CN111723523B - Estuary surplus water level prediction method based on cascade neural network - Google Patents

Estuary surplus water level prediction method based on cascade neural network Download PDF

Info

Publication number
CN111723523B
CN111723523B CN202010572944.1A CN202010572944A CN111723523B CN 111723523 B CN111723523 B CN 111723523B CN 202010572944 A CN202010572944 A CN 202010572944A CN 111723523 B CN111723523 B CN 111723523B
Authority
CN
China
Prior art keywords
water level
network
neural network
estuary
key data
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.)
Active
Application number
CN202010572944.1A
Other languages
Chinese (zh)
Other versions
CN111723523A (en
Inventor
任磊
袁家敏
潘广维
姬进财
傅林曦
杨清书
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sun Yat Sen University
Original Assignee
Sun Yat Sen University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Sun Yat Sen University filed Critical Sun Yat Sen University
Priority to CN202010572944.1A priority Critical patent/CN111723523B/en
Publication of CN111723523A publication Critical patent/CN111723523A/en
Application granted granted Critical
Publication of CN111723523B publication Critical patent/CN111723523B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • 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
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/10—Numerical modelling

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Operations Research (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a estuary surplus water level prediction method based on a cascade neural network, which comprises the following steps: acquiring key data influencing the change of the estuary water level, wherein the key data comprises first key data and second key data, and preprocessing the key data to obtain a data set; the data set is divided into a training set, a verification set and a test set; establishing a two-stage serial residual water level neural network prediction model, wherein a first-stage network is a GA-BP neural network, and a second-stage network is an RBF neural network; training the primary network by taking the first key data as input variables, and taking output values obtained by training and the second key data as input variables of the secondary network; training the secondary network, and outputting a predicted value of the estuary residual water level when the error meets the requirement. According to the invention, a plurality of factors influencing the residual water level are comprehensively considered, the prediction capability of the model is optimized, the respective advantages of the GA-BP network and the RBF network are fully exerted, and the accuracy and the efficiency of model prediction are improved.

Description

Estuary surplus water level prediction method based on cascade neural network
Technical Field
The invention relates to the technical field of measurement and control, in particular to a estuary surplus water level prediction method based on a cascade neural network.
Background
The residual water level, also called the average tide level (such as the average daily tide period water level, taking the average sea level for years as the elevation zero point), is a typical result of the nonlinear coupling action of two large powers of river mouth runoff and tide. Under the synergistic effect of the outside sea tide power and the river basin runoff power, the residual water level in the estuary area is influenced by multiple factors to present the characteristic of multiple time-space scale changes. The method has the advantages that the change rule of the residual water level is mastered, the residual water level is accurately predicted in real time, important support can be provided for estuary coastal treatment and ecological protection, and important reference can be provided for efficient utilization of estuary water resources and flood control safety.
Because the estuary area radial tide power process has obvious nonlinear coupling characteristics, the existing analysis and prediction of estuary residual water level are mostly based on methods such as an analytical solution method, a harmonic analysis method, a linear regression method, a numerical simulation method and the like. The existing method has limitations in residual water level prediction, an analysis solution simplifies or approximates the environment of residual water level/residual water level gradient of a research target variable through mathematical deduction, and a mathematical expression is obtained, wherein the use condition of the mathematical expression has limitations, namely the analysis formula can obtain more accurate results at certain estuaries and can not be applicable to other types of estuaries; harmonic analysis typically selects a representative harmonic constant to characterize its varying characteristics, while ignoring other signal information often results in an incomplete analysis; the linear regression method is suitable for describing and applying a linear process, and the residual water level with obvious nonlinear characteristics is difficult to reach the high-accuracy requirement by regression prediction; the given initial conditions and boundary conditions in the numerical simulation method directly affect the simulation accuracy of the residual water level, and small errors in the initial conditions and boundary conditions can cause large errors in the residual water level of the simulation target variable.
Therefore, a prediction method capable of accurately predicting the estuary residual water level by comprehensively considering the nonlinear characteristics of the residual water level and the change rule of multiple influencing factors is urgently needed.
Disclosure of Invention
The invention aims to provide a cascade neural network-based estuary residual water level prediction method, which aims to solve the technical problems that the existing method has limitation and larger prediction result error when predicting the estuary residual water level.
The aim of the invention can be achieved by the following technical scheme:
A estuary surplus water level prediction method based on a cascade neural network comprises the following steps:
Acquiring key data influencing the change of the estuary water level, wherein the key data comprises first key data and second key data, and preprocessing the key data to obtain a data set; wherein the data set is divided into a training set, a verification set and a test set;
Establishing a two-stage serial residual water level neural network prediction model, wherein the prediction model comprises a primary network and a secondary network, the primary network is a network combining a BP neural network with a genetic algorithm, and the secondary network is an RBF neural network;
Training the primary network by taking the first key data in the training set as input variables, and taking output values obtained by training and the second key data in the training set as input variables of the secondary network;
Training the secondary network, and outputting a predicted value of the estuary residual water level when the error meets the requirement.
Optionally, the first key data comprises the outside sea tide level, the water depth, the river width, the flow speed, the river basin runoff and the water level, and the second key data comprises the runoff factor, the tide factor and the tide action factor.
Optionally, preprocessing the key data further includes: and carrying out space-time unification, time interval division and standardization processing on the key data.
Optionally, establishing the two-stage serial residual water level neural network prediction model further includes: and respectively establishing a two-stage series-connection residual water level neural network prediction model for a plurality of characteristic stages, wherein the characteristic stages are divided according to residual water level characteristics during pretreatment.
Optionally, the characteristic stage specifically includes: flood period big tide stage, flood period little tide stage, dead water period big tide stage, dead water period little tide stage, level water period big tide stage and level water period little tide stage.
Optionally, training the primary network specifically includes: initializing parameters of the BP neural network, training the BP neural network by utilizing an optimal weight and a threshold value obtained by a genetic algorithm, calculating input and output values of each hidden layer, and outputting the output values to a secondary network when the error of the output values meets an error condition.
Optionally, initializing parameters of the BP neural network further comprises: maximum training times, learning accuracy, initial weight, initial threshold, and initial learning rate.
Optionally, training the secondary network specifically includes: initializing an RBF network, adopting a K-means clustering algorithm to determine a center, calculating the total mean square error of the RBF network, judging whether the error meets the requirement, correcting the weight if the error does not meet the requirement, and ending training if the error does not meet the requirement.
Optionally, the specific process of obtaining the optimal weight and the threshold value by using the genetic algorithm is as follows: and encoding the initial weight and the initial threshold in the BP neural network, repeatedly executing the operations of population initialization, fitness evaluation, selection, crossover and mutation until the constraint condition is met, and decoding to obtain the optimal weight and the threshold.
Optionally, outputting the predicted value of the estuary residual water level further includes: and carrying out real-time visual display on the predicted value.
The invention provides a estuary surplus water level prediction method based on a cascade neural network, which comprises the following steps: acquiring key data influencing the change of the estuary water level, wherein the key data comprises first key data and second key data, and preprocessing the key data to obtain a data set; wherein the data set is divided into a training set, a verification set and a test set; establishing a two-stage serial residual water level neural network prediction model, wherein the prediction model comprises a primary network and a secondary network, the primary network is a network combining a BP neural network with a genetic algorithm, and the secondary network is an RBF neural network; training the primary network by taking the first key data in the training set as input variables, and taking output values obtained by training and the second key data in the training set as input variables of the secondary network; training the secondary network, and outputting a predicted value of the estuary residual water level when the error meets the requirement.
According to the cascade neural network-based estuary residual water level prediction method, a two-stage cascade neural network prediction model is established based on nonlinear characteristics of estuary residual water levels and multi-factor change rules affecting the formation change of the estuary residual water levels, a first-stage network adopts a GA-BP neural network, a second-stage network adopts an RBF neural network, and the output of the first-stage network, a runoff factor, a tide factor and a tide interaction factor are taken as the input of the second-stage network; the input variables influencing the change of the estuary residual water level are classified, the comprehensive effect of influencing the residual water level is comprehensively considered, the structures of all levels of neural networks in the prediction model are optimized, the prediction capability of the model can be optimized, the advantages of the GA-BP neural network and the RBF neural network are fully exerted, and the accuracy and the efficiency of model prediction are improved.
Drawings
FIG. 1 is a schematic flow chart of a cascade neural network-based estuary water level prediction method;
FIG. 2 is a schematic diagram of a model for predicting the water level of estuary water based on a cascade neural network;
FIG. 3 is a schematic diagram of a BP calculation flow of a first-stage GA-BP network based on a cascade neural network estuary water level prediction method of the present invention;
FIG. 4 is a schematic diagram of a GA algorithm flow of a first-stage GA-BP network of the estuary water level prediction method based on a cascade neural network;
FIG. 5 is a schematic flow chart of a cascade neural network-based two-stage RBF neural network for a estuary water level prediction method.
Detailed Description
The embodiment of the invention provides a cascade neural network-based estuary residual water level prediction method, which aims to solve the technical problems that the existing method has limitation and larger prediction result error when predicting the estuary residual water level.
In order that the invention may be readily understood, a more complete description of the invention will be rendered by reference to the appended drawings. Preferred embodiments of the present invention are shown in the drawings. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and/or" as used herein includes any and all combinations of one or more of the associated listed items.
The analytical solution is to calculate the residual water level gradient (namely the first derivative of the residual water level along the x-axis direction) based on the momentum conservation equation to quantitatively describe the along-way change rate and the characteristics of the residual water level, and in order to simplify the calculation, the residual water level gradient term is assumed to be mainly balanced with the nonlinear friction term in the equation, and the residual water level gradient analytical solution can be obtained by a one-dimensional momentum equation:
wherein t is time; u is the cross-sectional average flow velocity; z is the water level; h is the water depth; g is gravity acceleration; k is the inverse of the Manning coefficient; e is the diagonal pressure term caused by density variation.
The average of the above moisture sampling period is obtained:
wherein the left term is positive pressure term, oblique pressure term and convection term respectively. The third term in the left equation may be further rewritten as:
Introduction of dimensionless Froude number The above can be changed to:
in the method, in the process of the invention, And/>The friedel and flow values at the initial moment of integration are respectively.
Based on the above conversion, a residual water level gradient calculation formula considering the influence of the inclined pressure can be obtained:
The secondary flow rate term in the nonlinear friction term is subjected to linearization treatment by using Chebyshev polynomial, so that the method is obtained:
Where v 'is the maximum flow rate possible and ε 0,ε1,ε2 is the dimensionless flow rate amplitude based on v', namely:
v′＝|u0|+v1+vz；
expanding the power of the cos function according to a trigonometric function formula and extracting the simple harmonic frequencies omega and 2 omega to obtain the following steps:
U|U|＝v’2[F0+F1cos(ωt+ψ1)+F2cos(2ωt+ψ2)]；
Wherein:
Thus, the analytical expression along the path Yu Shuiwei gradient can be simplified as:
wherein, the residual water level caused by the nonlinear friction term can be decomposed into the following 3 factors:
tidal current factor:
Runoff factor:
diameter tide interaction factor:
Hydrodynamic numerical models such as SCHISM (Semi-immediate Cross-scale Hydroscience INTEGRATED SYSTEM Model) can be obtained by generalizing the landform shoreline, setting initial conditions corresponding to simulation stages including a temperature field, a salinity field and boundary conditions such as open sea harmony and moisture separation, upstream river basin river flow and the like, and then performing simulation analysis on residual water level changes; the harmonic analysis method is to calculate a harmonic constant based on a least square method, and study the influence of different moisture characteristics and the combined effect thereof on the residual water level; the linear regression method is to process data (such as daily average, monthly average, etc.) of different types of residual water levels (positive residual water level or water increasing data sequence, negative residual water level or water decreasing data sequence and residual water level data sequence) according to different time periods, then establish a regression model, and estimate the variation trend of each data sequence.
The embodiment of the invention provides a estuary residual water level prediction method based on a cascade neural network, which comprises the following steps:
S1: acquiring key data influencing the change of the estuary water level, wherein the key data comprises first key data and second key data, and preprocessing the key data to obtain a data set; wherein the data set is divided into a training set, a verification set and a test set;
s2: establishing a two-stage serial residual water level neural network prediction model, wherein the prediction model comprises a primary network and a secondary network, the primary network is a network combining a BP neural network with a genetic algorithm, and the secondary network is an RBF neural network;
S3: training the primary network by taking the first key data in the training set as input variables, and taking output values obtained by training and the second key data in the training set as input variables of the secondary network;
s4: training the secondary network, and outputting a predicted value of the estuary residual water level when the error meets the requirement.
Obtaining key data: obtaining or calculating through a formula to obtain key data influencing the residual water level change, wherein the key data specifically comprise river width, water depth, upstream river basin runoff, open sea tide wave water level, flow rate and the like of a research area; and calculating runoff factors, tide factors and tide interaction factors (the calculation formula is shown in the above). The river width and the water depth of the research area can be obtained through satellite remote sensing data or inversion after unmanned aerial vehicle image shooting, the outside sea tide wave water level can be obtained through a tide level station which is located nearest to the research area, and the upstream river basin runoff, the water level and the flow velocity can be obtained through a hydrological institution which is governed by the research area.
Unifying data space time: because the obtained various data come from different data sources, the spatial point positions and time steps among the variables are not unified, and the spatial consistency and time synchronism of data information are required to be ensured for establishing a prediction model, the various data are required to be unified in space and time. In space, interpolating variables influencing the change of the residual water level to corresponding residual water level analysis areas/points by adopting an inverse distance weighted average interpolation method; and (3) taking the time of the existing residual water level historical data as a reference, and interpolating variable data influencing the residual water level change to the time corresponding to the residual water level by adopting a linear interpolation method.
Dividing time periods: because the residual water level has the characteristic of complex space-time variation, in order to improve the accuracy and efficiency of prediction, the embodiment of the invention is divided into three typical periods, and residual water level prediction models are respectively established, so that the module divides the data obtained in the previous module into 6 stages according to the historical hydrologic characteristics of a research area: flood period big tide stage, flood period little tide stage, dead water period big tide stage, dead water period little tide stage, level water period big tide stage, level water period little tide stage. Taking the mouth of the pearl river as an example, the flood period is generally 5 months to 8 months, the dead water period is generally 11 months to 2 months in the coming year, and the flat water period is generally 3 months, 4 months, 9 months and 10 months. The size tide phase division can be judged based on the historical tide level process line of the research estuary.
Data standardization processing: because of the large difference among variable change intervals, in order to effectively improve the precision of the cascade neural network model adopted in the embodiment of the invention and ensure the equality of the contribution effect of each variable in the neural network model, the adoption of dispersion standardization is to carry out linear transformation on original data so that the result falls into the [0,1] interval, and the following conversion function is used for transforming the sequence x 1,x2,…,xn:
The fluctuation interval of each variable is unified between [0,1] through the dispersion normalization processing of each variable time sequence.
Data set classification: the data subjected to standardization and space-time unification are selected and grouped and reconstructed according to 6 phase types (flood period big tide phase, flood period small tide phase, withered period big tide phase, withered period small tide phase, flat period big tide phase and flat period small tide phase) by taking time data as reference, and then the obtained 6 phase data sets are further classified according to 6:2:2, dividing all data sets of the 6 stages into a training set, a verification set and a test set (6:2:2) by adopting a random selection method.
The training set is used for fitting the model, and the classifying model is trained by setting parameters of the classifier. When the verification set is combined subsequently, different values of the same parameter are selected, and a plurality of prediction models are fitted; the verification set function is that after a plurality of models are trained through a training set, in order to find out the model with the best effect, each model is used for predicting verification set data, and model accuracy is recorded. Selecting parameters corresponding to the model with the best effect, namely, adjusting the model parameters; and after the test set obtains an optimal model through the training set and the verification set, model prediction is carried out by using the test set. To measure the performance and predictive power of the optimal model. The test set can be regarded as a never existing data set, and after model parameters have been determined, the test set is used for model performance evaluation. Thus, the data set obtained by this module is 6 feature phases, each phase containing 3 classes of data sets.
When the estuary residual water level prediction model is established based on the cascade neural network, the residual water level prediction model of each characteristic stage is established by using the cascade neural network based on 6 characteristic stages and 3 types of data sets in each stage, namely, the residual water level prediction model finally obtained contains 6 corresponding stage models, when the estuary residual water level prediction model is used, the stage (month and tide level characteristic stage) where the residual water level to be predicted is definitely needed is located, and then the corresponding stage model is used for prediction. The model building process comprises parameter adjustment, comparison, optimization, preference and the like.
The model prediction precision is evaluated by adopting Root Mean Square Error (RMSE), efficiency coefficient (NSE), average absolute error (MAE) and determination coefficient (R 2), and the calculation method comprises the following steps:
Wherein X i、Yi is the measured value and the estimated value of the ith period, and X' i、Y′i is the average value of the measured value and the estimated value of the model of the ith period; n is the time period. And respectively selecting models corresponding to 6 stages when the RMSE, NSE, MAE values corresponding to the 6 stages are minimum and the square of R is maximum as optimal models through comparison under different parameter conditions.
Visualization of the prediction results: in order to display the residual water level prediction results obtained based on the cascade neural network in real time, together with GIS and other software, the residual water level prediction values in different prediction window periods (such as 6 hours and 24 hours in the future) are displayed on line in real time in the river mouth and river basin map, so that response decision makers or using institutions can intuitively judge the residual water level prediction results conveniently, and the public can know the future residual water level change trend conveniently.
The prediction model of the single neural network is either a prediction model with insufficient number of hidden layer neurons or a prediction model with insufficient generalization level and no ideal prediction result can be obtained because the functions in the algorithm are complex and the samples have limited expressive power, so that the prediction model of the residual water level in the invention is based on a cascade neural network. From the aspect of information utilization, a single neural network prediction model can only describe a data sequence rule from one side, namely only can utilize effective information of a data part, and has certain limitation; different prediction methods can provide different useful information, the advantages of a single model are complemented by a combined model, data information can be mined to a greater extent, and a better prediction effect is obtained. Each level of subsystems of the multi-level neural network structure is generally composed of a feature extraction module and a neural network module. The feature extraction module extracts feature vectors from input data, and based on the feature vectors, the neural network module learns all training data in a learning stage, and classifies and judges the data input to the stage in a working stage, wherein the classifying and judging comprises two conditions of recognition output and refusal output. It is not a trivial assumption: the features extracted by the feature extraction module of one stage can effectively represent the essential features of a part of input samples, so that the part of samples are successfully identified in the stage, the rest of samples are rejected due to unclear features, and the samples are sent to the next stage for classification and discrimination based on another feature extraction and neural network structure. A good multi-stage neural network structure should satisfy two conditions:
(1) The feature extraction module of each stage is effectively designed, so that the feature extraction methods of each stage are mutually independent and mutually complementary as far as possible;
(2) The neural network modules at all levels are reasonably designed, so that the neural network can effectively reject samples which are obscured under the characteristics of the levels, and the false recognition rate is reduced as much as possible. The overall performance of the multi-stage neural network structure is greatly dependent on the level of the false recognition rate of each stage subsystem, and the false recognition rate of the first stages is particularly critical.
The embodiment of the invention adopts a cascade neural network with two stages connected in series, wherein the first stage adopts a GA-BP network, and the second stage adopts an RBF network. Because the BP neural network algorithm is calculated based on a gradient descent method, when the complex nonlinear function problem is processed, the BP algorithm can generate a local minimum solution instead of a global minimum solution, so that the embodiment of the invention adopts the BP neural network and Genetic Algorithm (GA) combined network in the first-stage network of the cascade neural network, the GA algorithm simulates survival rules of survival and superior and inferior survival of biological fittest in nature, and superior genes are transmitted backwards through hybridization, mutation and other reproduction operations. In the actual algorithm calculation, the GA algorithm is utilized to find the optimal values of the threshold value and the weight value, then the optimal values are assigned to the BP neural network, the received signal strength is used as an input value, and the distance value is used as an output value for training.
Referring to fig. 5, the second level RBF network has 3 parameters to learn, and the steps are as follows:
(1) Learning center t i (i=1, 2, …, I is the number of hidden units). The self-organizing learning process uses a clustering algorithm, and a K-means clustering algorithm is commonly used.
(2) The variance σ i (i=1, 2, …, I) is determined. When the RBF network uses a gaussian function, namely:
(i＝1,2,…,I)
the variance can be calculated using the following equation:
Where X k is the input in the RBF network and d max is the maximum distance between the centers taken.
(3) Learning weights W ij (i=1, 2, …, I; j=1, 2, …, J).
The weights may be learned using a Least Mean Square (LMS) algorithm. The actual output of the RBF network is Y (n) =g (n) W (n),
Y(n)＝[ykj(n)]，(k＝1,2,…,N；j＝1,2,…,J)
Y kj is the output value of the kth neuron of the RBF network output layer; n is the number of neurons of the RBF network output layer; j is the number of neurons of the hidden layer of the RBF network; j is the j-th neuron of the hidden layer of the RBF neural network; n is the nth neuron of the output layer of the RBF neural network.
Referring to fig. 2-5, based on the law of change of the residual water level, the embodiment of the invention divides the residual water level period into 6 stages, establishes a two-stage cascade neural network prediction model in series, and takes a data set of a large tide period in a flood period as an example, a specific implementation flow of the prediction model establishment and training comprises the following steps:
(1) First-order network BP: taking the outside sea tide level, water depth, river width, flow rate, river basin runoff and water level in the training set of the large tide period of the historical flood period of the research area as input variables of the GA-BP neural network model;
(2) First-order network BP: initializing BP neural network parameters, including maximum training times, learning accuracy, initial weight, initial threshold value, initial learning rate and the like;
(3) First-order network GA: coding the weight and the threshold initialized in the previous step according to the GA algorithm requirement;
(4) First-order network GA: and (3) encoding the initial total group by adopting a real number encoding method, wherein if the number of nodes at an input layer is t, the number of nodes at an output layer is m, the number of nodes at an hidden layer is h, and the encoding length S is as follows:
S＝t*h+h*m+m+h；
(5) First-order network GA: calculating fitness value, wherein the final result is influenced by the selection of fitness function, the individual selects according to the fitness value, the individual with high fitness is reserved, the derivative of error function E in BP neural network is used as fitness function, and the calculation formula is as follows:
Wherein T j is a true value, and O j is a predicted value.
(6) First-order network GA: selecting by adopting a roulette method, wherein the size of a sector area in each chromosome corresponding to the disc is in direct proportion to the fitness value of the chromosome corresponding to the sector area, and the probability P i of the i-th individual being selected is as follows:
(7) First-order network GA: through the above selection operation, a function with high fitness is selected, and then the crossing operation is performed between individuals with high fitness by using a real crossing method, thereby producing excellent individuals. If the p-th chromosome is Ap and the q-th chromosome is Aq, the chromosomes Ap and Aq are crossed at the position j, and the calculation formula is as follows:
Apj＝Apj*(1-γ)+Aqj*γ，
Aqj＝Aqj*(1-γ)+Apj*γ；
Wherein A qj、Apj is the crossover value of chromosome A q and A p at position j, and the value range of gamma is (0, 1).
(8) First-order network GA: then, the jth gene A pj of the p chromosome is selected for mutation operation, and the calculation formula is as follows:
Wherein A min is the lower bound of the gene; a max is an upper bound; the value range of r 1 is (0, 1); r 2 is 1 random number; g max is an iteration coefficient; g is the current iteration number.
(9) First-order network GA: repeating the operations of (3) - (8) until the constraint condition is judged to be met, otherwise, continuing training;
the constraints here are: when the difference between the output value and the true value is smaller than a predetermined value, the value of the predetermined value is usually set smaller, for example, 0.01 m, in order to improve the prediction accuracy of the model.
(10) First-order network GA: decoding the target individual based on the meeting of the judging result, and giving a weight and a threshold value to the BP neural network;
(11) First-order network BP: training the BP neural network by utilizing the optimal weight and the threshold value obtained by the GA, and calculating the input and output values of each layer;
(12) First-order network BP: calculating an output layer error;
(13) First-order network BP: judging whether an error condition is met, if so, outputting a result, otherwise, continuing training until the error condition is met;
(14) Secondary network RBF: taking a predicted value and second key data (specifically, a runoff factor, a tide factor and a tide interaction factor) output by the GA-BP in the primary network as input variables of a secondary network RBF neural network model;
And taking the output of the primary network as the input of the secondary network, and taking the output of the primary network, the runoff factor, the tide factor and the tide interaction factor as the input of the secondary network together. The influence of two types of key data is considered separately, the prediction capability of the model can be optimized, and the respective advantages of the GA-BP neural network and the RBF neural network are fully exerted.
(15) Secondary network RBF: initializing an RBF network;
(16) Secondary network RBF: determining a center by adopting a K-means clustering method;
(17) Secondary network RBF: calculating the total mean square error between the output value and the actual residual water level value in the RBF network by using the following steps:
Where X i is a single value in the sequence of values, Is the average of the set of values, and N is the number of overall values.
(18) Secondary network RBF: the weight in RBF network is corrected, and the least mean square algorithm is adopted, wherein the algorithm comprises the following steps:
s1: setting variables and parameters;
X (n) is an input vector (or called training sample), W (n) is a weight vector, b (n) is a deviation, d (n) is a desired output, y (n) is an actual output, eta is a learning rate, and n is the number of iterations;
s2: initializing, namely giving w (0) a smaller random non-zero value respectively, and enabling n=0;
S3: for a set of input samples x (n) and corresponding expected outputs d, calculate:
e(n)＝d(n)-XT(n)W(n)；
W(n+1)＝W(n)+ηX(n)e(n)；
S4: judging whether the condition is met, and ending the algorithm if the condition is met; if not, increasing n by 1, and proceeding to S3 for continuous execution.
(19) Secondary network RBF: judging whether the error meets the requirement, if so, stopping calculation of the prediction model, outputting a result, otherwise, continuing training until the error condition is met;
Y(n)＝G(n)*W(n)；
Wherein Y (N) = [ Y kj (N) ], (k=1, 2, …, N; j=1, 2, …, J);
Wherein y kj is the output value of the kth neuron of the RBF neural network output layer; n is the number of neurons of the RBF neural network output layer; j is the number of neurons in the hidden layer of the RBF neural network; j is the j-th neuron of the hidden layer of the RBF neural network; n is the nth neuron of the output layer of the RBF neural network.
(20) Secondary network RBF: and visually displaying the predicted residual water level value.
According to the same method, a residual water level cascade neural network prediction model can be established for other 5 stages.
According to the estuary residual water level prediction method based on the cascade neural network, residual water levels are respectively predicted according to the residual water level change characteristics in 6 characteristic stages; based on the nonlinear characteristics of the estuary residual water level and the multi-factor change rule affecting the formation change of the estuary residual water level, classifying and considering the factor characteristics affecting the residual water level change, establishing a two-stage cascade neural network prediction model of the residual water level, wherein the first-stage network adopts a GA-BP neural network, and the second-stage network adopts an RBF neural network.
According to the embodiment of the invention, the input variables are classified, the first key data are used as the input variables of the first-level network, the output value of the first-level network, the runoff factor, the tide factor and the tide interaction factor are used as the input of the second-level network, the comprehensive effect of influencing factors on the residual water level is comprehensively considered, the structures of all levels of neural networks in the prediction model are optimized, the prediction capability of the model can be optimized, the respective advantages of the GA-BP neural network and the RBF neural network are fully exerted, and the accuracy and the efficiency of model prediction are improved.
It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.
In the embodiments provided in the present application, it should be understood that the disclosed system, apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.
The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.
The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A cascade neural network-based estuary residual water level prediction method is characterized by comprising the following steps of:
Acquiring key data influencing the change of the estuary residual water level, and preprocessing the key data to obtain a data set; wherein the data set is divided into a training set, a verification set and a test set; the key data comprise first key data and second key data, the first key data comprise outside sea tide level, water depth, river width, flow speed, river basin runoff and water level, and the second key data comprise runoff factors, tide factors and tide action factors;
Establishing a two-stage serial residual water level neural network prediction model, wherein the prediction model comprises a primary network and a secondary network, the primary network is a network combining a BP neural network with a genetic algorithm, and the secondary network is an RBF neural network;
the establishing of the two-stage serial residual water level neural network prediction model further comprises the following steps: respectively establishing a two-stage serial residual water level neural network prediction model for a plurality of characteristic stages, wherein the characteristic stages are divided according to residual water level characteristics during pretreatment;
Training the primary network by taking the first key data in the training set as input variables, and taking output values obtained by training and the second key data in the training set as input variables of the secondary network;
Training the secondary network, and outputting a predicted value of the estuary residual water level when the error meets the requirement.
2. The cascade neural network-based estuary water level prediction method according to claim 1, wherein preprocessing the key data further comprises: and carrying out space-time unification, time interval division and standardization processing on the key data.
3. The cascade neural network-based estuary water level prediction method according to claim 1, wherein the characteristic stage is specifically: flood period big tide stage, flood period little tide stage, dead water period big tide stage, dead water period little tide stage, level water period big tide stage and level water period little tide stage.
4. The cascade neural network-based estuary water level prediction method according to claim 1, wherein training the primary network is specifically: initializing parameters of the BP neural network, training the BP neural network by utilizing an optimal weight and a threshold value obtained by a genetic algorithm, calculating input and output values of each hidden layer, and outputting the output values to a secondary network when the error of the output values meets an error condition.
5. The cascade neural network-based estuary water level prediction method according to claim 4, wherein initializing parameters of the BP neural network further comprises: maximum training times, learning accuracy, initial weight, initial threshold, and initial learning rate.
6. The cascade neural network-based estuary water level prediction method according to claim 1, wherein training the secondary network is specifically: initializing an RBF network, adopting a K-means clustering algorithm to determine a center, calculating the total mean square error of the RBF network, judging whether the error meets the requirement, correcting the weight if the error does not meet the requirement, and ending training if the error does not meet the requirement.
7. The cascade neural network-based estuary water level prediction method according to claim 1, wherein the specific process of obtaining the optimal weight and the threshold value by using the genetic algorithm is as follows: and encoding the initial weight and the initial threshold in the BP neural network, repeatedly executing the operations of population initialization, fitness evaluation, selection, crossover and mutation until the constraint condition is met, and decoding to obtain the optimal weight and the threshold.
8. The method for predicting estuary water level based on cascade neural network according to claim 1, wherein outputting the predicted value of estuary water level further comprises: and carrying out real-time visual display on the predicted value.
CN202010572944.1A 2020-06-22 2020-06-22 Estuary surplus water level prediction method based on cascade neural network Active CN111723523B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010572944.1A CN111723523B (en) 2020-06-22 2020-06-22 Estuary surplus water level prediction method based on cascade neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010572944.1A CN111723523B (en) 2020-06-22 2020-06-22 Estuary surplus water level prediction method based on cascade neural network

Publications (2)

Publication Number Publication Date
CN111723523A CN111723523A (en) 2020-09-29
CN111723523B true CN111723523B (en) 2024-04-30

Family

ID=72569917

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010572944.1A Active CN111723523B (en) 2020-06-22 2020-06-22 Estuary surplus water level prediction method based on cascade neural network

Country Status (1)

Country Link
CN (1) CN111723523B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113158556B (en) * 2021-03-31 2023-08-08 山东电力工程咨询院有限公司 Short-time high-precision forecasting method for regional water level
CN113141272B (en) * 2021-04-20 2022-03-15 浙江大学 Network security situation analysis method based on iteration optimization RBF neural network
CN114489182B (en) * 2022-02-11 2023-05-23 无锡市肯恩科技有限公司 PID temperature control method with double temperature detection function

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107883931A (en) * 2017-10-25 2018-04-06 深圳大学 Tide rectification method and device, storage medium based on structure instantaneous water level model
JP2019095240A (en) * 2017-11-20 2019-06-20 日本無線株式会社 Water level prediction method, water level prediction program, and water level prediction device
CN110443418A (en) * 2019-07-31 2019-11-12 西安科技大学 Urban water consumption prediction technique based on GA-BP neural network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107883931A (en) * 2017-10-25 2018-04-06 深圳大学 Tide rectification method and device, storage medium based on structure instantaneous water level model
JP2019095240A (en) * 2017-11-20 2019-06-20 日本無線株式会社 Water level prediction method, water level prediction program, and water level prediction device
CN110443418A (en) * 2019-07-31 2019-11-12 西安科技大学 Urban water consumption prediction technique based on GA-BP neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于级联神经网络的年降雨量预测;任刚红 等;土木建筑与环境工程;第38卷(第S2期);基于级联神经网络的年降雨量预测 *
基于集合经验模态分解与BP组合模型的短期余水位预测;屠泽杰 等;海洋通报;第39卷(第1期);第78-85页 *

Also Published As

Publication number Publication date
CN111723523A (en) 2020-09-29

Similar Documents

Publication Publication Date Title
Thomas et al. Likelihood-free inference by ratio estimation
CN109142171B (en) Urban PM10 concentration prediction method based on feature expansion and fusing with neural network
Qin et al. A dual-stage attention-based recurrent neural network for time series prediction
CN109063911B (en) Load aggregation grouping prediction method based on gated cycle unit network
CN111723523B (en) Estuary surplus water level prediction method based on cascade neural network
Lu et al. Generalized radial basis function neural network based on an improved dynamic particle swarm optimization and AdaBoost algorithm
CN108596327B (en) Seismic velocity spectrum artificial intelligence picking method based on deep learning
CN110059852A (en) A kind of stock yield prediction technique based on improvement random forests algorithm
Joseph et al. Rainfall prediction using data mining techniques
CN109143408B (en) Dynamic region combined short-time rainfall forecasting method based on MLP
CN109840587A (en) Reservoir reservoir inflow prediction technique based on deep learning
Jarábek et al. Energy load forecast using S2S deep neural networks with k-Shape clustering
Suryo et al. Improved time series prediction using LSTM neural network for smart agriculture application
Li et al. Leveraging LSTM for rapid intensifications prediction of tropical cyclones
Le et al. Application of gated recurrent unit (GRU) network for forecasting river water levels affected by tides
Kosana et al. Hybrid wind speed prediction framework using data pre-processing strategy based autoencoder network
Javaid et al. Employing a machine learning boosting classifiers based stacking ensemble model for detecting non technical losses in smart grids
CN113268929B (en) Short-term load interval prediction method and device
Han et al. A Hybrid BPNN-GARF-SVR PredictionModel Based on EEMD for Ship Motion.
Gumilar et al. Performance analysis of hybrid machine learning methods on imbalanced data (rainfall classification)
Giang et al. Monthly precipitation prediction using neural network algorithms in the Thua Thien Hue Province
Alam Recurrent neural networks in electricity load forecasting
Zeyad et al. A Comparative Analysis of Data Mining Methods for Weather Prediction
Khan et al. Forecasting renewable energy for environmental resilience through computational intelligence
CN115952906A (en) Short-term photovoltaic power prediction method, system, equipment and medium based on LSGAN-GRU

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
GR01 Patent grant
GR01 Patent grant