CN107180261A - Based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network - Google Patents
Based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network Download PDFInfo
- Publication number
- CN107180261A CN107180261A CN201710426571.5A CN201710426571A CN107180261A CN 107180261 A CN107180261 A CN 107180261A CN 201710426571 A CN201710426571 A CN 201710426571A CN 107180261 A CN107180261 A CN 107180261A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- neural network
- greenhouse
- particle
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Abstract
The present invention proposes a kind of based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, according to predicted time, each moment builds a BP neural network, ultimately form the BP neural network group of a rolling, this method operation includes two stages, unsupervised learning is carried out using autocoder first and obtains good initial network parameter, recycles improved localized particle group optimizing method to optimize the network parameter, sets up initial BP neural network;Then on the basis of initial BP neural network, carry out rolling training and prediction using the output of previous network as the part input of latter network.The present invention can accurately predict long-term Trend of Environmental Change in the greenhouse under Various Seasonal different geographical, and effectively improve the precision of prediction of Greenhouse grape.
Description
Technical field
It is especially a kind of stingy based on the greenhouse for rolling BP neural network the invention belongs to industrialized agriculture environmental forecasting field
Wait medium- and long-term forecasting method.
Background technology
The efficient production in greenhouse depends on suitable greenhouse micro-climate, sets up long-term in high-precision Greenhouse grape
Forecast model is to realizing that greenhouse Optimum Regulation is significant.Although the threshold value control method commonly used in current greenhouse is simple
It is easy, but high energy consumption, the stability of a system are poor.Based on proportional-integral-differential (Proportion-Integral-Derivative,
PID) the autocontrol method, reliability such as controller and Model Predictive Control (Model Predictive Control, MPC)
High, energy consumption is relatively low, but needs the ambient parameter of look-ahead multiple periods.Greenhouse grape simulation model is broadly divided into two classes:
One is mechanism model, and its parameter is more difficult to be determined, is not suitable for greenhouse flower.Two be experimental model, and also referred to as System Discrimination can
To carry out on-line tuning to model parameter, to meet the requirement of control.What is commonly used in experimental model is artificial nerve network model,
Because BP neural network is simple and fault-tolerant ability strong, it is most widely used in Greenhouse grape prediction.
Current domestic and foreign scholars establish the miniclimate simulation model based on BP neural network for different greenhouses, take
Good effect was obtained, research shows that artificial neural network is practical in terms of greenhouse micro-climate prediction, but these are pre-
Single-step Prediction, i.e. short-term forecast can only be carried out by surveying model majority, it is impossible to realize medium- and long-term forecasting, it is impossible to meet wanting for Optimum Regulation
Ask.In addition, there is certain advantage using BP neural network modeling, but it also has some shortcomings and deficiencies, is such as easily absorbed in office
The problems such as portion's minimum value, the undue selection for relying on initial weight and poor generalization ability, therefore the precision of BP neural network prediction
Still have greatly improved space.Conventional Many researchers do not propose improved method for the defect of BP neural network, only
Choose optimal result to show, in fact these results are not convincing to a certain extent.How BP nerve is improved
The precision of prediction of network, and the medium- and long-term forecasting of Greenhouse grape is realized, it is worth further research and inquires into.
The content of the invention
Technical problem solved by the invention is to provide a kind of based on long in the Greenhouse grape for rolling BP neural network
Phase Forecasting Methodology, the BP neural network group of a rolling is built according to predicted time, using the output of previous network as latter
The part input of individual network carries out the training and prediction of roller, effectively improves the precision of prediction of Greenhouse grape.
The technical solution for realizing the object of the invention is:
Based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, comprise the following steps:
Step 1:Set up initial BP neural network f1If current time is t, the inside greenhouse humiture of t is inputted, it is defeated
Go out the inside greenhouse humiture at the t+1 moment of prediction, and obtain f1Network parameter;
Step 2:Set up the BP neural network group rolled, including n-1 neutral net fn, each neutral net fnInclude instruction
Practice collection train_XnWith test set test_Xn, when being separated by one between the training set and test set of two neighboring neutral net
Carve, wherein, train_XnRepresent the training set at t+n-1 moment, test_XnRepresent the test set at t+n-1 moment, n >=2;
Step 3:Utilize train_XnF is trained with network parameter combination gradient descent methodnModel, after the completion of training, then will
train_XnIt is input to fnIn model, analog result train_Y is exportedn;By test_XnIt is input to fnIn model, output prediction knot
Fruit test_Yn;
Step 4:N=n+1 is made, step 3 is gone to.
Further, of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, step 1 has
Body includes:
Step 1-1:Pre-training is carried out to the inside greenhouse humiture of t based on unsupervised learning model, input is extracted
Exported after the feature of data, and reconstruct;
Step 1-2:Using the feature of data as the initiation parameter of BP neural network, have the target of supervision to learn,
Using the weight and threshold parameter of the improved localized particle group optimizing method combination genetic algorithm optimization BP neural network;
Step 1-3:Initial BP neural network f is set up using optimal weights and threshold parameter1, export the t+1 moment of prediction
Inside greenhouse humiture.
Further, it is of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, step 1-1
It is described to be specially to the method that input data is reconstructed:By the weight between input layer and hidden layer and threshold value { W(1),b(1)Make
For encoder, coding function uses sigmoid functions;By the weight between hidden layer and output layer and threshold value { W(2),b(2)Make
For decoder, decoding function uses tanh functions.
Further, it is of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, step 1-2
Concretely comprise the following steps:
Step 1-2-1:Population is divided into two subgroups, calculated simultaneously in spmd parallel organizations, population is initialized
Speed and position, learning rate C1And C2, inertia weight;
Step 1-2-2:The sub-average particle number of times of number of times is reset, global optimum is assigned to subgroup global optimum, i.e.,
BadNum [N]=0, PLg=Pg, badNum is particle number of times, and N is the particle numbering that number of times is below the average, PLgIt is global for subgroup
It is optimal, PgFor global optimum;
Step 1-2-3:The speed of more new particle and position:
vi(t+1)=ω vi(t)+c1r1(pavg-xi(t))+c2r2(pLg-xi(t)), xi(t+1)=xi(t)+vi(t+1),
Wherein, i=1,2 ..., N, t are current iteration number of times, and ω is inertia weight, and c1, c2 are accelerated factors, and r1, r2 are [0,1] areas
Between random number, vi(t) it is the former speed of particle, vi(t+1) it is the particle rapidity after updating, pavgFor individual extreme value central point, pLg
For the global optimum position of each subgroup, xi(t) put for particles in-situ, xi(t+1) it is the particle position after updating;
Step 1-2-4:Crossover operator is introduced, if the random number produced is less than crossover probability PC, then two subgroups perform friendship
Fork operation:xik=pLg1k, xjl=pLg2l, wherein, xikElement, p are tieed up for the kth of i-th of particle position in first subgroupLg1kFor
The kth dimension element of first subgroup global optimum position, xjlElement is tieed up for the l of j-th of particle position in second subgroup,
pLg2lElement, i, j=1,2 ..., N/2 and i ≠ j, k ∈ [(IN+1) * HN+ are tieed up for the l of second subgroup global optimum position
1, D], l ∈ [1, (IN+1) * HN], IN are the input layer number of neutral net, and HN is hidden layer neuron number, and D is
The dimension of particle, and the fitness J (i) of each particle is calculated, if random number is more than crossover probability PC, then without any behaviour
Make;
Step 1-2-5:Update local optimum Pi, will be new if the particle position after updating is better than original particle position
Particle position as the particle Pi, and it is used as the global optimum P in current iterationLg, update individual mechanism center point Pavg, meter
Each subgroup average fitness fit_avg is calculated, if the particle position after updating is not better than original particle position, without appointing
What is operated;
Step 1-2-6:Mutation operator is introduced, if J (i) < fit_avg, make badNum (i)+1, if badNum (i) >=
BadNumLimit, the then position of random initializtion particle and speed:xid=a+ (b-a) * rand, vid=m+ (n-m) * rand,
Wherein d=1,2 ..., D, a and b are the minimum and maximum positions for limiting particle, and m and n are the minimum and maximum speed for limiting particle
Degree, rand for [0,1) between uniform random number;
Step 1-2-7:Judge whether to reach default inner iterative number of times, if so, the subgroup for then comparing two subgroups is optimal,
Global optimum is obtained, if it is not, then going to step 1-2-3;
Step 1-2-8:Judge whether to reach maximum iteration or meet gbest (n)-gbest (n-4)<=0.0001,
If so, then stopping iteration, if it is not, then brick step 1-2-2.
Further, it is of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, step 1
Predictor formula is:
Wherein, (P)tFor t ambient parameter, (Tin)tFor t greenhouse observed temperature, (Hin)tIt is real for t greenhouse
Measuring moisture,For the t+1 moment greenhouse temperatures of prediction,For the t+1 moment chamber humidities of prediction.
Further, in the Greenhouse grape medium- and long-term forecasting method of the invention based on rolling BP neural network, step 2
A moment be 15min.
Further, it is of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, training set
train_XnGreenhouse external environment influence factor (P) including the t+n-1 momentt+n-1With neutral net fn-1Training set simulation knot
Fruit train_Yn-1, wherein train_Yn-1The inside greenhouse humiture at the t+n-1 moment including prediction.
Further, it is of the invention based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, rolling
The predictor formula of BP neural network is:
Wherein, (P)t+n-1For the ambient parameter at t+n-1 moment,For the t+n-1 moment greenhouse temperatures of prediction,For the t+n-1 moment chamber humidities of prediction,For the t+n moment greenhouse temperatures of prediction,For prediction
T+n moment chamber humidities.
The present invention uses above technical scheme compared with prior art, with following technique effect:
1st, method of the invention can continuously predict following 6-12 hours Greenhouse grape, test result indicates that, with tradition
BP single steps shift to an earlier date on line rolling prediction model and compare, more than 50% can be reduced by rolling the following 6 hours humiture errors of BP model predictions,
The cumulative errors of medium-term and long-term rolling forecast are greatly reduced, can accurately be predicted in the greenhouse under Various Seasonal different geographical
Long-range circumstances variation tendency, foundation is provided to formulate rational miniclimate regulation and control scheme.
2nd, the method first stage of the invention uses improved BP neural network, test result indicates that, with initial nerve net
Network model is compared using BP networks, and the following 6 hours humiture errors of rolling BP model predictions proposed by the present invention can be reduced
9.3%~45%, it is highly effective to illustrate improved BP neural network, and the overall prediction essence for rolling BP models can be improved conscientiously
Degree.
3rd, unsupervised learning model is used in Greenhouse grape medium- and long-term forecasting by method of the invention first, experimental result
Show, predicated error reduces 10% or so, and operational efficiency improves more than 20%.
4th, method of the invention is optimized using improved localized particle group optimizing method to BP neural network, with standard
Localized particle group optimizing method compare, prediction humiture error reduction by 10%~30%.
Brief description of the drawings
Fig. 1 is the model structure based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network of the present invention
Figure;
Fig. 2 is the improved localized particle group optimizing method flow chart of the present invention;
Fig. 3 is the god of n-th of BP based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network of the present invention
Through e-learning and prediction flow chart.
Embodiment
Embodiments of the present invention are described below in detail, the example of the embodiment is shown in the drawings, wherein from beginning
Same or similar element or element with same or like function are represented to same or similar label eventually.Below by ginseng
The embodiment for examining accompanying drawing description is exemplary, is only used for explaining the present invention, and is not construed as limiting the claims.
Structure chart based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network is as shown in figure 1, model point
For two stages, that is, the BP neural network group for setting up initial BP neural network and rolling.The initial neutral net bag of first stage
Include two steps, i.e. AE unsupervised learnings and BP neural network supervised learning.Second stage builds the BP neural network group rolled,
fn-1(n>=2) simulation output of model will be used as fnThe part input of model, fn-1After the completion of training, by fn-1Network parameter
It is used as fnInitial network parameter, because the predicted time interval of continuous two models is shorter (15 minutes), network parameter difference
It is smaller, and BP neural network has stronger reverse fine-tuning capability, therefore f2~fnBP neural network is used, further reduces pre-
Survey the error of model.
The function model of Greenhouse grape is set up according to indoor and outdoor factor of influence.If current time is t, t+1 is continuously predicted
In the data at~t+n moment, following formula, f1Initial BP neural network is represented, is trained using the True Data of t, network
It is output as the inside greenhouse humiture at t+1 moment;fnRepresent to roll the n-th (n in BP neural network>=2) individual model, using t+
The external environment parameters and f at n-1 momentn-1The model indoor temperature and humidity analogue value is trained, and network is output as the temperature at t+n moment
Chamber interior humiture.
Trained after n network, preserved the weight and threshold parameter of each BP neural network, then in carrying out it is long-term roll it is pre-
Survey.Prediction process is by fn-1Predict the outcome as fnPart input carry out continuous rolling forecast.
Wherein, P represents external environment parameters and inside greenhouse equipment state, includes [Tout, Hout, Ws, Sr, Fs, Vs]
In any number of parameters, (P)tFor t ambient parameter, (Tin)tFor t greenhouse observed temperature, (Hin)tFor t temperature
Humidity is surveyed in room,For the t+1 moment greenhouse temperatures of prediction,For the t+1 moment chamber humidities of prediction;(P)t+n-1
For the ambient parameter at t+n-1 moment,For the t+n-1 moment greenhouse temperatures of prediction,For prediction t+n-1 when
Carve chamber humidity,For the t+n moment greenhouse temperatures of prediction,For the t+n moment chamber humidities of prediction.
The detailed process of BP neural network Greenhouse grape medium- and long-term forecasting is rolled for implementation below.
1st, initial BP neural network is built
Initial BP neural network predicts the outcome as the part input of second model and just with network parameter
Beginning network parameter, and then influence to roll predicting the outcome for BP models entirety.To improve precision of prediction, initial BP neural network is first
Unsupervised learning is carried out using unsupervised learning model AE, the feature of data is extracted;Then it regard AE feature representation as BP god
Initiation parameter through network, then have the target of supervision to learn, and using improved localized particle group optimizing method come excellent
Change the network weight and threshold value.Because PSO algorithm has easily precocious, stability, therefore the present invention is proposed
A kind of improved localized particle group optimizing method (IPSO).Test set is finally inputted to checking network in the model trained and completed
Generalization ability.
(1) the BP neural network initial parameter optimization based on AE
Three layers of autocoder network are initially set up, input vector is equal with output vector each element.Input layer is with hiding
Weight and threshold value { W between layer(1),b(1)It is encoder, coding function uses sigmoid functions;Hidden layer and output layer it
Between weight and threshold value { W(2),b(2)It is decoder, decoding function uses tanh functions, then had:
Hi=sigmoid (W(1)Xi+b(1))
Yi=tanh (W(2)Hi+b(2))
AE is that a kind of unsupervised learning model, i.e. training data are no labels, is output as the reconstruct of input, by calculating
Reconstructed error obtains AE weight parameter, obtains the feature representation of input data.Shown in reconstructed error function J (θ) following formula.
M is the quantity of training sample in formula, and n is the network number of plies, and θ is the parameter of neutral net, including weight and bias term.
Section 1 is the mean square deviation between model output valve and desired value in braces, and Section 2 L2 is regular terms, to reduce weight
Amplitude of variation, it is to avoid over-fitting.
(2) the BP neural network parameter optimization based on IPSO algorithms
Traditional BP neural network is trained using gradient descent method, and Local uniqueness is stronger, but is easily absorbed in local optimum
Point.And because network parameter is more, the dimension of particle is higher, the performance of PSO algorithms can be with certain particle populations quantity
The increase of optimised problem dimension and reduce, in order to not increase the complexity of algorithm and improve precision, the present invention is proposed
IPSO algorithms, including:
(A) localized particle group optimizing method is taken.Population is divided into by multiple subgroups, the speed of particle by parallel algorithm
Updated based on individual optimal and subgroup global optimum, to strengthen ability of searching optimum, while improving the efficiency of algorithm.Due to nerve
Network structure is complex, and dimensionality of particle is higher, therefore particle rapidity more new formula is based on individual extreme value by the method for the present invention
Central point and global extremum.Individual extreme value central point is pavg=[pavg1,pavg2,…,pavgD], whereinChange
The particle group velocity entered updates formula and is shown below.
vi(t+1)=ω vi(t)+c1r1(pavg-xi(t))+c2r2(pLg-xi(t))
Wherein, pLgFor the global optimum position of each subgroup.
(B) crossover operator of genetic algorithm is introduced.Crossover operation is performed to particle position, to increase population diversity,
Avoid algorithm Premature Convergence.Network parameter is divided into two parts during intersection, Part I is neural network input layer to hiding
Parameter { the W of layer(1),b(1), Part II is parameter { W of the neutral net hidden layer to output layer(2),b(2)}.If crossover probability
For Pc, the individual x of first subgroupi=[xi1,xi2,…,xiD] with Pc probability and the global optimum position of first subgroup
pLg1Part II parameter is intersected;The individual x of second subgroupj=[xj1,xj2,…,xjD] with Pc probability and second subgroup
Global optimum position pLg2Part I parameter is intersected, and formula is as follows.
xik=pLg1k
xjl=pLg2l
Wherein, i, j=1,2 ..., N/2 and i ≠ j, k ∈ [(IN+1) * HN+1, D], l ∈ [1, (IN+1) * HN], IN is god
Input layer number through network, HN is hidden layer neuron number, and D is the weight of the dimension, i.e. neutral net of particle
With threshold parameter number sum, if output layer neuron number is ON, then D=IN*HN+HN*ON+HN+ON.
(C) mutation operator is introduced.If the adaptive value of some particle is repeatedly averagely fitted less than colony during Evolution of Population
It should be worth, then show that the Evolutionary direction of particle much deviates optimal solution, no longer adapt to current search environment, therefore introduce something lost
The mutation operator of propagation algorithm performs mutation operation to the particle, jumps out the particle for being absorbed in local value and continually looks for optimal solution,
Other particles, which then maintain the original state, to be continued to evolve, until convergence.Shown in variation mode following formula, i.e., change particle by initialization mode
Position and speed.
xid=a+ (b-a) * rand
vid=m+ (n-m) * rand
Wherein, d=1,2 ..., D, a and b be limit particle minimum and maximum position, namely neural network parameter model
Enclose;M and n are the minimum and maximum speed for limiting particle, determine the amplitude of particle seat change;Rand is [between 0,1)
Uniform random number.
IPSO algorithm flow charts are as shown in Figure 2.Specific algorithm flow is as follows:
Step 1:Population is divided into two subgroups, calculated simultaneously in spmd parallel organizations, initialization kind group velocity and position
Put, initialize learning rate c1, c2, the parameter such as inertia weight ω;
Step 2:The each sub-average number of times of particle clear 0 of statistics, i.e. badNum [N]=0;Global optimum is assigned to
Subgroup global optimum PLg=Pg;
Step 3:The position of more new particle and speed;
Step 4:Introduce crossover operator.If the random number produced is less than crossover probability Pc, two subgroups perform intersection respectively
Operation;
Step 5:Calculate the fitness J (i) of each particle;
Step 6:Local optimum Pi is updated, if particle position is better than original particle after updating, by new particle
Position as the particle Pi;
Step 7:1. subgroup global optimum is updated, if particle is better than original subgroup global optimum position after location updating
Put, then regard the position of the particle as the global optimum P in current iterationLg;2. more new individual extreme value center point Pavg。
Step 8:Calculate each subgroup average fitness fit_avg;
Step 9:Introduce mutation operator:If J (i)<Fit_avg, then make badNum (i)+1;If badNum (i)>=
BadNumLimit, the then position of random initializtion particle and speed.
Step 10:Reach after inner iterative number of times, mutually, that is, it is optimal to compare subgroup for two sub- flock-mates, thus obtain it is global most
It is excellent;Not up to then jump procedure 3;
Step 11:Reach maximum iteration or meet gbest (n)-gbest (n-4)<=0.0001 (i.e. fitness letter
Several continuous 5 times are constant) it is to stop iteration.Step 2 is circulated back to, until meeting end condition.
2nd, the BP neural network group rolled is built
It is to set up the BP neural network group rolled to roll BP model second stage, i.e., continuously set up n Single-step Prediction model,
Each network model has corresponding training set train_xn(n>=2) and test set test_xn, train_xnAnd test_xnGeneration
The data at table t+n-1 moment, are included (P)t+n-1、WithThree parameters.The training of two neighboring network model
A moment is only differed between collection and test set, data also postpone takes a moment downwards.train_ynAnd test_ynRespectively
N-th of network training collection train_xnWith test set test_xnAnalog result, comprisingWithTwo parameters.
N-th of BP neural network study and prediction process are as shown in figure 3, using t+n-1 (n>=2) moment training set
train_xnTrain fnModel, the training set includes the greenhouse internal and external environment influence factor (P) at t+n-1 momentt+n-1, and fn-1
The analog result train_y of the training set of modeln-1.Network is output as the measured data at t+n moment, is instructed using gradient descent method
Practice network.By training set train_x after the completion of trainingnF is inputted againnIn model, the analog result train_ of training set is obtained
yn, i.e. the inside greenhouse humiture analog result collection at t+n moment will be used as train_xn+1A part be used for train fn+1Mould
Type.Then by t+n-1 moment test sets test_xnT+n moment indoor temperature and humidities are obtained in input model to predict the outcome test_yn,
And it is used as test_xn+1A part, the indoor temperature and humidity for predicting the t+n+1 moment.So roll training and predict, realize
The medium- and long-term forecasting of Greenhouse grape.The purpose for training multiple networks is in order that training set is consistent with test set source, to improve
The precision of forecast model, i.e. fn(n>=indoor temperature and humidity the data 2) in the training sample and test sample of model are all from fn-1
The analog result of model.
Described above is only some embodiments of the present invention, it is noted that for the ordinary skill people of the art
For member, under the premise without departing from the principles of the invention, some improvement can also be made, these improvement should be regarded as the guarantor of the present invention
Protect scope.
Claims (8)
1. based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, it is characterised in that comprise the following steps:
Step 1:Set up initial BP neural network f1If current time is t, the inside greenhouse humiture of t is inputted, output is pre-
The inside greenhouse humiture at the t+1 moment of survey, and obtain f1Network parameter;
Step 2:Set up the BP neural network group rolled, including n-1 neutral net fn, each neutral net fnInclude training set
train_XnWith test set test_Xn, a moment is separated by between the training set and test set of two neighboring neutral net, its
In, train_XnRepresent the training set at t+n-1 moment, test_XnRepresent the test set at t+n-1 moment, n >=2;
Step 3:Utilize train_XnF is trained with network parameter combination gradient descent methodnModel, after the completion of training, then by train_
XnIt is input to fnIn model, analog result train_Y is exportedn;By test_XnIt is input to fnIn model, the test_ that predicts the outcome is exported
Yn;
Step 4:N=n+1 is made, step 3 is gone to.
2. it is according to claim 1 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, step 1 is specifically included:
Step 1-1:Pre-training is carried out to the inside greenhouse humiture of t based on unsupervised learning model, input data is extracted
Feature, and reconstruct after export;
Step 1-2:Using the feature of data as the initiation parameter of BP neural network, have the target of supervision to learn, use
The weight and threshold parameter of the improved localized particle group optimizing method combination genetic algorithm optimization BP neural network;
Step 1-3:Initial BP neural network f is set up using optimal weights and threshold parameter1, export the greenhouse at the t+1 moment of prediction
Internal humiture.
3. it is according to claim 2 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, described in step 1-1 is specially to the method that input data is reconstructed:By the weight and threshold value between input layer and hidden layer
{W(1),b(1)As encoder, coding function uses sigmoid functions;By the weight and threshold value between hidden layer and output layer
{W(2),b(2)As decoder, decoding function uses tanh functions.
4. it is according to claim 2 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, step 1-2's concretely comprises the following steps:
Step 1-2-1:Population is divided into two subgroups, calculated simultaneously in spmd parallel organizations, initialization kind group velocity
With position, learning rate C1And C2, inertia weight;
Step 1-2-2:The sub-average particle number of times of number of times is reset, global optimum is assigned to subgroup global optimum, i.e.,
BadNum [N]=0, PLg=Pg, badNum is particle number of times, and N is the particle numbering that number of times is below the average, PLgIt is global for subgroup
It is optimal, PgFor global optimum;
Step 1-2-3:The speed of more new particle and position:
vi(t+1)=ω vi(t)+c1r1(pavg-xi(t))+c2r2(pLg-xi(t)), xi(t+1)=xi(t)+vi(t+1), wherein, i
=1,2 ..., N, t are current iteration number of times, and ω is inertia weight, and c1, c2 are accelerated factors, r1, r2 be [0,1] it is interval with
Machine number, vi(t) it is the former speed of particle, vi(t+1) it is the particle rapidity after updating, pavgFor individual extreme value central point, pLgTo be each
The global optimum position of subgroup, xi(t) put for particles in-situ, xi(t+1) it is the particle position after updating;
Step 1-2-4:Crossover operator is introduced, if the random number produced is less than crossover probability PC, then two subgroups, which are performed, intersects behaviour
Make:xik=pLg1k, xjl=pLg2l, wherein, xikElement, p are tieed up for the kth of i-th of particle position in first subgroupLg1kFor first
The kth dimension element of individual subgroup global optimum position, xjlElement, p are tieed up for the l of j-th of particle position in second subgroupLg2lFor
The l dimension elements of second subgroup global optimum position, i, j=1,2 ..., N/2 and i ≠ j, k ∈ [(IN+1) * HN+1, D], l
∈ [1, (IN+1) * HN], IN are the input layer number of neutral net, and HN is hidden layer neuron number, and D is particle
Dimension, and the fitness J (i) of each particle is calculated, if random number is more than crossover probability PC, then without any operation;
Step 1-2-5:Update local optimum PiIf the particle position after updating is better than original particle position, by new particle position
Put the P as the particlei, and it is used as the global optimum P in current iterationLg, update individual mechanism center point Pavg, calculate each
Subgroup average fitness fit_avg, if the particle position after updating is not better than original particle position, without any behaviour
Make;
Step 1-2-6:Mutation operator is introduced, if J (i) < fit_avg, make badNum (i)+1, if badNum (i) >=
BadNumLimit, the then position of random initializtion particle and speed:xid=a+ (b-a) * rand, vid=m+ (n-m) * rand,
Wherein d=1,2 ..., D, a and b are the minimum and maximum positions for limiting particle, and m and n are the minimum and maximum speed for limiting particle
Degree, rand for [0,1) between uniform random number;
Step 1-2-7:Judge whether to reach default inner iterative number of times, if so, the subgroup for then comparing two subgroups is optimal, obtain
Global optimum, if it is not, then going to step 1-2-3;
Step 1-2-8:Judge whether to reach maximum iteration or meet gbest (n)-gbest (n-4)<=0.0001, if
It is then to stop iteration, if it is not, then brick step 1-2-2.
5. it is according to claim 1 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, the predictor formula of step 1 is:
<mrow>
<mo>&lsqb;</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>T</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>H</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>&rsqb;</mo>
<mo>=</mo>
<msub>
<mi>f</mi>
<mn>1</mn>
</msub>
<mo>&lsqb;</mo>
<msub>
<mrow>
<mo>(</mo>
<mi>P</mi>
<mo>)</mo>
</mrow>
<mi>t</mi>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<msub>
<mi>T</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<mi>t</mi>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<msub>
<mi>H</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<mi>t</mi>
</msub>
<mo>&rsqb;</mo>
</mrow>
Wherein, (P)tFor t ambient parameter, (Tin)tFor t greenhouse observed temperature, (Hin)tSurveyed for t greenhouse wet
Degree,For the t+1 moment greenhouse temperatures of prediction,For the t+1 moment chamber humidities of prediction.
6. it is according to claim 1 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, in step 2 a moment is 15min.
7. it is according to claim 1 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, training set train_XnGreenhouse external environment influence factor (P) including the t+n-1 momentt+n-1With neutral net fn-1Instruction
Practice collection analog result train_Yn-1, wherein train_Yn-1The inside greenhouse humiture at the t+n-1 moment including prediction.
8. it is according to claim 1 based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network, its feature
It is, the predictor formula of the BP neural network of rolling is:
<mrow>
<mo>&lsqb;</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>T</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>H</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>&rsqb;</mo>
<mo>=</mo>
<msub>
<mi>f</mi>
<mi>n</mi>
</msub>
<mo>&lsqb;</mo>
<msub>
<mrow>
<mo>(</mo>
<mi>P</mi>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>T</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mrow>
<mo>(</mo>
<mover>
<msub>
<mi>H</mi>
<mrow>
<mi>i</mi>
<mi>n</mi>
</mrow>
</msub>
<mo>^</mo>
</mover>
<mo>)</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>+</mo>
<mi>n</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>&rsqb;</mo>
</mrow>
Wherein, (P)t+n-1For the ambient parameter at t+n-1 moment,For the t+n-1 moment greenhouse temperatures of prediction,
For the t+n-1 moment chamber humidities of prediction,For the t+n moment greenhouse temperatures of prediction,For the t+n moment of prediction
Chamber humidity.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710426571.5A CN107180261B (en) | 2017-06-08 | 2017-06-08 | Greenhouse microclimate medium-long term prediction method based on rolling BP neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710426571.5A CN107180261B (en) | 2017-06-08 | 2017-06-08 | Greenhouse microclimate medium-long term prediction method based on rolling BP neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107180261A true CN107180261A (en) | 2017-09-19 |
CN107180261B CN107180261B (en) | 2020-03-17 |
Family
ID=59835322
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710426571.5A Active CN107180261B (en) | 2017-06-08 | 2017-06-08 | Greenhouse microclimate medium-long term prediction method based on rolling BP neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107180261B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909149A (en) * | 2017-10-26 | 2018-04-13 | 西北农林科技大学 | A kind of Temperature in Greenhouse Forecasting Methodology based on Genetic BP Neutral Network |
CN110414045A (en) * | 2019-06-18 | 2019-11-05 | 东华大学 | Short-term wind speed forecasting method based on VMD-GRU |
CN115358475A (en) * | 2022-08-29 | 2022-11-18 | 河南农业大学 | Disaster prediction method and system based on support vector machine and gray BP neural network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0980034A1 (en) * | 1998-08-13 | 2000-02-16 | C.S.E.M. Centre Suisse D'electronique Et De Microtechnique Sa | Building heating control system |
CN103105246A (en) * | 2012-12-31 | 2013-05-15 | 北京京鹏环球科技股份有限公司 | Greenhouse environment forecasting feedback method of back propagation (BP) neural network based on improvement of genetic algorithm |
CN103235620A (en) * | 2013-04-19 | 2013-08-07 | 河北农业大学 | Greenhouse environment intelligent control method based on global variable prediction model |
CN104715282A (en) * | 2015-02-13 | 2015-06-17 | 浙江工业大学 | Data prediction method based on improved PSO-BP neural network |
-
2017
- 2017-06-08 CN CN201710426571.5A patent/CN107180261B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0980034A1 (en) * | 1998-08-13 | 2000-02-16 | C.S.E.M. Centre Suisse D'electronique Et De Microtechnique Sa | Building heating control system |
CN103105246A (en) * | 2012-12-31 | 2013-05-15 | 北京京鹏环球科技股份有限公司 | Greenhouse environment forecasting feedback method of back propagation (BP) neural network based on improvement of genetic algorithm |
CN103235620A (en) * | 2013-04-19 | 2013-08-07 | 河北农业大学 | Greenhouse environment intelligent control method based on global variable prediction model |
CN104715282A (en) * | 2015-02-13 | 2015-06-17 | 浙江工业大学 | Data prediction method based on improved PSO-BP neural network |
Non-Patent Citations (4)
Title |
---|
朱春侠等: "BP神经网络在日光温室湿度预测中的应用", 《农机化研究》 * |
杜世强: "一种新的优化神经网络权值算法及其应用", 《西北民族大学学报(自然科学版)》 * |
杨德平等: "《经济预测方法及MATLAB实现》", 31 March 2012, 北京:机械工业出版社 * |
谷悦等: "基于BP神经网络群结构的风电场短期风速预测", 《农村电气化》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909149A (en) * | 2017-10-26 | 2018-04-13 | 西北农林科技大学 | A kind of Temperature in Greenhouse Forecasting Methodology based on Genetic BP Neutral Network |
CN110414045A (en) * | 2019-06-18 | 2019-11-05 | 东华大学 | Short-term wind speed forecasting method based on VMD-GRU |
CN110414045B (en) * | 2019-06-18 | 2023-08-11 | 东华大学 | Short-term wind speed prediction method based on VMD-GRU |
CN115358475A (en) * | 2022-08-29 | 2022-11-18 | 河南农业大学 | Disaster prediction method and system based on support vector machine and gray BP neural network |
Also Published As
Publication number | Publication date |
---|---|
CN107180261B (en) | 2020-03-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110458443A (en) | A kind of wisdom home energy management method and system based on deeply study | |
CN105913151A (en) | Photovoltaic power station power generation amount predication method based on adaptive mutation particle swarm and BP network | |
CN109146121A (en) | The power predicating method stopped in the case of limited production based on PSO-BP model | |
CN112733462B (en) | Ultra-short-term wind power plant power prediction method combining meteorological factors | |
CN108448610A (en) | A kind of short-term wind power prediction method based on deep learning | |
CN104636801A (en) | Transmission line audible noise prediction method based on BP neural network optimization | |
CN101315544A (en) | Greenhouse intelligent control method | |
CN104636985A (en) | Method for predicting radio disturbance of electric transmission line by using improved BP (back propagation) neural network | |
CN110222883A (en) | Load Prediction In Power Systems method based on wind Drive Optimization BP neural network | |
CN103793887B (en) | Short-term electric load on-line prediction method based on self-adaptive enhancement algorithm | |
CN103235620A (en) | Greenhouse environment intelligent control method based on global variable prediction model | |
CN107180261A (en) | Based on the Greenhouse grape medium- and long-term forecasting method for rolling BP neural network | |
CN105447509A (en) | Short-term power prediction method for photovoltaic power generation system | |
CN107466816A (en) | A kind of irrigation method based on dynamic multilayer extreme learning machine | |
Yue et al. | The prediction of greenhouse temperature and humidity based on LM-RBF network | |
CN106447133A (en) | Short-term electric load prediction method based on deep self-encoding network | |
CN107121926A (en) | A kind of industrial robot Reliability Modeling based on deep learning | |
CN110276472A (en) | A kind of offshore wind farm power ultra-short term prediction method based on LSTM deep learning network | |
CN108399470A (en) | A kind of indoor PM2.5 prediction techniques based on more example genetic neural networks | |
CN105160441A (en) | Real-time power load forecasting method based on integrated network of incremental transfinite vector regression machine | |
CN107400935A (en) | Adjusting method based on the melt-spinning technology for improving ELM | |
CN116522795A (en) | Comprehensive energy system simulation method and system based on digital twin model | |
CN113705922B (en) | Improved ultra-short-term wind power prediction algorithm and model building method | |
Tao et al. | On comparing six optimization algorithms for network-based wind speed forecasting | |
CN111401659A (en) | Ultra-short-term or short-term photovoltaic power generation power prediction method based on case reasoning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20230327 Address after: 201906 8312, building 8, No. 4361, Hutai Road, Baoshan District, Shanghai Patentee after: SHANGHAI LANCHANG AUTOMATION TECHNOLOGY CO.,LTD. Address before: Weigang Xuanwu District of Nanjing Jiangsu province 210095 No. 1 Patentee before: NANJING AGRICULTURAL University |
|
TR01 | Transfer of patent right |