CN110069801B - Pick distance estimation method based on artificial neural network - Google Patents
Pick distance estimation method based on artificial neural network Download PDFInfo
- Publication number
- CN110069801B CN110069801B CN201811414532.4A CN201811414532A CN110069801B CN 110069801 B CN110069801 B CN 110069801B CN 201811414532 A CN201811414532 A CN 201811414532A CN 110069801 B CN110069801 B CN 110069801B
- Authority
- CN
- China
- Prior art keywords
- data
- neural network
- input
- output
- samples
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
- Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping
Abstract
The invention discloses a pick distance estimation method based on an artificial neural network, which comprises the following steps: (1) getH、θ、qThe input quantity and the nappe choosing distance are output quantities, an integrated neural network system with proper input and output is constructed, and a forward neural network with an input layer, a hidden layer and an output layer structure is selected to realize the modeling of the nappe choosing distance; (2) collecting data and preprocessing a sample; (3) on the basis of the neural network model obtained in the step (1) and the learning sample obtained in the step (2), respectively sampling BP algorithm training neural networks for the omics learning samples to obtain the optimal model parameters and the weights of the individual networks, and integrating the individual networks; (4) measuredH、θ、qAnd (4) inputting the input quantity into the integrated neural network trained in the step (3) to estimate the magnitude of the pick distance. The method estimates the magnitude of the picking distance based on the artificial neural network, has the characteristics of simple and effective realization, and effectively improves the estimation precision.
Description
Technical Field
The invention relates to a pick distance estimation method based on an artificial neural network, and belongs to the field of hydraulic engineering.
Background
The trajectory jet energy dissipation has wide application in water conservancy and hydropower engineering due to simple structure, high energy dissipation rate and convenient construction. The position of the downstream pit of the trajectory energy dissipation is closely related to the water tongue elevation distance thereof, and the elevation distance is generally required to be more than 3-4 times of the maximum pit depth in engineering design and is an important hydraulic parameter for the operation safety of a water release building; the method for estimating the offset mainly comprises theoretical derivation, an empirical formula (nonlinear regression analysis is carried out on data of a model test), numerical simulation and the like. Although researchers have carried out a lot of research work on the estimation of the nappe picking distance from a plurality of angles, various formulas are also proposed, but the formulas carry out similar idealization, averaging, approximation and the like on prototype conditions under wide variation when the formulas are used for estimation, so that the result and the prototype have large deviation. Therefore, a more accurate method for estimating the water tongue picking distance is needed.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides the pick distance estimation method based on the artificial neural network, the size of the pick distance is estimated based on the artificial neural network, the method has the characteristics of simple and effective realization, and the estimation precision is effectively improved.
The technical scheme is as follows: in order to solve the technical problem, the distance picking estimation method based on the artificial neural network comprises the following steps:
(1) taking H, theta and q input quantities, taking the water tongue picking distance as an output quantity, constructing an integrated neural network system with proper input and output, and selecting a forward neural network with an input layer, a hidden layer and an output layer structure to realize the modeling of the water tongue picking distance;
(2) collecting data and preprocessing a sample;
(3) on the basis of the neural network model obtained in the step (1) and the learning sample obtained in the step (2), respectively sampling BP algorithm training neural networks for the omics learning samples to obtain the optimal model parameters and the weights of the individual networks, and integrating the individual networks;
(4) and (4) inputting the measured H, theta and q serving as input quantities into the integrated neural network trained in the step (3), and then estimating the size of the pick distance.
Preferably, the step (2) pretreatment comprises the following steps:
a. sample data correlation analysis: carrying out correlation analysis on the sample data, and combining parameter data with larger correlation;
b. pickRemoving redundant data: analyzing all collected data by statistical method, analyzing whether it accords with normal distribution, and rejecting
Redundancy data outside the range to improve the generalization performance of the neural network model, wherein: x is the mean and s is the standard deviation.
Preferably, the input amount in the step (1) is determined by the following steps:
a. establishing a combination type of input parameters: assuming that the number of the input parameters preliminarily determined in the step (2) is a, establishing parameter combinations consisting of a parameters and a parameter combinations consisting of any (a-1) parameters;
b. carrying out normalization processing on the sample according to the step (2); training the different input parameter combination types and the output data by adopting a normalization function premmx, performing inverse normalization on the trained matrix, and comparing the matrix with the output matrix, wherein the output matrix after the inverse normalization is a training result, and the required time is training time;
c. analyzing the sensitivity; carrying out sensitivity analysis on the output data and the data in the sample to obtain correlation R values reflecting training results under different combination types;
d. and comprehensively selecting input parameters. And comprehensively selecting a combination type of the input parameters according to the training time and the result of the sensitivity analysis.
Preferably, the number of samples in the step (2) is determined by:
a. the number of the samples is closely related to the neural network training result, and the specific relation between the number of the samples and the topological structure is calculated according to the following formula:
in the formula: n is the number of input variables; m is the number of output variables; h is the number of hidden layer nodes; p is the number of learning samples that need to be input.
b. And supplementing sample data: according to the requirement on the number of samples, checking whether the number of samples meets the requirement after the redundant data is removed, and continuing the next step if the number of samples meets the requirement; otherwise, continuing to collect and remove the data until the number of samples meets the requirement.
In the invention, the combination type of input parameters is comprehensively selected according to the training time and the result of sensitivity analysis, and H, theta and R are input0Q, the training time is longer when the four parameters are used; and when three input parameters are used, the time is shortened by several times compared with the training time of four parameters, and the difference between the obtained prediction result and the sample data correlation degree is smaller, so that three input data with the highest correlation degree between the prediction data and the sample data in the three input data are selected, namely H, theta and q.
Has the advantages that: according to the choosing distance estimation method based on the artificial neural network, the low error and the high correlation effectively indicate that the artificial neural network method is obviously superior to a regression analysis method in forecasting the choosing distance of the water tongue of the drift; in addition, compared with the traditional theoretical formula, the prediction result of the artificial neural network is more reasonable, and for the sample conditions of the embodiment of the invention, the correlation degree of the prediction result of the BPGL model is about 30% higher than that of the traditional theoretical formula.
Drawings
FIG. 1 is a trajectory diagram of trajectory of jet flow nappe motion of trajectory dissipater.
FIG. 2a is a positive-Taiwan distribution curve of frequency number and single-width flow q of incoming flow; FIG. 2b is a positive Tailored distribution curve of upstream head versus frequency; FIG. 2c is a positive power distribution curve of flip bucket angle and frequency; fig. 2d shows a positive power distribution of the correlation R value with frequency.
FIG. 3a is a graph of sensitivity analysis of estimated and measured values; FIG. 3b is another sensitivity analysis of estimates versus measurements.
Fig. 4 is a three-layer neural network topology structure diagram.
FIG. 5 is a graph of neural network training results.
Fig. 6 is an error analysis chart.
FIG. 7 is a diagram of the validation results of an example of the engineering.
Detailed Description
The present invention will be further described with reference to the accompanying drawings.
The invention discloses a pick distance estimation method based on an artificial neural network, which comprises the following steps of:
step 1: taking H, theta and q input quantities, taking the water tongue picking distance as an output quantity, constructing an integrated neural network system with proper input and output, and selecting a forward neural network with an input layer, a hidden layer and an output layer structure to realize the modeling of the water tongue picking distance;
and step 11, establishing a trajectory diagram (shown in figure 1) of trajectory of the trajectory diagram of the trajectory energy dissipater nappe, summarizing and summarizing the current methods for estimating the nappe lifting distance and related formulas related to the methods, wherein the methods comprise theoretical analysis, model test and numerical simulation.
And step 12, preliminarily determining influence factors influencing the water tongue picking distance (the picking distance of the center point of the circular arc-shaped flip bucket). According to a theoretical formula of the choosing distance (formula 2) given by the water conservancy industry standard and a general most common empirical formula of the choosing distance (formula 3) of the jet flow nappe at the deepest position of the downstream pit, theoretical analysis is carried out through the pi theorem, and influence factors influencing the choosing distance (central point) of the nappe and a functional relation between the influence factors are preliminarily determined (formula 4).
xc＝f(q，H，h0，ht，ρ，μ，g，R0，θ，st，a，B) (4)
And step 13, optimizing the functional relation. Aiming at the specific arrangement of the engineering trajectory energy dissipater, the channel width, the ridge height, the height from the bottom of the energy dissipater outlet to the surface of the channel and the like are fixed, the viscosity and the downstream water depth are neglected, and the functional relation influencing the water tongue trajectory distance (central point) is simplified according to the principle, so that a formula (5) can be obtained.
xc＝f(q，H，h0，g，ρ，R0，θ)
In the embodiment, g and rho are respectively gravity acceleration and water density, are regarded as constants in the actual calculation process and are not considered in the actual training process; meanwhile, the depth of the incoming flow is directly related to the flow, variables can be combined, and the number of input parameters can be 4.
14. And establishing a combination type of the input parameters. The number of the input parameters determined in the step 13 is 4, and a parameter combination consisting of 4 parameters and four parameter combinations consisting of any 3 parameters are established, and the total number of the parameter combinations is 5, wherein the specific combination forms are as follows: (H, theta, q, R0)、(θ，q，R0)、(H，q，R0)、(H，θ，R0)、(H，θ，q)。
b. And carrying out normalization processing on the sample. And training the different input parameter combination types and the output data by adopting a normalization function premmx, performing inverse normalization on the trained matrix, comparing the matrix with the output matrix, wherein the output matrix after inverse normalization is a training result, and the required time is training time.
15. And (4) sensitivity analysis. Sensitivity analysis is carried out on the output data and the data in the sample, and H, theta and R are input0Q four parameters, the distance x is chosencThe training result has the highest correlation degree, and the R value is 0.9853; when any three parameters are selected, the correlation of the three input parameters H, θ and q is optimal, and the R value is 0.9757 (fig. 3a and 3 b).
16. And comprehensively selecting input parameters. Comprehensively selecting the combination type of input parameters according to the results of the training time and the sensitivity analysis, and inputting H, theta and R0Q, the training time is longer when the four parameters are used; and when three input parameters are used, the time is shortened by several times compared with the training time of four parameters, and the difference between the obtained prediction result and the sample data correlation degree is smaller, so that three input data with the highest correlation degree between the prediction data and the sample data in the three input data are selected, namely H, theta and q.
Step 2: and collecting data and preprocessing the sample data. Before neural network training, the preprocessing of sample data mainly comprises three processes:
a. and removing redundant data. Analyzing all collected data by statistical method from more than 500 groups of results of monomer model test of related engineering at home and abroad, analyzing whether the data accords with normal distribution, and rejecting
Redundant data outside the range is used for improving the induction performance of the neural network model, and finally an effective data 304 group is obtained, wherein the normal distribution of the input parameters is shown in figures 2 a-2 d.
b. According to
According to the situation of acquiring the collected data, the number of samples required for each training is finally determined to be 40, and the valid data 304 group processed in the step 4(b) meets the requirement of the number of samples.
And step 3: and establishing a neural network model related to the picking distance.
a. The artificial neural network model selected in this embodiment is a multilayer feedforward network model for error back propagation training and three improved BP models thereof to predict the picking distance, including: BP momentum model (BPGD), conjugate gradient model (BPGG) and Levenberg-Marquardt model (BPGL).
b. Neural network topologies. And estimating the pick distance by using the selected artificial neural network model, wherein the artificial neural network forms different networks according to different connection modes, and the different networks comprise three-layer topological structures of an input layer, a hidden layer and an output layer. Selection [3331]Modes, including three-layer topologies; three input parameters, q, H, theta; three hidden layers; three unit numbers and 1 output parameter, namely the picking distance xcAs in fig. 4. Therefore, the water tongue distance raising result can be obtained.
c. And (5) training results. For the model test with the scale range of 1: 50-1: 100, based on the regression analysis result that the prediction accuracy of the artificial neural network on the distance is far higher than that of the artificial neural network, in the four BP models selected by the case, the correlation degree of BPGD is the highest, the R value is as high as 0.9862, meanwhile, the correlation degree of a common BP algorithm (BPM) is the lowest, the R value is 0.9757, but the correlation degrees among BPGL, BPGD and BPGG are not very different. As shown in particular in fig. 5.
Parameters and weights thereof can be obtained by training 304 groups of samples by adopting a BPGD neural network model.
d. And (5) error analysis. The prediction correlation degrees of the BPM and the improved models thereof are all larger than the regression analysis result, wherein the BPGD with the highest correlation degree has a correlation coefficient higher than about 9.53 percent of the regression analysis result, and the average error and the root mean square error of the BPGD are also the smallest in the five models, which are respectively as follows: 0.515 and 0.094, as shown in particular in fig. 6.
And 4, comparing with the relevant concrete engineering example, and further verifying the accuracy of the calculation result.
In this embodiment, 16 typical engineering examples are selected, and the results of regression analysis, BPM, BPGL and measured values are compared with the calculation results of the conventional formulas (1) and (2), and it is found that the results of BPM and BPGL are both significantly better than the calculation results of the theoretical formulas, wherein the correlation of the calculation results of BPGL is the highest, and the correlation of the calculation results of BPGL is about 34.5% and 33.2% higher than that of the actually measured BPGL16 group examples. (see fig. 7)
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.
Claims (1)
1. A pick distance estimation method based on an artificial neural network is characterized by comprising the following steps:
(1) taking an upstream water head H, a flip bucket lifting angle theta and an incoming flow single-width flow q as input quantities, taking a water tongue lifting distance as an output quantity, constructing an integrated neural network system with proper input and output, and selecting a forward neural network with an input layer, a hidden layer and an output layer structure to realize modeling of the water tongue lifting distance;
(2) collecting data and preprocessing a sample;
(3) on the basis of the neural network model obtained in the step (1) and the learning sample obtained in the step (2), respectively sampling BP algorithm training neural networks for the omics learning samples to obtain the optimal model parameters and the weights of the individual networks, and integrating the individual networks;
(4) inputting the measured H, theta and q serving as input quantities into the integrated neural network trained in the step (3), and estimating the magnitude of the picking distance;
the pretreatment in the step (2) is as follows: sample data correlation analysis: carrying out correlation analysis on the sample data, and combining parameter data with larger correlation; analyzing all collected data by using a statistical method, analyzing whether the collected data accords with normal distribution or not, deleting partial deviation data if the collected data does not accord with the normal distribution, expanding new data to see whether the collected data meets the normal distribution or not, and rejecting the data
Redundant data outside of the range, wherein:
is the sample mean, s is the standard deviation;
the input quantity in the step (1) is determined by the following steps:
a. establishing a combination type of input parameters: assuming that the number of the input parameters preliminarily determined in the step (2) is a, establishing parameter combinations consisting of a parameters and a parameter combinations consisting of any (a-1) parameters;
b. carrying out normalization processing on the sample according to the step (2); training the different input parameter combination types and the output data by adopting a normalization function premmx, performing inverse normalization on the trained matrix, and comparing the matrix with the output matrix, wherein the output matrix after the inverse normalization is a training result, and the required time is training time;
c. analyzing the sensitivity; carrying out sensitivity analysis on the output data and the data in the sample to obtain correlation R values reflecting training results under different combination types;
d. comprehensively selecting input parameters, namely comprehensively selecting a combination type of the input parameters according to the training time and the result of sensitivity analysis;
the number of samples in the step (2) is determined by the following steps:
a. the number of the samples is closely related to the neural network training result, and the specific relation between the number of the samples and the topological structure is calculated according to the following formula:
in the formula, n is the number of input variables; m is the number of output variables; h is the number of hidden layer nodes; p is the number of learning samples needing to be input;
b. and supplementing sample data: according to the requirement on the number of samples, checking whether the number of samples meets the requirement after the redundant data is removed, and continuing the next step if the number of samples meets the requirement; otherwise, continuing to collect and remove the data until the number of samples meets the requirement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811414532.4A CN110069801B (en) | 2018-11-26 | 2018-11-26 | Pick distance estimation method based on artificial neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811414532.4A CN110069801B (en) | 2018-11-26 | 2018-11-26 | Pick distance estimation method based on artificial neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110069801A CN110069801A (en) | 2019-07-30 |
| CN110069801B true CN110069801B (en) | 2022-05-24 |
Family
ID=67365851
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811414532.4A Active CN110069801B (en) | 2018-11-26 | 2018-11-26 | Pick distance estimation method based on artificial neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110069801B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104404926A (en) * | 2014-10-08 | 2015-03-11 | 四川大学 | Overflow dam with dam face cantilever sills for current diversion and energy dissipation |
| WO2015172560A1 (en) * | 2014-05-16 | 2015-11-19 | 华南理工大学 | Central air conditioner cooling load prediction method based on bp neural network |
| CN107727355A (en) * | 2017-09-27 | 2018-02-23 | 中国科学院、水利部成都山地灾害与环境研究所 | A kind of fluid chooses horizontal range measuring method and its application |
| CN108535434A (en) * | 2018-04-09 | 2018-09-14 | 重庆交通大学 | Method based on Neural Network model predictive building site surrounding body turbidity |
| CN108755618A (en) * | 2018-06-25 | 2018-11-06 | 广东省水利电力勘测设计研究院 | A kind of medium and small reservoirs Spillway stream dissipation and scouring method |
-
2018
- 2018-11-26 CN CN201811414532.4A patent/CN110069801B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015172560A1 (en) * | 2014-05-16 | 2015-11-19 | 华南理工大学 | Central air conditioner cooling load prediction method based on bp neural network |
| CN104404926A (en) * | 2014-10-08 | 2015-03-11 | 四川大学 | Overflow dam with dam face cantilever sills for current diversion and energy dissipation |
| CN107727355A (en) * | 2017-09-27 | 2018-02-23 | 中国科学院、水利部成都山地灾害与环境研究所 | A kind of fluid chooses horizontal range measuring method and its application |
| CN108535434A (en) * | 2018-04-09 | 2018-09-14 | 重庆交通大学 | Method based on Neural Network model predictive building site surrounding body turbidity |
| CN108755618A (en) * | 2018-06-25 | 2018-11-06 | 广东省水利电力勘测设计研究院 | A kind of medium and small reservoirs Spillway stream dissipation and scouring method |
Non-Patent Citations (5)
| Title |
|---|
| Ski-jump trajectory based on take-off velocity;Jian-hua Wu 等;《Journal of Hydrodynamics》;20160201;166-169 * |
| 侧壁齿坎窄缝消能工的流态及消能;姚莉 等;《第二十七届全国水动力学研讨会》;20151106;172-177 * |
| 挑流水舌特性及其影响;宁景昊等;《黑龙江大学工程学报》;20171225(第04期);5-10 * |
| 挑流泄洪雾化的人工神经网络模型初探;彭新民等;《中国农村水利水电》;20060125(第01期);63-64 * |
| 泄洪雾化预测的人工神经网络方法初步研究;柳海涛 等;《中国水利学会2005学术年会》;20051001;206-213 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110069801A (en) | 2019-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103761423B (en) | A kind of hot rolled plate microstructure and property prediction method based on PSO-ELM | |
| CN102032935B (en) | Soft measurement method for sewage pumping station flow of urban drainage converged network | |
| WO2020248228A1 (en) | Computing node load prediction method in a hadoop platform | |
| CN103577710B (en) | Aviation Power Converter faults Forecasting Methodology based on fractional order UPF | |
| CN111160520A (en) | BP neural network wind speed prediction method based on genetic algorithm optimization | |
| CN108051035A (en) | The pipe network model recognition methods of neural network model based on gating cycle unit | |
| CN109460621A (en) | Based on the gas turbine blades fault detection method for improving neural network | |
| CN103336867B (en) | Proton Exchange Membrane Fuel Cells model optimization disposal route | |
| CN109190839B (en) | Wind direction-fused intelligent rolling prediction method for wind speed along railway | |
| CN103942434A (en) | Hydropower station plant dam structure vibration response predication method based on SSPSO-GRNN | |
| CN107885928A (en) | Consider the stepstress acceleration Degradation Reliability analysis method of measurement error | |
| CN109034225A (en) | A kind of combination stochastic variable ash and the modified uncertain parameters estimation method of Bayesian model | |
| CN112364560A (en) | Intelligent prediction method for working hours of mine rock drilling equipment | |
| CN114580283B (en) | Mountain stream strong tide estuary branched section medium-long term dynamic relief evolution numerical simulation method | |
| CN109840335A (en) | Based on the radial forging pit prediction optimization method for strengthening T-S fuzzy neural network | |
| CN111709454A (en) | Multi-wind-field output clustering evaluation method based on optimal copula model | |
| CN113642259B (en) | River flow field reconstruction method based on multilayer flow velocity measurement | |
| CN114858207A (en) | Soft measurement-based gridding source tracing investigation method for drain outlet of river channel | |
| CN110069801B (en) | Pick distance estimation method based on artificial neural network | |
| CN113849943A (en) | Water supply network node water demand amount checking method coupled with pressure prior information | |
| CN112307536B (en) | Dam seepage parameter inversion method | |
| CN106599541B (en) | A kind of structure and parameter on-line identification method of dynamic power load model | |
| CN110555566B (en) | B-spline quantile regression-based photoelectric probability density prediction method | |
| CN103489037A (en) | Forecasting method of wind resource capable of generating power | |
| CN110807511A (en) | Coarse set theory-cyclic neural network model structure |
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 |