CN115860231A - MCR _ BilSTM-based intelligent flood forecasting method - Google Patents
MCR _ BilSTM-based intelligent flood forecasting method Download PDFInfo
- Publication number
- CN115860231A CN115860231A CN202211570315.0A CN202211570315A CN115860231A CN 115860231 A CN115860231 A CN 115860231A CN 202211570315 A CN202211570315 A CN 202211570315A CN 115860231 A CN115860231 A CN 115860231A
- Authority
- CN
- China
- Prior art keywords
- bilstm
- model
- forecast
- hydrological
- station
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses An intelligent flood forecasting method based on MCR _ BilSTM (Anensemble model based on CNN, resNet and BilSTM, MCR _ BilSTM), which comprises the following steps: (1) the forecasting factors are preferably: carrying out preprocessing steps such as missing value interpolation, data normalization and data set division on hydrological data of a target water level station and a related rainfall station in a forecast section, and then constructing a correlation coefficient matrix optimal forecast factor; (2) weight configuration: different weights are distributed to the optimal forecasting factors by adopting an attention mechanism, the long-distance dependency relationship of the input and output sequences is established, and the correlation between input and output is improved; (3) forecasting model construction: constructing a BilSTM network module, extracting time sequence characteristics between forecast input factors (hydrological data) from the two aspects of forward and reverse, and adopting a regularization structure to avoid overfitting of a model and improve the generalization capability of model parameters to data; constructing an enhanced CNN network module, extracting local spatial features of hydrological data by using CNN, deepening the depth of the CNN network by using ResNet, and mining more latent spatial feature information; (4) model integration and fusion: the prediction results of the BilSTM network module and the enhanced CNN network module are subjected to linear fusion after passing through a full connection layer respectively, and are integrated into a terminal-terminal hydrological integrated prediction model (MCR _ BilSTM), so that high-precision intelligent flood prediction is realized; and (5) outputting the result: and performing inverse normalization processing on the MCR _ BilSTM model fusion result, and outputting a final prediction result of the model.
Description
Technical Field
The invention relates to a hydrological prediction technology, in particular to An intelligent flood forecasting method based on MCR _ BilSTM (Anensembleble model based on CNN, resNet and BilSTM and MCR _ BilSTM).
Background
Accurate flood forecasting can effectively support flood early warning, dispatching, flood control and disaster reduction in a drainage basin, and has important significance for area planning, irrigation water taking, sediment transport and other hydrological applications and the like. With the rapid development of information technology, the water conservancy business department has already formed the water and rain condition observation data of the integration of sky, sky and ground, and how to apply the widely applied machine learning algorithm in the fields of automatic control, image recognition and the like to the field of hydrologic forecast, construct a data-driven forecast model with physical meaning, and improve the intelligent business capability and forecast accuracy of flood forecast, so that the method becomes the driving force for solving the problem of hydrology and promoting the progress of hydrology at present.
The flood process is a nonlinear relation with high uncertainty and complexity, and although a physical model can obtain a better prediction result in the hydrologic prediction process, the establishment of the model needs to consider the physical mechanism and variables of a basin, such as soil water content and the like, so that the parameter calibration of the model is difficult in the physical model establishment process; due to the limitations of single machine learning models such as SVM and BP neural network models, it is difficult to comprehensively extract the space-time characteristics of the flood process, so that the effect in hydrologic prediction is often unsatisfactory, and both models have limitations.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the defects in the prior art and provides an intelligent flood forecasting method based on MCR _ BilSTM, which integrates a BilSTM model and machine learning models such as CNN, resNet and the like into an end-to-end integrated intelligent flood forecasting model, thereby not only effectively excavating the time sequence characteristics of a hydrological sequence, but also giving consideration to the local spatial characteristics of the hydrological sequence and providing a new research idea and application support for intelligent flood forecasting.
The technical scheme is as follows: the invention discloses an MCR _ BilSTM-based intelligent flood forecasting method, which comprises the following steps:
the S1 forecasting factor is preferably: carrying out preprocessing steps such as missing value interpolation, data normalization and data set division on hydrological data of a target water level station and a related rainfall station in a forecast section, and then constructing a correlation coefficient matrix optimal forecast factor;
s2, weight configuration: different weights are distributed to the optimal forecasting factors by adopting an attention mechanism, the long-distance dependency relationship of the input and output sequences is established, and the correlation between input and output is improved;
s3, construction of a forecasting model: constructing a BilSTM network module, extracting time sequence characteristics between forecast input factors (hydrological data) from the two aspects of forward and reverse, and adopting a regularization structure to avoid model overfitting and improve the generalization capability of model parameters to data; constructing an enhanced CNN network module, extracting local spatial features of hydrological data by using CNN, deepening the depth of the CNN network by using ResNet, and mining more latent spatial feature information;
s4, model integration and fusion: the predictions of the BilSTM network module and the enhanced CNN network module are subjected to linear fusion after passing through a full connection layer respectively, and are integrated into a terminal-terminal hydrological integrated prediction model (MCR _ BilSTM), so that high-precision intelligent flood prediction is realized;
and S5, performing inverse normalization processing on the MCR _ BilSTM model fusion result, and outputting a final model prediction result.
In the deep learning-based integrated hydrologic forecasting method, the detailed steps of step S1 are as follows:
s11, normalizing the input variable by adopting a Min-Max formula to accelerate the convergence speed and the training process, wherein the calculation method comprises the following steps:
wherein X max 、X min Maximum and minimum values of the original sequence, respectively; x i Is an element in the original sequence; x' i Is the result of normalization, and has a value of 0,1]Within the range;
s12, mining and extracting the topographic features, the spatial position relations and other spatial features of the forecast section, constructing a hydrological spatial relation topological graph, and determining rainfall stations and hydrological station observation elements which have influence on the hydrological process of the forecast section as forecasting factors of the model;
s13, mining and extracting time sequence characteristics of the forecast section, calculating correlation coefficients among the m correlation rainfall stations, the hydrological stations and the forecast section determined in the step S12 by adopting Pearson correlation coefficients, and determining a correlation station X k ＝{x 1 ，x 2 ，x 3 ，.....x n } (k =1.. M) pair predicted stations Y = { Y = 1 ，y 2 ，y 3 ，.....y n The forecasting elements (water level and flow) influence the time sequence length, and the correlation coefficient calculation formula is as follows:
wherein n is the length of the sequence,
represents a mean value of a sequence>
In order to be the covariance,
is the standard deviation;
s14 for the associated station X in the hydrological space relationship topological graph in the step S12 k Sequencing the spatial distance from the forecast station Y, and calculating according to the formula (2) to obtain the associated station X k Influencing the timing length T of the forecast station Y corr Extracting T-T corr Sequence value of t time is used as association station X k The forecast input factor of (1); influencing the timing length T corr The determination method of (2) is as follows:
1) Selecting the maximum correlation coefficient of the associated station with the minimum spatial distance with the forecast station Y in turn, and taking the time corresponding to the coefficient as the influence time sequence length T of the associated station with the minimum spatial distance corr ；
2) If the influence of two associated stations affects the timing length T corr If the spatial distances are the same but have larger difference, selecting the secondary large correlation number from the associated stations with larger distances, and taking the time corresponding to the coefficient as the length T of the influence time sequence corr (ii) a If the spatial distances of two associated stations are close, the same T can be adopted corr As the length of the impact timing of the two associated stations.
In the integrated hydrologic forecast method based on deep learning, the step S2 specifically includes:
taking a forecast period of 1h (time t + 1) as an example, the forecast factor sequence preferred by the step S1 is
(k =1.. M) the input is an attention mechanism, and internal relations among input data are better mined, so that characteristic information hidden in a sequence is screened out. The attention mechanism calculation formula is as follows:
s i ＝tanh(w 1 y t+1 +w 2 X i +b) (3)
wherein, X i A sequence of predictor factors representing an ith associated station; y is t+1 Represents t +1 forecasting elements of a forecasting station at a moment; s i The attention weight score corresponding to the forecast factor sequence Xi of the ith associated station; w is a 1 、w 2 Representing a weight matrix; b is a bias matrix; a is i Representing the assigned attention weight of the current input; v i Is the output vector of the attention-driven layer.
In the deep learning-based integrated hydrologic forecasting method, the step S3 uses the new vector V obtained in the step S2 i As an input, an optimal deep learning model is constructed to train a mapping function f (x) of the model, and the steps include:
s31: bi LSTM module f for constructing bidirectional long-short term memory network 1 (x) Acquiring time sequence characteristics of hydrological data through three gate structures of forward LSTM and backward LSTM, avoiding overfitting of the network by using two regularization structures of BatchNormalization and Dropout, and reducing errors so as to obtain a prediction result y pred，1 ＝f 1 (x) The calculation formula is as follows:
wherein h is t Is a hidden state vector of the BilSTM;
hidden state vectors of positive and negative LSTM respectively;
s32: construction of CNN network module f based on ResNet enhancement 2 (x) Performing convolution operation on the hydrological time sequence by adopting a one-dimensional convolution kernel of the CNN network, adding a plurality of residual error units ResNet to deepen the CNN network level, extracting the characteristic component of hydrological data on a local space structure, and obtaining a prediction result y pred，2 ＝f 2 (x) The calculation formula is as follows:
X (l+1) ＝X l +g(X l ；θ l ) (8)
wherein the content of the first and second substances,
b (l) respectively is a weight matrix and a bias matrix of the l layer; g is a residual function; theta l Are parameters that can be learned by layer i.
In the deep learning-based integrated hydrologic forecasting method, the step S4 includes:
bi-directional long-short time memory network BilSTM module f 1 (x) And a CNN network module f based on ResNet enhancement 2 (x) Predicted result y of pred，1 、y pred，2 Respectively carrying out linear fusion through a full connection layer, and integrating into a terminal-terminal hydrological integrated prediction model (MCR _ BilSTM) to realize high-precision intelligent flood prediction; the calculation formula for linear fusion is as follows:
y pred (t)＝ω 1 y pred，1 (t)+ω 2 y pred，2 (t) (9)
ω 1 +ω 2 ＝1 (10)
wherein, y pred (t) is the final forecast result at time t; y is pred，1 (t)、y pred，2 (t) the prediction results of the two network modules at the moment t are respectively obtained; omega 1 、ω 2 The weight assignment for the model is determined by the least squares method.
In the deep learning-based integrated hydrologic forecasting method, the step S5 includes:
performing inverse normalization processing on the MCR _ BilSTM model fusion prediction result, and outputting a final model prediction result, wherein an inverse normalization formula is as follows:
y′ pred (t)＝y pred (t)*(X max -X min )+X min (11)
wherein, X max 、X min The maximum value and the minimum value are kept consistent during normalization; y is pred (t) is the normalized prediction result at time t; y' pred (t) is transAnd predicting the result at the normalized t moment.
The invention has the beneficial effects that: compared with the prior art, the invention has the following advantages:
the integrated hydrological forecasting method based on deep learning is mainly applied to medium and small watersheds, not only captures local spatial features of hydrological sequences, but also extracts hydrological time sequence features, and constructs an integrated forecasting model through a deep learning network so as to fully learn the nonlinear response relation of a flood process, so that forecasting results of different networks are mutually complemented, and uncertain factors in the forecasting process are effectively reduced.
Drawings
FIG. 1 is a model framework diagram of an MCR _ BilSTM-based intelligent flood forecasting method of the present invention;
FIG. 2 is a topological diagram of the spatial relationship of the east mountain station in the flow area of the Qinhuai river;
FIG. 3 is a comparison graph of the future 2h model prediction results of different models in the embodiment;
FIG. 4 is a comparison graph of the future 4h model prediction results of different models in the embodiment;
FIG. 5 is a comparison graph of the future 6h model prediction results of different models in the embodiment.
Detailed Description
The invention will be described in further detail with reference to the drawings and the following detailed description, but the scope of protection of the invention is not limited to the embodiments.
Referring to fig. 1, the intelligent flood forecasting method based on MCR _ BiLSTM of the present embodiment includes the following steps:
the S1 forecasting factor is preferably: selecting a east mountain hydrological station in a Yangtze river basin of Yangtze river as a target hydrological station, and connecting the east mountain hydrological station 2014/05/28: 00-2019/01/23: 40845 time-interval water level data are organized into a hydrological time sequence data set in 00 hours, and after preprocessing steps such as missing value interpolation, data normalization and data set division are carried out on hydrological data of a target water level station and a related rainfall station in a forecast section, a related coefficient matrix optimal forecast factor is constructed;
s11, normalizing the input variable by adopting a Min-Max formula to accelerate the convergence rate and the training process, wherein the calculation method comprises the following steps:
wherein, X max 、X min Maximum and minimum values of the original sequence; x i Is an element in the original sequence; x' i Is the result of normalization, and has a value of 0,1]Within the range;
s12, excavating and extracting topographic features, spatial position relations and other spatial features of a forecast section, constructing a hydrological spatial relationship topological graph shown in the graph 2, and determining 14 rainfall stations and hydrological station observation elements which have influence on the hydrological process of the forecast section as forecasting factors of the model, wherein the 14 rainfall stations are respectively a convenience station, a soil bridge station, a crouchun station, a Zhao village, a clock mountain, an east mountain, an astronomical bridge station, a preweld village, a sentence capacity station, a sentence capacity reservoir, a Chishan sluice, a Bisheng bridge, an Tiansi temple and a Beishan reservoir;
s13, mining and extracting time sequence characteristics of the forecast section, calculating correlation coefficients between the 14 correlation rainfall stations, the hydrological stations and the forecast section determined in the step S12 by adopting Pearson correlation coefficients, and determining a correlation station X k ＝{x 1 ，x 2 ，x 3 ，.....x n M for predicted station Y = { Y = (k =1.. M) } 1 ，y 2 ，y 3 ，.....y n Forecasting elements (water level and flow) influence the time sequence length;
s14, for the relation station X in the hydrological space relation topological graph in the step S12 k Sequencing the spatial distance from the forecast station Y, and calculating according to the formula (2) to obtain the associated station X k Influencing the timing length T of the forecast station Y corr Extracting T-T corr Sequence value of t time is used as association station X k The forecast input factor of (1); influencing the timing length T corr The determination method of (2) is as follows:
1) Selecting the maximum correlation coefficient of the associated station with the minimum spatial distance with the forecast station Y in turn, and taking the time corresponding to the coefficient as the influence time sequence length T of the associated station with the minimum spatial distance corr ；
2) If the influence of two associated stations affects the timing length T corr If the spatial distances are the same but have larger difference, selecting the secondary large correlation number from the associated stations with larger distances, and taking the time corresponding to the coefficient as the length T of the influence time sequence corr (ii) a If the spatial distances of two associated stations are close, the same T can be adopted corr As the length of the affected time sequence of the two associated stations.
In this embodiment, the correlation coefficients of each association station and the forecast target station in the flow field of the qinhuai river and the length of the influence time sequence (black body) corresponding to the correlation coefficients are shown in table 1:
TABLE 1
In table 1, the east mountain water level (t + 1) represents the water level 1 hour in the future of the current time t of the east mountain station, and the east mountain (t-1) represents the precipitation amount 1 hour before the current time t of the east mountain station. If the optimum correlation coefficient of the natural bridge is at the time T-5, the corresponding T corr And 5, selecting the rainfall data of the station at the time of (t-5) -t as a forecasting factor of the forecasting model at the time of (t + 1), and the like.
S2, weight configuration: allocating different weights for the optimal forecasting factors by adopting an attention mechanism, establishing a long-distance dependency relationship of input and output sequences, and improving the correlation between input and output, wherein the method specifically comprises the following steps:
taking a forecast period of 1h (t +1 time) as an example, the forecast factor sequence optimized in the step S1 is
(k =1.. M) input to the attention mechanism, internal relations between input data are better mined, and hidden data are screened outCharacteristic information in the sequence. The attention mechanism calculation formula is as follows:
s i ＝tanh(w 1 y t+1 +w 2 X i +b) (2)
wherein, X i A sequence of predictor factors representing an ith associated station; y is t+1 A forecast element representing a forecast station at time t + 1; s i The attention weight score corresponding to the forecast factor sequence Xi of the ith associated station; w is a 1 、w 2 Representing a weight matrix; b is a bias matrix; a is i Representing the assigned attention weight of the current input; v i Is the output vector of the attention-driven layer.
S3, construction of a forecasting model: constructing a BilSTM network module, extracting time sequence characteristics between forecast input factors (hydrological data) from the two aspects of forward and reverse, and adopting a regularization structure to avoid model overfitting and improve the generalization capability of model parameters to data; constructing an enhanced CNN network module, extracting local spatial features of hydrological data by using CNN, deepening the depth of the CNN network by using ResNet, and mining more latent spatial feature information;
specifically, the S3 uses the new vector V obtained in the step S2 i As input, an optimal deep learning model is constructed to train a mapping function f (x) of the model, which includes the steps of:
s31: bi LSTM module f for constructing bidirectional long-time and short-time memory network 1 (x) Acquiring time sequence characteristics of hydrological data through three gate structures of forward LSTM and backward LSTM, avoiding overfitting of the network by using two regularization structures of BatchNormalization and Dropout, and reducing errors so as to obtain a prediction result y pred，1 ＝f 1 (x)；
S32: construction ofCNN network module f based on ResNet enhancement 2 (x) Performing convolution operation on the hydrological time sequence by adopting a one-dimensional convolution kernel of the CNN network, adding a plurality of residual error units ResNet to deepen the CNN network level, extracting the characteristic component of hydrological data on a local space structure, and obtaining a prediction result y pred，2 ＝f 2 (x)。
In this embodiment, the parameters of the model mainly include a learning rate lr, a batch size batch _ size, the number of hidden layers and nodes of hidden layers, an optimizer, an activation function, the number of filters, a convolution kernel size, dropout, a step size, the number of residual error units, and the like. Performing multiple experiments by using a random search algorithm, performing comparative analysis on errors and precision of experimental results, and finally determining each model parameter as shown in table 2:
TABLE 2
And S4, model integration and fusion: bi-directional long-short time memory network BilSTM module f 1 (x) And a CNN network module f based on ResNet enhancement 2 (x) Predicted result y of pred，1 、y pred，2 And linear fusion is respectively carried out through full connection layers, and an end-end hydrological integrated prediction model (MCR _ BilSTM) is integrated, so that high-precision intelligent flood prediction is realized.
In this embodiment, the bidirectional long-and-short time memory network BilSTM module f 1 (x) And a CNN network module f based on ResNet enhancement 2 (x) Weight assignment ω for linear fusion 1 、ω 2 The method specifically comprises the following steps: when the forecast period is 2h, omega 1 ＝0.5387，ω 2 =0.4613; when the forecast period is 4h, omega 1 ＝0.5513，ω 2 =0.4487; when the forecast period is 6h, omega 1 ＝0.5977，ω 2 ＝0.4023。
S5, carrying out inverse normalization processing on the MCR _ BilSTM model fusion result, and outputting a final model prediction result, wherein an inverse normalization formula is as follows:
y′ pred (t)＝y pred (t)*(X max -X min )+X min (11)
wherein, X max 、X min The maximum value and the minimum value are kept consistent during normalization; y is pred (t) is the normalized prediction result at time t; y' pred (t) is the prediction result at time t after denormalization.
The integrated hydrological forecasting method based on deep learning is used for forecasting water level data of the east mountain hydrological station of the Qinhuai river, SVM, CNN, bilSTM and CNN _ BilSTM models are adopted for comparison, and the selected time period is 2018/5/11: 00-2018/5/21 12:00, and 5, FIG. 3, FIG. 4, and FIG. 5 show the prediction process lines of the 5 models at 2 hours, 4 hours, and 6 hours in the future, respectively, and the evaluation indexes of the prediction results are RMSE and R 2 MAE and NSE, and the calculation formula is as follows:
wherein, y i 、
And &>
Respectively represent an observed value, a predicted value and an average value>
Is the average of the predicted values; n represents the number of samples.
Table 3 is a relevant comparison in this example.
TABLE 3
The results in the table show that the method has the best effect, and the four evaluation indexes are superior to other four models, which shows that the method has higher precision and stronger robustness.
According to the embodiment, the bidirectional long-time and short-time memory network BilSTM module and the CNN module based on ResNet enhancement are integrated, the time sequence characteristics of the hydrological sequence are mined, the local spatial characteristics of the hydrological sequence are considered, the forecasting effect better than that of a single deep learning model is obtained to a certain extent, the forecasting precision is improved, and the application value is good.
Claims (6)
1. An MCR _ BilSTM-based intelligent flood forecasting method is characterized by comprising the following steps:
the S1 forecasting factor is preferably: performing preprocessing steps such as missing value interpolation, data normalization, data set division and the like on hydrological data of a target water level station and a related rainfall station in a forecast section, and then constructing a correlation coefficient matrix optimal forecast factor;
s2, weight configuration: different weights are distributed to the optimal forecasting factors by adopting an attention mechanism, the long-distance dependency relationship of the input and output sequences is established, and the correlation between input and output is improved;
s3, construction of a forecasting model: constructing a BilSTM network module, extracting time sequence characteristics between forecast input factors (hydrological data) from the two aspects of forward and reverse, and adopting a regularization structure to avoid model overfitting and improve the generalization capability of model parameters to data; constructing an enhanced CNN network module, extracting local spatial features of hydrological data by using CNN, deepening the depth of the CNN network by using ResNet, and mining more latent spatial feature information;
and S4, model integration and fusion: the predictions of the BilSTM network module and the enhanced CNN network module are subjected to linear fusion after passing through a full connection layer respectively, and are integrated into a terminal-terminal hydrological integrated prediction model (MCR _ BilSTM), so that high-precision intelligent flood prediction is realized;
and S5, carrying out inverse normalization processing on the MCR _ BilSTM model fusion result, and outputting a final model prediction result.
2. An MCR _ BiLSTM-based intelligent flood forecasting method as claimed in claim 1, wherein the detailed steps of the step S1 are:
s11, normalizing the input variable by adopting a Min-Max formula to accelerate the convergence rate and the training process, wherein the calculation method comprises the following steps:
wherein, X max 、X min Maximum and minimum values of the original sequence; x i Is an element in the original sequence; x' i Is the result of normalization, and has a value of 0,1]Within the range;
s12, mining and extracting spatial features such as topographic features and spatial position relations of the forecast section, constructing a hydrological spatial relation topological graph, and determining rainfall stations and hydrological station observation elements which have influence on the hydrological process of the forecast section as forecasting factors of the model;
s13, mining and extracting time sequence characteristics of the forecast section, calculating correlation coefficients among the m correlation rainfall stations, the hydrological stations and the forecast section determined in the step S12 by adopting Pearson correlation coefficients, and determining a correlation station X k ＝{x 1 ，x 2 ，x 3 ，.....x n M for predicted station Y = { Y = (k =1.. M) } 1 ，y 2 ，y 3 ，.....y n The forecasting elements (water level and flow) influence the time sequence length, and the correlation coefficient calculation formula is as follows:
wherein n is the length of the sequence,
represents the mean value of the sequence, is>
Is covariance->
Is the standard deviation;
s14 for the associated station X in the hydrological space relationship topological graph in the step S12 k Sequencing the spatial distance from the forecast station Y, and calculating according to the formula (2) to obtain the associated station X k Influencing the timing length T of the forecast station Y corr Extracting T-T corr Sequence value of t time as related station X k The forecast input factor of (1); influencing the timing length T corr The determination method of (2) is as follows:
1) Selecting the maximum correlation coefficient of the associated station with the minimum spatial distance with the forecast station Y in turn, and taking the time corresponding to the coefficient as the influence time sequence length T of the associated station with the minimum spatial distance corr ；
2) If the influence of two associated stations affects the timing length T corr If the spatial distances are the same but have larger difference, selecting the secondary large correlation number from the associated stations with larger distances, and taking the time corresponding to the coefficient as the length T of the influence time sequence corr (ii) a If the spatial distances of two associated stations are close, the same T can be adopted corr As the length of the impact timing of the two associated stations.
3. The deep learning-based integrated hydrological forecasting method according to claim 1, wherein the step S2 comprises the following steps:
taking a forecast period of 1h (t +1 time) as an example, the forecast factor sequence X optimized in the step S1 k ＝{x t-Tcorr ，x t-Tcorr+1 ，.....，x t-1 ，x t In (k =1.. M) an attention mechanism is inputted, and moreAnd internal relations among the input data are well mined, so that characteristic information hidden in the sequence is screened out. The attention mechanism calculation formula is as follows:
s i ＝tanh(w 1 y t+1 +w 2 X i +b) (3)
wherein, X i A sequence of predictor factors representing an ith associated station; y is t+1 A forecast element representing a forecast station at time t + 1; s i The attention weight score corresponding to the forecast factor sequence Xi of the ith associated station; w is a 1 、w 2 Representing a weight matrix; b is a bias matrix; a is i Representing the assigned attention weight of the current input; v i Is the output vector of the attention-driven layer.
4. An MCR _ BiLSTM-based intelligent flood forecasting method as claimed in claim 1, wherein said step S3 uses the new vector V obtained in said step S2 i As an input, an optimal deep learning model is constructed to train a mapping function f (x) of the model, and the steps include:
s31: bi LSTM module f for constructing bidirectional long-time and short-time memory network 1 (x) Acquiring time sequence characteristics of hydrological data through three gate structures of forward LSTM and backward LSTM, avoiding overfitting of the network by using two regularization structures of BatchNormalization and Dropout, and reducing errors so as to obtain a prediction result y pred，1 ＝f 1 (x) The calculation formula is as follows:
wherein h is t Is a hidden state vector of the BilSTM;
hidden state vectors of positive and negative LSTM respectively;
s32: construction of CNN network module f based on ResNet enhancement 2 (x) Performing convolution operation on the hydrological time sequence by adopting a one-dimensional convolution kernel of the CNN network, adding a plurality of residual error units ResNet to deepen the CNN network level, extracting the characteristic component of hydrological data on a local space structure, and obtaining a prediction result y pred，2 ＝f 2 (x) The calculation formula is as follows:
X (l+1) ＝X l +g(X l ；θ l ) (8)
wherein the content of the first and second substances,
b (l) respectively is a weight matrix and a bias matrix of the l layer; g is a residual function; theta l Are parameters that can be learned by layer i.
5. The MCR _ BiLSTM-based intelligent flood forecasting method according to claim 1, wherein the step S4 is implemented by:
bi-directional long-short time memory network BilSTM module f 1 (x) And a CNN network module f based on ResNet enhancement 2 (x) Predicted result y of pred，1 、y pred，2 Linear fusion is respectively carried out through a full connection layer, and an end-end hydrological integrated prediction model (MCR _ BilSTM) is integrated, so that high-precision intelligent flood prediction is realized; the calculation formula for linear fusion is as follows:
y pred (t)＝ω 1 y pred，1 (t)+ω 2 y pred，2 (t) (9)
ω 1 +ω 2 ＝1 (10)
wherein, y pred (t) is the final forecast result at time t; y is pred，1 (t)、y pred，2 (t) the prediction results of the two network modules at the time t are respectively; omega 1 、ω 2 The weight assignment for the model is determined by the least squares method.
6. The MCR _ BiLSTM-based intelligent flood forecasting method according to claim 1, wherein the step S5 is implemented by:
performing inverse normalization processing on the MCR _ BilSTM model fusion prediction result, and outputting a final model prediction result, wherein an inverse normalization formula is as follows:
y′ pred (t)＝y pred (t)*(X max -X min )+X min (11)
wherein X max 、X min The normalization is kept consistent with the maximum value and the minimum value during normalization; y is pred (t) is the normalized prediction result at time t; y' pred (t) is the prediction result at time t after denormalization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211570315.0A CN115860231A (en) | 2022-12-08 | 2022-12-08 | MCR _ BilSTM-based intelligent flood forecasting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211570315.0A CN115860231A (en) | 2022-12-08 | 2022-12-08 | MCR _ BilSTM-based intelligent flood forecasting method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115860231A true CN115860231A (en) | 2023-03-28 |
Family
ID=85671020
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211570315.0A Pending CN115860231A (en) | 2022-12-08 | 2022-12-08 | MCR _ BilSTM-based intelligent flood forecasting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115860231A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116778395A (en) * | 2023-08-21 | 2023-09-19 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
-
2022
- 2022-12-08 CN CN202211570315.0A patent/CN115860231A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116778395A (en) * | 2023-08-21 | 2023-09-19 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
| CN116778395B (en) * | 2023-08-21 | 2023-10-24 | 成都理工大学 | Mountain torrent flood video identification monitoring method based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Falah et al. | Artificial neural networks for flood susceptibility mapping in data-scarce urban areas | |
| CN111612243B (en) | Traffic speed prediction method, system and storage medium | |
| Kumar et al. | Artificial neural network model for rainfall-runoff-A case study | |
| CN110766942B (en) | Traffic network congestion prediction method based on convolution long-term and short-term memory network | |
| Nikoo et al. | Flood-routing modeling with neural network optimized by social-based algorithm | |
| CN109583565B (en) | Flood prediction method based on attention model long-time and short-time memory network | |
| CN110047291B (en) | Short-term traffic flow prediction method considering diffusion process | |
| CN112132149B (en) | Semantic segmentation method and device for remote sensing image | |
| CN113673769B (en) | Traffic flow prediction method of graph neural network based on multivariate time sequence interpolation | |
| Yadav et al. | Prediction of suspended sediment yield by artificial neural network and traditional mathematical model in Mahanadi river basin, India | |
| CN116822382B (en) | Sea surface temperature prediction method and network based on space-time multiple characteristic diagram convolution | |
| CN112182063A (en) | Method for constructing hydrological forecasting model based on space-time characteristics | |
| CN107133398A (en) | A kind of river ethic Forecasting Methodology based on complex network | |
| CN115860231A (en) | MCR _ BilSTM-based intelligent flood forecasting method | |
| CN112561132A (en) | Water flow prediction model based on neural network | |
| CN112784479A (en) | Flood flow prediction method | |
| CN116910534A (en) | Space-time intelligent prediction method and device for ocean environmental elements in different sea areas | |
| CN115310532A (en) | Basin multipoint prediction early warning method based on space-time association mixed deep learning model | |
| Ouma et al. | Flood Susceptibility Mapping Using Image‐Based 2D‐CNN Deep Learning: Overview and Case Study Application Using Multiparametric Spatial Data in Data‐Scarce Urban Environments | |
| CN117233869B (en) | Site short-term wind speed prediction method based on GRU-BiTCN | |
| Kim et al. | Analysis of AI-based techniques for forecasting water level according to rainfall | |
| CN117636183A (en) | Small sample remote sensing image classification method based on self-supervision pre-training | |
| CN115797557A (en) | Self-supervision 3D scene flow estimation method based on graph attention network | |
| CN115762147A (en) | Traffic flow prediction method based on adaptive graph attention neural network | |
| CN112990041B (en) | Remote sensing image building extraction method based on improved U-net |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication |