CN117114190A - River runoff prediction method and device based on mixed deep learning - Google Patents
River runoff prediction method and device based on mixed deep learning Download PDFInfo
- Publication number
- CN117114190A CN117114190A CN202311096524.0A CN202311096524A CN117114190A CN 117114190 A CN117114190 A CN 117114190A CN 202311096524 A CN202311096524 A CN 202311096524A CN 117114190 A CN117114190 A CN 117114190A
- Authority
- CN
- China
- Prior art keywords
- memory network
- layer
- attention
- runoff
- long
- 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
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000013135 deep learning Methods 0.000 title claims abstract description 28
- 230000015654 memory Effects 0.000 claims abstract description 56
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 23
- 230000007246 mechanism Effects 0.000 claims abstract description 22
- 238000011144 upstream manufacturing Methods 0.000 claims abstract description 16
- 238000012795 verification Methods 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 19
- 230000008569 process Effects 0.000 claims description 15
- 238000013527 convolutional neural network Methods 0.000 claims description 14
- 238000004364 calculation method Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000009826 distribution Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 230000001364 causal effect Effects 0.000 abstract description 6
- 230000006403 short-term memory Effects 0.000 abstract description 6
- 230000007787 long-term memory Effects 0.000 abstract description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 2
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 241000088844 Nothocestrum Species 0.000 description 1
- 241001417524 Pomacanthidae Species 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Business, Economics & Management (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Databases & Information Systems (AREA)
- Human Resources & Organizations (AREA)
- Algebra (AREA)
- Marketing (AREA)
- Development Economics (AREA)
- Quality & Reliability (AREA)
- General Business, Economics & Management (AREA)
- Operations Research (AREA)
- Tourism & Hospitality (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of river runoff prediction, and discloses a river runoff prediction method and equipment based on mixed deep learning, wherein the method comprises the following steps: (1) Decomposing a daily runoff sequence of a target hydrologic station into modal components and residual errors by adopting variation modal decomposition, and constructing a convolution long and short memory network model based on an attention mechanism; (2) Constructing a data set by using the modal component and residual error of the daily runoff of the target hydrological station and the runoff sequence of the upstream hydrological station, and dividing the data set into a training set and a verification set according to time; (3) Training the convolution long-term and short-term memory network model by using a training set, and verifying the trained convolution long-term and short-term memory network model by using a verification set; and finally, predicting the daily runoff of the target hydrologic station by adopting a convolution long-short-term memory network model after verification is qualified. The invention can accurately consider the causal relation structure between the runoffs of the upstream and downstream stations and can obtain the runoff prediction result with high precision.
Description
Technical Field
The invention belongs to the technical field of river runoff prediction, and particularly relates to a river runoff prediction method and equipment based on mixed deep learning.
Background
Runoff is an important condition for supplying water for life, industry and agriculture, and is a limiting factor for the development scale of regional socioeconomic performance. The high-precision short-term flow forecast plays an extremely important role in preventing flood disasters, efficiently utilizing water resources, optimizing and dispatching reservoir groups and the like. The machine learning is introduced into the runoff prediction process, so that the runoff prediction method is expanded, and the runoff prediction precision is improved.
The actual load data is accompanied by irregular vacancies and noise interference, and in the case of insufficient data quantity, a combined prediction method is proposed for improving the prediction precision and can be classified into 2 types. The method is characterized in that a attention mechanism is introduced into the implicit state of the predicted LSTM and different weights are given to the implicit state by distributing weights and combining a plurality of algorithms, such as ' Interpretable spatio-temporal attention LSTM model for flood forecasting ' of NEUROCOMPUTING ' 2022, so that the influence of important information is enhanced, and the load prediction precision is improved.
The other is to decompose the load sequence and then predict, and common decomposition methods include wavelet transform (WA), empirical Mode Decomposition (EMD), and Variational Mode Decomposition (VMD). For example, in the meeting paper Research on streamflow forecast based on EEMD and long short-term memory Proceedings in the AIEA meeting of 2021, the load sequence is decomposed by adopting an empirical mode decomposition mode, then the obtained several eigenvalues are grouped and integrated, and finally the integrated eigenvalues are input into a mixed neural network for prediction. This approach reduces model complexity but ignores EEMD modal aliasing problems while not taking into account the effects of incoming flows.
The invention patent 'VMD-CNN-BiLSTM-ATT hybrid model-based weather drought prediction method and device' discloses a drought prediction method which introduces a variation modal decomposition and attention mechanism into a convolutional neural network and a long-term and short-term memory network, but can not predict the weather well when data such as weather, land utilization, vegetation coverage, topography and social water are lack as input.
In summary, the problems of the prior art are: (1) There is obvious causal correlation between upstream and downstream hydrologic stations of river channel runoff, and at present, few students introduce the upstream and downstream correlation into a machine learning model to predict the river channel runoff. How to utilize the causal relationship of the upstream and downstream stations to effectively extract the internal features of the runoff so as to better predict the target station is a technical problem existing in the current river runoff prediction. (2) The prior art has applied VMD techniques to radial flow prediction, but how to fully mine the decomposed modal features is a missing loop in the model framework.
Disclosure of Invention
Aiming at the defects or improvement demands of the prior art, the invention provides a river runoff prediction method and device based on mixed deep learning, which can accurately consider the causal relationship structure between upstream site runoffs and downstream site runoffs and can obtain a runoff prediction result with high precision.
To achieve the above object, according to one aspect of the present invention, there is provided a river runoff prediction method based on hybrid deep learning, the method further comprising the steps of:
(1) Decomposing a daily runoff sequence of a target hydrologic station into modal components and residual errors by adopting variation modal decomposition, and constructing a convolution long and short memory network model based on an attention mechanism;
(2) Constructing a data set by using modal components, residual errors and upstream hydrologic station runoff sequences of the target hydrologic station daily runoffs, and dividing the data set into a training set and a verification set according to time;
(3) Training a convolution long-period memory network model based on an attention mechanism by using a training set, and verifying the trained convolution long-period memory network model by using a verification set; and finally, predicting the daily runoff of the target hydrologic station by adopting a convolution long-short-term memory network model after verification is qualified.
Further, decomposing the daily runoff sequence of the target hydrologic station into modal components and residuals by adopting variational modal decomposition comprises the following sub-steps:
(1) Establishing constraint variation problems: assuming that the original signal f (t) is decomposed into K modal components with different frequency characteristics, the minimum sum of the estimated bandwidths of all the modalities is ensured, and meanwhile, the sum of all the modalities is equal to the original signal, a constraint variation expression is established as follows:
wherein: { u k -the kth modal component after signal decomposition; { omega k -the frequency center of the kth modal component; delta (t) represents dirac distribution; * Is a convolution operator;
(2) Solving an optimal solution of the constraint variation problem: and changing the constraint variation problem into the non-constraint variation problem through Lagrange transformation to solve the constraint variation problem:
wherein: alpha is a quadratic penalty term factor; lambda is the lagrangian multiplier.
Further, solving the optimal solution of the constraint variation problem comprises the steps of:
(2.1) initializationAnd->And the iteration number k; wherein,
for initial mode after signal decompositionComponent, & gt>Is the initial frequency center of the kth modal component,
is an initial Lagrangian multiplier;
(2.2) applying an alternate direction multiplier method to each modal component u k And its center frequency omega k Updating, wherein the formula is as follows:
(2.3) determining whether or not it satisfiesAnd k is<K, if so, outputting the final +.>And omega k If not, the step (2.2) is switched to update until the requirement is met.
Further, the convolution long-short memory network model consists of an input layer, a long-short memory network layer, a convolution neural network layer, an attention layer, a long-short memory network layer, a flattening layer, a full-connection layer and an output layer; the data flow process is as follows: the input layer receives the tag and the relevant factor feature set; (II) a long-period memory network layer, wherein two long-period memory network layers are added to learn and memorize the regularity of load data; (III) adding a one-dimensional convolutional neural network layer on the basis, extracting the internal characteristics of load data, and alternately building the internal characteristics by convolutional layers with multiplied convolutional kernel numbers and a maximum pooling layer; (iv) introducing an attention layer, applying attention weights to the outputs of the convolution layers, generating an attention weighted feature representation; (V) on the attention weighted feature representation, two long-short term memory network layers are applied again; (VI) adding a flattening layer and a full connecting layer, and outputting the features in a one-dimensional way; and (VII) outputting the prediction result by an output layer, wherein the output dimension corresponds to the dimension of the prediction target.
Further, a long-period memory network is adopted in the long-period memory network layer, and the calculation formula is as follows:
i (t) ＝σ(W i h (t-1) +U i x (t) +b i )
a (t) ＝tanh(W a h (t-1) +U a x (t) +b a )
C (t) ＝C (t-1) ⊙f (t) +i (t) ⊙a (t)
o (t) ＝sigmod(W o h (t-1) +U o x (t) +b o )
h (t) ＝o (t) ⊙tanh(C (t) )
wherein i is (t) A is an input door (t) Is forgetful door C (t) O is input unit (t) For outputting the door h (t) Is an output vector; x is x (t) Representing input vector, C (t-1) Representing the state of the cell at the last moment, W i 、W a 、W o Weight matrix of input gate, forget gate and output gate respectively, b i 、b a 、b o Bias terms of the input gate, the forget gate and the output gate, respectively, and the product of Hadamard is as follows.
