations (9) and (10):

[0106] E=k1E W +k2E D (9);

[0107]

[0108] E w is the mean square error of all weights of the IBP neural network; E D is the mean square error of the IBP neural network output, ω r′ is the weight of the IBP neural network, r′=1, 2, 3…N, N is a positive integer; k1 and k2 are regularization coefficients, and the sizes of these two coefficients determine the training effect of the improved performance function.

[0109] Step S33, using the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network, thereby realizing the transmission of information, thereby building a BiLSTM-IBP forecasting model based on time series;

[0110] S4. Optimizing the parameters of the BiLSTM-IBP prediction model: The multiple improved whale optimization algorithm (MIWOA) is used to optimize the number of epochs, batch size batch_size, number of hidden layer nodes, and IBP regularization coefficient of the BiLSTM-IBP prediction model to obtain the parameters of the optimal BiLSTM_IBP prediction model, specifically:

[0111] Step S41, initialize the number of iterations epoch, batch size batch_size, number of hidden layer nodes, and IBP regularization coefficient of the BiLSTM-IBP prediction model, and randomly generate a population using the chaotic initialization strategy of formula (11):

[0112] z(t m +1)=1-m(cos(ncos -1 z(t m ))) 2 ,z(t m )∈[-1,1] (11);

[0113] In the formula, z(t m )、z(t m +1) is at t m ,t m +1 moment chaotic mapping state, m and n are control variables, preferably m = 3, n = 5;

[0114] Step S42, introduce the Salp Swarm Algorithm (SSA) leader mechanism to improve the whale optimization algorithm WOA, and the improved population update formula is shown in formula (12):

[0115]

[0116] In the formula, x 1,s represents the first salp position in the s-dimensional search space, F s Indicates the location of food, ub s and lb s They represent the upper and lower bounds of the s-th dimension search space respectively; c2 and c3 represent two independent random numbers that obey the distribution, and c1 is the control factor in the algorithm. The calculation formula is shown in formula (13):

[0117]

[0118] Where t represents the current number of iterations, T represents the maximum number of iterations, and e is a natural constant.

[0119] Step S43, a new nonlinear decreasing strategy is used to calculate the convergence factor a, so as to better update the search range A and disturbance factor C of the new population. The calculation formula of a is shown in formula (14):

[0120]

[0121] Furthermore, in formula (11), π is the circumference of a circle, t is the current number of iterations, and T is the maximum number of iterations.

[0122] The calculation formulas of A and C are shown in formula (15) and formula (16):

[0123] A = 2 × a × ra (15);

[0124] C = 2 × r (16);

[0125] r is a random number between 0 and 1.

[0126] Step S44, fitness and position update: Update the new position and fitness of each individual according to the search range A, and use the nonlinear parameter c1 of SSA to optimize the attack hunting and predation process of WOA. The position update formula is shown in equations (17)-(19):

[0127] X(t+1)=c1X * (t)-A·|C·X * (t)-X(t)| (17)

[0128] X(t+1)=D t ·e bt ·cos(2πd)+c1X * (t) (18)

[0129] X(t+1)=X rand (t)-A·|C·X rand (t)-X(t)| (19)

[0130] In the above formula, t is the number of current iterations, X * (t) is the current optimal position, X(t) is the current position of the whale, and D t represents the distance between the prey and the whale, b is the logarithmic spiral coefficient, when the value of b changes, the shape of the spiral will also change; X rand (t) is a random position;

[0131] Step S45, using the above steps to obtain the optimal BiLSTM-IBP prediction model parameters (iteration number epoch, batch size batch_size, number of hidden layer nodes, IBP regularization coefficient), and determine whether the current iteration number t reaches the maximum number of loops T loop end condition; if the maximum number of loops T is reached, save the current model and its parameters, and jump to step S5; otherwise, reinitialize a set of parameters, repeat the steps in the above S4 until the optimal model parameters are obtained, and assign the optimal solution to the BiLSTM-IBP prediction model and save it;

[0132] S5. Model verification: Use the constructed BiLSTM-IBP prediction model and the optimal model parameters to predict the test set data divided in step S12 to obtain the output result of the prediction model.

[0133] The coefficient of determination R 2 , Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and other evaluation indicators are used to evaluate the output results of the forecast model.

[0134] Example 3

[0135] The embodiment of the present invention provides a flood forecasting method taking into account the flood peak propagation time, such as Figure 1 As shown, the following steps are included:

[0136] S1. Data collection and preprocessing: Perform missing value interpolation, data normalization, data set division and other preprocessing steps on the hydrological data of the target water level station and related rainfall stations collected within the forecast section to construct a hydrological time series data set that meets the user's quality requirements;

[0137] S2. Prediction factor optimization: correlation analysis and flood propagation time mining algorithm are used to optimize the prediction factors of the prediction model;

[0138] S3. Construction of forecast model: construct the BiLSTM_IBP forecast model, extract the time series characteristics between the forecast input factors (hydrological data) from both the forward and reverse aspects, and use the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network, so as to improve the prediction accuracy and generalization ability of the model;

[0139] S4. Model parameter optimization: The multiple improved whale optimization algorithm (MIWOA) is used to optimize the epoch number, batch size, number of hidden layer nodes, and IBP regularization coefficient of the BiLSTM_IBP prediction model to obtain the optimal BiLSTM_IBP prediction model parameters.

[0140] S5, model verification: forecast the input data set using the forecast model and its optimal model parameters determined in step S4, obtain the model forecast results and evaluate them.

[0141] In step S1 of the embodiment of the present invention, the collected hydrological data set is derived from a certain prediction section and its associated measuring stations, and is subjected to preprocessing steps such as missing value filling, consistency repair, data normalization, and data set division to form a hydrological time series data set;

[0142] In step S2 of the embodiment of the present invention, firstly, the correlation analysis method is used to extract the associated stations that have an impact on the current forecast section and their impact degree, so as to determine the data of the model training station; then Figure 2 The process shown in the figure uses the improved RFDTW algorithm to optimize the dynamic bending distance measurement method to extract the time lag of the impact of different associated stations on the forecast section, thereby constructing the input data set of the forecast model.

[0143] In step S3 of the embodiment of the present invention, the following is adopted: Figure 3 The process shown builds the BiLSTM_IBP forecast model, taking the forecast factor determined in the above step S3 as the input of the model. First, the BiLSTM model is built to mine the time series characteristics of the hydrological time series in both the forward and reverse directions. Then, the output of the hidden layer of the BiLSTM network model is used as the input of the IBP neural network to realize the information transmission, thereby building the time series prediction model BiLSTM-IBP.

[0144] In step S4 of the embodiment of the present invention, the following is adopted: Figure 4 The process shown implements the multiple improved whale optimization algorithm MIWOA method to parameterize the BiLSTM_IBP forecast model.

[0145] Figure 5 As shown in the figure, in the model with a forecast period of 12h, the BiLSTM_IBP forecast model is optimized by MIWOA, the hidden layer node number hidden_size is set to 64, the batch size batch_size is set to 50, the model iteration number epochs is set to 225, and the IBP regularization coefficients k1 and k2 are set to 0.02 and 0.015; in the SVM model, the kernel function parameter g is set to 0.08, and the penalty coefficient c is set to 0.5; in the CNN model, the number of filters filters is set to 32, and the number of filters filters is set to 32. Figure 5It can be seen that the prediction effects of LSTM, SVM and CNN models are mostly close to the original time series, but the fluctuations at the peaks and inflection points are more obvious; while the BiLSTM_IBP forecast model is more consistent with the original time series as a whole, with a shorter peak lag time at the peaks and troughs and other inflection points, and the overall effect is the best. Therefore, whether from the overall test set or a flood process, the accuracy of the BiLSTM_IBP forecast model is higher than that of the other four models.

Claims

1. A flood forecasting method considering flood peak propagation time, characterized in that: Follow the steps below to implement it: S1, data collection and preprocessing, and dividing the preprocessed hydrological data set into training set and test set; S2. Prediction factor optimization: The prediction factors are optimized by using correlation analysis method and flood propagation time mining method; S3. Build the BiLSTM_IBP prediction model, extract the temporal characteristics between prediction factors from both the forward and reverse aspects, and use the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network; S4. Optimizing the parameters of the BiLSTM_IBP prediction model; S5. Use the constructed BiLSTM-IBP prediction model and the optimal prediction model parameters to predict the test set data and obtain the output result of the prediction model.

2. The flood forecasting method considering flood peak propagation time as claimed in claim 1, characterized in that: In the step S1, specifically: S11. Collect hydrological data from target hydrological stations and related rainfall stations to form a hydrological data set; the hydrological data include water level time series data, flow time series data, rainfall time series data, hydrological station spatial location information, and rainfall station spatial location information; S12. Preprocess the hydrological data set, and divide the preprocessed hydrological data set into a training set and a test set, with the ratio of the training set to the test set being 7:

3.

3. The flood forecasting method considering flood peak propagation time as claimed in claim 2, characterized in that: In the step S2, specifically: S21. Extract the upstream and downstream hydrological stations and rainfall stations that have an impact on the current forecast section and their impact on the current forecast section based on the spatial relationship distance, geographic attribute distance and time series correlation distance between the rainfall station, hydrological station and the forecast section; S22. Use the improved RFDTW algorithm to optimize the dynamic bending distance measurement method to perform similarity matching on the flood processes between the upstream and downstream related hydrological stations extracted in step S21, build a flood propagation law analysis model, use the least squares method for curve fitting, and mine the time propagation relationship of the hydrological flood processes related to the upstream and downstream stations. Use the mined time series propagation relationship to extract the time series data of the related stations as the input of the BiLSTM_IBP forecasting model.

4. The flood forecasting method considering flood peak propagation time as claimed in claim 3, characterized in that: In the step S22, specifically: Step S22.1, using the sliding window technology to extract the characteristic points of the peak and trough of the flow process of the two upstream and downstream hydrological stations, and using the improved RFDTW dynamic time warping algorithm to match the peak flow and trough flow characteristics of the two upstream and downstream stations, to obtain the time difference between the characteristic points in the water flow process of the two time series, and then to obtain the propagation time of the water flow between the two upstream and downstream stations; RFDTW uses the improved LB_FR lower bound function to train the training set sequence X=(x1,x2,…,x i ,…,x n ) and Y=(y1,y2,…,y,…,y n ) to filter and obtain a candidate training set with better similarity; The upper bound value U in the LB_FR lower bound function i Using formula (1) to calculate, the lower limit value L i Calculated using formula (2); In the formula, and is the sequence point x i Respectively with point x i-1 and x i+1 The first derivative of ; Furthermore, RFDTW uses a weighted dynamic time warping method to calculate the distance of the candidate training set; wherein the weighted dynamic time warping mathematical model is shown in equations (3) to (5); WDTW(1,1)=dis W (x1,y1) (3); dis W (x i ,y i )=|w |i-j| (x i -y i )| p (5); Where dis represents the distance between two sequence points; WDTW and DTW represent the weighted dynamic time warping and dynamic time warping measurement methods; || p represents the Lp distance metric, w |i-j| Represents sequence point x i ,y i The weight of the distance measure between them; Let l = ij, l represents the distance between the corresponding points of two time series, and the weight w l The calculation formula is shown in formula (6): Among them, w max Yes l The maximum value, n mid is the coordinate of the middle point of the time series length n, and g is a constant parameter; Step S22.2, based on the peak flow and propagation time obtained in step S22.1, the least squares method is used to perform curve fitting, and the time series data of the associated sites are extracted based on the mined time series propagation relationship as the input of the forecast model.

5. The flood forecasting method considering flood peak propagation time as claimed in claim 4, characterized in that: In the step S3, specifically: Step S31, constructing a BiLSTM network model, and obtaining the temporal characteristics of the associated site input data extracted in step S22 through the three-gate structure of the forward and reverse LSTMs; And use BatchNormalization and Dropout two regularization structures to avoid network overfitting, so as to obtain the prediction results, as shown in formula (7) and formula (8): is the hidden state vector of BiLSTM; are the hidden state vectors of the forward and reverse LSTM respectively; h t It is the forecast output result after the fully connected layer; Step S32, constructing a BP neural network module improved by Bayesian regularization, adding a penalty term to the Bayesian regularization algorithm, adjusting the training performance function of the BP neural network, and finally obtaining the mathematical model of the IBP neural network, as shown in equations (9) and (10): E=k1E W +k2E D (9); E w is the mean square error of all weights of the IBP neural network; E D is the mean square error of the IBP neural network output, ω r is the weight of the IBP neural network, r′=1,2,3…N, N is a positive integer; k1 and k2 are regularization coefficients; Step S33, using the output of the hidden layer of the BiLSTM network model as the input of the IBP neural network to build a BiLSTM-IBP prediction model based on time series.

6. The flood forecasting method considering flood peak propagation time as claimed in claim 5, characterized in that: In the step S4, specifically: Step S41, initialize the number of iterations epoch, batch size batch_size, number of hidden layer nodes, and IBP regularization coefficient of the BiLSTM-IBP prediction model, and randomly generate a population using the chaotic initialization strategy of formula (11); z(t m +1)=1-m(cos(ncos -1 z(t m ))) 2 ,z(t m )∈[-1,1] (11); In the formula, z(t m )、z(t m +1) is at t m ,t m +1 moment chaos map state, m and n are control variables; Step S42, introduce the salp group leader mechanism to improve the whale optimization algorithm WOA, and the improved population update formula is shown in formula (12): In the formula, x 1,s represents the first salp position in the s-dimensional search space, F s Indicates the location of food, ub s and lb s They represent the upper and lower bounds of the s-th dimension search space respectively; c2 and c3 represent two independent random numbers that obey the distribution, and c1 is the control factor in the algorithm. The calculation formula is shown in formula (13): In the formula, t represents the current number of iterations, T represents the maximum number of iterations, and e is a natural constant; Step S43, using a new nonlinear decreasing strategy to calculate the convergence factor a, so as to better update the search range A and disturbance factor C of the new population; the calculation formula of a is shown in formula (14): The calculation formulas of A and C are shown in formula (15) and formula (16): A = 2 × a × ra (15); C=2×r (16); r is a random number between 0 and 1; Step S44, fitness and position update: update the new position and fitness of each individual according to the search range A, and use the nonlinear parameter c1 of SSA to optimize the attack hunting and predation process of WOA; the position update formula is shown in equations (17)-(19): X(t+1)=c1X * (t)-A·|C·X * (t)-X(t)| (17) X(t+1)=D t ·e bt ·cos(2πd)+c1X * (t) (18) X(t+1)=X rand (t)-A·|C·X rand (t)-X(t)| (19) Where, X * (t) is the current optimal position, X(t) is the current position of the whale D t represents the distance between the prey and the whale, b is the logarithmic spiral coefficient, X rand (t) is a random position; Step S45, using the above steps to obtain the optimal prediction model parameters, including the number of iterations epoch, batch size batch_size, number of hidden layer nodes, IBP regularization coefficient, and determine whether the current number of iterations t reaches the maximum number of cycles T loop end condition; if the maximum number of cycles T is reached, save the current model and its parameters, and jump to step S5; Otherwise, a set of parameters is reinitialized and the above steps are repeated until the optimal prediction model parameters are obtained, and the optimal solution is assigned to the BiLSTM-IBP prediction model.

Citation Information

Patent Citations

  • Vehicle information identification system and method

    CN110491141A

  • Aircraft ground air conditioner energy consumption prediction method based on IALO-AM-BiLSTM model and storage medium

    CN114065635A

  • Urban rail transit network service toughness measurement method and application

    CN114493201A

  • Intelligent flood forecasting method based on MCRBiLSTM

    CN115860231A

  • Flood forecasting method and device based on BOA-DA-BiLSTM

    CN116341738A