CN106250980A - The sliding squid cental fishing ground Forecasting Methodology of a kind of Argentina - Google Patents

Abstract

The sliding squid cental fishing ground Forecasting Methodology of a kind of Argentina, arranges including spatial and temporal scales setting, envirment factor, sets up cental fishing ground forecast model, it is characterized in that spatial and temporal scales arranges the space scale using three ranks, week and the time scale of two ranks of the moon；It is the main environment factor that envirment factor arranges employing table temperature (SST), it is aided with sea level height (SSH), two kinds of envirment factors of chlorophyll a (Chl a) again, when setting up cental fishing ground forecast model, envirment factor is divided into four kinds of situations: according to spatial and temporal scales and envirment factor facilities, set up the sample plan collection of 24 kinds of situations；Cental fishing ground forecast model uses classical error back propagation BP neural network model, BP neural network model is three-decker, i.e. input layer, hidden layer and output layer, the temp-spatial factors in input layer input fishing ground and envirment factor, output layer output CPUE or the fishing ground grading index being transformed by CPUE are for forecast.

Description

The sliding squid cental fishing ground Forecasting Methodology of a kind of Argentina

Technical field

The present invention relates to the sliding squid Predicting Center Fishery method of Predicting Center Fishery method, especially Argentina.

Background technology

Predicting Center Fishery is the one of fishing condition speed report, and Predicting Center Fishery can be that fishing improves fishery harvesting accurately Yield also reduces fuel cost, and fishing condition speed report is to cental fishing ground position, shoal of fish trend and prosperous in following 24h or several day Probability is predicted, fishing news commander's unit timing every day forecast content be quickly and accurately broadcast to by telecommunication system Produce ship, reach to command the purpose of produced on-site.

The cental fishing ground of the most existing multiple method forecast economic squids of ocean property, the basis of these methods is Fish row Moving the relation between biological condition and environmental condition and rule thereof, essence is all to obtain according to production leadtime data sample " Heuristics " is used for forecasting, but the most not furthers investigate the spatial and temporal scales of sample and the selection of envirment factor, substantially in the past It is rule of thumb to set (such as little fishing zone, big fishing zone etc.), does not accounts for different time and space scales and envirment factor to Predicting Center Fishery The impact of model；In the selection of model, seldom consider the real time problems of the marine environment factor.The ocean ring of different waters Border condition is different, and the mechanism that cental fishing ground is formed is the most different, and therefore cental fishing ground is predicted by its time and spatial resolution Also there were significant differences in the impact of model.Therefore, for understanding the sample spatial and temporal scales of the economic squids of ocean property and envirment factor Select impact on its cental fishing ground forecasting model, establish during businessization is run under optimal spatial and temporal scales and envirment factor Heart fishing ground forecasting model.

Summary of the invention

The present invention have studied the sample spatial and temporal scales of the economic squids of ocean property and the selection of envirment factor to its center fishing The impact of field prediction model, proposes a kind of sliding squid Predicting Center Fishery method of Argentina.

Technical scheme includes that spatial and temporal scales is arranged, envirment factor arranges, sets up cental fishing ground forecast model, its Feature be spatial and temporal scales arrange use three ranks space scale, longitude and latitude be respectively 0.25 ° × 0.25 °, 0.5 ° × 0.5 °, 1.0 ° × 1.0 °, week and the time scale of two ranks of the moon；It is the main environment factor that envirment factor arranges employing SST, then is aided with Two kinds of envirment factors of SSH, Chl-a, are divided into four kinds of situations when setting up cental fishing ground forecast model by envirment factor: I SST；Ⅱ SST, SSH；III SST, Chl-a；IV SST, SSH, Chl-a；According to spatial and temporal scales and envirment factor facilities, set up 24 kinds of feelings The sample plan collection of condition；Cental fishing ground forecast model uses classical error back propagation BP neural network model, BP nerve net Network model is three-decker, i.e. input layer, hidden layer and output layer, the temp-spatial factors in input layer input fishing ground and envirment factor, Output layer output CPUE or the fishing ground grading index being transformed by CPUE；During BP neural network model forward-propagating, sample Enter from input layer, process through hidden layer activation primitive, be transmitted to output layer, as output layer actual output and desired output not Coincidence loss requirement, then proceed to the back-propagation phase of error, and back propagation is by error by hidden layer to input layer successively Back propagation, gives all node of each layer by error distribution, thus obtains the error signal of each node layer, and this error signal is as repairing Positive foundation, the forward-propagating of this signal is carried out again and again with the back propagation of error, and weights constantly adjust, this process Be performed until network output error be reduced to acceptable degree or proceed to study number of times set in advance till, pass through Learning training acquires best model, for forecast.

The present invention considers the selection of different time and space scales and the envirment factor impact on cental fishing ground forecast model, uses warp The error backward propagation method (Error Backpropagation Network, BP) of allusion quotation, BP neutral net belongs to multilamellar Feedforward neural network, uses the supervision algorithm of error back propagation, and BP neutral net can learn and store substantial amounts of pattern and reflect Emission mode, for main (high yield) activity duration, optimal operation marine site scope, squid perches suitable SST scope, SSH model Enclosing, Chl-a scope is forecast, improves fishery harvesting yield for fishing and reduces fuel cost offer technical support.

Accompanying drawing explanation

Fig. 1 is BP Artificial Neural Network Structures figure.

Detailed description of the invention

In order to compare the suitableeest spatial and temporal scales of the Predicting Center Fishery model of the economic squids of ocean property, arrange three The space scale of rank, longitude and latitude is respectively 0.25 ° × 0.25 °, 0.5 ° × 0.5 °, 1.0 ° × 1.0 °, the time of two ranks Yardstick is week and the moon.

The resource abundance in the fishing ground of the economic squids of ocean property is not only affected by temp-spatial factors, and by the ring of habitat Border Effects of Factors.Wherein, SST is to be widely studied and of paramount importance factor of influence, therefore, selected SST be main environment because of Son, then it is aided with two kinds of envirment factors of SSH, Chl-a, so envirment factor being divided into four kinds when setting up Predicting Center Fishery model Situation (table 1).

Table 1 envirment factor is arranged

Therefore, according to spatial and temporal scales and the envirment factor facilities of sample, set up the economic squids center fishing of ocean property The sample plan collection of field prediction model has following 24 kinds of situations:

Table 2BP forecasting model sample set scheme

Cental fishing ground forecast model uses classical error backward propagation method (Error Backpropagation Network, BP), BP neutral net belongs to multilayer feedforward neural network, uses the supervision algorithm of error back propagation, and BP is neural Network can learn and store substantial amounts of mode map pattern.

BP model uses three-decker, i.e. input layer, hidden layer and output layer (Fig. 1).Input layer be fishing ground space-time because of Son and the marine environment factor, output layer is CPUE or the fishing ground grading index being transformed by CPUE, different fishing grounds grade Division methods is with reference to the domain knowledge of fisheries expert.Hidden layer node number is determined by empirical equation:

P_num=2N_num+1

In formula: P_numFor hidden layer node number, N_numFor input layer number.

BP algorithm mainly includes the forward-propagating of learning process signal and two process compositions of back propagation of error.Forward During propagation, sample from input layer enter, through hidden layer activation primitive process, be transmitted to output layer, as output layer actual output with Desired output does not meets error requirements, then proceed to the back-propagation phase of error.Back propagation is by error with some form By hidden layer to input layer successively back propagation, give all node of each layer by error distribution, thus obtain the mistake of each node layer Difference signal, this error signal is as the foundation revised.The forward-propagating of this signal is to go round and begin again with the back propagation of error Ground is carried out, and weights constantly adjust, namely the process of e-learning.This process is performed until the error of network output and is reduced to Acceptable degree or till proceeding to study number of times set in advance.

Training method uses steepest descent method.Assuming that input neuron number is M, hidden layer neuron number is I, output Layer neuron number is J.Input layer m-th neuron is designated as x_m, hidden layer i-th neuron is designated as k_i, output layer jth god It is designated as y through unit_j.From x_mTo k_iLink weights be w_mi, from k_iTo y_jConnection weights be w_ij.Hidden layer transmission function is Sigmoid function, output layer transmission function is linear function.U and v represents input and the output of each layer respectively, asRepresent I The input of layer (hidden layer) first neuron.The actual output of network is represented by:

$Y\left(n\right)=\[v_J^1,v_J^2,\mathrm{...},v_J^J\]$

The desired output of network is:

D (n)=[d₁, d₂..., d_J]

N is iterations.The definitions for error signals of nth iteration is:

e_j(n)=d_j(n)-Y_j(n)

Error energy is defined as:

$e\left(n\right)=\frac{1}{2}\underset{j=1}{\overset{J}{\Σ}}{e}_j^2\left(n\right)$

Training process is i.e. the process reduced by error energy.

In the weighed value adjusting stage, successively it is reversed adjustment along network.First adjust between hidden layer and output layer Weight w_ij, according to steepest descent method, error should be calculated to w_ijGradientOpposite direction further along the direction is adjusted Whole:

${\mathrm{\Δw}}_{ij}\left(n\right)=-\mathrm{\η}\frac{\∂e\left(n\right)}{\∂w_{ij}\left(n\right)}$

w_ij(n+1)=Δ w_ij(n)+w_ij(n)

Gradient can be obtained by seeking local derviation, according to the chain type rule of differential, has

$\frac{\∂e\left(n\right)}{\∂w_{ij}\left(n\right)}=\frac{\∂e\left(n\right)}{\∂e_j\left(n\right)}\·\frac{\∂e_j\left(n\right)}{\∂v_J^j\left(n\right)}.\frac{\∂v_J^j\left(n\right)}{\∂u_J^j\left(n\right)}\·\frac{\∂u_J^j\left(n\right)}{\∂w_{ij}\left(n\right)}$

Owing to e (n) is e_jThe quadratic function of (n), its differential is linear function:

$\frac{\∂e\left(n\right)}{\∂e_j\left(n\right)}=e_j\left(n\right)$

$\frac{\∂e_j\left(n\right)}{\∂v_J^j\left(n\right)}=-1$

Output layer transmission function derivative:

$\frac{\∂v_J^j\left(n\right)}{\∂u_J^j\left(n\right)}=g^{\′}{u}_J^j\left(n\right)$

$\frac{\∂u_J^j\left(n\right)}{\∂w_{ij}\left(n\right)}=v_I^i\left(n\right)$

Therefore, Grad is

$\frac{\∂e\left(n\right)}{\∂w_{ij}\left(n\right)}=-e_j\left(n\right)g^{\′}\left(u_J^j\right(n\left)\right)v_I^i\left(n\right)$

The correction of weights is

${\mathrm{\Δw}}_{ij}\left(n\right)={\mathrm{\ηe}}_j\left(n\right)g^{\′}\left(u_J^j\right(n\left)\right)v_J^i\left(n\right)$

The definition of introducing partial gradient:

${\mathrm{\δ}}_J^j=-\frac{\∂e\left(n\right)}{\∂u_J^j\left(n\right)}=-\frac{\∂e\left(n\right)}{\∂e_j\left(n\right)}\·\frac{\∂e_j\left(n\right)}{\∂v_J^j\left(n\right)}\·\frac{\∂v_J^j\left(n\right)}{\∂u_J^j\left(n\right)}=e_j\left(n\right)g^{\′}\left(u_J^j\right(n\left)\right)$

So the correction of weights is:

${\mathrm{\Δw}}_{ij}\left(n\right)={\mathrm{\η\δ}}_J^j{v}_I^i\left(n\right)$

At output layer, transmission function is linear function, and therefore its derivative is 1, i.e.

$g^{\′}\left(u_J^j\right(n\left)\right)=1$

So can obtain

${\mathrm{\Δw}}_{ij}\left(n\right)={\mathrm{\ηe}}_j\left(n\right)v_I^i\left(n\right)$

Error signal is propagated forward, to the weight w between input layer and hidden layer_miIt is adjusted, seemingly should with previous step class Have

${\mathrm{\Δw}}_{mi}\left(n\right)={\mathrm{\η\δ}}_J^j{v}_M^m\left(n\right)$

For the output of input neuron, i.e.

For partial gradient, it is defined as

${\mathrm{\δ}}_I^i=--\frac{\∂e\left(n\right)}{\∂u_I^i\left(n\right)}=-\frac{\∂e\left(n\right)}{\∂v_I^i\left(n\right)}\·\frac{\∂v_I^i\left(n\right)}{\∂u_I^i\left(n\right)}=-\frac{\∂e\left(n\right)}{\∂v_I^i\left(n\right)}{f}^{\′}\left(u_I^i\right(n\left)\right)$

F (g) is sigmoid function, and previous step calculates visible the most again,

$\frac{\∂e\left(n\right)}{\∂v_I^i\left(n\right)}=\underset{j=1}{\overset{J}{\Σ}}{\mathrm{\δ}}_J^j{w}_{ij}$

Therefore have

${\mathrm{\δ}}_I^i=f^{\′}\left(u_I^i\right(n\left)\right)\underset{j=1}{\overset{J}{\Σ}}{\mathrm{\δ}}_J^j{w}_{ij}$

Arriving this, the study weighed value adjusting process of three layers of BP network terminates, and can be attributed to:

Weighed value adjusting amount Δ w=learning rate η partial gradient δ last layer output signal v.As for learning rate η, error model The setting enclosed etc., carries out progressively tuning when not over-fitting.

The process of setting up of BP neutral net completes in matlab (2010b) software, uses the plan of Neural Network Toolbox Conjunction instrument, is divided into training sample, checking sample and test sample three part by sample set.The parameter of network design is: study speed Rate 0.1, momentum parameter 0.5, the transmission function between input layer and hidden layer, hidden layer and output layer neuron is S type respectively Tan tansig, linear function purelin；The terminal parameter of network training is: maximum frequency of training is 1000, and maximum is by mistake Difference is given as 0.001.Model acquires best model by repeatedly training, and weighting is heavily for forecast.

BP forecasting model evaluated in terms of forecast precision, stability and interpretability three by model:

(1) forecast precision evaluation

When model is output as CPUE grade, the correct grade percentage ratio gone out according to model prediction, relatively various models Precision；When model is output as CPUE numerical value, the mean square error (MSE) of computation model, compares the precision of each model.

$MSE=\frac{1}{N}{\mathrm{\Σ}}_{k=1}^N{(y_k-{\hat{y}}_k)}^2$

Wherein, y_kFor the actual value of CPUE,Predicted value for CPUE.

(2) estimation of stability

Evaluate the stability of the BP model accuracy of different Sample Establishing, calculate average relative variation value (Average Relative Variance, ARV), it is defined as

$ARV=\frac{{\Σ}_{i=1}^N{\[x\left(i\right)-\hat{x}\left(i\right)\]}^2}{{\Σ}_{i=1}^N{\[x\left(i\right)-\stackrel{\‾}{x}\left(i\right)\]}^2}$

Wherein, N is the number comparing data, and x (i) is fishing ground grade actual value,For fishing ground grade actual mean value,For fishing ground grade forecast value.Average relative variation value ARV is the least, shows that prediction effect is the best, and ARV=0 represents and reaches Perfect forecast effect, as ARV=1, shows that model has only reached the prediction effect of meansigma methods.

(3) interpretability evaluation

With correlation of variables (Independent variable relevance) and sensitive analysis (Sensitivity Analyses) interpretability of the forecasting model set up on different time and space scales and envirment factor sample is evaluated.

Variable is correlated with for relatively each input variable contribution rate to CPUE, and computational methods are that input variable connects with hidden layer The ratio of the weight quadratic sum connect and all input layer variablees to hidden layer connection weight quadratic sum.

Sensitive analysis is to probe into the relation between input variable change and output variable, and its process is: first calculate each The maximum of individual input variable, minima, intermediate value, meansigma methods, mode particular value；Then select one of them input variable, make It gradually changes from minima to maximum, and other input variables are all defined as in four particular values, change change in turn The input variable changed, observes the situation of change of output variable.

Table 3 Fishing Ground of lllex argentinus based on CPUE grade

Correlation of variables analysis shows: under time-of-week yardstick and month yardstick, the SST contribution to fishing ground forecasting model Rate is maximum, next to that " latitude " variable.

Table 4 forecasting model correlation of variables is analyzed

The sliding squid Predicting Center Fishery model of Argentina of different time and space scales and envirment factor is set up by this Forecasting Methodology, Forecast precision more than 90%, ARV value, about 0.2, has the highest precision and minimum ARV value.

Claims (2)

1. the sliding squid cental fishing ground Forecasting Methodology of Argentina, arranges including spatial and temporal scales setting, envirment factor, sets up center Fishing ground forecast model, it is characterized in that spatial and temporal scales arrange use three ranks space scale, longitude and latitude be respectively 0.25 ° × 0.25 °, 0.5 ° × 0.5 °, 1.0 ° × 1.0 °, week and the time scale of two ranks of the moon；It is main that envirment factor arranges employing SST Want envirment factor, then be aided with two kinds of envirment factors of SSH, Chl-a, when setting up cental fishing ground forecast model, envirment factor is divided into Four kinds of situations: I SST；II SST, SSH；III SST, Chl-a；IV SST, SSH, Chl-a；According to spatial and temporal scales and environment because of Sub-facilities, sets up the sample plan collection of 24 kinds of situations；Cental fishing ground forecast model uses classical error back propagation BP Neural network model, BP neural network model is three-decker, i.e. input layer, hidden layer and output layer, input layer input fishing ground Temp-spatial factors and envirment factor, output layer output CPUE or the fishing ground grading index being transformed by CPUE；BP nerve net During network model forward-propagating, sample enters from input layer, processes through hidden layer activation primitive, is transmitted to output layer, such as output layer Actual output does not meets error requirements with desired output, then proceed to the back-propagation phase of error, and back propagation is by error By hidden layer to input layer successively back propagation, give all node of each layer by error distribution, thus obtain the mistake of each node layer Difference signal, this error signal is as the foundation revised, and the forward-propagating of this signal is with the back propagation of error again and again Carrying out, weights constantly adjust, and this process is performed until the error of network output and is reduced to acceptable degree or proceeds to pre- Till the study number of times first set, acquire best model by learning training, for forecast.

The sliding squid cental fishing ground Forecasting Methodology of a kind of Argentina the most according to claim 1, is characterized in that hidden layer node Number is by empirical equation P_num=2N_num+ 1 determines, in formula: P_numFor hidden layer node number, N_numFor input layer number.