Further, a convolutional neural network is adopted in the convolutional neural network layer, and the calculation formula is as follows:
in the method, in the process of the invention,the output representing the kth feature map, f is the activation function, W k Representing weights linked to the kth feature map, x representing input data, i, j representingSize of convolution kernel, b k Representing the deviation of the kth feature map.
Further, the attention mechanism applies attention weights to the output of the convolution layer, generates an attention weighted feature representation, and the attention mechanism corresponds to the steps of:
(1) Let the input sequence at time t beThe weight calculation process comprises the following steps:
in the method, in the process of the invention,weights representing the ith feature of the tth time input sequence, +.>
(2) Combining the time weights of each input feature to obtain an attention matrix of each input feature:
(3) Setting a time sequence of input features asThe attention weight calculation process of the time series of the input features is as follows:
in the middle ofWeights representing individual times, +.>
(4) Combining the time weights of each input feature to obtain an attention matrix of each time:
(5) Performing hadamard product on the obtained feature attention weight matrix A and the time attention weight matrix B to obtain a time and feature attention weight matrix C:
on the basis of dividing the data set, training a convolution long-short-period memory network model based on an attention mechanism by utilizing a training set, and correcting model parameters of the convolution long-period memory network model by adopting a self-adaptive moment estimation algorithm with mean square error as a target; and outputting a training value of a daily runoff decomposition mode and a residual error of the target hydrologic station by the trained convolution long-short-term memory network model.
The invention also provides a river runoff prediction system based on the mixed deep learning, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the river runoff prediction method based on the mixed deep learning when executing the computer program.
The present invention also provides a computer readable storage medium storing machine executable instructions that, when invoked and executed by a processor, cause the processor to implement a mixed deep learning based river runoff prediction method as described above.
In general, compared with the prior art, the river runoff prediction method and the river runoff prediction equipment based on mixed deep learning have the following advantages:
1. aiming at the defect that the current river runoff prediction method for deep learning rarely introduces an upstream-downstream relation in a machine learning model to perform river runoff prediction, the method adopts the runoffs which are introduced into the upstream hydrologic station related to the hydrologic station to jointly perform the runoff prediction of the target hydrologic station, fully considers the characteristic of space-time variation of the runoffs, and ensures that the runoffs of the target hydrologic station simulated by the model are more accurate.
2. On the basis of utilizing the causal relationship between upstream and downstream flow of a river channel, the variation mode demarcation is adopted to filter the runoff sequence interference information, the convolutional neural network and the attention mechanism focus on key influencing factors, the accuracy and the stability of the model to the runoff simulation of a target site are improved, and technical support is provided for water resource planning management.
Drawings
FIG. 1 is a flow diagram of a river runoff prediction method based on mixed deep learning;
FIG. 2 is a schematic diagram of a long-short-term memory convolutional neural network model based on an attention mechanism according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an attention mechanism provided by an embodiment of the present invention;
fig. 4 is a schematic diagram of comparison between a predicted curve and an actual measured curve of a prediction method and a prediction method using a comparison model according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
The invention provides a river runoff prediction method based on mixed deep learning, which utilizes the causal relation between upstream and downstream flows of a river and adopts a variation mode demarcation to filter runoff sequence interference information, a convolutional neural network and an attention mechanism focus on key influencing factors, so that the accuracy and stability of a model to runoff simulation of a target site are improved, and technical support is provided for water resource planning management.
Referring to fig. 1, 2 and 3, the prediction method mainly includes the following steps:
and step one, decomposing a daily runoff sequence of the target hydrologic station into a modal component and a residual error by adopting variation modal decomposition, and constructing a convolution long and short memory network model based on an attention mechanism.
Decomposing the daily runoff sequence of the target hydrologic station into modal components and residual errors by adopting variational modal decomposition comprises the following substeps:
(1) Establishing constraint variation problems: assuming that the original signal f (t) is decomposed into K modal components with different frequency characteristics, the minimum sum of the estimated bandwidths of all the modalities is ensured, and meanwhile, the sum of all the modalities is equal to the original signal, a constraint variation expression is established as follows:
wherein: { u k -the kth modal component after signal decomposition; { omega k -the frequency center of the kth modal component; delta (t) represents dirac distribution; * Is a convolution operator.
(2) Solving an optimal solution of the constraint variation problem: and changing the constraint variation problem into the non-constraint variation problem through Lagrange transformation to solve the constraint variation problem:
wherein: alpha is a quadratic penalty term factor; lambda is the lagrangian multiplier.
Solving the constraint variation problem optimal solution comprises the following steps:
(2.1) initializationAnd->And the iteration number k; wherein (1)>For the kth initial modal component after signal decomposition, < ->Is the initial frequency center of the kth modal component, < ->Is an initial Lagrangian multiplier;
(2.2) applying an alternate direction multiplier method to each modal component u k And its center frequency omega k Updating is performed according to the following formula:
(2.3) determining whether or not it satisfiesAnd k is<K, if so, outputting the final +.>And omega k If not, returning to the step (2.2) for updating until the requirement is met.
(3) Calculating a decomposition residual error: and calculating the residual error as the difference between the actual runoff sequence of the predicted target site and each mode sequence after decomposition.
The convolution long-short memory network model consists of an input layer, a long-short memory network layer, a convolution neural network layer, a attention layer, a long-short memory network layer, a flattening layer, a full connection layer and an output layer.
The data flow process is as follows: the input layer receives the tag and the relevant factor feature set; (II) a long-period memory network layer, wherein two long-period memory network layers are added to learn and memorize the regularity of load data; (III) adding a one-dimensional convolutional neural network layer on the basis, extracting the internal characteristics of load data, and alternately building the internal characteristics by convolutional layers with multiplied convolutional kernel numbers and a maximum pooling layer; (iv) introducing an attention layer, applying attention weights to the outputs of the convolution layers, generating an attention weighted feature representation; (V) on the attention weighted feature representation, two long-short term memory network layers are applied again; (VI) adding a flattening layer and a full connecting layer, and outputting the features in a one-dimensional way; and (VII) outputting the prediction result by an output layer, wherein the output dimension corresponds to the dimension of the prediction target.
The long-period memory network layer adopts a long-period memory network, and the calculation definition is as follows:
i (t) ＝σ(W i h (t-1) +U i x (t) +b i )
a (t) ＝tanh(W a h (t-1) +U a x (t) +b a )
C (t) ＝C (t-1) ⊙f (t) +i (t) ⊙a (t)
o (t) ＝sigmod(W o h (t-1) +U o x (t) +b o )
h (t) ＝o (t) ⊙tanh(C (t) )
wherein i is (t) A is an input door (t) Is forgetful door C (t) O is input unit (t) For outputting the door h (t) Is an output vector; x is x (t) Representing input vector, C (t-1) Representing the state of the cell at the last moment, W i 、W a 、W o Weight matrix of input gate, forget gate and output gate respectively, b i 、b a 、b o Bias terms of the input gate, the forget gate and the output gate, respectively, and the product of Hadamard is as follows.
The convolutional neural network layer adopts a convolutional neural network, and the calculation definition is as follows:
in the method, in the process of the invention,the output representing the kth feature map, f is the activation function, W k Representing weights linked to the kth feature map, x representing the input data, i, j representing the size of the convolution kernel, b k Representing the deviation of the kth feature map.
The attention mechanism applies attention weights to the output of the convolution layer, generates attention weighted feature representations, focuses on important features by calculating different input feature probability distributions, ignores irrelevant information, and accordingly improves the capability of the model in utilizing the key features. For m input feature quantities of n time nodes, forming a two-dimensional matrix X epsilon R of m rows and n columns m×n The attention mechanism mainly comprises the following steps:
(1) Let the input sequence at time t beThe weight calculation process is as follows:
wherein:weights representing the ith feature of the tth time input sequence, +.>
(2) Combining the time weights of each input feature to obtain an attention matrix of each input feature:
(3) Setting a time sequence of input features asThe attention weight calculation process of the time series of the input features is as follows:
in the middle ofWeights representing individual times, +.>
(4) Combining the time weights of each input feature to obtain an attention matrix of each time:
(5) Performing hadamard product on the obtained feature attention weight matrix A and the time attention weight matrix B to obtain a time and feature attention weight matrix C:
and secondly, constructing a data set by using the modal component, residual error and upstream hydrologic station runoff sequence of the target hydrologic station daily runoff, and dividing the data set into a training set and a verification set according to time.
The target hydrologic station refers to a hydrologic station needing daily runoff prediction; an upstream hydrologic station refers to a hydrologic station located in the river upstream of the target hydrologic station and having hydrologic correlation therewith.
Training the convolution long-period memory network model based on the attention mechanism by using a training set, and verifying the trained convolution long-period memory network model by using a verification set; and finally, predicting the daily runoff of the target hydrologic station by adopting a convolution long-short-term memory network model after verification is qualified.
On the basis of dividing a data set, training a convolution long-short-period memory network model based on an attention mechanism by utilizing a training set, and correcting model parameters of the convolution long-period memory network model by adopting a self-adaptive moment estimation algorithm with mean square error as a target; and outputting a training value of a daily runoff decomposition mode and a residual error of the target hydrologic station by the trained convolution long-short-term memory network model.
The self-adaptive moment estimation algorithm obtains a first moment estimation m by calculating a mean square error gradient of a training set t And second moment estimate v t To calculate the individual adaptive learning rate of the parameter. Setting the objective function as the mean square error f t (θ) represents a random function of the parameter θ at the t-th time step. To reduce f t The expectation of (θ) requires the use of small sample noise described by randomness, the gradient of the objective function with respect to the variable θ is calculated as follows:
calculating a first moment estimate m t And second moment estimate v t The formula of (2) is as follows:
m t ＝β 1 m t-1 +(1-β 1 )g t
wherein the parameter beta 1 、β 2 E [0, 1), control m t And v t Is not less than a predetermined threshold. To solve for early moment estimate values of 0, m is generally required to be equal to t And v t Respectively correcting and estimating the first moment after correctionAnd second moment estimates are:
every iteration of a cycle, the parameter θ is updated t The updating method comprises the following steps:
wherein alpha is the learning rate; epsilon=10 -8 Is a constant parameter. And (5) converging the objective function to the optimal solution direction through parameter updating iteration.
And then verifying the trained convolution long-short-period memory network model by using the verification set, and evaluating the verification result of the convolution long-period memory network model by using the evaluation index.
And finally, predicting the daily runoff of the target hydrologic station by adopting the verified convolution long-term and short-term memory network model. Firstly, performing variation modal decomposition on measured runoff data of a target hydrologic station in front of T days to obtain a measured runoff decomposition mode and residual error of the target hydrologic station, and secondly, forming an input data set by the measured runoff decomposition mode and residual error of the target hydrologic station and the measured runoff data of an upstream hydrologic station in front of T-3 days to serve as input of a convolution long-term and short-term memory network model based on an attention mechanism; outputting a predicted value of a real-time daily runoff decomposition mode and a residual error of the target hydrologic station on the T+1th day by the model; and finally, carrying out linear addition on the actual measurement daily runoff decomposition mode and the predicted value of the residual error, thereby obtaining the actual measurement daily runoff on the T+1th day of the target hydrologic station.
The invention will be further described in detail with reference to the following examples.
The invention takes Yichang, zhicheng, shashi and Guanli four hydrological site runoff data of Yichang to Ling Angeles section of Yangtze river basin as objects, the data adopts 1 day of data time step from 10 months 1 to 10 months 1 year 2020 in 2006, the total time period is 5145, 8/9 is a training set before division, and 1/9 is a test set after division. And selecting 3 time periods before the prison, the upstream hydrologic station, such as the sandy city, the branch city and the Yichang, as characteristics, and taking the prison station as a target hydrologic station.
To verify the predictive performance of the mixed deep learning based river runoff prediction method, a model (CNN-LSTM) for prediction by combining CNN and LSTM was constructed to predict the average flow rate and compared. Table 1 lists the evaluation index of the predicted average runoff for the two models. The evaluation index adopts Root Mean Square Error (RMSE), mean Absolute Error (MAE) and determination coefficient (R) 2 ) The smaller the value, the higher the prediction accuracy, the average absolute error percent (MAPE). As can be seen from Table 1, the prediction accuracy of VMD-CNN-AM-LSTM (the method provided by the present invention) is higher than that of CNN-LSTM model, which indicates that the method of the present invention is more excellent. The absolute difference between the two models and the measured value is calculated in fig. 4, and the difference of the prediction accuracy of the two models can be seen more clearly.
TABLE 1 runoff prediction index comparison Table
| Models | RMSE | MAE | R 2 | MAPE(％) |
| CNN-LSTM | 895.3777 | 535.3072 | 0.9915 | 2.96 |
| VMD-CNN-AM-LSTM | 585.3610 | 382.4830 | 0.9944 | 2.37 |
The invention also provides a river runoff prediction system based on the mixed deep learning, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the river runoff prediction method based on the mixed deep learning when executing the computer program.
The present invention also provides a computer readable storage medium storing machine executable instructions that, when invoked and executed by a processor, cause the processor to implement a mixed deep learning based river runoff prediction method as described above.
It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.
Claims (10)
1. The river runoff prediction method based on mixed deep learning is characterized by further comprising the following steps of:
(1) Decomposing a daily runoff sequence of a target hydrologic station into modal components and residual errors by adopting variation modal decomposition, and constructing a convolution long and short memory network model based on an attention mechanism;
(2) Constructing a data set by using modal components, residual errors and upstream hydrologic station runoff sequences of the target hydrologic station daily runoffs, and dividing the data set into a training set and a verification set according to time;
(3) Training a convolution long-period memory network model based on an attention mechanism by using a training set, and verifying the trained convolution long-period memory network model by using a verification set; and finally, predicting the daily runoff of the target hydrologic station by adopting a convolution long-short-term memory network model after verification is qualified.
2. The river runoff prediction method based on mixed deep learning as claimed in claim 1, wherein: decomposing the daily runoff sequence of the target hydrologic station into modal components and residual errors by adopting variational modal decomposition comprises the following substeps:
(1) Establishing constraint variation problems: assuming that the original signal f (t) is decomposed into K modal components with different frequency characteristics, the minimum sum of the estimated bandwidths of all the modalities is ensured, and meanwhile, the sum of all the modalities is equal to the original signal, a constraint variation expression is established as follows:
wherein: { u k -the kth modal component after signal decomposition; { omega k -the frequency center of the kth modal component; delta (t) represents dirac distribution; * Is a convolution operator;
(2) Solving an optimal solution of the constraint variation problem: and changing the constraint variation problem into the non-constraint variation problem through Lagrange transformation to solve the constraint variation problem:
wherein: alpha is a quadratic penalty term factor; lambda is the lagrangian multiplier.
3. The river runoff prediction method based on mixed deep learning as claimed in claim 2, wherein: solving the optimal solution of the constraint variation problem comprises the following steps:
(2.1) initializationAnd->And the iteration number k; wherein (1)>For the initial modal component after signal decomposition, +.>Is the initial frequency center of the kth modal component, < ->Is an initial Lagrangian multiplier;
(2.2) applying an alternate direction multiplier method to each modal component u k And its center frequency omega k Updating, wherein the formula is as follows:
(2.3) determining whether or not it satisfiesAnd k is<K, if so, outputting the final +.>And omega k If not, the step (2.2) is switched to update until the requirement is met.
4. The river runoff prediction method based on mixed deep learning as claimed in claim 1, wherein: the convolution long-short memory network model consists of an input layer, a long-short memory network layer, a convolution neural network layer, a attention layer, a long-short memory network layer, a flattening layer, a full-connection layer and an output layer; the data flow process is as follows: (I) the input layer receives the tag and the set of relevant factor characteristics; (II) a long-period memory network layer, wherein two long-period memory network layers are added, and regularity between load data is learned and memorized; (III) adding a one-dimensional convolutional neural network layer on the basis, and extracting the internal characteristics of load data, wherein the internal characteristics are formed by alternately building a convolutional layer with multiplied convolutional kernel number and a maximum pooling layer; (IV) introducing an attention layer, applying attention weights to the outputs of the convolution layers, generating an attention weighted feature representation; (V) on the attention weighted feature representation, two long-short term memory network layers are applied again; (VI) adding a flattening layer and a full connection layer, and carrying out one-dimensional output on the characteristics; and (VII) outputting the prediction result by the output layer, wherein the output dimension corresponds to the dimension of the prediction target.
5. The river runoff prediction method based on mixed deep learning as claimed in claim 4, wherein: the long-period memory network layer adopts a long-period memory network, and the calculation formula is as follows:
i (t) ＝σ(W i h (t-1) +U i x (t) +b i )
a (t) ＝tanh(W a h (t-1) +U a x (t) +b a )
C (t) ＝C (t-1) ⊙f (t) +i (t) ⊙a (t)
o (t) ＝sigmod(W o h (t-1) +U o x (t) +b o )
h (t) ＝o (t) ⊙tanh(C (t) )
wherein i is (t) A is an input door (t) Is forgetful door C (t) O is input unit (t) For outputting the door h (t) Is an output vector; x is x (t) Representing input vector, C (t-1) Indicating the last time sheetMeta state, W i 、W a 、W o Weight matrix of input gate, forget gate and output gate respectively, b i 、b a 、b o Bias terms of the input gate, the forget gate and the output gate, respectively, and the product of Hadamard is as follows.
6. The river runoff prediction method based on mixed deep learning as claimed in claim 5, wherein: the convolutional neural network layer adopts a convolutional neural network, and the calculation formula is as follows:
in the method, in the process of the invention,the output representing the kth feature map, f is the activation function, W k Representing weights linked to the kth feature map, x representing the input data, i, j representing the size of the convolution kernel, b k Representing the deviation of the kth feature map.
7. The river runoff prediction method based on mixed deep learning as claimed in claim 5, wherein: the attention mechanism applies attention weights to the output of the convolution layer, generates an attention weighted feature representation, and corresponds to the steps of:
(1) Let the input sequence at time t beThe weight calculation process comprises the following steps:
in the method, in the process of the invention,weights representing the ith feature of the tth time input sequence, +.>
(2) Combining the time weights of each input feature to obtain an attention matrix of each input feature:
(3) Setting a time sequence of input features asThe attention weight calculation process of the time series of the input features is as follows:
in the middle ofWeights representing individual times, +.>
(4) Combining the time weights of each input feature to obtain an attention matrix of each time:
(5) Performing hadamard product on the obtained feature attention weight matrix A and the time attention weight matrix B to obtain a time and feature attention weight matrix C:
8. the river runoff prediction method based on mixed deep learning as claimed in claim 1, wherein: on the basis of dividing a data set, training a convolution long-short-period memory network model based on an attention mechanism by utilizing a training set, and correcting model parameters of the convolution long-period memory network model by adopting a self-adaptive moment estimation algorithm with mean square error as a target; and outputting a training value of a daily runoff decomposition mode and a residual error of the target hydrologic station by the trained convolution long-short-term memory network model.
9. River runoff prediction system based on mixed deep learning, which is characterized in that: the system comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the mixed deep learning-based river runoff prediction method according to any one of claims 1-8 when executing the computer program.
10. A computer-readable storage medium, characterized by: the computer-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the hybrid deep learning-based river runoff prediction method of any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311096524.0A CN117114190A (en) | 2023-08-29 | 2023-08-29 | River runoff prediction method and device based on mixed deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311096524.0A CN117114190A (en) | 2023-08-29 | 2023-08-29 | River runoff prediction method and device based on mixed deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117114190A true CN117114190A (en) | 2023-11-24 |
Family
ID=88807264
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311096524.0A Pending CN117114190A (en) | 2023-08-29 | 2023-08-29 | River runoff prediction method and device based on mixed deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117114190A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117744884A (en) * | 2023-12-29 | 2024-03-22 | 南方电网调峰调频发电有限公司鲁布革水力发电厂 | Reservoir water flow prediction model construction method and reservoir water flow prediction method |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401529A (en) * | 2020-04-21 | 2020-07-10 | 中国水利水电科学研究院 | River ecological flow supervision method based on remote sensing technology |
| CN113361777A (en) * | 2021-06-08 | 2021-09-07 | 淮阴工学院 | Runoff prediction method and system based on VMD decomposition and IHHO optimization LSTM |
| CN113705864A (en) * | 2021-08-13 | 2021-11-26 | 华北水利水电大学 | Weather drought prediction method and device based on VMD-CNN-BilSTM-ATT mixed model |
| CN114186734A (en) * | 2021-12-10 | 2022-03-15 | 华中科技大学 | River flood prediction method capable of explaining research strategy based on deep learning model |
| CN114218870A (en) * | 2021-12-22 | 2022-03-22 | 大连理工大学 | Wind speed prediction method based on variational modal decomposition and attention mechanism |
| CN114692976A (en) * | 2022-04-01 | 2022-07-01 | 重庆大学 | Travel intention prediction method based on semi-supervised graph embedded frame |
| CN115347571A (en) * | 2022-10-17 | 2022-11-15 | 国网江西省电力有限公司电力科学研究院 | Photovoltaic power generation power short-term prediction method and device based on transfer learning |
| CN115455817A (en) * | 2022-09-06 | 2022-12-09 | 云南大学 | Short-term runoff hydrological prediction method based on deep learning |
| CN115496263A (en) * | 2022-08-24 | 2022-12-20 | 苏州中立志科技有限公司 | Short-term photovoltaic power generation prediction method based on space-time genetic attention network |
| US20220405864A1 (en) * | 2019-11-19 | 2022-12-22 | Zhejiang University | Crop yield estimation method based on deep temporal and spatial feature combined learning |
| CN116050601A (en) * | 2022-12-31 | 2023-05-02 | 湖南大唐先一科技有限公司 | Load prediction method based on variational modal decomposition and long-term and short-term memory mixed model |
-
2023
- 2023-08-29 CN CN202311096524.0A patent/CN117114190A/en active Pending
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220405864A1 (en) * | 2019-11-19 | 2022-12-22 | Zhejiang University | Crop yield estimation method based on deep temporal and spatial feature combined learning |
| CN111401529A (en) * | 2020-04-21 | 2020-07-10 | 中国水利水电科学研究院 | River ecological flow supervision method based on remote sensing technology |
| CN113361777A (en) * | 2021-06-08 | 2021-09-07 | 淮阴工学院 | Runoff prediction method and system based on VMD decomposition and IHHO optimization LSTM |
| CN113705864A (en) * | 2021-08-13 | 2021-11-26 | 华北水利水电大学 | Weather drought prediction method and device based on VMD-CNN-BilSTM-ATT mixed model |
| CN114186734A (en) * | 2021-12-10 | 2022-03-15 | 华中科技大学 | River flood prediction method capable of explaining research strategy based on deep learning model |
| CN114218870A (en) * | 2021-12-22 | 2022-03-22 | 大连理工大学 | Wind speed prediction method based on variational modal decomposition and attention mechanism |
| CN114692976A (en) * | 2022-04-01 | 2022-07-01 | 重庆大学 | Travel intention prediction method based on semi-supervised graph embedded frame |
| CN115496263A (en) * | 2022-08-24 | 2022-12-20 | 苏州中立志科技有限公司 | Short-term photovoltaic power generation prediction method based on space-time genetic attention network |
| CN115455817A (en) * | 2022-09-06 | 2022-12-09 | 云南大学 | Short-term runoff hydrological prediction method based on deep learning |
| CN115347571A (en) * | 2022-10-17 | 2022-11-15 | 国网江西省电力有限公司电力科学研究院 | Photovoltaic power generation power short-term prediction method and device based on transfer learning |
| CN116050601A (en) * | 2022-12-31 | 2023-05-02 | 湖南大唐先一科技有限公司 | Load prediction method based on variational modal decomposition and long-term and short-term memory mixed model |
Non-Patent Citations (2)
| Title |
|---|
| DONGYANG HAN: "An attention-based LSTM model for long-term runoff forecasting and factor recognition", 《ENVIRONMENTAL RESEARCH》, 17 January 2023 (2023-01-17), pages 1 - 13 * |
| LING KANG 等: "deep learning framework with time series analysis methods for runoff prediction", 《WATER》, 23 February 2021 (2021-02-23), pages 1 - 16 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117744884A (en) * | 2023-12-29 | 2024-03-22 | 南方电网调峰调频发电有限公司鲁布革水力发电厂 | Reservoir water flow prediction model construction method and reservoir water flow prediction method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tian | Modes decomposition forecasting approach for ultra-short-term wind speed | |
| CN109902801B (en) | Flood collective forecasting method based on variational reasoning Bayesian neural network | |
| Aly | An intelligent hybrid model of neuro Wavelet, time series and Recurrent Kalman Filter for wind speed forecasting | |
| Araghinejad | Data-driven modeling: using MATLAB® in water resources and environmental engineering | |
| CN109583565B (en) | Flood prediction method based on attention model long-time and short-time memory network | |
| CN110942194A (en) | Wind power prediction error interval evaluation method based on TCN | |
| CN111563706A (en) | Multivariable logistics freight volume prediction method based on LSTM network | |
| CN111767517B (en) | BiGRU multi-step prediction method, system and storage medium applied to flood prediction | |
| CN105868853B (en) | Method for predicting short-term wind power combination probability | |
| Agarwal et al. | Runoff modelling through back propagation artificial neural network with variable rainfall-runoff data | |
| McAdam et al. | Forecasting inflation with thick models and neural networks | |
| CN107153874A (en) | Water quality prediction method and system | |
| Heng et al. | Probabilistic and deterministic wind speed forecasting based on non-parametric approaches and wind characteristics information | |
| CN110210656B (en) | Shared bicycle flow prediction method and system based on station behavior analysis | |
| CN113592144A (en) | Medium-and-long-term runoff probability forecasting method and system | |
| CN112598248A (en) | Load prediction method, load prediction device, computer equipment and storage medium | |
| CN109637196A (en) | En-route sector traffic probability density prediction technique | |
| CN117114190A (en) | River runoff prediction method and device based on mixed deep learning | |
| Lv et al. | An improved long short-term memory neural network for stock forecast | |
| CN110826774A (en) | Bus load prediction method and device, computer equipment and storage medium | |
| Loucks et al. | An introduction to probability, statistics, and uncertainty | |
| CN111461786A (en) | Goods sales prediction method and device based on Prophet-CEEMDAN-ARIMA | |
| CN115587666A (en) | Load prediction method and system based on seasonal trend decomposition and hybrid neural network | |
| CN114694379B (en) | Traffic flow prediction method and system based on self-adaptive dynamic graph convolution | |
| CN114723188A (en) | Water quality prediction method, device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |