CN104424507A - Prediction method and prediction device of echo state network - Google Patents
Prediction method and prediction device of echo state network Download PDFInfo
- Publication number
- CN104424507A CN104424507A CN201310379811.2A CN201310379811A CN104424507A CN 104424507 A CN104424507 A CN 104424507A CN 201310379811 A CN201310379811 A CN 201310379811A CN 104424507 A CN104424507 A CN 104424507A
- Authority
- CN
- China
- Prior art keywords
- prediction
- unit
- data
- neuron
- forecast model
- 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.)
- Granted
Links
Abstract
The embodiment of the invention provides a prediction method and a prediction device of an echo state network (ESN) and relates to the technical field of a communication service network so that the prediction performance of the ESN can be improved effectively and the prediction speed and the prediction precision can be improved. The method is as follows: through establishment of a prediction model of the ESN, wherein a dynamic pool of the prediction model is a mixed neuron dynamic pool which includes wavelet neurons, a training set is obtained from input data and prediction training is carried out on the prediction model according to the training set and then the prediction model which completes training is obtained; and through prediction input data, the prediction model which completes training is predicted and then prediction output data in a prediction process is obtained. The prediction method and the prediction device are used for prediction of nonlinear chaotic time sequence data.
Description
Technical field
The present invention relates to communication service network technical field, particularly relate to a kind of Forecasting Methodology and prediction unit of echo state network.
Background technology
Neural network is used nonlinear system to be predicted to have good effect, and be used widely.ESN(Echo State Network, Echo State Networks) be a kind of novel recirculating network structure and recirculating network learning method.ESN has large-scale intrinsic nerve unit, makes this ESN possess good short term memory capacity, in the prediction task of chaos time sequence, well more a lot of than traditional neural network.
At use ESN echo state network in the prediction of nonlinear system, the predetermined speed in traditional ESN echo state network is slow, and precision of prediction can not reach requirement.
Summary of the invention
Embodiments of the invention provide a kind of Forecasting Methodology and prediction unit of echo state networking, effectively can promote the estimated performance of ESN network, improve predetermined speed and precision of prediction.
For achieving the above object, embodiments of the invention adopt following technical scheme:
First aspect, provides a kind of Forecasting Methodology of echo state network, comprising:
Set up the forecast model of echo state network, the dynamic pond of described forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit;
From input data, obtain training set, according to described training set, prediction training is carried out to described forecast model, obtain the forecast model after having trained;
Predicted the forecast model that described training completes by prediction input data, the prediction obtained in forecasting process exports data.
In conjunction with first aspect, in the first mode in the cards, the described forecast model setting up echo state network, the dynamic pond of described forecast model is that the composite nerve unit comprising wavelet neural unit dynamically comprises in pond:
Set up the forecast model of echo state network, described forecast model comprises input layer, middle layer and output layer;
The neuron in described middle layer is interconnected to form described dynamic pond, described dynamic pond is the first dynamically pond of composite nerve, and in the dynamic pond of quantity and described composite nerve unit of wavelet neural unit described in the dynamic pond of described composite nerve unit, the ratio value of neuron population amount is preset ratio value.
Wherein, neuron in described dynamic pond is that annular connects topological structure, described annular connects the unit of wavelet neural described in topological structure and is alternately connected with the general neural unit except described wavelet neural unit, and described wavelet neural unit is Sigmoid type wavelet neural unit.
In conjunction with the first mode in the cards of first aspect, in the second mode in the cards, described from input data obtain training set, according to described training set to described forecast model carry out prediction training, acquisition trained after forecast model comprise:
According to time sequencing, obtain the described training set in the predetermined number unit interval, described sampling input data obtained;
After initialization is carried out to described echo state network, using the input data of described training set as dynamic pond neuron state more new formula, obtain dynamic pond neuron state value;
The neuronic united state matrix of described input layer, described middle layer neuron and described output layer is obtained according to described dynamic pond neuron state value;
Described united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula.
The second in conjunction with first aspect mode in the cards, in the third mode in the cards, described by predicting that input data are predicted the forecast model that described training completes, the prediction obtained in forecasting process exports data and comprises:
Using the prediction input data as data prediction formula of selected data, input data according to described prediction and described data prediction formula is predicted the forecast model that described training completes;
The prediction obtained in forecasting process exports data.
In conjunction with the third mode in the cards of first aspect, in the 4th kind of mode in the cards, described dynamic pond neuron state more new formula comprises:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is described training set input data, and x (n) is the dynamic pond neuron state value before upgrading, and x (n+1) is the dynamic pond neuron state value after upgrading, y (n) is described training set output data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer;
Described output weight matrix formula comprises:
Wherein, (W
out)
t=M
-1t, wherein, W
outfor described output weight matrix, M is described united state matrix, and T is the teacher's data matrix obtained according to described training set.
Described data prediction formula comprises:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is described prediction input data, and y (n) is prediction output data, and the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor described output weight matrix.
Second aspect, provides a kind of prediction unit, comprising:
Set up unit, for setting up the forecast model of echo state network, the dynamic pond of described forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit;
Training unit, for obtaining training set from input data, carrying out prediction training according to described training set to described forecast model, obtaining the forecast model after having trained;
Predicting unit, for being predicted the forecast model that described training completes by prediction input data, the prediction obtained in forecasting process exports data.
In conjunction with second aspect, in the first mode in the cards, described set up unit specifically for:
Set up the forecast model of echo state network, described forecast model comprises input layer, middle layer and output layer;
The neuron in described middle layer is interconnected to form described dynamic pond, described dynamic pond is the first dynamically pond of composite nerve, and in the dynamic pond of quantity and described composite nerve unit of wavelet neural unit described in the dynamic pond of described composite nerve unit, the ratio value of neuron population amount is preset ratio value.
Wherein, neuron in described dynamic pond is that annular connects topological structure, described annular connects the unit of wavelet neural described in topological structure and is alternately connected with the general neural unit except described wavelet neural unit, and described wavelet neural unit is Sigmoid type wavelet neural unit.
In conjunction with the first mode in the cards of second aspect, in the second mode in the cards, described training unit comprises:
Sampling subelement, for according to time sequencing, obtains the described training set obtained described sampling input data in the predetermined number unit interval;
First obtains subelement, for after carrying out initialization to described echo state network, using the input data of described training set as dynamic pond neuron state more new formula, obtains dynamic pond neuron state value;
Second obtains subelement, for obtaining the neuronic united state matrix of described input layer, described middle layer neuron and described output layer according to described dynamic pond neuron state value;
3rd obtains subelement, for described united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula.
The second in conjunction with second aspect mode in the cards, in the third mode in the cards, described predicting unit is used for:
Using the prediction input data as data prediction formula of selected data, input data according to described prediction and described data prediction formula is predicted the forecast model that described training completes;
The prediction obtained in forecasting process exports data.
In conjunction with the third mode in the cards of second aspect, in the 4th kind of mode in the cards, described dynamic pond neuron state more new formula comprises:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is described training set input data, and x (n) is the dynamic pond neuron state value before upgrading, and x (n+1) is the dynamic pond neuron state value after upgrading, y (n) is described training set output data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer;
Described output weight matrix formula comprises:
Wherein, (W
out)
t=M
-1t, wherein, W
outfor described output weight matrix, M is described united state matrix, and T is the teacher's data matrix obtained according to described training set.
Described data prediction formula comprises:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is described prediction input data, and y (n) is prediction output data, and the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor described output weight matrix.
The embodiment of the present invention provides a kind of Forecasting Methodology and prediction unit of echo state network, by setting up the forecast model of echo state network, the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit, training set is obtained from input data, according to training set, prediction training is carried out to forecast model, obtain the forecast model after having trained, by prediction input data, the forecast model of having trained is predicted, the prediction obtained in forecasting process exports data, effectively can promote the estimated performance of ESN network, improve predetermined speed and precision of prediction.
Accompanying drawing explanation
In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.
Fig. 1 provides a kind of Forecasting Methodology schematic flow sheet of echo state network for the embodiment of the present invention;
Fig. 2 provides a kind of Forecasting Methodology schematic flow sheet of echo state network for the embodiment of the present invention;
Fig. 3 provides a kind of dynamic pond neuron ring topology figure of echo state network for the embodiment of the present invention;
A kind of prediction unit structural representation that Fig. 4 provides for the embodiment of the present invention;
A kind of prediction unit structural representation that Fig. 5 provides for the embodiment of the present invention.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is only the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.
The embodiment of the present invention provides a kind of Forecasting Methodology of echo state network, as shown in Figure 1, comprising:
101, prediction unit sets up the forecast model of echo state network, and the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit.
Exemplary, this prediction unit can be the subscriber equipmenies such as computer.This prediction unit comprises input layer, middle layer and output layer.Wherein, input layer is made up of some input neurons, and middle layer is made up of some intrerneurons, and output layer is made up of some output neurons.
Here neuron also can be understood as unit, and also namely input layer is made up of input block, and middle layer is made up of temporary location, and output layer is made up of output unit.
Wherein, the neuron of inside, middle layer is interconnected to form dynamic pond.Dynamic pond in the present embodiment is the dynamic pond of composite nerve unit, and this composite nerve unit dynamically pond comprises wavelet neural general neural unit that is first and that think except wavelet neural is first composition.Here wavelet neural unit can be Sigmoid type wavelet neural unit.
102, prediction unit obtains training set from input data, carries out prediction training, obtain the forecast model after having trained according to training set to forecast model.
Exemplary, this training set can extract Network Central Node IP(Internet Protocol, Internet protocol) packet.
103, prediction unit is predicted the forecast model of having trained by prediction input data, and the prediction obtained in forecasting process exports data.
The embodiment of the present invention provides a kind of Forecasting Methodology of echo state network, by setting up the forecast model of echo state network, the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit, training set is obtained from input data, according to training set, prediction training is carried out to forecast model, obtain the forecast model after having trained, by prediction input data, the forecast model of having trained is predicted, the prediction obtained in forecasting process exports data, effectively can promote the estimated performance of ESN network, improve predetermined speed and precision of prediction.
The embodiment of the present invention provides a kind of Forecasting Methodology of echo state network, as shown in Figure 2, comprising:
201, prediction unit sets up the forecast model of echo state network, and the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit.
Exemplary, ESN(Echo State Network, echo state network) main application fields can be time series forecasting, inverse mode, sequential classification, nonlinear Control, image procossing etc., be a kind of novel artificial neural network.It has large-scale intrinsic nerve unit, makes it possess good short term memory capacity.
Wherein, the forecast model of this ESN comprises input layer, middle layer and output layer.Input layer is made up of input neuron, and as the input of this ESN network, if input layer has K input neuron, then the input dimension of this ESN network is K; Output layer is made up of output neuron, and for exporting predicting the outcome of ESN network, if the output neuron of output layer is L, then the output dimension of this ESN network is L; Middle layer is made up of intrerneuron, if middle layer is by N number of intrerneuron, input layer and middle layer neuron are interconnected to constitute the input weight matrix W of K*N
in, whole three layers are interconnected with output layer neuron the output weight matrix W forming (K+N+L) * L
out, output layer and middle layer neuron are interconnected and form (L*N) feedback weight matrix W
back.The intrerneuron in middle layer is interconnected to constitute the inner weight matrix W in middle layer of N*N, and namely middle layer neuron is interconnected to form dynamic pond.
Concrete, W
in, W
backbe all changeless with W, dynamic pond is here composite nerve unit dynamically pond, composite nerve unit dynamically in pond the neuronic quantity of small echo be preset ratio value with the ratio value of neuron population amount in composite nerve unit dynamic pond.Wherein, wavelet neural unit is Sigmoid type wavelet neural unit, and wavelet neural unit is replaced the neuronic partial nerve unit in middle layer, also namely composite nerve unit is dynamically made up of wavelet neural unit and the general neural unit except wavelet neural unit in pond.Wherein, wavelet neural unit can be 40% ~ 60% of middle layer neuron population amount, such as, generally get wavelet neural unit and account for 50% of middle layer neuron population amount.
Concrete, the activation function of this Sigmoid type wavelet neural unit can adopt as minor function:
Wherein,
r
mixbe mixed into ratio for this little nodal point, be such as that 50%, N represents all neuronic numbers in middle layer, j represents the number of middle layer wavelet neural unit, and the activation function of general neural unit can be tanh function.
After setting up three layers of forecast model, the topological structure in the dynamic pond of ESN is changed.Concrete, the connected mode that the Stochastic sum adopted by the neuron in the dynamic pond of echo state network is sparse, changes into simple ring topology, and omits the neuronic feedback link of output neuron alignment layer.For the composite nerve unit dynamically pond being mixed into wavelet neural unit, the neuron in the dynamic pond of composite nerve unit is that annular connects topological structure.
Concrete, with wavelet neural, unit accounts for 50% of middle layer neuron population amount, can first random selecting general neural unit, this general neural unit is replaced with wavelet neural unit, then order replaces with wavelet neural unit at interval of a general neural unit, final wavelet neural unit is alternately connected with general neural unit, forms the ring topology alternately connected.As shown in Figure 3, such as, in the dynamic pond of composite nerve unit, neuronic total quantity is 10, and wavelet neural unit is 5, and general neural unit is 5, and pond which simplify echo state network forecast model, thus can improve predetermined speed.
In addition, optionally, the neuronic connection weight of input layer and middle layer all can be decided to be v, the interval of v is [0.1,1], generally chooses 0.1; The connection weight of dynamic pond neuron self is all decided to be r, the interval of r is [0.1,1], generally choose 0.5, connection weight between neuron all can be set to a designated value by the way, thus further simplify echo state network forecast model, improve predetermined speed.
202, prediction unit is according to time sequencing, obtains the training set obtained sampling input data in the predetermined number unit interval.
Exemplary, can be by choosing certain Network Central Node IP packet, according to time sequencing, the packet obtained in its unit interval is input data, the input data of getting in turn in the predetermined number unit interval according to time sequencing like this obtain training set, such as predetermined number here can be 1000 chronomeres, can be also other value, not limit here.Unit interval can be one minute, can be also other value, not limit here.Wherein, inputting data is network True Data.
203, prediction unit is after carrying out initialization to echo state network, using the input data of training set as dynamic pond neuron state more new formula, obtains dynamic pond neuron state value.
Concrete, first, initialization is carried out to ESN network, namely to the dynamic pond of mixing, W
in, W, W
outcarry out initialization, and make middle layer internal state quantitative change be x (0)=0, wherein x is that a N*1 ties up state vector.Then, by the training set of input, ESN network is trained, the input/output signal of this training set is corresponded to input block and the output unit of this ESN network.Wherein, ESN network dynamic pond neural state more new formula be:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is training set input data, and x (n) is the dynamic pond neuron state value before renewal, and x (n+1) is the dynamic pond neuron state value after upgrading, and y (n) is that training set exports data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer, and represent time point, f is the neuronic activation function in dynamic pond, when dynamic pond neuron is general neural unit, and f=tanh (x); When dynamic pond neuron is Sigmoid type wavelet neural unit, f=ψ
d,t(x).
Here x (n+1) is vector value, i.e. x (n+1)=(x1 (n), x2 (n) ... x_N (n)), each dynamic pond neuron oneself corresponding state value corresponding.
204, prediction unit obtains the neuronic united state matrix of input layer, middle layer neuron and output layer according to dynamic pond neuron state value.
Concrete, because the neuronic state value in dynamic pond may be subject to the impact of original state value, front T0 section time series can be given up, then in the T-T0 section time, according to the dynamic pond neuron state value of each time point that step 203 obtains, and each state value x (n) of each time point is formed as independent a line the united state matrix M that scale is (T-T0+1) * (K+N+L) together with input layer, output layer neuron.Wherein, K is the number of input neuron, and N is the neuronic number in dynamic pond, and L is the number of output neuron.
205, united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula by prediction unit.
Optionally, exporting weight matrix formula can be:
(W
out)
t=M
-1t ', wherein, W
outfor exporting weight matrix, (W
out)
tfor exporting weight matrix W
outturn order matrix, namely t represents transposition, and M is united state matrix, and T ' is the teacher's data matrix obtained according to training set.Wherein, scale can be obtained for (T-T according to output data y (n) of training set
0) teacher's data matrix T ' of * L.
Like this, the forecast model of having trained can just be obtained.
206, prediction unit is using the prediction input data as data prediction formula of selected data, predicts the forecast model of having trained according to prediction input data and data prediction formula.
Concrete, after completing ESN network training, the ESN trained can input data by input prediction, predicts the forecast model of having trained.
In whole step operational process, this ESN network, by constantly updating dynamic pond neuron state, obtains and exports weight matrix, just can predict according to prediction input data and data prediction formula.
Wherein, data prediction formula is:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is prediction input data, and y (n) exports data for predicting, the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor exporting weight matrix.
207, the prediction that prediction unit obtains in forecasting process exports data.
In forecasting process, prediction can be obtained according to above-mentioned data prediction formula and export data.Such as, for network traffics, prediction exports the estimated value that data are future time point uninterrupted.
Like this, illustrated by above-mentioned steps, by adding the chaotic property in the dynamic pond of ESN network, add the neuron kind in ESN middle layer, alternately be connected to form simple annular to the wavelet neural in dynamic pond unit and the first interval of general neural to be connected, effectively can improve the estimated performance of ESN network, raising precision of prediction, such as, when this prediction unit is used for the prediction to network traffics, the needs of network traffics aspect prediction can be met better.
The True Data that the present embodiment adopts data center to gather, effectively can be solved the realistic problem of predicting network flow, well be applied by prediction.
In addition, the present embodiment, except may be used for predicting network flow, also can be applied to network participle or clustering algorithm, implementation and the embodiment of the present invention similar, repeat no more here.
The embodiment of the present invention provides a kind of Forecasting Methodology of echo state network, by setting up the forecast model of echo state network, the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit, training set is obtained from input data, according to training set, prediction training is carried out to forecast model, obtain the forecast model after having trained, by prediction input data, the forecast model of having trained is predicted, the prediction obtained in forecasting process exports data, effectively can promote the estimated performance of ESN network, improve predetermined speed and precision of prediction.
The embodiment of the present invention provides a kind of prediction unit 01, as shown in Figure 4, comprising:
Set up unit 011, for setting up the forecast model of echo state network, the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit.
Training unit 012, for obtaining training set from input data, carrying out prediction training according to training set to forecast model, obtaining the forecast model after having trained.
Predicting unit 013, for being predicted the forecast model of having trained by prediction input data, the prediction obtained in forecasting process exports data.
Optionally, setting up unit 011 can be specifically for:
Set up the forecast model of echo state network, forecast model comprises input layer, middle layer and output layer;
The neuron in middle layer is interconnected to form dynamic pond, and dynamic pond is the first dynamically pond of composite nerve, and in the dynamic pond of composite nerve unit, in the dynamic pond of the neuronic quantity of small echo and composite nerve unit, the ratio value of neuron population amount is preset ratio value.
Optionally, the neuron in dynamic pond is that annular connects topological structure, and annular connects small echo neuron in topological structure and is alternately connected with the general neural unit except wavelet neural unit.
Optionally, wavelet neural unit is Sigmoid type wavelet neural unit.
Optionally, as shown in Figure 5, training unit 012 can comprise:
Sampling subelement 0121, for according to time sequencing, obtains the training set obtained sampling input data in the predetermined number unit interval;
First obtains subelement 0122, for after carrying out initialization to echo state network, using the input data of training set as dynamic pond neuron state more new formula, obtains dynamic pond neuron state value;
Second obtains subelement 0123, for obtaining the neuronic united state matrix of input layer, middle layer neuron and output layer according to dynamic pond neuron state value;
3rd obtains subelement 0124, for united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula.
Optionally, predicting unit 013 may be used for:
Using the prediction input data as data prediction formula of selected data, according to prediction input data and data prediction formula, the forecast model of having trained is predicted;
The prediction obtained in forecasting process exports data.
Wherein, dynamic pond neuron state more new formula comprise:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is training set input data, and x (n) is the dynamic pond neuron state value before renewal, and x (n+1) is the dynamic pond neuron state value after upgrading, and y (n) is that training set exports data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer;
Export weight matrix formula to comprise:
Wherein, (W
out)
t=M
-1t, wherein, W
outfor exporting weight matrix, M is united state matrix, and T is the teacher's data matrix obtained according to training set.
Data prediction formula comprises:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is prediction input data, and y (n) exports data for predicting, the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor exporting weight matrix.
By above-mentioned explanation, in this prediction unit 01, input neuron may be used for the input of training set and prediction input data, and output neuron exports the output of data for teacher's data and prediction.Wherein, dynamic pond intrerneuron can be made up of several storeies, can store the state of each each storer of time point, thus obtains the output weight matrix after training.
In addition, each input of input neuron connects and composes input connection matrix with each storer in middle layer, and connection matrix in the middle of intrerneuron is formed according to cross feedback Topology connection, whole neuron and output neuron are formed and export weight matrix.
Like this, in this prediction unit 01, training set is inputted by input neuron, this device is trained, this device can record the input of each time point, output and intermediateness value, and combine the total state matrix of formation, obtain the output weight matrix of having trained by inverting, line item of going forward side by side.After training completes, use and export weight matrix, applied forcasting Formula Input Technology sample sequence, be also input prediction data, obtain predicting the outcome of this sample sequence, namely prediction exports data.
The embodiment of the present invention provides a kind of prediction unit, by setting up the forecast model of echo state network, the dynamic pond of forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit, training set is obtained from input data, according to training set, prediction training is carried out to forecast model, obtain the forecast model after having trained, by prediction input data, the forecast model of having trained is predicted, the prediction obtained in forecasting process exports data, effectively can promote the estimated performance of ESN network, improve predetermined speed and precision of prediction.
In several embodiments that the application provides, should be understood that disclosed apparatus and method can realize by another way.Such as, apparatus embodiments described above is only schematic, such as, the division of described unit, be only a kind of logic function to divide, actual can have other dividing mode when realizing, such as multiple unit or assembly can in conjunction with or another system can be integrated into, or some features can be ignored, or do not perform.Another point, shown or discussed coupling each other or direct-coupling or communication connection can be by some interfaces, and the indirect coupling of device or unit or communication connection can be electrical, machinery or other form.
In addition, in the equipment in each embodiment of the present invention and system, each functional unit can be integrated in a processing unit, also can be that the independent physics of unit comprises, also can two or more unit in a unit integrated.And above-mentioned each unit both can adopt the form of hardware to realize, the form that hardware also can be adopted to add SFU software functional unit had realized.
The all or part of step realizing said method embodiment can have been come by the hardware that programmed instruction is relevant, and aforesaid program can be stored in a computer read/write memory medium, and this program, when performing, performs the step comprising said method embodiment; And aforesaid storage medium comprises: USB flash disk, portable hard drive, ROM (read-only memory) (Read Only Memory, be called for short ROM), random access memory (Random Access Memory, be called for short RAM), magnetic disc or CD etc. various can be program code stored medium.
The above; be only the specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; change can be expected easily or replace, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of described claim.
Claims (10)
1. a Forecasting Methodology for echo state network, is characterized in that, comprising:
Set up the forecast model of echo state network, the dynamic pond of described forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit;
From input data, obtain training set, according to described training set, prediction training is carried out to described forecast model, obtain the forecast model after having trained;
Predicted the forecast model that described training completes by prediction input data, the prediction obtained in forecasting process exports data.
2. method according to claim 1, is characterized in that, the described forecast model setting up echo state network, and the dynamic pond of described forecast model is that the composite nerve unit comprising wavelet neural unit dynamically comprises in pond:
Set up the forecast model of echo state network, described forecast model comprises input layer, middle layer and output layer;
The neuron in described middle layer is interconnected to form described dynamic pond, described dynamic pond is the first dynamically pond of composite nerve, and in the dynamic pond of quantity and described composite nerve unit of wavelet neural unit described in the dynamic pond of described composite nerve unit, the ratio value of neuron population amount is preset ratio value;
Wherein, neuron in described dynamic pond is that annular connects topological structure, described annular connects the unit of wavelet neural described in topological structure and is alternately connected with the general neural unit except described wavelet neural unit, and described wavelet neural unit is Sigmoid type wavelet neural unit.
3. method according to claim 2, is characterized in that, described from input data obtain training set, according to described training set to described forecast model carry out prediction training, acquisition trained after forecast model comprise:
According to time sequencing, obtain the described training set in the predetermined number unit interval, described sampling input data obtained;
After initialization is carried out to described echo state network, using the input data of described training set as dynamic pond neuron state more new formula, obtain dynamic pond neuron state value;
The neuronic united state matrix of described input layer, described middle layer neuron and described output layer is obtained according to described dynamic pond neuron state value;
Described united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula.
4. method according to claim 3, is characterized in that, described by predicting that input data are predicted the forecast model that described training completes, the prediction obtained in forecasting process exports data and comprises:
Using the prediction input data as data prediction formula of selected data, input data according to described prediction and described data prediction formula is predicted the forecast model that described training completes;
The prediction obtained in forecasting process exports data.
5. method according to claim 4, is characterized in that, described dynamic pond neuron state more new formula comprises:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is described training set input data, and x (n) is the dynamic pond neuron state value before upgrading, and x (n+1) is the dynamic pond neuron state value after upgrading, y (n) is described training set output data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer;
Described output weight matrix formula comprises:
Wherein, (W
out)
t=M
-1t, wherein, W
outfor described output weight matrix, M is described united state matrix, and T is the teacher's data matrix obtained according to described training set;
Described data prediction formula comprises:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is described prediction input data, and y (n) is prediction output data, and the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor described output weight matrix.
6. a prediction unit, is characterized in that, comprising:
Set up unit, for setting up the forecast model of echo state network, the dynamic pond of described forecast model is the composite nerve unit dynamically pond comprising wavelet neural unit;
Training unit, for obtaining training set from input data, carrying out prediction training according to described training set to described forecast model, obtaining the forecast model after having trained;
Predicting unit, for being predicted the forecast model that described training completes by prediction input data, the prediction obtained in forecasting process exports data.
7. device according to claim 6, is characterized in that, described set up unit specifically for:
Set up the forecast model of echo state network, described forecast model comprises input layer, middle layer and output layer;
The neuron in described middle layer is interconnected to form described dynamic pond, described dynamic pond is the first dynamically pond of composite nerve, and in the dynamic pond of quantity and described composite nerve unit of wavelet neural unit described in the dynamic pond of described composite nerve unit, the ratio value of neuron population amount is preset ratio value;
Wherein, neuron in described dynamic pond is that annular connects topological structure, described annular connects the unit of wavelet neural described in topological structure and is alternately connected with the general neural unit except described wavelet neural unit, and described wavelet neural unit is Sigmoid type wavelet neural unit.
8. device according to claim 7, is characterized in that, described training unit comprises:
Sampling subelement, for according to time sequencing, obtains the described training set obtained described sampling input data in the predetermined number unit interval;
First obtains subelement, for after carrying out initialization to described echo state network, using the input data of described training set as dynamic pond neuron state more new formula, obtains dynamic pond neuron state value;
Second obtains subelement, for obtaining the neuronic united state matrix of described input layer, described middle layer neuron and described output layer according to described dynamic pond neuron state value;
3rd obtains subelement, for described united state matrix and teacher's data are obtained output weight matrix as the input exporting weight matrix formula.
9. device according to claim 8, is characterized in that, described predicting unit is used for:
Using the prediction input data as data prediction formula of selected data, input data according to described prediction and described data prediction formula is predicted the forecast model that described training completes;
The prediction obtained in forecasting process exports data.
10. device according to claim 9, is characterized in that, described dynamic pond neuron state more new formula comprises:
x(n+1)=f(W
inu(n+1)+Wx(n)+W
backy(n))
Wherein, u (n+1) is described training set input data, and x (n) is the dynamic pond neuron state value before upgrading, and x (n+1) is the dynamic pond neuron state value after upgrading, y (n) is described training set output data, W
infor input weight matrix, W is the inside weight matrix in dynamic pond, and n is nonnegative integer;
Described output weight matrix formula comprises:
Wherein, (W
out)
t=M
-1t, wherein, W
outfor described output weight matrix, M is described united state matrix, and T is the teacher's data matrix obtained according to described training set;
Described data prediction formula comprises:
y(n+1)=f
out(W
out(u(n+1),x(n+1),y(n)))
Wherein, u (n+1) is described prediction input data, and y (n) is prediction output data, and the prediction that y (n+1) is next unit interval exports data, and x (n+1) is the dynamic pond neuron state value after upgrading, W
outfor described output weight matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310379811.2A CN104424507B (en) | 2013-08-28 | 2013-08-28 | Prediction method and prediction device of echo state network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310379811.2A CN104424507B (en) | 2013-08-28 | 2013-08-28 | Prediction method and prediction device of echo state network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104424507A true CN104424507A (en) | 2015-03-18 |
| CN104424507B CN104424507B (en) | 2020-03-03 |
Family
ID=52973415
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310379811.2A Active CN104424507B (en) | 2013-08-28 | 2013-08-28 | Prediction method and prediction device of echo state network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104424507B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777752A (en) * | 2016-12-30 | 2017-05-31 | 华东交通大学 | A kind of bullet train follows the trail of operation curve Optimal Setting method |
| CN106846173A (en) * | 2016-12-30 | 2017-06-13 | 国网新疆电力公司电力科学研究院 | Short-term wind power forecast method based on EWT ESN |
| CN108113665A (en) * | 2017-12-14 | 2018-06-05 | 河北大学 | A kind of automatic noise-reduction method of electrocardiosignal |
| CN109727218A (en) * | 2018-12-30 | 2019-05-07 | 杭州铭智云教育科技有限公司 | A kind of full graphics extracting method |
| CN109726691A (en) * | 2018-12-30 | 2019-05-07 | 杭州铭智云教育科技有限公司 | A kind of monitoring method |
| CN110118986A (en) * | 2019-04-18 | 2019-08-13 | 广州供电局有限公司 | Localization method, control method, computer equipment and computer readable storage medium |
| CN111563767A (en) * | 2020-04-26 | 2020-08-21 | 中国银行股份有限公司 | Stock price prediction method and device |
| CN113689699A (en) * | 2021-08-25 | 2021-11-23 | 浙江数智交院科技股份有限公司 | Traffic flow prediction method and device, electronic equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102147273A (en) * | 2010-01-29 | 2011-08-10 | 大连理工大学 | Data-based blast-furnace gas dynamic predication method for metallurgical enterprises |
| CN102663493A (en) * | 2012-03-23 | 2012-09-12 | 天津工业大学 | Delaying nerve network used for time sequence prediction |
| JP5087879B2 (en) * | 2006-08-03 | 2012-12-05 | 住友金属工業株式会社 | Powder injection method |
-
2013
- 2013-08-28 CN CN201310379811.2A patent/CN104424507B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP5087879B2 (en) * | 2006-08-03 | 2012-12-05 | 住友金属工業株式会社 | Powder injection method |
| CN102147273A (en) * | 2010-01-29 | 2011-08-10 | 大连理工大学 | Data-based blast-furnace gas dynamic predication method for metallurgical enterprises |
| CN102663493A (en) * | 2012-03-23 | 2012-09-12 | 天津工业大学 | Delaying nerve network used for time sequence prediction |
Non-Patent Citations (2)
| Title |
|---|
| 王卓群: "基于回声状态网络的移动话务量预测方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 王瑟: "基于改进的回声状态神经网络的非线性预测", 《中国优秀博硕士学位论文全文数据库 (硕士) 信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777752A (en) * | 2016-12-30 | 2017-05-31 | 华东交通大学 | A kind of bullet train follows the trail of operation curve Optimal Setting method |
| CN106846173A (en) * | 2016-12-30 | 2017-06-13 | 国网新疆电力公司电力科学研究院 | Short-term wind power forecast method based on EWT ESN |
| CN106777752B (en) * | 2016-12-30 | 2019-04-02 | 华东交通大学 | A kind of bullet train tracking operation curve optimal setting method |
| CN108113665A (en) * | 2017-12-14 | 2018-06-05 | 河北大学 | A kind of automatic noise-reduction method of electrocardiosignal |
| CN109727218A (en) * | 2018-12-30 | 2019-05-07 | 杭州铭智云教育科技有限公司 | A kind of full graphics extracting method |
| CN109726691A (en) * | 2018-12-30 | 2019-05-07 | 杭州铭智云教育科技有限公司 | A kind of monitoring method |
| CN109727218B (en) * | 2018-12-30 | 2021-01-08 | 南方电网物资有限公司 | Complete graph extraction method |
| CN110118986A (en) * | 2019-04-18 | 2019-08-13 | 广州供电局有限公司 | Localization method, control method, computer equipment and computer readable storage medium |
| CN110118986B (en) * | 2019-04-18 | 2021-06-22 | 广东电网有限责任公司广州供电局 | Positioning method, control method, computer device, and computer-readable storage medium |
| CN111563767A (en) * | 2020-04-26 | 2020-08-21 | 中国银行股份有限公司 | Stock price prediction method and device |
| CN113689699A (en) * | 2021-08-25 | 2021-11-23 | 浙江数智交院科技股份有限公司 | Traffic flow prediction method and device, electronic equipment and storage medium |
| CN113689699B (en) * | 2021-08-25 | 2023-04-18 | 浙江数智交院科技股份有限公司 | Traffic flow prediction method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104424507B (en) | 2020-03-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104424507A (en) | Prediction method and prediction device of echo state network | |
| CN104346517A (en) | Echo state network based prediction method and prediction device | |
| CN106471526B (en) | For handling the method and system of image | |
| CN108648095A (en) | A kind of nodal information hidden method accumulating gradient network based on picture scroll | |
| Zheng et al. | Improving the efficiency of multi-objective evolutionary algorithms through decomposition: An application to water distribution network design | |
| CN102567391A (en) | Method and device for building classification forecasting mixed model | |
| Eagle et al. | Interaction Networks: Generating High Level Hints Based on Network Community Clustering. | |
| CN106875004A (en) | Composite mode neuronal messages processing method and system | |
| Reddy et al. | Investigation of reconfiguration effect on makespan with social network method for flexible job shop scheduling problem | |
| CN112541584B (en) | Deep neural network model parallel mode selection method | |
| CN107133181A (en) | A kind of construction method of difference wavelet neural network software fault prediction technology | |
| CN106803135A (en) | The Forecasting Methodology and device of a kind of photovoltaic power generation system output power | |
| CN113780002A (en) | Knowledge reasoning method and device based on graph representation learning and deep reinforcement learning | |
| CN106874506A (en) | community mining method and system based on statistical model | |
| CN104462773A (en) | River network construction method used for biodiversity conversation and application thereof | |
| CN102779142A (en) | Quick community discovery method based on community closeness | |
| CN109583588A (en) | A kind of short-term wind speed forecasting method and system | |
| CN104361462B (en) | Social network influence maximization approach based on cultural gene algorithm | |
| CN104915714A (en) | Predication method and device based on echo state network (ESN) | |
| CN109681165A (en) | For the water filling policy optimization method and device in oil extraction in oil field | |
| CN109446213A (en) | A kind of learning materials intelligently pushing method and device | |
| CN109800517A (en) | Improved reverse modeling method for magnetorheological damper | |
| CN104331738A (en) | Network reconfiguration algorithm based on game theory and genetic algorithm | |
| CN107423811A (en) | The streamflow change attribution recognition methods combined based on BP artificial neural networks and Scene Simulation | |
| Deal et al. | A use-driven approach to large-scale urban modelling and planning support |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |