CN105894090A - Tide intelligent real-time forecasting method based on adaptive variation particle swarm optimization - Google Patents

Abstract

The invention discloses a tide intelligent real-time forecasting method based on adaptive variation particle swarm optimization (PSO). The tide intelligent real-time forecasting method comprises the following steps: loading the tide actual measured data; constructing an SAPSO-BP network forecasting model; calculating an error function; performing loop iteration optimizing; and setting the network parameters of a BP (Back Propagation) neural network. As the tide intelligent real-time forecasting method based on adaptive variation particle swarm optimization utilizes the variation idea in a genetic algorithm GA and introduces the variation operation in the PSO algorithm, the tide intelligent real-time forecasting method expands the continuously-reduced population searching space during the iteration process so as to enable the particles to jump out of the previously searched optimal position and to search in a larger searching space and can maintain the diversity of population at the same time, thus improving the possibility of searching a superior value by means of the algorithm. Therefore, relative to a traditional PSO-BP algorithm, the tide intelligent real-time forecasting method based on adaptive variation particle swarm optimization is higher in the searching accuracy and searching efficiency, and is higher in prediction accuracy, compared with a traditional PSO-BP mode and a harmonic analysis model.

Description

A kind of tide intelligence Real-time Forecasting Method based on TSP question particle group optimizing

Technical field

The present invention relates to the forecasting technique of a kind of tide, particularly a kind of tide intelligence Real-time Forecasting Method.

Background technology

The learning process of BP (Back Propagation) neutral net, i.e. error backpropagation algorithm, by information Forward-propagating and two processes of backpropagation composition of error.The each neuron of input layer is responsible for receiving from outward The input information on boundary, and pass to each neuron in intermediate layer；Intermediate layer is that internal information processes layer, is responsible for letter Breath conversion, according to the demand of information change ability, intermediate layer can be designed as single hidden layer or many hidden layer configurations； Last hidden layer is delivered to the information of each neuron of output layer, after further treatment after, complete once to learn Forward-propagating processing procedure, output layer outwardly export information processing result.When reality output and expectation When output is not inconsistent, enter the back-propagation phase of error.Error passes through output layer, declines by error gradient Mode correction each layer weights, to the successively anti-pass of hidden layer, input layer.The information forward-propagating gone round and begun again and mistake Difference back-propagation process, is the processes that constantly adjust of each layer weights, is also the process of neural network learning training, This process is performed until the error of network output and is reduced to acceptable degree, or set in advance Till practising number of times.Although BP network is widely used, but self there is also some defects and deficiency: by Being fixing in learning rate, therefore the convergence rate of network is slow, needs the longer training time；Network is hidden The number of plies and the selection of unit number containing layer there is no theoretic guidance；The learning and memory of network has instability Property.

Particle swarm optimization algorithm (Particle Swarm Optimization, PSO) is a kind of based on intelligence side of colony The evolutionary computing of method, is mainly used to seek globally optimal solution.Its basic thought is, each optimization problem Potential solution is all a particle in search volume, and all of particle has an optimised function to determine Adaptive value, each particle also have velocity vector to determine direction that they circle in the air and distance, then particles Just follow the search in solution space of the current optimal particle.PSO is initialized as a group random particles, then Optimal solution is found by iteration.In each iteration, particle updates oneself by following the tracks of two extreme values, First is exactly the preferably solution that particle itself finds to current time, and this solves and is referred to as individual best values, Another extreme value is exactly the preferably solution that whole population is found to current time, this value be global best fitness (i.e. Optimal solution).

PSO-BP particle swarm optimization algorithm is a kind of evolutionary computation technique, and PSO is similar to genetic algorithm, is A kind of optimization tool based on iteration.System initialization is one group of RANDOM SOLUTION, by iterated search optimal value. But it does not has intersection and the variation of genetic algorithm, but the particle that particle follows optimum in solution space is carried out Search.PSO optimization neural network includes 2 aspects: the 1) topological structure of optimization neural network；2) instruction Practice the connection weights between neural network model, the mainly each layer of optimization neural network model and threshold values.Tradition PSO-BP forecast model utilize the connection weights between PSO layer each to neutral net to carry out excellent with threshold values Change, to accelerate the training speed of neutral net and the training quality improving neutral net.

System of harmonic analysis: tide is sea level cyclical upturn and downturn campaign.The motive power that tide produces is The vector sum of the centrifugal inertial force needed for the gravitation of the power to lead tide of celestial body, i.e. celestial body and the earth-celestial body relative motion. Actual application has many kind methods can with the lifting information of observational record tidal level, but the most general one The method of kind is exactly system of harmonic analysis, and complicated tide is resolved into some periodically variable parts, wherein by it Each part is the partial tide ripple produced by a Between Celestial Tide-generating Forces supposed.Tidal level according to actual observation Data are analyzed, can be in the hope of each constant in tide mediation model.Then further according to the tide obtained Harmonic constant, it will be appreciated that the size of partial tide wave component, and can be used to calculate tide and for tidal wave numerical value meter Calculate and foundation is provided.In theory, the part of tide is complicated, and the partial tide quantity of tide is numerous, But in engineering calculation, relatively small when the mean amplitude of tide of partial tide nighttide and partial tide Greenwich delay angle (divides The angle value corresponding to partial tide high water time is there is to somewhere in tide celestial body through somewhere meridian upper transit) relatively Time long, it is negligible for having significant component of tide partial tide composition.In actual computation, somewhere Actual tide level is represented by:

$h\left(t\right)=H_0+\underset{k=1}{\overset{n}{\Σ}}{h}_{k}cos({\mathrm{\ω}}_{k}t-{\mathrm{\φ}}_k)$

Wherein: H₀For the mean sea level height during analyzing, n is tide partial tide quantity, h_kFor the amplitude of partial tide, ω_kAnd φ_kBeing frequency and the phase place of partial tide respectively, t is any time.

System of harmonic analysis is the technology that tide prediction aspect is the most traditional, and it is with tidal static and dynamics as base Plinth, through improvement for many years and development, substantially having been carried out can be to the stabilization forecast of tidal level.But, System of harmonic analysis needs substantial amounts of long-term observation tide level data to analyze, and can obtain being in harmonious proportion the most accurately Analyze model.But owing to long-term field data observational record spends cost the highest, therefore typically it is difficult to The long-term observation data needed to these.And the average prediction error of harmonic analysis model is about 20-30cm. It addition, harmonic analysis model only only accounts for the impact in terms of Between Celestial Tide-generating Forces, but the generation of tide is by all Impact such as many time-varying factors such as wind-force, wind direction, air pressure, ocean temperatures.Therefore harmonic analysis model is neglected Having omited the time-varying some effects that tide produces, the factor generated due to tide is complicated and changeable, so tide is whole Body change show certain non-linear and uncertainty, traditional static structure harmonic analysis model be difficult to into Row high accuracy and the forecast of real-time tidal level.

Traditional PSO optimized algorithm also exists that easy Premature Convergence, search precision be relatively low, later stage iteration efficiency The shortcoming such as the highest.Neutral net based on error back propagation (BP neutral net) is now the most popular, Practical a kind of neutral net, but owing to BP neutral net is highly susceptible to topology of networks and network The impact of size, the simulation result fluctuation thus resulting in BP network changes greatly.Meanwhile, BP is neural The convergence rate of network can be selected to be affected by network initial weight threshold value, if initial network parameter choosing It is improper to select, it is possible to cause the simulation training of BP neutral net can be absorbed in local optimum.It is proposed to this end that base PSO optimized algorithm forecasting model in BP neutral net.

Traditional PSO-BP mixing forecasting model.The most improved traditional PS O-BP forecast model is by grain Subgroup optimized algorithm (particle swarm optimism algorithm) is applied to BP (back propagation) The optimization of the network parameter of neutral net, i.e. in order to weights and the threshold value of Optimized BP Neural Network.PSO-BP Model using the weight threshold of BP neutral net as the population particle position of PSO optimized algorithm carry out random at the beginning of Beginningization, then calculates the fitness value of each particle by error function (fitness function).Use and calculate Fitness value judge whether iterative algorithm meets the requirement of iteration optimizing, finally obtain through PSO iteration optimizing Best initial weights threshold value, and best initial weights threshold value is assigned to BP neutral net, carries out emulation experiment.

But traditional PSO-BP forecasting model yet suffers from, and search precision is low, search efficiency is low and forecast precision is low Problem.

Document related to the present invention is as follows:

[1]C.P.Tsai and T.L.Lee,Back-propagation neural network in tidal-level forecasting.Journal of Waterways,Port,Coastal,and Ocean Engineering,125(4): 195-202,1999。

[2]Y.Guo,J.P.Zhang and R.Dai,Marine Navigation,Dalian:Dalian Maritime University Press,2014。

[3]J.Kennedy and R.Eberhart,Particle Swarm Optimization,Proceedings of IEEE International Conference on Neural Networks,1942-1948,1995。

[4]J.C.Yin,J.Z.Zou,and F.Xu,Sequential learning radial basis function network for real-time tidal level predictions,Ocean Engineering,57:49-55,2013。

[5]Y.Guo,J.P.Zhang and R.Dai,Marine Navigation,Dalian:Dalian Maritime University Press,2014。

[6]G.Li.Y.L.Hao,and Y.X.Zhao,Research of neural network to tidal prediction,Proceedings of International Joint Conference on Computational Science and Optimization,282-284,2009。

[7]S.Yu,K.Zhu,F.Diao,A dynamic all parameters adaptive BP neural networks model and its application on oil reservoir prediction,Applied Mathematic and Computation,195:66-75,2008。

Summary of the invention

For solving the problems referred to above that prior art exists, the present invention to design a kind of search precision height, search effect The tide intelligence Real-time Forecasting Method based on TSP question particle group optimizing that rate is high and forecast precision is high.

To achieve these goals, technical scheme is as follows: a kind of based on TSP question population The tide intelligence Real-time Forecasting Method optimized, comprises the following steps:

A, loading tide measured data

Described tide measured data derives from the real-time monitored record value of each tidal observation point, and by tide Measured data normalized.

B, structure SAPSO-BP network forecasting model

Based on the tide measured data after normalized, create BP neural network model and SAPSO is set Optimized algorithm, builds SAPSO-BP network forecasting model: will introduce biography by adaptive mutation operator SA PSO in the particle swarm optimization algorithm of system, then includes weights and threshold by the network parameter of BP neural network model Value, is initialized as the population particle position of adaptive particle swarm optimization algorithm SAPSO.Changed by SAPSO Obtain optimum particle position for optimizing and be weights and the threshold value of BP neural network model, optimizing is obtained Optimal network parameter is assigned to BP neural network model and carries out final network simulation forecast.

BP neutral net hidden layer output valve is calculated by equation below:

$H_j=f(\underset{i=1}{\overset{n}{\Σ}}{w}_{ij}{x}_i-{\mathrm{\θ}}_j)j=1,2,\mathrm{...},l---\left(1\right)$

Wherein: θ_jFor the threshold value of hidden layer, f () is the nonlinear transfer function of hidden layer node.w_ikFor input Weights between layer and hidden layer, n is input layer number, and l is node in hidden layer, x_iFor input data. BP network output layer output valve is calculated by equation below:

$O_k=\underset{j=1}{\overset{t}{\Σ}}{H}_j{w}_{jk}-a_k,k=1,2,...,m---\left(2\right)$

Wherein: a_kOutput layer threshold value, w_jkFor the weights between hidden layer and output layer, m is output layer node Number, H_jFor hidden layer output valve.

The error calculation formula of SAPSO-BP network forecasting model is as follows:

$e_k=\frac{1}{2}\underset{k}{\overset{m}{\Σ}}{(Y_k-O_k)}^2,k=1,2,\mathrm{...},m---\left(3\right)$

Wherein Y_kFor the tide measured data of SAPSO-BP network forecasting model, O_kFor SAPSO-BP network The simulation data data of forecasting model, m is output layer nodes.

SAPSO-BP network forecasting model in iterative process each time, the position of particle and speed more new formula As follows:

v_i(t+1)=ω * v_i(t)+c₁*r₁*(p_i-x_i(t))+c₂*r₂*(p_g-x_i(t)) (4)

x_i(t+1)=x_i(t)+v_i(t+1) j=1,2 ..., n (5)

Wherein ω is inertia weight, and k is current iteration number of times, x_iFor particle position, v_iFor particle rapidity, P_i For individual extreme value, P_gFor colony's extreme value, c₁And c₂For nonnegative constant, r₁And r₂For between 0 and 1 And random number.For preventing particle blind search, initial position and speed to particle limit.SAPSO-BP The network parameter c1=c2=1.55 of network forecasting model, iteration optimizing number of times is 200, and population scale is 20, The initial velocity of each particle is limited between [-3,3], and the initial position of each particle is limited between [-5,5], Adaptive mutation rate formula in SAPSO-BP network forecasting model is as follows:

Pop (j, pos)=λ * rands (1,1) (6)

Wherein: j is number of particles, pos is a uniform Discrete Stochastic integer.Pop is particle populations quantity, λ is the maximum of particle populations quantity.

C, the network parameter of BP neutral net is initialized as the particle populations position of SAPSO optimized algorithm, The network parameter of BP neural network includes: weights between input layer and hidden layer, hidden layer threshold value, implicit Weights between layer and output layer and output layer threshold value.Calculate particle according to error function formula initially to adapt to Degree functional value.Error function computing formula is as follows:

Error=| Y_k-O_k| (7)

Wherein: Y_kFor the actual observation data of forecasting model, O_kSimulation data data for forecasting model.

D, in iterative process each time, particle updates the speed of self according to more new formula (4) and (5) And position.And calculate new fitness function value according to error function computing formula (7).Then heredity is used for reference Variation thought in algorithm, introduces TSP question operation, then according to TSP question in PSO algorithm Formula (6) calculates and more new particle individuality extreme value and colony's extreme value.

E, judge that the optimum individual fitness function value error function value that i.e. error function formula (7) calculates is No meet error requirement be set, or whether iteration optimizing number of times reaches to arrange requirement, requires if met, Perform step F, otherwise return step D and proceed loop iteration optimizing.

F, terminate SAPSO optimized algorithm iteration optimizing, the optimal network parameter assignment that optimizing is obtained to BP neutral net carries out emulation experiment.

G, the network parameter of BP neutral net is set: iterative cycles number of times is arranged between [1,500], study Rate and learning objective are all disposed between [0,1], then optimal network parameter are assigned to BP neutral net and carry out tide Nighttide real-time prediction emulation experiment.

Iterative cycles number of times described in step G of the present invention is 100, learning rate is 0.1, learning objective is 0.00001。

Compared with prior art, the method have the advantages that

Particle swarm optimization algorithm PSO convergence is fast, has the strongest versatility, but there is easy precocity simultaneously The shortcomings such as convergence, search precision is low, later stage iteration is inefficient.The present invention uses for reference Genetic Algorithms (Genetic Algorithm) the variation thought in, introduces mutation operation, i.e. to some variable with necessarily in PSO algorithm Probability reinitialize.Mutation operation has expanded the population search volume the most constantly reduced, So that particle can jump out the optimal location that prior searches arrives, bigger search volume is launched search, Maintain again the diversity of population simultaneously, improve algorithm and search out the more figure of merit and obtain possibility.Therefore, general Introducing mutation operator on the basis of logical particle cluster algorithm, basic thought is after particle updates every time, with Certain probability reinitializes particle.Select the real-time observed data at same harbour as network forecasting model Training input data, and use harmonic analysis model, PSO-BP network model, and SAPSO-BP mould Type carries out emulation experiment.Can show that by the com-parison and analysis of Fig. 2 the adaptation of SAPSO-BP forecast model is write music Line decrease speed is substantially fast than PSO-BP, and fall is the most relatively large.Therefore, SAPSO-BP calculates Method has higher search precision and search efficiency relative to traditional PSO-BP algorithm.By the analysis ratio of Fig. 3 Relatively can show that SAPSO-BP model has more relative to traditional PSO-BP model and harmonic analysis model High precision of prediction.

Accompanying drawing explanation

The present invention has 3, accompanying drawing, wherein:

Fig. 1 is the schematic flow sheet of the present invention.

Fig. 2 be the fitness function value curve ratio of two kinds of Forecasting Methodologies relatively.

Fig. 3 is the precision of prediction contrast of three kinds of different Forecasting Methodologies.

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is further described through.

As in figure 2 it is shown, calculate according to the flow process shown in Fig. 1, SAPSO-BP network forecasting model adapts to The PSO-BP network forecasting model that the fall off rate of degree functional value is substantially more improved than tradition is fast, say, that The error reduction speed of SAPSO-BP network forecasting model is very fast, later stage iteration efficiency is high.And it tends to steady The error amount of timing is also significantly less than the error amount of the most improved PSO-BP network forecasting model of tradition.In addition From the fitness value curve of two kinds of network forecasting models it can be seen that the particle swarm optimization algorithm of band mutation operator The local optimum that before can jumping out rapidly, iteration optimizing obtains, thus obtain more excellent result.

As shown in Figure 3: three kinds of heterogeneous networks forecasting models use identical training data at identical emulation ring Precision of prediction contrast under border.The prediction error of system of harmonic analysis is can get between-0.05 to-0.25 by Fig. 3 Change fluctuation, and its change fluctuating range the most greatly and also mean error the most greatly, traditional harmonic analysis is described Method forecasting model is difficult to carry out high-precision tide prediction smoothly.Traditional PSO-BP network forecasting model Prediction error changes fluctuation between 0.1 to-0.05, and relative fluctuation Amplitude Ratio system of harmonic analysis has had the biggest Declining, and its fluctuating error scope there has also been certain reducing, mean error is gradually to 0 convergence.SAPSO-BP The predicated error of network forecasting model changes fluctuation between 0.05 to-0.05, and its fluctuating range is relative to front two Kind of network forecasting model has had and has significantly reduced, and fluctuating range is compared more steady, and mean error becomes substantially In 0.It can be inferred that: the SAPSO-BP network forecasting model of SAPSO-BP network forecasting model Prediction error is minimum in three kinds of forecasting models, and its prediction error amplitude of variation is also minimum, because of This TSP question particle group optimizing BP neural network model improved can carry out stable high-precision tide Intelligence real-time prediction.

Claims (2)

1. a tide intelligence Real-time Forecasting Method based on TSP question particle group optimizing, it is characterised in that: Comprise the following steps:

A, loading tide measured data:

Described tide measured data derives from the real-time monitored record value of each tidal observation point, and by tide Measured data normalized；

B, structure SAPSO-BP network forecasting model:

Based on the tide measured data after normalized, create BP neural network model and SAPSO is set Optimized algorithm, builds SAPSO-BP network forecasting model: will introduce biography by adaptive mutation operator SA PSO in the particle swarm optimization algorithm of system, then includes weights and threshold by the network parameter of BP neural network model Value, is initialized as the population particle position of adaptive particle swarm optimization algorithm SAPSO；Changed by SAPSO Obtain optimum particle position for optimizing and be weights and the threshold value of BP neural network model, optimizing is obtained Optimal network parameter is assigned to BP neural network model and carries out final network simulation forecast；

BP neutral net hidden layer output valve is calculated by equation below:

$H_j=f(\underset{i=1}{\overset{n}{\Σ}}{w}_{ij}{x}_i-{\mathrm{\θ}}_j),j=1,2,...,l---\left(1\right)$

Wherein: θ_jFor the threshold value of hidden layer, f () is the nonlinear transfer function of hidden layer node；w_ikFor input Weights between layer and hidden layer, n is input layer number, and l is node in hidden layer, x_iFor input data； BP network output layer output valve is calculated by equation below:

$O_k=\underset{j=1}{\overset{t}{\Σ}}{H}_j{w}_{jk}-a_k,k=1,2,...,m---\left(2\right)$

Wherein: a_kOutput layer threshold value, w_jkFor the weights between hidden layer and output layer, m is output layer node Number, H_jFor hidden layer output valve；

The error calculation formula of SAPSO-BP network forecasting model is as follows:

$e_k=\frac{1}{2}\underset{k}{\overset{m}{\Σ}}{(Y_k-O_k)}^2,k=1,2,...,m---\left(3\right)$

Wherein Y_kFor the tide measured data of SAPSO-BP network forecasting model, O_kFor SAPSO-BP network The simulation data data of forecasting model, m is output layer nodes；

SAPSO-BP network forecasting model in iterative process each time, the position of particle and speed more new formula As follows:

v_i(t+1)=ω * v_i(t)+c₁*r₁*(p_i-x_i(t))+c₂*r₂*(p_g-x_i(t)) (4)

x_i(t+1)=x_i(t)+v_i(t+1) j=1,2 ..., n (5)

Wherein ω is inertia weight, and k is current iteration number of times, x_iFor particle position, v_iFor particle rapidity, P_i For individual extreme value, P_gFor colony's extreme value, c₁And c₂For nonnegative constant, r₁And r₂For between 0 and 1 And random number；For preventing particle blind search, initial position and speed to particle limit；SAPSO-BP The network parameter c1=c2=1.55 of network forecasting model, iteration optimizing number of times is 200, and population scale is 20, The initial velocity of each particle is limited between [-3,3], and the initial position of each particle is limited between [-5,5], Adaptive mutation rate formula in SAPSO-BP network forecasting model is as follows:

Pop (j, pos)=λ * rands (1,1) (6)

Wherein: j is number of particles, pos is a uniform Discrete Stochastic integer；Pop is particle populations quantity, λ is the maximum of particle populations quantity；

C, the network parameter of BP neutral net is initialized as the particle populations position of SAPSO optimized algorithm, The network parameter of BP neural network includes: weights between input layer and hidden layer, hidden layer threshold value, implicit Weights between layer and output layer and output layer threshold value；Calculate particle according to error function formula initially to adapt to Degree functional value；Error function computing formula is as follows:

Error=| Y_k-O_k| (7)

Wherein: Y_kFor the actual observation data of forecasting model, O_kSimulation data data for forecasting model；

D, in iterative process each time, particle updates the speed of self according to more new formula (4) and (5) And position；And calculate new fitness function value according to error function computing formula (7)；Then heredity is used for reference Variation thought in algorithm, introduces TSP question operation, then according to TSP question in PSO algorithm Formula (6) calculates and more new particle individuality extreme value and colony's extreme value；

E, judge that the optimum individual fitness function value error function value that i.e. error function formula (7) calculates is No meet error requirement be set, or whether iteration optimizing number of times reaches to arrange requirement, requires if met, Perform step F, otherwise return step D and proceed loop iteration optimizing；

F, terminate SAPSO optimized algorithm iteration optimizing, the optimal network parameter assignment that optimizing is obtained to BP neutral net carries out emulation experiment；

G, the network parameter of BP neutral net is set: iterative cycles number of times is arranged between [1,500], study Rate and learning objective are all disposed between [0,1], then optimal network parameter are assigned to BP neutral net and carry out tide Nighttide real-time prediction emulation experiment.

A kind of tide intelligence based on TSP question particle group optimizing the most according to claim 1 is real Time forecasting procedure, it is characterised in that: in step G, described iterative cycles number of times is 100, learning rate is 0.1, learning objective is 0.00001.