CN106779148B - A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion - Google Patents
A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion Download PDFInfo
- Publication number
- CN106779148B CN106779148B CN201611020632.XA CN201611020632A CN106779148B CN 106779148 B CN106779148 B CN 106779148B CN 201611020632 A CN201611020632 A CN 201611020632A CN 106779148 B CN106779148 B CN 106779148B
- Authority
- CN
- China
- Prior art keywords
- measuring station
- air measuring
- data
- target air
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000004927 fusion Effects 0.000 title claims abstract description 9
- 238000001914 filtration Methods 0.000 claims abstract description 12
- 238000012545 processing Methods 0.000 claims abstract description 7
- 230000007935 neutral effect Effects 0.000 claims description 89
- 238000013528 artificial neural network Methods 0.000 claims description 53
- 238000004422 calculation algorithm Methods 0.000 claims description 48
- 238000012549 training Methods 0.000 claims description 46
- 238000000354 decomposition reaction Methods 0.000 claims description 34
- 210000002364 input neuron Anatomy 0.000 claims description 34
- 241001269238 Data Species 0.000 claims description 19
- 238000003062 neural network model Methods 0.000 claims description 19
- 239000002245 particle Substances 0.000 claims description 19
- 238000010276 construction Methods 0.000 claims description 14
- 210000004205 output neuron Anatomy 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000007689 inspection Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 210000004218 nerve net Anatomy 0.000 claims description 4
- 238000007476 Maximum Likelihood Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 230000015572 biosynthetic process Effects 0.000 claims description 2
- 241001608638 Nycteribiinae Species 0.000 claims 1
- 230000002452 interceptive effect Effects 0.000 abstract 1
- 238000005457 optimization Methods 0.000 description 25
- 210000002569 neuron Anatomy 0.000 description 12
- 238000013277 forecasting method Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 230000035772 mutation Effects 0.000 description 4
- 210000005036 nerve Anatomy 0.000 description 4
- 239000003016 pheromone Substances 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000000053 physical method Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 235000008694 Humulus lupulus Nutrition 0.000 description 1
- 241000406668 Loxodonta cyclotis Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000009022 nonlinear effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
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
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- 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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/40—Business processes related to the transportation industry
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Business, Economics & Management (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- Medical Informatics (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Development Economics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Primary Health Care (AREA)
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
Abstract
The invention discloses a kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion, this method comprises the following steps:1. 5 auxiliary air measuring stations are installed around the air measuring station position;2. original air speed data is handled with Interactive Multiple-Model Kalman filtering method;Step 3:Small echo processing is carried out using filtered data, and prediction submodel is built to the low-frequency data after small echo processing;Step 4:The advanced multi-step prediction value of target air measuring station and weather forecast target air measuring station wind speed value input BAYESIAN combined model that the advanced multi-step Predictive Model of extraterrestrial target air measuring station, self advanced multi-step Predictive Model of target air measuring station and the meteorological advanced multi-step Predictive Model of target air measuring station are obtained, obtain final target air measuring station predicted value;The present invention can not only be avoided that the data outage that single air measuring station hardware fault is caused, and the longer emergency processing time can be provided to the high-speed railway safe driving under bad wind environment.
Description
Technical field
The invention belongs to railway forecasting wind speed field, more particularly to a kind of line of high-speed railway of multi-model multiple features fusion
Wind speed forecasting method.
Background technology
With the economic sustainable and stable development of China, railway construction in China enters high-speed development period.And with railway operation
The increase of circuit and the lifting of train speed, security, stationarity, the comfortableness of train operation are received more and more attention.
High wind is to cause one of Major Natural Disasters of train accident, and high wind weather occurs often in some Along Railways of China area, this
Huge challenge is brought to the operation of train safety and steady.In order to prevent the generation of train accident, it is necessary to set up railway gale monitoring
Early warning system, so as to railway department forward scheduling commander, Along Railway forecasting wind speed technology be exactly the system core technology it
One.
Wind velocity signal is a kind of random, nonlinear properties, is predicted very difficult.At present, the research field of forecasting wind speed is more
Concentrate on wind-powered electricity generation field prediction, method is divided into statistical method, physical method and learning method, conventional model include neutral net,
SVMs, Kalman filtering, time series, wavelet decomposition, empirical mode decomposition etc., also have some mixed models.
It is little that wind power plant generally build wind speed difference in the area that physical features is flat, wind direction is stable, smaller range in.Different from wind
Electric field, Along Railway terrain environment is complicated, diverse location point wind speed significant difference, and sufficiently strong transient state wind may trigger row
Car accident, therefore, Along Railway each position point forecasting wind speed must be accurate.Simultaneously as hardware device exist integrity problem,
Single forecast result of model is unstable, and Along Railway forecasting wind speed data are impermissible for interrupt output, so, Along Railway wind speed
The necessary performance of prediction is stable, uninterruptedly can export high-precision forecast data under a variety of unusual conditions.Therefore, in the urgent need to setting up
A kind of precision of prediction is high, performance is stable, can incorporate the Along Railway wind speed forecasting method of many factors and a variety of forecast models.
The content of the invention
It is an object of the invention to overcome deficiency present in existing Along Railway wind speed forecasting method, there is provided a kind of multimode
The Along Railway wind speed forecasting method of type multiple features fusion.This method has merged 4 basic models, includes space, time, gas
As multiple elements such as, physics, during prediction, each model data, which exists, interweaves, and last built-up pattern can be adaptively adjusted 4 bases
The weights of this forecast model, prediction stability is high, it is possible to achieve advanced multi-step prediction, with engineering application value.
A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion, comprises the following steps:
Step 1:N number of auxiliary air measuring station is at least installed around target air measuring station position, adopted in real time using air measuring station is aided in
The air speed data of collection auxiliary air measuring station, obtains the wind speed sample set of target air measuring station and auxiliary air measuring station;
Wherein, N is the integer more than or equal to 5;
Step 2:Auxiliary air measuring station data and target air measuring station data are filtered and 1 layer depth wavelet decomposition successively,
Extract low-frequency data part;
Step 3:The auxiliary air measuring station obtained using step 2 and the low-frequency data part of target air measuring station build space-mesh
Mark the advanced multi-step Predictive Model of air measuring station, meanwhile, using the low-frequency data part of target air measuring station build self-target air measuring station
Advanced multi-step Predictive Model;
Step 4:Low-frequency data part and air pressure, humidity and the temperature of target air measuring station present position using target air measuring station
Degrees of data builds meteorology-advanced multi-step Predictive Model of target air measuring station;
Step 5:Utilize numerical weather forecast acquisition of information target air measuring station wind speed value;
Step 6:By space-advanced multi-step Predictive Model of target air measuring station, self-the advanced multi-step prediction mould of target air measuring station
The advanced multi-step prediction value of target air measuring station and step 5 that type and the advanced multi-step Predictive Model of meteorology-target air measuring station are obtained are obtained
The target air measuring station wind speed value input BAYESIAN combined model obtained, obtains final target air measuring station predicted value;
The advanced multi-step prediction is referred under being obtained using current time T air speed data input correspondence forecast model for the moment
T+1 forecasting wind speed is carved, then, correspondence forecast model is inputted again using subsequent time T+1 wind speed value, obtains T+2
The wind speed value at moment, reciprocal iteration obtains advanced multi-step prediction value.
Advanced multi-step prediction value is obtained using BAYESIAN combined model to 4 forecast models without simple equality to add
Power, but the forecast model output to higher precision assigns bigger proportionality coefficient, that is, allows the forecast model institute of higher precision defeated
The wind speed value gone out occupies greater proportion in weighting.
Further, the space in the step 2-advanced multi-step Predictive Model of target air measuring station and self-target survey wind
Stand advanced multi-step Predictive Model construction step it is as follows:
Step 2.1:FEEMD decomposition is carried out to the low-frequency data part of auxiliary air measuring station and target air measuring station, obtains each
The corresponding low frequency of air measuring station, intermediate frequency and high frequency subsequence;
Step 2.2:Each IMF components of auxiliary air measuring station and the high frequency subsequence of target air measuring station are based on respectively
The training of PSO LVQ neutral nets, builds the LVQ god of auxiliary air measuring station and self respective high frequency of target air measuring station based on PSO
Through network model, the advanced multi-step prediction value of self high frequency of auxiliary air measuring station and target air measuring station is obtained;
The limit based on CBA is set up respectively to each IMF components of auxiliary air measuring station and the intermediate frequency subsequence of target air measuring station
Learning machine is trained, and builds the auxiliary extreme learning machine mould of air measuring station and self respective intermediate frequency of target air measuring station based on CBA
Type, obtains the advanced multi-step prediction value of self intermediate frequency of auxiliary air measuring station and target air measuring station;
Each IMF components of low frequency subsequence to aiding in air measuring station and target air measuring station are set up RARIMA and instructed respectively
Practice, build auxiliary air measuring station and self respective low frequency RARIMA model of target air measuring station, obtain auxiliary air measuring station and mesh
Mark the advanced multi-step prediction value of self low frequency of air measuring station;
Step 2.3:By LVQ neural network model, the intermediate frequency limit based on CBA of the high frequency of target air measuring station based on PSO
Learning machine model and low frequency RARIMA model combinations, formed self-the advanced multi-step Predictive Model of target air measuring station;
Step 2.4:To aid in the low frequency, intermediate frequency and high frequency subsequence data of air measuring station as input, target air measuring station
Low frequency, intermediate frequency and high frequency subsequence data build each BP neural network mould of the frequency range subsequence based on Adaboost as output
Type, is based on using aiding in the corresponding advanced multi-step prediction value of each frequency range subsequence of air measuring station to input each frequency range subsequence
Adaboost BP neural network model, obtains the advanced multi-step prediction value of each frequency range of target air measuring station correspondence, with each frequency range
BP neural network model formation space-target air measuring station advanced multi-step Predictive Model of the sequence based on Adaboost;
The BP neural network model of each frequency range subsequence based on Adaboost refers to, using Adaboost algorithm to many
The mean absolute error for the predicted value that individual BP neural network is exported according to each BP neural network carries out not decile weighted sum, each
BP neural network is using 3 layer network structures of multiple input neurons and 1 output neuron, the input data for training
In, at the acquisition time previous moment super than previous group input data of every group of data, the acquisition time of output data is than last
One group of input data super previous moment.
In the mixed model, Adaboost BP neural network model includes 10 BP neural networks as weak study
Device, the BP neural network with the mean absolute relative error of forecasting wind speed more than 5% is considered as the object for needing to strengthen study, BP god
Greatest iteration step number through network takes 300 steps.The prediction of 10 BP neural networks by Adaboost algorithm to being included is defeated
Go out and carried out not decile weighted sum, i.e. Adaboost algorithm in an iterative process, it is defeated according to the prediction of each BP neural network
The mean absolute relative error gone out, adjusts the BP neural network in next iteration meter in whole neural network inverse systems in real time
Shared weight ratio.The better single BP neural network of performance will be endowed higher weights.
Adaboost-BP Neural network mixed models are divided into training study and prediction calculates two aspects.In training study
Stage, each BP neural network is administrative by the low frequency, intermediate frequency and high frequency subsequence of the auxiliary air measuring station inputted in being screened
Whole IMF components, each BP neural network is output as the low frequency, intermediate frequency and the administrative whole of high frequency subsequence of target air measuring station
IMF components, each BP neural network learns the training of itself is completed, then by Adaboost algorithm to whole BP nerves
The weight ratio of network is adjusted, and to the last obtains the optimal weighting ratio value of this 10 BP neural networks.In the present invention
In, in each BP neural network itself neuron connection weight and hidden layer threshold value be randomly assigned.
Single BP neural network uses 3 layer network structures of multiple input neurons and 1 output neuron, i.e.,:It is single
Individual BP neural network needs that synchronously network is learnt and trained with multigroup input air speed data and 1 group of output wind speed data;
After 10 BP neural networks that Adaboost algorithm optimizes all complete training, as long as multiple inputs to each BP neural network
Neuron port inputs the air speed data of multigroup equal length, and each BP neural network can just export 1 group automatically with inputting wind speed etc.
The prediction of wind speed sequence of length, then carries out not decile using BP neural network weight ratio value determined by Adaboost algorithm
Weighted sum, obtains the final prediction of wind speed sequence of Adaboost-BP Neural network mixed models.
After Adaboost-BP Neural network mixed models complete global learning training, it is possible to wait and utilize newly defeated
Enter the forecasting wind speed output data that air speed data obtains equal length.Namely each BP neural network input trained is obtained
Each frequency layer obtained is used after the advanced multi-step prediction value of auxiliary air measuring station that various combination forecast model is obtained, the BP nerve nets
Network can just predict the advanced multi-step prediction value of target air measuring station.When the advanced multistep for completing all 10 BP neural networks is pre-
It is final to obtain target air measuring station after measured value output, it is possible to which the weight ratio provided with Adaboost algorithm is summed
Advanced multistep wind speed value.
Adaboost BP neural network model includes 10 BP neural networks as weak learner, with forecasting wind speed
BP neural network of the mean absolute relative error more than 5% is considered as the object for needing to strengthen study, and the maximum of BP neural network changes
Number of riding instead of walk takes 300 steps.
Further, the specific construction step of LVQ neural network model of self high frequency based on PSO is as follows:
(1) particle populations of the connection weight of LVQ models are randomly generated using PSO, each particle represents one group of LVQ model
Initial connection weight;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of PSO algorithms, is completed
The training learning process of PSO algorithms so that the particle in particle populations is constantly brought near optimal particle, is exported optimal
The initial connection weight of LVQ neutral nets;
(3) the obtained initial connection weight of LVQ neutral nets will be trained to LVQ god by PSO in above two steps
It is configured, input and output data using each IMF components of the high frequency subsequence at wind station as neutral net, completes through network
Study and training of the LVQ neutral nets itself to wind speed, form the PSO- of self high frequency subsequence air speed data high-precision forecast
LVQ mixed models;
For in the input data of training, when the acquisition time of every group of data is super than previous group input data previous
Carve, acquisition time than last group input data super previous moment of output data.
Utilize the LVQ neutral nets that set up PSO optimizes complete to each IMF wind speed component in wind speed high frequency subsequence
Calculated into advanced multi-step prediction, to obtain the advanced multi-step prediction value of the corresponding wind speed of respective IMF components.
The LVQ neutral nets set up use the structures of 3 input neurons and 1 output neuron, i.e.,:LVQ god
Need that synchronously network is learnt and trained with 3 groups of input air speed datas and 1 group of output wind speed data through network;When PSO is excellent
The LVQ neutral nets of change are completed after training, as long as inputting 3 corresponding wind to 3 input neuron ports of LVQ neutral nets
Fast data, LVQ neutral nets can just export 1 wind speed value automatically.
The initial population quantity generated using PSO sets iterations when taking 50, PSO optimum choice LVQ neutral nets
Take 200 times, particle speed of service maximum is 0.5, the minimum value of the particle speed of service is 0.01;The iteration of LVQ neutral nets
Target mean absolute relative error takes 5%;LVQ neutral nets obtain PSO assign the initial connection weight of optimal network after from
The maximum study iterative steps of body take 100 steps.
The essence that the LVQ neutral nets that each IMF components of high frequency subsequence set up PSO optimizations respectively are trained is to utilize
PSO selects the initial connection weight of LVQ neutral nets so that the LVQ neutral nets of optimization can preferably follow the trail of the sub- sequence of high frequency
The extreme mutation rule of each administrative IMF component air speed data of row, realizes the prediction of high-precision intermediate frequency subsequence.
Further, the construction step of extreme learning machine model of self intermediate frequency based on CBA is as follows:
(1) bat algorithm is initialized;
The location parameter of every bat represents the one group of output weights and hidden layer node threshold value of extreme learning machine, every bat
The initial loudness and pulse rate of bat take 0.5, and bat number 50, iterations is 100, and the flight step-length of every bat takes
It is [0.001,0.05] to be worth scope;
(2) fitness of the setting value as bat algorithm is reached using the output wind speed mean absolute relative error of extreme learning machine
Guidance function, the loudness and pulse rate of every bat of real-time update so that bat fly to extreme learning machine input weights and
The corresponding optimal solution of hidden layer node threshold value;
In preceding 50 iteration, the flight step-length of bat takes maximum, since the 51st iteration, the flight step-length of bat
Take minimum value;
(3) the input weights and the corresponding optimal solution of hidden layer node threshold value of the extreme learning machine of (2) acquisition are utilized, by wind
Input and output data of each IMF components of the intermediate frequency subsequence stood as extreme learning machine, carry out extreme learning machine model instruction
Practice, obtain the extreme learning machine model based on CBA;
For in the input data of training, when the acquisition time of every group of data is super than previous group input data previous
Carve, acquisition time than last group input data super previous moment of output data.
Utilize the extreme learning machine that set up CBA optimizes complete to each IMF wind speed component in wind speed intermediate frequency subsequence
Calculated into advanced multi-step prediction, to obtain the advanced multi-step prediction value of the corresponding wind speed of respective IMF components.
The extreme learning machine set up uses the structures of 3 input neurons and 1 output neuron, i.e.,:The limit
Habit machine needs that synchronously network is learnt and trained with 3 groups of input air speed datas and 1 group of output wind speed data;When CBA optimizations
Extreme learning machine complete after training, as long as inputting 3 corresponding wind speed numbers to 3 of extreme learning machine input neuron ports
According to extreme learning machine can just export 1 wind speed value automatically.
The iterative target mean absolute relative error of extreme learning machine model takes 5%, in CBA optimum choice extreme learning machines
Initial input weights and hidden layer node threshold stage and extreme learning machine assignment phase, pole after CBA algorithm optimizing is obtained
The greatest iteration step number of limit learning machine takes 200 steps.
It is profit to set up the essence that the extreme learning machines of CBA optimizations is trained respectively to each IMF components of intermediate frequency subsequence
With the initial input weights and hidden layer node threshold value of variable-step self-adaptive bat algorithm optimization extreme learning machine so that optimization
Extreme learning machine can preferably follow the trail of the mutation rule of each administrative IMF component air speed data of intermediate frequency subsequence, realize high
The prediction of the intermediate frequency subsequence of precision.
The flight step-length of bat automatically changes with being incremented by for iterations, it is ensured that bat algorithm is to extreme learning machine
Ability of searching optimum on forecasting wind speed, makes to take maximum during the 50 step iteration before algorithm optimizing, it is therefore an objective to avoid calculating
Method too early be absorbed in locally optimal solution;With the increase (i.e. since 50 steps untill 100 steps) of iterations, the flight of bat
Step-length actively reduces, and minimum value is taken, so that bat algorithm is more accurately solved in later stage Fast Convergent.
Further, the construction step of self low frequency RARIMA model is as follows:
(1) whole IMF components wind speed numbers of acquisition are decomposed to each low frequency subsequence using nonparametric pleasure boat method of inspection
According to progress data stationarity inspection;
Such as run into certain IMF components air speed data and be presented non-stationary, then difference meter is carried out to this section of IMF components air speed data
Calculate untill it shows stationarity;
(2) acquisition is decomposed to each low frequency subsequence to whole IMF components wind speed after step 1 stationary test
Data carry out sample auto-correlation and sample partial correlation is calculated, and true according to the auto-correlation and partial correlation value of respective component air speed data
Determine the optimal type and Optimal order of RARIMA models;
(3) the optimal type and Optimal order of the RARIMA models obtained to step 2, are solved each using maximum-likelihood method
The equation coefficient of individual IMF components air speed data correspondence RARIMA models, using each IMF components of the low frequency subsequence at wind station as
The input data of RARIMA models, forms self low frequency RARIMA model.
During the prediction of advanced multistep, the equation of newest predicted value real-time update RARIMA models is constantly utilized
Coefficient.
RARIMA model equations are actually an auto-correlation expression formula, that is, describe current air speed value and history wind
The relation of speed value.Such as, RARIMA models are all set up to whole IMF components inside low frequency subsequence, are also to use each IMF
Component data fits the RARIMA models of different parameters.As described in following instance, certain IMF component has 500 data, just
Set up a RARIMA model.So being exactly to certain section it is believed that RARIMA models do not have so-called input and output data
IMF component datas set up an auto-correlation expression formula, so but need prediction when, just input the historical juncture data, obtain
The data at current time, by that analogy, until untill moment corresponding wind speed value is wanted in acquisition.
The reason for setting up RARIMA models to each IMF component of low frequency subsequence be:Relative to medium, high frequency subsequence institute
Each IMF component after decomposition, each administrative IMF component of low frequency subsequence is more steady.By being used before this patent
Unscented kalman filtering method and 1 layer of wavelet method denoising after, the air speed data of low frequency subsequence does not include wind speed number of hops
According to.It is wise that process of fitting treatment is now realized using the RARIMA models in statistical mathematics field, can accomplish fitting precision and
Export taking into account for real-time.
In the step content, the calculation procedure of the excellent wheel method of inspection of nonparametric, Difference Calculation and RARIMA models belongs to
Existing algorithm in statistical mathematics.
Further, the construction step of the meteorology-advanced multi-step Predictive Model of target air measuring station is as follows:
(1) population of the connection weight of Elman Model of Neural Network is randomly generated using ACO;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of ACO algorithms, is carried out
The training learning process of ACO algorithms so that the ant in initialization ant population is by iteration constantly close to optimal solution, output
The optimal initial connection weight of Elman neutral nets and hidden layer threshold value;
(3) the initial connection weight obtained and threshold value is trained to carry out Elman neutral nets by ACO above-mentioned steps
Set, carried out the second layer low frequency component data of air pressure, temperature, humidity and target air measuring station as Elman neutral nets
Input and output data to study and the training of wind speed, form meteorology-advanced multi-step Predictive Model of target air measuring station;
For in the input data of training, when the acquisition time of every group of data is super than previous group input data previous
Carve, acquisition time than last group input data super previous moment of output data.
The Elman neutral nets that set up ACO optimizes are utilized to the of air pressure, temperature, humidity and target air measuring station
This 4 kinds of objects of two layers of low frequency wind speed component data complete advanced multi-step predictions and calculated, corresponding advanced many to obtain various objects
Walk predicted value.
The Elman neutral nets set up use the structures of 4 input neurons and 1 output neuron, i.e.,:Elman
Neutral net needs that synchronously network is learnt and trained with the data and 1 group of output wind speed data of 4 groups of above-mentioned 4 kinds of objects;
After the Elman neutral nets that ACO optimizes complete training, as long as distinguishing 4 input neuron ports of Elman neutral nets
The data of 4 above-mentioned 4 kinds of objects are inputted, Elman neutral nets can just export 1 wind speed value automatically.
Initial population ant quantity takes 100, and the ant pheromones significance level factor takes 0.5, the important journey of ant heuristic function
The degree factor takes 1, and ant pheromones volatilization factor takes 0.1.Iterations during ACO optimum choice Elman neutral nets takes 50 times.
The iterative target mean absolute relative error of Elman neutral nets takes 5%.Elman neutral nets are obtaining the optimal of ACO impartings
The maximum study iterative steps of itself take 100 steps after the initial connection weight of network.
Further, the filtering process described in step 2 uses Unscented kalman filtering method.
Further, the auxiliary air measuring station for being used in step 3 build model refers to m selected in accordance with the following methods
Aid in air measuring station:
First, to being entered by the filtered wind speed sample set of step 2 using adaptive noise completely integrated empirical modal
Row is decomposed;
Secondly, processing is filtered to the data after decomposition;
Then, filtered data signal reconstruction will be carried out again, and will obtain the wind speed reconstruct data of each air measuring station;
The wind speed reconstruct data of each auxiliary air measuring station and the wind speed reconstruct data of target air measuring station are subjected to correlation test,
Sorted from high to low by the degree of correlation, select and reconstruct m groups auxiliary air measuring station before Data mutuality degree ranking with target air measuring station wind speed
Wind speed reconstructs data and corresponding m auxiliary air measuring station;
Wherein, m is integer, and span is [3,60%N].
Beneficial effect
Compared with prior art, the present invention has advantages below:
1. having incorporated the multiple elements such as space, time, meteorology, physics, auxiliary air measuring station data, target air measuring station make use of
A variety of data such as data, history meteorological data, numerical weather forecast data, it is ensured that the diversity of data.
2. statistical method, physical method, learning method are organically combined, forecasting reliability is improved.
There is data interlacing in 3.4 basic models, reduce amount of calculation during prediction;4 kinds of models pass through standard
BAYESIAN combined model is weighted fusion, is adaptively adjusted weights, improves the stability of prediction.Ensure in any one
In the case of the embedded forecast model timeliness in inherence, the wind speed forecasting method proposed by the invention based on multi-model multiple features strategy
Remain able to export the advanced multi-step prediction value of final wind speed exactly, for the Railway Traffic Dispatching Control under harsher wind conditions.
In the 1st kind of model proposed, by Unscented kalman filtering method and wavelet method to target air measuring station and auxiliary
The original air speed data of whole of air measuring station has carried out filtering and noise reduction processing so that below for correlation calculations and the wind speed of prediction
Data maintain pure property and rule potentiality to greatest extent.In addition, using FEEMD algorithms to the wind speed number after filtering and noise reduction
According to multilayer decomposition has been carried out, generating more air speed data samples is used for the modeling in later stage.Air speed data after decomposition according to
The difference of their jump frequency has been divided into low frequency, intermediate frequency and the gear of high frequency three, and by different combination forecastings to this three
Class is predicted respectively, has taken into account the precision and real-time of final forecast model.For example, being built to the IMF components of low frequency subsequence
Simple RARIMA models are found, the extreme learning machine mould set up to the IMF components of intermediate frequency subsequence after the CBA optimizations of medium accuracy
Type, the IMF decomposition to high frequency subsequence then establishes the LVQ neural network prediction models of high-precision PSO optimizations.Only with regard to this
Three classes are for the hybrid prediction model of different frequency, and technical staff is just difficult to what is obtained without innovation.And this patent is not
To being directly realized by final prediction with this three classes hybrid prediction model, but their predicted value is input to again newly-built
Adaboost-BP hybrid neural networks forecast models in, dexterously auxiliary air measuring station and target air measuring station contact one
Rise, realize high-precision forecasting wind speed.
2nd kind of model does not set up the spatial prediction model of auxiliary air measuring station and target air measuring station, but precision of prediction is put
Target air measuring station itself has been arrived, it is final directly to realize prediction output by the forecast model group of built IMF components.
3rd kind of model dexterously introduces air pressure, three objects of humidity and temperature so that build the robustness of forecast model
It can get a promotion.
The foundation of 4th kind of model has incorporated the technology for the numerical weather forecast that Modern High-Speed Along Railway is generally used.
Brief description of the drawings
Fig. 1 is the principle flow chart of the method for the invention;
Fig. 2 schemes for the BP neural network training of the Adaboost optimizations of model 1 in the present invention;
Fig. 3 is the advanced multi-step Predictive Model figure of target air measuring station of model 1 in the present invention;
Fig. 4 is the advanced multi-step Predictive Model figure of target air measuring station of model 2 in the present invention;
1 layer of wavelet decomposition figure that Fig. 5 is target air measuring station A in the present invention;
Fig. 6 is 2 layers of wavelet decomposition figure of target air measuring station of model 3 in the present invention;
Fig. 7 is the railway forecasting wind speed result schematic diagram that is obtained using Forecasting Methodology proposed by the invention;
Fig. 8 is the railway forecasting wind speed result schematic diagram that is obtained using traditional single ELMAN neural network models;
Fig. 9 is the railway forecasting wind speed result schematic diagram obtained using tradition single ARIMA models.
Embodiment
Below in conjunction with drawings and examples, the present invention is described further:
As shown in figure 1, a kind of forecasting wind speed of high speed railway line side of multi-model multiple features fusion, comprises the following steps:
Step 1:N number of auxiliary air measuring station is at least installed around target air measuring station position, adopted in real time using air measuring station is aided in
The air speed data of collection auxiliary air measuring station, obtains the wind speed sample set of target air measuring station and auxiliary air measuring station;
Wherein, N is the integer more than or equal to 5;
Step 2:Auxiliary air measuring station data and target air measuring station data are filtered and 1 layer depth wavelet decomposition successively,
Extract low-frequency data part;
Step 3:Using low-frequency data build space-advanced multi-step Predictive Model of target air measuring station, self-target air measuring station
Advanced multi-step Predictive Model and meteorology-advanced multi-step Predictive Model of target air measuring station;
Step 4:By space-advanced multi-step Predictive Model of target air measuring station, self-the advanced multi-step prediction mould of target air measuring station
The advanced multi-step prediction value of target air measuring station and weather forecast that type and the advanced multi-step Predictive Model of meteorology-target air measuring station are obtained
Target air measuring station wind speed value inputs BAYESIAN combined model, obtains final target air measuring station predicted value;
Wherein, the step of space-target air measuring station advanced multi-step Predictive Model prediction of wind speed value is as follows:
1st, it is that the following wind speed of railway of certain target air measuring station position is realized to predict, 5 is installed around the air measuring station position
Individual auxiliary air measuring station.Obtain the original air speed data of same session target air measuring station and 5 auxiliary air measuring stations, every group of air speed data
Comprising 600 data, first 500 in 600 data are used to model, the 501st~600 data are used to verify.
Target air measuring station is designated as A, 5 auxiliary air measuring stations be designated as respectively before B, C, D, E, F, each air measuring station 500 it is original
Air speed data is expressed as follows:
Target air measuring station A original air speed data:{a1,a2,a3...,a499,a500}
Aid in air measuring station B original air speed data:{b1,b2,b3...,b499,b500}
Aid in air measuring station C original air speed data:{c1,c2,c3...,c499,c500}
Aid in air measuring station D original air speed data:{d1,d2,d3...,d499,d500}
Aid in air measuring station E original air speed data:{e1,e2,e3...,e499,e500}
Aid in air measuring station F original air speed data:{f1,f2,f3...,f499,f500}
2nd, with Unscented kalman filtering method to air measuring station A, B, C, D, E, F original air speed data is filtered processing, gone
Except potential error in air speed data, following filtered air speed datas are obtained:
The filtered air speed datas of target air measuring station A:{a1′,a′2,a3′...,a′499,a5′00}
Aid in the filtered air speed datas of air measuring station B:{b1′,b2′,b3′...,b4′99,b5′00}
Aid in the filtered air speed datas of air measuring station C:{c1′,c′2,c3′...,c′499,c5′00}
Aid in the filtered air speed datas of air measuring station D:{d1′,d2′,d3′...,d4′99,d5′00}
Aid in the filtered air speed datas of air measuring station E:{e1′,e′2,e3′...,e′499,e5′00}
Aid in the filtered air speed datas of air measuring station F:{f1′,f2′,f3′...,f4′99,f5′00}
3rd, high frequency Jump is removed with the Mallat tower algorithms of 1 layer of wavelet decomposition respectively to filtered each data,
Take low-frequency data:
Low-frequency data after target air measuring station A wavelet decompositions:{a1″,a″2,a3″...,a″499,a5″00}
Aid in the low-frequency data after air measuring station B wavelet decompositions:{b1″,b2″,b3″...,b″499,b″500}
Aid in the low-frequency data after air measuring station C wavelet decompositions:{c1″,c″2,c3″...,c″499,c″500}
Aid in the low-frequency data after air measuring station D wavelet decompositions:{d1″,d2″,d3″...,d″499,d″500}
Aid in the low-frequency data after air measuring station E wavelet decompositions:{e1″,e″2,e3″...,e″499,e″500}
Aid in the low-frequency data after air measuring station F wavelet decompositions:{f″1,f2″,f″3...,f″499,f″500}
4th, each auxiliary air measuring station low-frequency data is shown with target air measuring station low-frequency data respectively by Frechet distances
Work property is examined, and 5 groups of data of air measuring station will be aided in carry out conspicuousness sequence by group, selects conspicuousness maximum and suitable preceding 3
Group data and its corresponding 3 auxiliary air measuring station.This 3 auxiliary air measuring stations such as selected are respectively auxiliary air measuring station B, auxiliary survey
Wind station C, auxiliary air measuring station D.
5th, aid in the low-frequency data of air measuring station to carry out FEEMD decomposition respectively target air measuring station and B, C, D for selecting, obtain
Following components:
Target air measuring station A:AIMF1,AIMF2,...AIMFn,AR
Aid in air measuring station B:BIMF1,BIMF2,...BIMFn,BR
Aid in air measuring station C:CIMF1,CIMF2,...CIMFn,CR
Aid in air measuring station D:DIMF1,DIMF2,...DIMFn,DR
6th, each IMF components after being decomposed to 3 auxiliary air measuring stations are divided into low frequency, intermediate frequency, the Seed Sequences of high frequency 3 by frequency,
RARIMA is set up respectively to each IMF components of low frequency subsequence to be trained, and each IMF components of intermediate frequency subsequence are set up respectively
The extreme learning machine of CBA optimizations is trained, and each IMF components of high frequency subsequence set up the LVQ neutral nets of PSO optimizations respectively
It is trained, the advanced multi-step prediction value of each IMF components is obtained eventually through iteration.
The specific construction step of LVQ neural network model of the high frequency based on PSO is as follows:
The initial population quantity generated using PSO sets iterations when taking 50, PSO optimum choice LVQ neutral nets
Take 200 times, particle speed of service maximum is 0.5, the minimum value of the particle speed of service is 0.01;The iteration of LVQ neutral nets
Target mean absolute relative error takes 5%;LVQ neutral nets obtain PSO assign the initial connection weight of optimal network after from
The maximum study iterative steps of body take 100 steps.
(1) particle populations of the connection weight of LVQ models are randomly generated using PSO, each particle represents one group of LVQ model
Initial connection weight;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of PSO algorithms, is completed
The training learning process of PSO algorithms so that the particle in particle populations is constantly brought near optimal particle, is exported optimal
The initial connection weight of LVQ neutral nets;
(3) the obtained initial connection weight of LVQ neutral nets will be trained to LVQ god by PSO in above two steps
It is configured through network, each IMF components of the high frequency subsequence of air measuring station and target air measuring station will be aided in be used as neutral net
Input data, completes study and training of the LVQ neutral nets itself to wind speed, forms self high frequency subsequence air speed data high-precision
Spend the PSO-LVQ mixed models of prediction.
The essence that the LVQ neutral nets that each IMF components of high frequency subsequence set up PSO optimizations respectively are trained is to utilize
PSO selects the initial connection weight of LVQ neutral nets so that the LVQ neutral nets of optimization can preferably follow the trail of the sub- sequence of high frequency
The extreme mutation rule of each administrative IMF component air speed data of row, realizes the prediction of high-precision intermediate frequency subsequence.
For example, have 2 groups of intermediate frequency IMF wind speed components (if being referred to as IMF1 and IMF2) inside wind speed high frequency subsequence,
Their air speed data length is 500.IMF1 components have 500 wind speed sample datas, then by the 1-497 wind speed sample
The input air speed data of neuron is inputted as the 1st of LVQ neutral nets, the 2-498 air speed data is regard as LVQ nerves
The input air speed data of the 2nd of network input neuron, regard the 3-499 air speed data as the 3rd of LVQ neutral nets
The input air speed data of neuron is inputted, the 4-500 air speed data is regard as unique output neuron of LVQ neutral nets
These air speed data groups, are then synchronously loaded on LVQ neural network models, according to described PSO by output wind speed data
The step of algorithm optimization LVQ neutral nets, completes whole study and training.
It is defeated to the 1st input neuron of LVQ neutral nets after PSO-LVQ Neural network mixed models complete to learn
Enter the 498th air speed data of IMF1 components, to the 499th of the 2nd input neuron input IMF1 component of LVQ neutral nets the
Individual air speed data, the 500th air speed data of IMF1 components is inputted to the 3rd input neuron of LVQ neutral nets, then
LVQ neutral nets will export 1 air speed value automatically, the value be IMF1 components the 501st wind speed value (that is,
Above-mentioned steps are realized obtains advanced 1 step predicted value using 1-500 existing decomposition wind speed samples of IMF1 components, i.e.,
501st wind speed value).
By that analogy, multi-Step Iterations calculating is carried out, to the 1st input neuron input IMF1 component of LVQ neutral nets
The 499th air speed data, the 500th wind speed number of IMF1 components is inputted to the 2nd of LVQ neutral nets input neuron
According to the 501st wind speed that above advanced 1 step prediction has been obtained is passed through in the 3rd input neuron input to LVQ neutral nets
Predicted value, then LVQ neutral nets will export 1 air speed value automatically again, then the value is the 502nd wind of IMF1 components
(namely above-mentioned steps predict institute to fast predicted value using 1-500 existing wind speed samples of IMF1 components by advanced 2 step
The 502nd wind speed value obtained).Like this, the LVQ neutral nets after PSO algorithm optimizations can be completed to IMF1 components
The advanced multistep forecasting wind speed of required any step number is calculated.Other decomposed components (such as IMF2 components) in high frequency wind speed subsequence
The step of advanced multistep wind speed value obtained by PSO-LVQ Neural network mixed models and above-mentioned IMF1 components
Step is the same.
The construction step of extreme learning machine model of the intermediate frequency based on CBA is as follows:
The iterative target mean absolute relative error of extreme learning machine model takes 5%, in CBA optimum choice extreme learning machines
Initial input weights and hidden layer node threshold stage and extreme learning machine assignment phase, pole after CBA algorithm optimizing is obtained
The greatest iteration step number of limit learning machine takes 200 steps.
(1) bat algorithm is initialized;
The location parameter of every bat represents the one group of output weights and hidden layer node threshold value of extreme learning machine, every bat
The initial loudness and pulse rate of bat take 0.5, and bat number 50, iterations is 100, and the flight step-length of every bat takes
It is [0.001,0.05] to be worth scope;
(2) fitness of the setting value as bat algorithm is reached using the output wind speed mean absolute relative error of extreme learning machine
Guidance function, the loudness and pulse rate of every bat of real-time update so that bat fly to extreme learning machine input weights and
The corresponding optimal solution of hidden layer node threshold value;
In preceding 50 iteration, the flight step-length of bat takes maximum, since the 51st iteration, the flight step-length of bat
Take minimum value;
(3) the input weights and the corresponding optimal solution of hidden layer node threshold value of the extreme learning machine of (2) acquisition are utilized, by wind
Input and output data of each IMF components of the intermediate frequency subsequence stood as extreme learning machine, carry out extreme learning machine model instruction
Practice, obtain the extreme learning machine model based on CBA;
For in the input data of training, when the acquisition time of every group of data is super than previous group input data previous
Carve, acquisition time than last group input data super previous moment of output data.
For example, have 2 groups of intermediate frequency IMF wind speed components (if being referred to as IMF1 and IMF2) inside wind speed intermediate frequency subsequence,
Their air speed data length is 500.IMF1 components have 500 wind speed sample datas, then by the 1-497 wind speed sample
The input air speed data of neuron is inputted as the 1st of extreme learning machine, the 2-498 air speed data is learnt as the limit
The input air speed data of the 2nd of machine input neuron, it is defeated using the 3-499 air speed data as the 3rd of extreme learning machine
Enter the input air speed data of neuron, regard the 4-500 air speed data as the defeated of unique output neuron of extreme learning machine
Go out air speed data, then synchronously these air speed data groups are loaded on extreme learning machine model, according to described CBA algorithms
The step of optimizing extreme learning machine model completes whole study and training.
After CBA- extreme learning machines mixed model completes to learn, the 1st input neuron input to extreme learning machine
498th air speed data of IMF1 components, to the 499th of the 2nd input neuron input IMF1 component of extreme learning machine
Air speed data, inputs the 500th air speed data of IMF1 components, then the limit to the 3rd input neuron of extreme learning machine
Learning machine will export 1 air speed value automatically, and the value is the 501st wind speed value of IMF1 components (that is, above-mentioned step
Rapid realize obtains advanced 1 step predicted value, i.e., the 501st using 1-500 existing decomposition wind speed samples of IMF1 components
Individual wind speed value).
By that analogy, multi-Step Iterations calculating is carried out, IMF1 components are inputted to the 1st input neuron of extreme learning machine
499th air speed data, the 500th air speed data of IMF1 components is inputted to the 2nd input neuron of extreme learning machine, right
3rd input neuron input of extreme learning machine predicts the 501st forecasting wind speed obtained by above advanced 1 step
Value, then extreme learning machine will export 1 air speed value automatically again, then the value is that the 502nd wind speed of IMF1 components is pre-
(namely above-mentioned steps are obtained measured value using 1-500 existing wind speed samples of IMF1 components by the prediction of advanced 2 step
The 502nd wind speed value).Like this, the extreme learning machine after CBA algorithm optimizations is appointed needed for being completed to IMF1 components
The advanced multistep forecasting wind speed for step number of anticipating is calculated.Other decomposed components (such as IMF2 components) in intermediate frequency wind speed subsequence pass through
The step of the step of advanced multistep wind speed value that CBA- extreme learning machine mixed models are obtained, is with above-mentioned IMF1 components
Equally.
It is profit to set up the essence that the extreme learning machines of CBA optimizations is trained respectively to each IMF components of intermediate frequency subsequence
With the initial input weights and hidden layer node threshold value of variable-step self-adaptive bat algorithm optimization extreme learning machine so that optimization
Extreme learning machine can preferably follow the trail of the mutation rule of each administrative IMF component air speed data of intermediate frequency subsequence, realize high
The prediction of the intermediate frequency subsequence of precision.
The flight step-length of bat automatically changes with being incremented by for iterations, it is ensured that bat algorithm is to extreme learning machine
Ability of searching optimum on forecasting wind speed, makes to take maximum during the 50 step iteration before algorithm optimizing, it is therefore an objective to avoid calculating
Method too early be absorbed in locally optimal solution;With the increase (i.e. since 50 steps untill 100 steps) of iterations, the flight of bat
Step-length actively reduces, and minimum value is taken, so that bat algorithm is more accurately solved in later stage Fast Convergent.
The construction step of low frequency RARIMA models is as follows:
(1) whole IMF components wind speed numbers of acquisition are decomposed to each low frequency subsequence using nonparametric pleasure boat method of inspection
According to progress data stationarity inspection;
Such as run into certain IMF components air speed data and be presented non-stationary, then difference meter is carried out to this section of IMF components air speed data
Calculate untill it shows stationarity;
(2) acquisition is decomposed to each low frequency subsequence to whole IMF components wind speed after step 1 stationary test
Data carry out sample auto-correlation and sample partial correlation is calculated, and true according to the auto-correlation and partial correlation value of respective component air speed data
Determine the optimal type and Optimal order of RARIMA models;
(3) the optimal type and Optimal order of the RARIMA models obtained to step 2, are solved each using maximum-likelihood method
The equation coefficient of individual IMF components air speed data correspondence RARIMA models, forms self low frequency RARIMA model;
During the prediction of advanced multistep, the equation of newest predicted value real-time update RARIMA models is constantly utilized
Coefficient.
7th, as shown in Fig. 2 being modeled respectively to high frequency series, middle frequency sequence, low frequency sequence, each sequence is aided in B, C, D
Each IMF components of air measuring station are input, using each IMF components of target air measuring station as output, are optimized using Adaboost
BP neural network is trained.
8th, as shown in figure 3, B, C, D the auxiliary advanced multi-step prediction value of air measuring station obtained to prediction brings what is trained into
The BP neural network of Adaboost optimizations, then signal reconstruction is carried out, finally give the advanced multistep forecasting wind speed of target air measuring station
Value.
Adaboost BP neural network model includes 10 BP neural networks as weak learner, with forecasting wind speed
BP neural network of the mean absolute relative error more than 5% is considered as the object for needing to strengthen study, and the maximum of BP neural network changes
Number of riding instead of walk takes 300 steps.
Such as, there are 3 by the auxiliary air measuring station filtered out, after FEEMD is decomposed, they, which amount to, possesses 20 groups of IMF
Component, each component has 500 air speed datas, while to be divided into low frequency, intermediate frequency different with three groups of high frequency for this 20 group component
Subsequence in.In order to input this totally 20 groups of IMF component, each BP neural network requires 20 input neuron numbers.
In other words, the input data of each BP neural network through the auxiliary air measuring station being screened out possessed it is whole
IMF components.
Target air measuring station possesses 5 groups of IMF components after FEEMD is decomposed, then they will be used as each BP neural network
Output data, to each BP neural network carry out learning training.Because to different wind speed sample datas, FEEMD decomposes production
Raw IMF component numbers are different, so the input neuron and output neuron number of BP neural network are also dynamic
, it is respectively depending on the auxiliary air measuring station filtered out and target air measuring station finally possesses several IMF components numbers.But to same group
For identical wind speed sample data, no matter they are from auxiliary air measuring station or from target air measuring station, by FEEMD points
IMF component numbers produced by after solution are fixed, that is to say, that the input neuron of corresponding 10 BP neural networks and defeated
It is also fixed to go out neuron number.
In Adaboost-BP Neural network mixed models, the structure of each BP neural network is input neuron and defeated
Go out what neuron number was just as, but for the air speed data sample different to Along Railway, the BP nerves set up every time
Network is different again.
Self-the advanced multi-step Predictive Model of target air measuring station carries out comprising the following steps that for forecasting wind speed:
1st, target air measuring station use with the 1st forecast model identical initial data, by the original wind speed number of target air measuring station
Handled according to Unscented kalman filtering method, remove potential error in air speed data.
Target air measuring station A original air speed data:{a1,a2,a3...,a499,a500}
The filtered air speed datas of target air measuring station A:{a1′,a′2,a3′...,a′499,a5′00}
2nd, high frequency Jump is removed with the Mallat tower algorithms of 1 layer of wavelet decomposition to filtered data, takes low frequency
Data:
Low-frequency data after target air measuring station A wavelet decompositions:{a1″,a″2,a3″...,a″499,a″500}
3rd, FEEMD decomposition is carried out to the low-frequency data of target air measuring station, obtains following components:
Target air measuring station A:AIMF1,AIMF2,...AIMFn,AR
4th, as shown in figure 4, being divided into low frequency, intermediate frequency, the Seed Sequences of high frequency 3 by frequency to each IMF components after decomposition, to low
Each IMF components of frequency subsequence are set up RARIMA and are trained respectively, and CBA is set up respectively to each IMF components of intermediate frequency subsequence
The extreme learning machine of optimization is trained, and the LVQ neutral nets that each IMF components of high frequency subsequence set up PSO optimizations respectively are entered
Row training, the advanced multi-step prediction value of each IMF components is obtained eventually through iteration.
5th, to each component weighted calculation, reconstruction signal obtains the advanced multistep wind speed value of target air measuring station.
The advanced multi-step Predictive Model of meteorology-target air measuring station carries out comprising the following steps that for forecasting wind speed:
1st, target air measuring station use with the 1st forecast model identical initial data, by the original wind speed number of target air measuring station
Handled according to Unscented kalman filtering method, remove potential error in air speed data.
Target air measuring station A original air speed data:{a1,a2,a3...,a499,a500}
The filtered air speed datas of target air measuring station A:{a1′,a′2,a3′...,a′499,a5′00}
2nd, as shown in figure 5, removing high frequency jump spy with the tower algorithms of Mallat of 1 layer of wavelet decomposition to filtered data
Levy, take low-frequency data A1:
Low-frequency data A after target air measuring station A wavelet decompositions1:{a1″,a″2,a3″...,a″499,a″500}
3rd, as shown in fig. 6, to low-frequency data A1The wavelet decomposition that the tower algorithms of Mallat carry out 2 layer depths is reused, is taken
2nd layer of low frequency component A111:
Low-frequency data A after target air measuring station A wavelet decompositions111:{a1″′,′a″′2,a3″′...,a″′499,a″′500}
4th, the Elman of ACO optimizations is set up to historical barometric, humidity, temperature and wavelet decomposition is obtained twice low frequency component
Neutral net is trained, and obtains jump ahead prediction.
Initial population ant quantity takes 100, and the ant pheromones significance level factor takes 0.5, the important journey of ant heuristic function
The degree factor takes 1, and ant pheromones volatilization factor takes 0.1.Iterations during ACO optimum choice Elman neutral nets takes 50 times.
The iterative target mean absolute relative error of Elman neutral nets takes 5%.Elman neutral nets are obtaining the optimal of ACO impartings
The maximum study iterative steps of itself take 100 steps after the initial connection weight of network.
(1) population of the connection weight of Elman Model of Neural Network is randomly generated using ACO;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of ACO algorithms, is carried out
The training learning process of ACO algorithms so that the ant in initialization ant population is by iteration constantly close to optimal solution, output
The optimal initial connection weight of Elman neutral nets and hidden layer threshold value;
(3) the initial connection weight obtained and threshold value is trained to carry out Elman neutral nets by ACO above-mentioned steps
Set, carried out the second layer low frequency component data of air pressure, temperature, humidity and target air measuring station as Elman neutral nets
Input and output data to study and the training of wind speed, form meteorology-advanced multi-step Predictive Model of target air measuring station;
For in the input data of training, when the acquisition time of every group of data is super than previous group input data previous
Carve, acquisition time than last group input data super previous moment of output data.
The Elman neutral nets that set up ACO optimizes are utilized to the of air pressure, temperature, humidity and target air measuring station
This 4 kinds of objects of two layers of low frequency wind speed component data complete advanced multi-step predictions and calculated, corresponding advanced many to obtain various objects
Walk predicted value.
The Elman neutral nets set up use the structures of 4 input neurons and 1 output neuron, i.e.,:Elman
Neutral net needs that synchronously network is learnt and trained with the data and 1 group of output wind speed data of 4 groups of above-mentioned 4 kinds of objects;
After the Elman neutral nets that ACO optimizes complete training, as long as distinguishing 4 input neuron ports of Elman neutral nets
The data of 4 above-mentioned 4 kinds of objects are inputted, Elman neutral nets can just export 1 wind speed value automatically.
Such as, certain moment use state air pressure, temperature, humidity and target air measuring station this 4 kinds of second layer low frequency wind speed it is right
The sample data of elephant, their data length is 500.It regard the 1-499 sample of air pressure as the of Elman neutral nets
The input data of 1 input neuron, is inputted the 1-499 sample of temperature data as the 2nd of Elman neutral nets
The input data of neuron, neuron is inputted using the 1-499 sample of humidity data as the 3rd of Elman neutral nets
Input data, regard the 1-499 sample of the second layer low frequency wind speed of target air measuring station as the 4th of Elman neutral nets
The input data of individual input neuron, regard the 2-500 sample of the second layer low frequency wind speed of target air measuring station as Elman god
Then these data groups are synchronously loaded into Elman nerve nets by the output data of unique 1 output neuron through network
On network, according to the study and training that whole is completed the step of described ACO algorithm optimization Elman neural network models.
After ACO-Elman Neural network mixed models complete to learn, to the 1st input nerve of Elman neutral nets
500th sample data of first input air pressure object, neuron input temp object is inputted to the 2nd of Elman neutral nets
The 500th sample data, the 500th sample number of humidity object is inputted to the 3rd of Elman neutral nets input neuron
According to the 500th of the second layer low frequency wind speed object of the 4th input neuron input target air measuring station of Elman neutral nets the
Individual sample data, then Elman neutral nets will export 1 air speed value automatically, the value is that the second layer of target air measuring station is low
501st wind speed value of frequency wind speed object is (that is, above-mentioned steps realize the using above-mentioned 4 kinds different objects
1-500 sample data obtains the advanced 1 step predicted value of target air measuring station wind speed object, i.e., the 501st wind speed value).
By that analogy, multi-Step Iterations calculating is carried out, continues the 1st, the 2nd and the 3rd input to Elman neutral nets
Neuron difference input air pressure, the newest measured data of temperature and humidity, i.e. their corresponding 501st measured data, then
The 501st wind speed value that the 4th input neuron input to Elman neutral nets is obtained using above-mentioned steps simultaneously,
So Elman neutral nets will export 1 air speed value, the of the second layer low frequency wind speed object of target air measuring station automatically again
502 wind speed values are (that is, above-mentioned steps realize the 1- using air pressure, the different objects of this 3 kinds of temperature and humidity
500 sample datas and their corresponding newest 501st measured value, obtain advanced the 2 of target air measuring station wind speed object
Walk the wind speed value of predicted value, i.e., the 502nd).Like this, the Elman neutral nets after ACO algorithm optimizations can be surveyed to target
Wind speed object advanced multistep forecasting wind speed of any step number needed for completing in wind station is calculated.
It is noted that in the calculating process of the advanced multi-Step Iterations of above-mentioned Elman neutral nets, only wind speed
The step of object has constantly carried out incremental input, other 3 meteorological objects (i.e. air pressure, temperature and humidity) are all to employ him
Measured value.So the present invention not only effectively utilizes the rule of wind speed itself, and efficiently utilizes other 3 meteorologies
The data potential contribution of object, therefore obtain the success to target air measuring station wind speed object high-precision forecast.
5th, it is the corresponding moment air pressure of numerical weather forecast, humidity, the predicted value of temperature and jump ahead predicted value band is superb
Through network, final iteration obtains the advanced multi-step prediction value of target air measuring station.
4th forecast model comprises the following steps:
1st, the wind speed value of target air measuring station is obtained by numerical weather forecast.
With BAYESIAN combined model predicting the outcome come 4 fundamental forecasting models of summary.
Wherein, when calculating t+1 moment forecasting wind speeds, the weights of 4 fundamental forecasting models change in BAYESIAN combined model
In generation, is since preceding t.This example takes t=500.
Realize that the result of forecasting wind speed is as shown in Figure 7 using Forecasting Methodology proposed by the invention.Utilize existing ELMAN god
Realize that the result of forecasting wind speed is as shown in Figure 8 through network model.Using the result of existing ARIMA model realizations forecasting wind speed as schemed
Shown in 9.Using formula (1-3) to the progress precision index calculating that predicts the outcome shown in Fig. 7-Fig. 9,1 and table 2 the results are shown in Table.
Mean absolute error:
Mean absolute relative error:
Root-mean-square error:
In above-mentioned formula, n is to take n to be 100 in air speed data number for model testing, this example.X (i) is actual measurement
Air speed data,For prediction of wind speed data.
Table 1:Utilize the precision of prediction of Forecasting Methodology proposed by the invention
Mean absolute error | 0.2940m/s |
Mean absolute relative error | 1.51% |
Root-mean-square error | 0.2821m/s |
Table 2:Utilize the precision of prediction of existing ELMAN neural network models
Mean absolute error | 1.6961m/s |
Mean absolute relative error | 7.99% |
Root-mean-square error | 2.2125m/s |
Table 3:Utilize the precision of prediction of existing ARIMA models
Mean absolute error | 1.7237m/s |
Mean absolute relative error | 8.54% |
Root-mean-square error | 1.8426m/s |
From the point of view of Fig. 7, Fig. 8 and Fig. 9, and combination Tables 1 and 2, method of the present invention from mean absolute error, is put down
From the point of view of equal absolute relative error and root-mean-square error, hence it is evident that better than prior art, show that the method for the invention has preferable
Application effect.
Claims (8)
1. a kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion, it is characterised in that comprise the following steps:
Step 1:N number of auxiliary air measuring station is at least installed around target air measuring station position, it is auxiliary using aiding in air measuring station to gather in real time
The air speed data of air measuring station is helped, the wind speed sample set of target air measuring station and auxiliary air measuring station is obtained;
Wherein, N is the integer more than or equal to 5;
Step 2:Auxiliary air measuring station data and target air measuring station data are filtered and 1 layer depth wavelet decomposition successively, extracted
Low-frequency data part;
Step 3:The auxiliary air measuring station obtained using step 2 and the low-frequency data part of target air measuring station are built space-target and surveyed
The advanced multi-step Predictive Model in wind station, meanwhile, using the low-frequency data part of target air measuring station build self-target air measuring station is advanced
Multi-step Predictive Model;
Step 4:Low-frequency data part and air pressure, humidity and the temperature number of target air measuring station present position using target air measuring station
According to structure meteorology-advanced multi-step Predictive Model of target air measuring station;
Step 5:Utilize numerical weather forecast acquisition of information target air measuring station wind speed value;
Step 6:By space-advanced multi-step Predictive Model of target air measuring station, self-the advanced multi-step Predictive Model of target air measuring station with
And the advanced multi-step prediction value of target air measuring station that obtains of the advanced multi-step Predictive Model of meteorology-target air measuring station and step 5 are obtained
Target air measuring station wind speed value inputs BAYESIAN combined model, obtains final target air measuring station predicted value;
The advanced multi-step prediction refers to obtain subsequent time T+ using current time T air speed data input correspondence forecast model
1 forecasting wind speed, then, inputs correspondence forecast model using subsequent time T+1 wind speed value, obtains the T+2 moment again
Wind speed value, reciprocal iteration obtains advanced multi-step prediction value.
2. according to the method described in claim 1, it is characterised in that space-advanced multistep of target air measuring station in the step 2
Forecast model and self-construction step of the advanced multi-step Predictive Model of target air measuring station is as follows:
Step 2.1:FEEMD decomposition is carried out to the low-frequency data part of auxiliary air measuring station and target air measuring station, each survey wind is obtained
Stand corresponding low frequency, intermediate frequency and high frequency subsequence;
Step 2.2:Each IMF components of auxiliary air measuring station and the high frequency subsequence of target air measuring station are carried out based on PSO's respectively
The training of LVQ neutral nets, builds the auxiliary LVQ nerve nets of air measuring station and self respective high frequency of target air measuring station based on PSO
Network model, obtains the advanced multi-step prediction value of self high frequency of auxiliary air measuring station and target air measuring station;
Limit study based on CBA is set up respectively to each IMF components of auxiliary air measuring station and the intermediate frequency subsequence of target air measuring station
Machine is trained, and is built the auxiliary extreme learning machine model of air measuring station and self respective intermediate frequency of target air measuring station based on CBA, is obtained
Take the advanced multi-step prediction value of self intermediate frequency of auxiliary air measuring station and target air measuring station;
Each IMF components of low frequency subsequence to aiding in air measuring station and target air measuring station are set up RARIMA and are trained respectively, structure
Auxiliary air measuring station and self respective low frequency RARIMA model of target air measuring station is built, auxiliary air measuring station and target air measuring station is obtained
The advanced multi-step prediction value of self low frequency;
Step 2.3:By the limit study of LVQ neural network model of the high frequency of target air measuring station based on PSO, intermediate frequency based on CBA
Machine model and low frequency RARIMA model combinations, formed self-the advanced multi-step Predictive Model of target air measuring station;
Step 2.4:To aid in the low frequency, intermediate frequency and high frequency subsequence data of air measuring station as input, the low frequency of target air measuring station,
Intermediate frequency and high frequency subsequence data build each BP neural network model of the frequency range subsequence based on Adaboost, profit as output
Each frequency range subsequence is inputted based on Adaboost's with the corresponding advanced multi-step prediction value of each frequency range subsequence of auxiliary air measuring station
BP neural network model, is obtained the advanced multi-step prediction value of each frequency range of target air measuring station correspondence, is based on each frequency range subsequence
Adaboost BP neural network model formation space-advanced multi-step Predictive Model of target air measuring station;
The BP neural network model of each frequency range subsequence based on Adaboost refers to, using Adaboost algorithm to multiple BP
The mean absolute error for the predicted value that neutral net is exported according to each BP neural network carries out not decile weighted sum, each BP god
Through network using 3 layer network structures of multiple input neurons and 1 output neuron, in the input data of training,
At the acquisition time of the every group of data previous moment super than previous group input data, the acquisition time of output data is than last group
The input data super previous moment.
3. method according to claim 2, it is characterised in that LVQ neural network model of self high frequency based on PSO
Specific construction step it is as follows:
(1) randomly generate the particle populations of the connection weight of LVQ models using PSO, each particle represent one group of LVQ model just
Beginning connection weight;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of PSO algorithms, completes PSO
The training learning process of algorithm so that the particle in particle populations is constantly brought near optimal particle, exports optimal LVQ god
Through the initial connection weight of network;
(3) the obtained initial connection weight of LVQ neutral nets will be trained to LVQ nerve nets by PSO in above two steps
Network is configured, input and output data using each IMF components of the high frequency subsequence at wind station as neutral net, completes LVQ
Study and training of the neutral net itself to wind speed, form the PSO-LVQ of self high frequency subsequence air speed data high-precision forecast
Mixed model;
It is the acquisition time previous moment super than previous group input data of every group of data, defeated in the input data of training
Go out acquisition time than last group input data super previous moment of data.
4. method according to claim 2, it is characterised in that extreme learning machine model of self intermediate frequency based on CBA
Construction step it is as follows:
(1) bat algorithm is initialized;
The location parameter of every bat represents the one group of output weights and hidden layer node threshold value of extreme learning machine, every bat
Initial loudness and pulse rate take 0.5, and bat number 50, iterations is 100, the flight step-length value model of every bat
Enclose for [0.001,0.05];
(2) reach that setting value is guided as the fitness of bat algorithm using the output wind speed mean absolute relative error of extreme learning machine
Function, the loudness and pulse rate of every bat of real-time update so that bat flies to the input weights of extreme learning machine and implied
The corresponding optimal solution of node layer threshold value;
In preceding 50 iteration, the flight step-length of bat takes maximum, since the 51st iteration, and the flight step-length of bat takes most
Small value;
(3) the input weights and the corresponding optimal solution of hidden layer node threshold value of the extreme learning machine of (2) acquisition are utilized, by wind station
Input and output data of each IMF components of intermediate frequency subsequence as extreme learning machine, carry out extreme learning machine model training, obtain
To the extreme learning machine model based on CBA;
It is the acquisition time previous moment super than previous group input data of every group of data, defeated in the input data of training
Go out acquisition time than last group input data super previous moment of data.
5. method according to claim 2, it is characterised in that the construction step of self low frequency RARIMA model is such as
Under:
(1) whole IMF components air speed datas that acquisition is decomposed to each low frequency subsequence using nonparametric pleasure boat method of inspection enter
Row data stationary test;
Such as run into certain IMF components air speed data and be presented non-stationary, then it is straight to carry out Difference Calculation to this section of IMF components air speed data
Untill it shows stationarity;
(2) acquisition is decomposed to each low frequency subsequence to whole IMF components air speed datas after step 1 stationary test
Carry out sample auto-correlation and sample partial correlation is calculated, and the auto-correlation and partial correlation value according to respective component air speed data are determined
The optimal type and Optimal order of RARIMA models;
(3) the optimal type and Optimal order of the RARIMA models obtained to step 2, each is solved using maximum-likelihood method
The equation coefficient of IMF components air speed data correspondence RARIMA models, using each IMF components of the low frequency subsequence at wind station as
The input data of RARIMA models, forms self low frequency RARIMA model.
6. the method according to claim any one of 1-5, it is characterised in that the meteorology-advanced multistep of target air measuring station
The construction step of forecast model is as follows:
(1) population of the connection weight of Elman Model of Neural Network is randomly generated using ACO;
(2) to realize that minimum windspeed mean absolute relative error reaches that setting value, as the training direction of ACO algorithms, carries out ACO
The training learning process of algorithm so that the ant in initialization ant population is by iteration constantly close to optimal solution, and output is most
The excellent initial connection weight of Elman neutral nets and hidden layer threshold value;
(3) the initial connection weight obtained and threshold value is trained to set Elman neutral nets by ACO above-mentioned steps
Put, regard the second layer low frequency component data of air pressure, temperature, humidity and target air measuring station as the progress pair of Elman neutral nets
The input of study and the training of wind speed and output data, form meteorology-advanced multi-step Predictive Model of target air measuring station;
It is the acquisition time previous moment super than previous group input data of every group of data, defeated in the input data of training
Go out acquisition time than last group input data super previous moment of data.
7. method according to claim 6, it is characterised in that the filtering process described in step 2 uses Unscented kalman
Filter method.
8. the method according to claim any one of 1-5, it is characterised in that described to be used in step 3 build the auxiliary of model
Air measuring station is helped to refer to the m auxiliary air measuring station selected in accordance with the following methods:
First, to being divided by the filtered wind speed sample set of step 2 using adaptive noise completely integrated empirical modal
Solution;
Secondly, processing is filtered to the data after decomposition;
Then, filtered data signal reconstruction will be carried out again, and will obtain the wind speed reconstruct data of each air measuring station;
The wind speed reconstruct data of each auxiliary air measuring station and the wind speed reconstruct data of target air measuring station are subjected to correlation test, by phase
Guan Du sorts from high to low, selects and reconstructs the wind speed that m groups before Data mutuality degree ranking aid in air measuring station with target air measuring station wind speed
Reconstruct data and corresponding m auxiliary air measuring station;
Wherein, m is integer, and span is [3,60%N].
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611020632.XA CN106779148B (en) | 2016-11-14 | 2016-11-14 | A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611020632.XA CN106779148B (en) | 2016-11-14 | 2016-11-14 | A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106779148A CN106779148A (en) | 2017-05-31 |
CN106779148B true CN106779148B (en) | 2017-10-13 |
Family
ID=58969599
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611020632.XA Active CN106779148B (en) | 2016-11-14 | 2016-11-14 | A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106779148B (en) |
Families Citing this family (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108416372B (en) * | 2018-02-11 | 2022-07-19 | 同济大学 | Wind power plant wind speed acquisition method based on multiple sensors |
CN108510737B (en) * | 2018-04-12 | 2020-04-10 | 中南大学 | Unmanned vehicle power supply real-time monitoring method and device integrating wind environment |
CN109034476B (en) * | 2018-07-27 | 2020-11-17 | 中南大学 | High-speed railway line extreme wind speed big data clustering prediction method |
CN109063907B (en) * | 2018-07-27 | 2020-06-05 | 中南大学 | Intelligent traversal large-step-length prediction method for maximum wind speed along high-speed railway |
CN109034478B (en) * | 2018-07-27 | 2021-09-24 | 中南大学 | High-precision prediction method for high-wind iterative competition along high-speed railway |
CN108985521A (en) * | 2018-08-02 | 2018-12-11 | 广东工业大学 | Output power predicting method, device, equipment and the medium of photovoltaic generating system |
CN109297713B (en) * | 2018-08-07 | 2019-12-31 | 浙江大学 | Steam turbine fault diagnosis method based on stable and non-stable vibration signal characteristic selection |
CN109297735B (en) * | 2018-09-11 | 2020-02-28 | 浙江大学 | Vibration signal fault diagnosis method for intelligent power plant coal mill |
CN109684771B (en) * | 2019-01-11 | 2021-09-03 | 西安电子科技大学 | Maneuvering target state prediction optimization method based on interactive multi-model |
CN110009037B (en) * | 2019-04-03 | 2020-10-27 | 中南大学 | Short-term engineering wind speed prediction method and system based on physical information coupling |
CN110210646B (en) * | 2019-04-25 | 2021-08-10 | 浙江大学 | High-speed railway strong wind alarm holding time dynamic adjustment method based on wind speed prediction |
CN112149230B (en) * | 2020-09-27 | 2022-06-10 | 中南大学 | Method for predicting comfort deterioration of wind-induced train of strong wind railway |
CN113191524B (en) * | 2021-01-29 | 2023-07-11 | 河北建投新能源有限公司 | Wind speed prediction method, device, storage medium and processor |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5646343A (en) * | 1993-07-02 | 1997-07-08 | Pritchard; Declan Nigel | System and method for monitoring wind characteristics |
CN102830446A (en) * | 2012-08-13 | 2012-12-19 | 国电南京自动化股份有限公司 | Intelligent meteorological station system capable of forecasting meteorological data |
CN103389388A (en) * | 2012-05-08 | 2013-11-13 | 华锐风电科技(集团)股份有限公司 | Method and device for predicting wind speed in wind power plant and method and system for predicting power in wind power plant |
CN103616734A (en) * | 2013-12-11 | 2014-03-05 | 山东大学 | System and method for large-range synchronous real-time meteorological data measurement and wind speed and direction prediction |
CN204330280U (en) * | 2015-01-18 | 2015-05-13 | 中南大学 | Train parting flow field on-site testing device |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2478600A (en) * | 2010-03-12 | 2011-09-14 | Vestas Wind Sys As | A wind energy power plant optical vibration sensor |
-
2016
- 2016-11-14 CN CN201611020632.XA patent/CN106779148B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5646343A (en) * | 1993-07-02 | 1997-07-08 | Pritchard; Declan Nigel | System and method for monitoring wind characteristics |
CN103389388A (en) * | 2012-05-08 | 2013-11-13 | 华锐风电科技(集团)股份有限公司 | Method and device for predicting wind speed in wind power plant and method and system for predicting power in wind power plant |
CN103389388B (en) * | 2012-05-08 | 2015-08-19 | 华锐风电科技(集团)股份有限公司 | Method for forecasting and device thereof and power forecasting method and system thereof |
CN102830446A (en) * | 2012-08-13 | 2012-12-19 | 国电南京自动化股份有限公司 | Intelligent meteorological station system capable of forecasting meteorological data |
CN103616734A (en) * | 2013-12-11 | 2014-03-05 | 山东大学 | System and method for large-range synchronous real-time meteorological data measurement and wind speed and direction prediction |
CN103616734B (en) * | 2013-12-11 | 2015-11-18 | 山东大学 | Synchronous real time meteorological data is measured and wind speed and direction prognoses system and method on a large scale |
CN204330280U (en) * | 2015-01-18 | 2015-05-13 | 中南大学 | Train parting flow field on-site testing device |
Also Published As
Publication number | Publication date |
---|---|
CN106779148A (en) | 2017-05-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106779148B (en) | A kind of method for forecasting wind speed of high speed railway line of multi-model multiple features fusion | |
CN106779151B (en) | A kind of line of high-speed railway wind speed multi-point multi-layer coupling prediction method | |
CN106372731B (en) | A kind of high wind line of high-speed railway wind speed spatial network structure forecast method | |
CN104765916B (en) | A kind of Dynamics Performance of High Speed Trains parameter optimization method | |
CN110007652A (en) | A kind of Hydropower Unit degradation trend interval prediction method and system | |
CN102129259B (en) | Neural network proportion integration (PI)-based intelligent temperature control system and method for sand dust environment test wind tunnel | |
CN106772695B (en) | A kind of Along Railway wind speed forecasting method for merging many air measuring station measured datas | |
CN106557840B (en) | A kind of high wind line of high-speed railway wind speed adaptive decomposition Forecasting Methodology | |
CN107909206A (en) | A kind of PM2.5 Forecasting Methodologies based on deep structure Recognition with Recurrent Neural Network | |
CN105243259A (en) | Extreme learning machine based rapid prediction method for fluctuating wind speed | |
Liu et al. | A fault diagnosis intelligent algorithm based on improved BP neural network | |
CN104636801A (en) | Transmission line audible noise prediction method based on BP neural network optimization | |
CN106529818A (en) | Water quality evaluation prediction method based on fuzzy wavelet neural network | |
CN107677473A (en) | A kind of GRNN rotating machinery fault Forecasting Methodologies based on FOA optimizations | |
CN108053052A (en) | A kind of oil truck oil and gas leakage speed intelligent monitor system | |
CN110363349A (en) | A kind of LSTM neural network hydrologic(al) prognosis method and system based on ASCS | |
CN109934422A (en) | Neural network wind speed prediction method based on time series data analysis | |
CN103793887A (en) | Short-term electrical load on-line predicting method based on self-adaptation enhancing algorithm | |
CN103927460A (en) | Wind power plant short-term wind speed prediction method based on RBF | |
CN109063388A (en) | The micro climate architecture design addressing design method of wind environment simulation | |
CN104732067A (en) | Industrial process modeling forecasting method oriented at flow object | |
CN109408896B (en) | Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production | |
CN115206444A (en) | Optimal drug dosage prediction method based on FCM-ANFIS model | |
CN110277783A (en) | A kind of online Active Splitting Control and system of electric system | |
CN107180261B (en) | Greenhouse microclimate medium-long term prediction method based on rolling BP neural network |
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 |