CN107578101A - A kind of data stream load Forecasting Methodology - Google Patents
A kind of data stream load Forecasting Methodology Download PDFInfo
- Publication number
- CN107578101A CN107578101A CN201710876477.XA CN201710876477A CN107578101A CN 107578101 A CN107578101 A CN 107578101A CN 201710876477 A CN201710876477 A CN 201710876477A CN 107578101 A CN107578101 A CN 107578101A
- Authority
- CN
- China
- Prior art keywords
- neuron
- som
- triumph
- load
- networks
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a kind of data stream load Forecasting Methodology, belong to Research On The Key Technology In Data Stream field, solve most load estimation algorithm in the prior art, continually request accesses server, the problem of reducing the income that load estimation is brought.The present invention includes calculating topological data set caused by initialization given data source, calculating topology, calculating task；According to the SOM networks after topological data set training optimization；The data characteristics of new calculating task is extracted, finds the triumph neuron of the SOM networks trained；Judge whether triumph neuron meets constraint；Step 5：Meet, load estimation, output loading prediction result are carried out by SOM networks, while the weights of the neighborhood neuron in SOM networks are adjusted, re -training SOM networks；Do not meet, neuron is added to SOM network dynamics；Load estimation is carried out by linear regression method, and obtains real load information, re -training SOM networks, re -training SOM networks are prepared to predict next time.The present invention is used for load estimation.
Description
Technical field
A kind of data stream load Forecasting Methodology, for load estimation, belong to Research On The Key Technology In Data Stream field.
Background technology
With the continuous development of technology, increasing computing resource is migrated to high in the clouds, is serviced and pressed using platform
The mode of charging is needed, how on the premise of system stable operation is ensured, reduces operating cost as far as possible, turns into a research heat
Point, load estimation technology can solve the above problems to a certain extent, therefore load estimation has become Data Stream Processing system
One of study hotspot and research emphasis in system.
According to prediction algorithm model, linear prediction and nonlinear prediction are can be largely classified into for the forecast model of load.
Wherein linear prediction mainly includes arma modeling, FARIMA models, and nonlinear prediction mainly includes neutral net, wavelet theory
And SVMs, due to the uncertainty of stream process fluctuation of load rule, therefore the method for nonlinear prediction is more by everybody
Attention.Most prediction algorithm, its general principle are advised based on existing load time sequence data and other history
Rule is predicted, such as autoregressive moving-average model, difference autoregressive moving-average model, Holter exponential smoothing.This
A little prediction algorithms are established on the basis of each moment system load truth is understood, and continually request accesses, and can give service
Device brings no small expense, reduces the income that load estimation is brought to a certain extent.And above-mentioned algorithm is not directed to
The characteristics of stream processing system, is studied.
The content of the invention
It is an object of the invention to：Solves most load estimation algorithm in the prior art, continually request accesses clothes
It is engaged in device, brings no small expense to server, the problem of reducing the income that load estimation is brought；A kind of data flow is provided to bear
Carry Forecasting Methodology.
The technical solution adopted by the present invention is as follows：
A kind of data stream load Forecasting Methodology, comprises the following steps：
Step 1：Initialize given data source, calculate the caused calculating topological data set of topology, calculating task；
Step 2：According to the SOM networks after topological data set training optimization；
Step 3：The data characteristics of new calculating task is extracted, finds the triumph neuron of the SOM networks trained；
Step 4：Judge whether triumph neuron meets constraint, if met, jump to step 5, otherwise jump to step
6；
Step 5：Load estimation is carried out by the SOM networks that train, output loading prediction result, while to SOM networks
In the weights of neighborhood neuron be adjusted, and jump to step 8；
Step 6：Neuron is added to SOM network dynamics；
Step 7：Load estimation is carried out by linear regression method, and obtains real load information；
Step 8：Load estimation result re -training SOM networks are added, are prepared for prediction next time.
Further, in the step 2, the SOM networks after optimization include,
Newly-increased neuron weight vector initialization strategy：In result using last load estimation, the nerve that newly increases
The output of member meets that the weight vector of the triumph neuron of constraint initializes the neuron node, weight vector profit as a part
In result with last load estimation, the output of the neuron newly increased meets that the vector value of constraint carries out partially-initialized,
And on the upper SOM network foundations once trained, adjust the weight vector of newly-increased neuron node；
SOM hit predicted mechanism:When triumph neuron meets constraint corresponding to input pattern, loaded using SOM
Prediction, i.e., use using SOM as grader, otherwise will carry out load estimation using the mode of linear regression；
SOM state machines：With online learning art, after prediction each time is good, current data can be also utilized, to network
Enter Mobile state adjustment.
Further, in the step 3, discriminant function is utilizedFind the triumph god of the SOM networks trained
Through member.
Further, the step 4 comprises the following steps that：
(41) triumph neuron is solved；
(42) the SOM network conditions used according to user configuration and step (41) generate neuron triumph threshold value H and nerve
First triumph threshold value penalty factor σ；
(43) judge whether the Euclidean distance of triumph neuron is more than H (i) (1- σ), if it does, being determined as that needs add
Add neuron；Otherwise it is assumed that the input pattern belongs to the classification of the triumph neuron.
5. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that in the step 5, utilize
wj(n+1)=wj(n)+η(n)hj,i(x)(n)(X(n)-wj(n)), n=1,2 ..., T carry out the weighed value adjusting of neighborhood neuron, its
In, wjIt is output neuron j weights, n is iterations, and T refers to total iterations, and X represents input, and η is learning rate, and h is
The neighborhood function of weights, it is the function to the distance of winning neuron.
Further, three phases are included in the step 6, it is specific as follows：
Initial phase：
1.1st, the situation of input pattern is predicted according to known load, initializes neuron number, random initializtion neuron
Weight vector w；
1.2nd, triumph neuron threshold value is calculated according to step 1.1 and load deviation tolerance；
Build phase：
2.1st, the given data source of input initialization is added after initial phase into SOM networks, topology is calculated, calculates and appoint
Topological data set is calculated caused by business；
2.2nd, arg min are utilizedj||x(n)-wj| |, j=1,2 ..., l finds neuron of being won in traditional SOM algorithms；
2.3rd, judge whether triumph neuron is more than triumph neuron threshold value, if it is lower, jumping to step 2.5；If
It is more than, jumps to step 2.4；
If the 2.4, triumph neuron is boundary node, increases neuron number, that is, add neuron, jump to step
2.6；If triumph neuron is not boundary node, step 2.5 is jumped to；
2.5th, the weight vector of neighborhood neuron in SOM networks is updated, that is, updates SOM network output layers, triumph neuron
Neighbouring neuron, go to step step 2.7；
If the 2.6, having newly increased neuron, the node newly increased is initialized using the weight vector of triumph neuron, and
It is initial value to reset to initial value and adjust neighborhood learning rate, and field refers to add the neck of the triumph neuron after new neuron
Domain neuron, go to step step 2.7；
2.7th, repeat step 2.2 arrives step 2.6, is tended towards stability until for existing data, Clustering Effect.
Sorting phase：
3.1st, learning rate is reduced after build phase, improves classification accuracy；
3.2nd, triumph neuron is found, all known loads are obtained from triumph neuron, it is calculated and is averaged, loaded
The result of prediction.
Further, in the step 7, when SOM network triumph neurons are judged to not meeting, the side of linear regression is utilized
Formula is predicted, and obtains real load information again afterwards, is added SOM networks and is trained.
In summary, by adopting the above-described technical solution, the beneficial effects of the invention are as follows：
1st, present invention incorporates the characteristics of the data source in stream process and computation schema, the same of the speed of prediction is being improved
When, it ensure that forecasting accuracy；
2nd, in the present invention when the input pattern occurred in data flow system is determined as new input by existing network, move
State adds neuron, and returns to prediction result using linear regression prediction mode, and the accuracy of this predicted value is calculated less than SOM
Method, but simple and fast, it can just be used linearly only when triumph neuron is determined incongruent using regression forecasting mode
The mode of recurrence is predicted, it is ensured that has result return, after waiting the adjustment of SOM networks stable, prediction next time, which is still, to be passed through
SOM networks.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of the present invention；
Fig. 2 is SOM network state transition diagrams in the present invention；
Fig. 3 is the schematic diagram that neuron is dynamically added in the present invention.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
SOM operation principle is made up of three processes：Competition, cooperation and adaptive.(1) competition process.In this process,
It can be calculated according to different input patterns using discriminant function, and select the maximum neuron of discriminant score as the god that wins
Through member.Input vector isCorresponding output neuron i weight vectorSentence
Other functionF (*) is linear function, to cause y maximum, it is necessary to the inner product of vectorValue is maximum, and
When input vector and weight vector are all normalized, the inner product value maximum is just equivalent to the European of input vector and weight vector
Distance is minimum.Therefore, when finding best match neuron, usually using the method for discrimination of minimum euclidean distanceAccord with
The neuron for closing following condition is best match neuron：Wherein, n is
The number of output neuron.(2) cooperative process.It can be cooperated with each other between adjacent neuron.In the neighbour of triumph neuron
In the range of domain, BMU determines the neighborhood position of triumph neuron.The tune of weight vectors will be carried out in the neighborhood of triumph neuron
It is whole.When training initial, contiguous range is generally large, and during training is carried out, the contiguous range can be gradually reduced.Neighborhood shape is more
For square, other regular polygons are alternatively sometimes.Neighborhood function embodies the effect of neighborhood, and the radius of neighbourhood can be used to count
Calculate the size of neighborhood.Neighborhood functionValue reflect triumph neuronInteraction to neuron j in neighborhood
(suppress or encourage).I.e. with triumph neuronFor the center of circle, excitability side feedback is had to similar neuron, and to farther out
Neuron, then show the feedback of inhibition.Conventional neighborhood function is Gaussian function： It is adjacent
Domain radius sigma (n) is dynamic change, is gradually reduced with time n increase.σ (n)=σ0e-n/τWherein σ0For initial neighborhood half
Footpath, initial neighborhood are typically set to the half of whole output plane, and τ is time constant.(3) cynapse adaptive process.Neighborhood function
, can be to the weight w of neuron j in neighborhood after it is determined thatjAdjust accordingly.The power of neuron in triumph neuron and its topological neighborhood
Value vector can changing form to do and update with Hebb learning rules.Neuron j adjustment formula is wn(n+1)=wj(n)+η(n)
hj,i(x)(n)(X(n)-wj(n) wherein, η is one and is more than 0 constant for being less than 1, referred to as learning efficiency by), n=1,2 ..., T, with
The conversion for the time gradually decreases to 0, is typically adjusted in the following way：Wherein T is total iterations, and η (0) is initial learning efficiency.
In the present invention, basic ideas are to load similar principle caused by utilizing similar computation schema, by similar load
Cluster, so as to reach the purpose of load estimation.
As Fig. 1 shows the flow of whole algorithm, comprise the following steps：
Step 1：Initialization given data source, calculate topology, subscribe to calculating topological data set caused by (calculating task).It is defeated
The type of incoming vector combination data source and all possible calculating topological operator, value are 0 or 1, and whether representative is calculated,
Form n-dimensional vector.On this basis, plus data amount check, calculated for reducing the load estimation of same data type, in this base
Cycle time factor is added on plinth, forms n+2 dimensional vectors；Given data source, topology is calculated, this is subscribed to and is three kinds and need calculating to open
The data of pin, data stream load is predicted by them.
Step 2：According to the SOM networks after topological data set training optimization, the SOM networks after optimization include newly-increased neuron
Weight vector initializes the tactful, mechanism such as SOM hit predicted mechanism, SOM state machines to improve operational efficiency.
Wherein, neuron weight vector initialization strategy is：The efficiency of e-learning is initial by SOM network connection weights
The influence of value is very big.If the weights of newly-increased neuron use consistent random initializtion strategy with whole network weights, can make
Still reached into successive ignition less than desired learning rate, in order to accelerate the training of network, use the result of last load estimation
In, the output of the neuron newly increased meets that the weight vector of the triumph neuron of constraint initializes the node as a part,
The vector value that weight vector also meets constraint using the output of the neuron in the result of last load estimation, newly increased is carried out
Partially-initialized, and on the basis of original training result, adjust the weight vector of new node.
SOM hit predicted mechanism：Load estimation is carried out using SOM algorithms, mainly can be by between input pattern
Similitude judges loading condition, and similar input pattern is judged as same category.So when input pattern does not meet triumph god
, it is necessary to add neuron when constraint through member, known load information, the opposing party are not on the one hand included in new neuron
Face, the complexity for clustering training process are higher.In order to solve the two problems, a kind of SOM hit predicteds mechanism is proposed, is being inputted
When triumph neuron meets constraint corresponding to pattern, the result just predicted using SOM, that is, make using SOM as grader
With, otherwise will using linear regression mode to carry out load estimation.Then, the real information of load is accessed, is added to nerve
In member.For input matrix X, regression coefficient is stored in vectorial w, and the result of prediction can pass through Y=XTW is provided.In order that
Prediction effect is best, is weighed using square error：It is expressed in matrix as (y-Xw)T(y-Xw), w is asked
Lead, make it be equal to zero, it is as follows to solve w：It is current to calculate using historic loadSo as to
Load is predicted.
Hit predicted mechanism：According to SOM hit predicted mechanism, if miss, SOM networks can be caused to be instructed again
Practice, and the time complexity of training process is higher, can not meet the requirement of data flow processing system real-time.It is proposed to this end that
The concept of SOM state machines, safeguard two SOM networks.
Fig. 2 is the state transition graph of SOM networks, and right side is the SOM graders for load estimation.
A) starting stage, left side are identical with right side；
B) when adding neuron, left side is transformed into the addition neuron stage；
C) training stage is subsequently entered；
D) after training is completed, into the stabilization sub stage, at this moment, the SOM neutral nets for replicating left side replace the classification on right side
Device
From above-mentioned steps, right side is always stable SOM graders, and the efficiency that it is calculated is higher, can reach several
According to the requirement of stream processing system, left side is the SOM networks being trained in iteration.
Three kinds of optimal ways above are provided to improve operational efficiency of the SOM networks in stream process load running.
Step 3：When new calculating task arrives, with step 1, propose that data characteristics (builds n+ in step 1
The process of 2 dimensional vectors is exactly the process of an extraction data characteristics.), utilize discriminant functionFind SOM networks
Triumph neuron,
Step 4：Judge whether triumph neuron meets constraint, if met, jump to step 5, otherwise jump to step
6.In decision process, triumph neuron threshold value and triumph neuron threshold value penalty factor are added.Define neuron triumph threshold value H
(i)：Represent when neuron i wins, judge the threshold value of load class accuracy.That is H (i)<arg minj||x(n)-wj||,j
=1,2 ..., l represent that load input pattern is similar, and cluster load estimation is effective, otherwise it is assumed that the type input pattern is new
Input pattern, neuron should be added, wherein, x (n) is input data, wjIt is output neuron j weights, j is output nerve
The sequence number of member.Define neuron triumph threshold value penalty factor σ：Punish=H (i) σ it is determined that neuron triumph threshold value when
Wait, it is necessary to dynamically adjusted according to current SOM network conditions and tolerable error, wherein, penalty factor is according to currently
Operation conditions and the value of tolerable load error dynamic adjustment, and tolerable error is closely related with system.Different is
Its value of uniting is different.If triumph neuron threshold value is too small, neuron can be frequently added, aggravates model training burden；If threshold value mistake
Greatly, the inaccuracy of load estimation can be caused, so according to current operation conditions and tolerable load error, dynamic adjustment punishment
The value of the factor, reach load estimation expected effect.It is defaulted as 0.Specific judgment mode is as follows：(1) triumph neuron is solved；
(2) according to user configuration, (user configuration includes SOM network structures, weighed value adjusting function (such as sombrero function), nerve of winning
First generation rule.) and Current Situation of Neural Network situation generation triumph neuron threshold value H and triumph neuron threshold value penalty factor σ；
(3) relation between the Euclidean distance of triumph neuron and H (i) (1- σ) is judged；(4) if triumph neuron it is European away from
From more than H (i) (1- σ), it is judged to needing to add neuron；Otherwise it is assumed that the input pattern belongs to the class of the triumph neuron
Not.
Step 5：Load estimation, output loading prediction result are carried out by SOM networks, while carries out neighborhood neuron and (refers to
SOM (Self-organizing Maps) network output layer, the peripheral nerve member of triumph neuron, the training process of SOM networks, distance is won refreshing
More remote through member, the adjustment of weights is more suppressed.) weighed value adjusting, and jump to step 8.The SOM network classifications contain one
The similar load information of class, the predicted value averaged as load, and utilize wj(n+1)=wj(n)+η(n)hj,i(x)(n)(X
(n)-wj), (n) n=1,2 ..., T carries out the weighed value adjusting of neighborhood neuron；
Step 6：Dynamic addition neuron (if not meeting constraint, illustrate the number of the output layer neuron of SOM networks,
Select very few, sorted thick, it is necessary to add new neuron, neuron here just refers to SOM output layer neurons.), divide again
For three phases；
1) initial phase
A) according to the situation of known load input pattern, initialization neuron number (being defaulted as 4), random initializtion nerve
First weight vector w；
B) triumph neuron threshold value constraint H is calculated according to load deviation tolerance, by inputting the minimum Euler with weights
Distance determines triumph neuron, but this minimum Euler's distance is sometimes still excessive, at this moment will add new neuron,
And this basis for estimation is exactly load deviation tolerance；
2) build phase
A) the given data source of input initialization is added after initial phase into network, calculates topology, calculating task production
Raw calculating topological data set；
B) arg min are utilizedj||x(n)-wj| |, j=1,2 ..., l finds neuron of being won in traditional SOM algorithms；
Whether the neuron for c) judging to win is more than threshold value, if it is lower, jumping to step e；If it does, jump to step
It is rapid d)；
If d) triumph neuron is boundary node, increases neuron number, jump to step f；If triumph neuron
It is not boundary node, continues；
E) weight vector of neighborhood neuron is updated；
If f) having newly increased neuron, the node newly increased is initialized using the weight vector of triumph neuron；
If g) having newly increased neuron, it is initial value to reset to initial value and adjust neighborhood learning rate；
H) repeat step b to step g, tends towards stability until for existing data, Clustering Effect；
3) sorting phase
A) learning rate is reduced after build phase, improves classification accuracy, learning rate is one and subtracts letter on iterations
Number.Energy function in similar simulated annealing.So learning rate can reduce after iteration each time, until the threshold less than setting
It is 0 after value.
B) triumph neuron is found, all known loads are obtained from triumph neuron, it is calculated and is averaged, it is pre- to represent load
The result of survey.
If initial phase is very suitable, then the neuron of build phase addition will be less, and training is faster.Only
After build phase is stable, sorting phase just has correctly stable classification.
Step 7：Pass through equation of linear regressionLoad is predicted, and obtains real load information,
When SOM network triumph neurons are judged to not meeting, it is predicted, is obtained again afterwards true negative using the mode of linear regression
Information carrying ceases, and adds SOM networks and is trained.；
Step 8：Re -training SOM networks, prepared for prediction next time.
Compared with prior art, a kind of data stream load Forecasting Methodology provided by the present invention, is combined in stream process
The characteristics of data source and computation schema, while the speed of prediction is improved, it ensure that forecasting accuracy.When in data flow system
When the input pattern of appearance is determined as new input by existing network, dynamic adds neuron, and uses linear regression prediction
Mode returns to prediction result, and the accuracy of this predicted value is less than SOM algorithms, but simple and fast, classifies in SOM miss
When can be used as one preferably supplement.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
All any modification, equivalent and improvement made within refreshing and principle etc., should be included in the scope of the protection.
Claims (7)
1. a kind of data stream load Forecasting Methodology, comprises the following steps：
Step 1：Initialize given data source, calculate the caused calculating topological data set of topology, calculating task；
Step 2：According to the SOM networks after topological data set training optimization；
Step 3：The data characteristics of new calculating task is extracted, finds the triumph neuron of the SOM networks trained；
Step 4：Judge whether triumph neuron meets constraint, if met, jump to step 5, otherwise jump to step 6；
Step 5：Load estimation is carried out by the SOM networks that train, output loading prediction result, while in SOM networks
The weights of neighborhood neuron are adjusted, and jump to step 8；
Step 6：Neuron is added to SOM network dynamics；
Step 7：Load estimation is carried out by linear regression method, and obtains real load information；
Step 8：Load estimation result re -training SOM networks are added, are prepared for prediction next time.
2. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that in the step 2, after optimization
SOM networks include,
Newly-increased neuron weight vector initialization strategy：In result using last load estimation, the neuron that newly increases
The weight vector of the triumph neuron of output satisfaction constraint initializes the neuron node as a part, and weight vector utilizes upper
In the result of load estimation, the vector value progress partially-initialized of the output satisfaction constraint of the neuron newly increased, and
On the SOM network foundations that last time trains, the weight vector of newly-increased neuron node is adjusted；
SOM hit predicted mechanism:When triumph neuron meets constraint corresponding to input pattern, load estimation is carried out using SOM,
Used using SOM as grader, otherwise will carry out load estimation using the mode of linear regression；
SOM state machines：With online learning art, after prediction each time is good, current data can be also utilized, network is carried out
Dynamic adjusts.
3. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that in the step 3, utilize differentiation
FunctionFind the triumph neuron of the SOM networks trained.
A kind of 4. data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that the specific steps of the step 4
It is as follows：
(41) triumph neuron is solved；
(42) the SOM network conditions generation neuron triumph threshold value H and neuron used according to user configuration and step (41) is obtained
Win threshold value penalty factor σ；
(43) judge whether the Euclidean distance of triumph neuron is more than H (i) (1- σ), if it does, being judged to needing to add god
Through member；Otherwise it is assumed that the input pattern belongs to the classification of the triumph neuron.
5. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that in the step 5, utilize wj(n+
1)=wj(n)+η(n)hj,i(x)(n)(X(n)-wj(n)), n=1,2 ..., T carry out the weighed value adjusting of neighborhood neuron, wherein, wj
It is output neuron j weights, n is iterations, and T refers to total iterations, and X represents input, and η is learning rate, and h is weights
Neighborhood function, it is the function to the distance of winning neuron.
6. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that three are included in the step 6
It is stage, specific as follows：
Initial phase：
1.1st, the situation of input pattern is predicted according to known load, initializes neuron number, the power of random initializtion neuron
It is worth vectorial w；
1.2nd, triumph neuron threshold value is calculated according to step 1.1 and load deviation tolerance；
Build phase：
2.1st, the given data source of input initialization is added after initial phase into SOM networks, calculates topology, calculating task production
Raw calculating topological data set；
2.2nd, arg min are utilizedj||x(n)-wj| |, j=1,2 ..., l finds neuron of being won in traditional SOM algorithms；
2.3rd, judge whether triumph neuron is more than triumph neuron threshold value, if it is lower, jumping to step 2.5；If it does,
Jump to step 2.4；
If the 2.4, triumph neuron is boundary node, increases neuron number, that is, add neuron, jump to step 2.6；
If triumph neuron is not boundary node, step 2.5 is jumped to；
2.5th, the weight vector of neighborhood neuron in SOM networks is updated, i.e. renewal SOM network output layers, near triumph neuron
Neuron, go to step step 2.7；
If the 2.6, having newly increased neuron, the node newly increased is initialized using the weight vector of triumph neuron, and will learn
It is initial value that habit rate, which resets to initial value and adjusts neighborhood, and field refers to add the field god of the triumph neuron after new neuron
Through member, step step 2.7 is gone to；
2.7th, repeat step 2.2 arrives step 2.6, is tended towards stability until for existing data, Clustering Effect.
Sorting phase：
3.1st, learning rate is reduced after build phase, improves classification accuracy；
3.2nd, triumph neuron is found, all known loads are obtained from triumph neuron, it is calculated and is averaged, obtain load estimation
Result.
7. a kind of data stream load Forecasting Methodology as claimed in claim 1, it is characterised in that in the step 7, when SOM nets
When network triumph neuron is judged to not meeting, it is predicted using the mode of linear regression, obtains real load information again afterwards,
SOM networks are added to be trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710876477.XA CN107578101B (en) | 2017-09-25 | 2017-09-25 | Data stream load prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710876477.XA CN107578101B (en) | 2017-09-25 | 2017-09-25 | Data stream load prediction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107578101A true CN107578101A (en) | 2018-01-12 |
| CN107578101B CN107578101B (en) | 2020-06-09 |
Family
ID=61039864
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710876477.XA Active CN107578101B (en) | 2017-09-25 | 2017-09-25 | Data stream load prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107578101B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635948A (en) * | 2018-12-19 | 2019-04-16 | 北京达佳互联信息技术有限公司 | On-line training method, apparatus, system and computer readable storage medium |
| CN111082997A (en) * | 2019-12-30 | 2020-04-28 | 西安电子科技大学 | Network function arrangement method based on service identification in mobile edge computing platform |
| CN111915102A (en) * | 2020-08-22 | 2020-11-10 | 武汉空心科技有限公司 | Load prediction-based work platform workload prediction method and system |
| CN111932396A (en) * | 2020-06-05 | 2020-11-13 | 国网江苏省电力有限公司 | Automatic identification method for power distribution network topology network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103984867A (en) * | 2014-05-20 | 2014-08-13 | 南京理工大学 | Recession analysis method for virtual machine monitor in manner of context awareness |
| US20160078368A1 (en) * | 2010-05-26 | 2016-03-17 | Automation Anywhere, Inc. | Artificial intelligence & knowledge based automation enhancement |
| CN106161120A (en) * | 2016-10-08 | 2016-11-23 | 电子科技大学 | The distributed meta-data management method of dynamic equalization load |
| CN106679880A (en) * | 2016-12-21 | 2017-05-17 | 华南理工大学 | Pressure sensor temperature compensating method based on FOA-optimized SOM-RBF |
-
2017
- 2017-09-25 CN CN201710876477.XA patent/CN107578101B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160078368A1 (en) * | 2010-05-26 | 2016-03-17 | Automation Anywhere, Inc. | Artificial intelligence & knowledge based automation enhancement |
| CN103984867A (en) * | 2014-05-20 | 2014-08-13 | 南京理工大学 | Recession analysis method for virtual machine monitor in manner of context awareness |
| CN106161120A (en) * | 2016-10-08 | 2016-11-23 | 电子科技大学 | The distributed meta-data management method of dynamic equalization load |
| CN106679880A (en) * | 2016-12-21 | 2017-05-17 | 华南理工大学 | Pressure sensor temperature compensating method based on FOA-optimized SOM-RBF |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635948A (en) * | 2018-12-19 | 2019-04-16 | 北京达佳互联信息技术有限公司 | On-line training method, apparatus, system and computer readable storage medium |
| CN111082997A (en) * | 2019-12-30 | 2020-04-28 | 西安电子科技大学 | Network function arrangement method based on service identification in mobile edge computing platform |
| CN111082997B (en) * | 2019-12-30 | 2021-05-14 | 西安电子科技大学 | Network function arrangement method based on service identification in mobile edge computing platform |
| CN111932396A (en) * | 2020-06-05 | 2020-11-13 | 国网江苏省电力有限公司 | Automatic identification method for power distribution network topology network |
| CN111932396B (en) * | 2020-06-05 | 2022-11-04 | 国网江苏省电力有限公司 | Automatic identification method for power distribution network topology network |
| CN111915102A (en) * | 2020-08-22 | 2020-11-10 | 武汉空心科技有限公司 | Load prediction-based work platform workload prediction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107578101B (en) | 2020-06-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kohler et al. | PSO+: A new particle swarm optimization algorithm for constrained problems | |
| CN107578101A (en) | A kind of data stream load Forecasting Methodology | |
| Zhang et al. | Typical wind power scenario generation for multiple wind farms using conditional improved Wasserstein generative adversarial network | |
| CN102567391B (en) | Method and device for building classification forecasting mixed model | |
| CN112329948B (en) | Multi-agent strategy prediction method and device | |
| CN104636801B (en) | A kind of prediction transmission line of electricity audible noise method based on Optimized BP Neural Network | |
| CN107316099A (en) | Ammunition Storage Reliability Forecasting Methodology based on particle group optimizing BP neural network | |
| WO2022083009A1 (en) | Customized product performance prediction method based on heterogeneous data error compensation fusion | |
| WO2021164118A1 (en) | Ac/dc bulk power grid dynamic security risk situation rolling prospective early warning method and system | |
| CN109413710B (en) | Clustering method and device of wireless sensor network based on genetic algorithm optimization | |
| Li et al. | Hybrid optimization algorithm based on chaos, cloud and particle swarm optimization algorithm | |
| CN103971160A (en) | Particle swarm optimization method based on complex network | |
| Kucuktezcan et al. | A new dynamic security enhancement method via genetic algorithms integrated with neural network based tools | |
| CN108594654A (en) | A kind of overhead crane anti-swing control method based on two type fuzzy sliding modes | |
| CN112990485A (en) | Knowledge strategy selection method and device based on reinforcement learning | |
| CN110414718A (en) | A kind of distribution network reliability index optimization method under deep learning | |
| Zhao et al. | Particle swarm optimization algorithm with self-organizing mapping for Nash equilibrium strategy in application of multiobjective optimization | |
| CN105844334B (en) | A kind of temperature interpolation method based on radial base neural net | |
| CN110163540A (en) | Electric power system transient stability prevention and control method and system | |
| CN115293052A (en) | Power system active power flow online optimization control method, storage medium and device | |
| CN114006370A (en) | Power system transient stability analysis and evaluation method and system | |
| Łapa et al. | Aspects of structure and parameters selection of control systems using selected multi-population algorithms | |
| Zhang et al. | Deep reinforcement learning for load shedding against short-term voltage instability in large power systems | |
| Ekárt et al. | Genetic programming with transfer learning for urban traffic modelling and prediction | |
| CN109697531A (en) | A kind of logistics park-hinterland Forecast of Logistics Demand method |
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 |