CN108446794A - One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques - Google Patents
One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques Download PDFInfo
- Publication number
- CN108446794A CN108446794A CN201810157943.3A CN201810157943A CN108446794A CN 108446794 A CN108446794 A CN 108446794A CN 201810157943 A CN201810157943 A CN 201810157943A CN 108446794 A CN108446794 A CN 108446794A
- Authority
- CN
- China
- Prior art keywords
- data
- convolutional neural
- neural networks
- input
- deep learning
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
- G06Q10/06315—Needs-based resource requirements planning or analysis
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W24/00—Supervisory, monitoring or testing arrangements
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Development Economics (AREA)
- General Business, Economics & Management (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Educational Administration (AREA)
- Health & Medical Sciences (AREA)
- Signal Processing (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to data identification and data representation technologies fields, disclose one kind and being based on multiple convolutional neural networks combination framework deep learning prediction techniques, and dimension transformation is carried out to data；Prepare training dataset, validation data set and test data set；Data after dimension is converted according to periodic state number it is corresponding input in different convolutional neural networks, through convolutional neural networks, treated that data obtain final result according to chronological order into the input full Connection Neural Network of depth after rearrangement；Training is terminated in advance using verification collection obtains model；Prediction test set obtains prediction result.The present invention handles the data of different cycles state targetedly mining data rule information respectively by multiple convolutional neural networks, reduce the network number of plies, the full Connection Neural Network of depth is inputted after convolutional neural networks treated data rearrangement sequence, substantially reduce data dimension, over-fitting is effectively alleviated, prediction accuracy is improved.
Description
Technical field
The invention belongs to data identification and data representation technologies fields, more particularly to one kind being based on multiple convolutional neural networks
In conjunction with framework deep learning prediction technique.
Background technology
Currently, the prior art commonly used in the trade is such：The prediction of cell flow is to alleviate the important step of cell pressure
Suddenly, predistribution of the situation of change of a cell flow for resource is predicted in advance, and the equilibrium of load has great significance, and leads to
After carrying out resource planning in advance, not only the pressure of base station side can reduce, and the usage experience of user also will be promoted significantly.Cell
Volume forecasting is a kind of time series forecasting problem, and time series is chronological set of number ordered series of numbers, is that reflection is a certain
The statistical indicator of phenomenon, time series forecasting problem is extremely common in practical applications, at present the time series forecasting side of mainstream
Method is mainly there are three major class, the respectively 1) method based on time series analysis, and 2) prediction technique based on statistical learning, 3) base
In the method for deep learning.The method of time series analysis is mainly started with from the angle of statistics, by sharp from historical series
Prediction of the suitable parameter realization to Future Data is found with modes such as sliding averages, typical Time series analysis method has
Autoregressive Integrated Moving Average (ARIMA) model, Three-exponential Smoothing algorithm etc.；Use system
The method for counting study mainly carries out model construction, traditional recurrence mould by the regression model that manual extraction feature is arranged in pairs or groups traditional again
Type includes linear regression, Gradient Boosting Decision Tree (GBDT) etc.；In time series problem, depth
It is Long Short-Term Memory (LSTM) neural network to learn common method, the network because of its memory characteristic, compared with
It is suitble to the modeling of time series problem, is widely applied in time series forecasting.Time series analysis method and depth
Requirement of the method for habit to data volume is relatively high, and the data of short time are difficult that the two models is made to reach satisfactory table
It is existing, therefore the data of short period generally learn to be modeled by the way of manual extraction feature using conventional statistics.And traditional system
The way of meter study manual extraction feature and the method for time series analysis, which all exist, needs a large amount of experiences and knowledge analysis data
The problem of, it needs constantly to be improved model and finds suitable parameter, this step needs a large amount of time.Use deep learning
Then can automatic learning characteristic, can be obtained good result in the case where not needing excessive artificial experience and participating in.In available letter
Cease it is more in the case of, can get the time series of a plurality of synchronization, i.e. multidimensional time-series, in cell flow forecasting problem
Cell flow data can not only be used, while the data such as number of users of cell can also be used for modeling and provide key message, for this
Kind various dimensions time series data, traditional Time series analysis method can not efficiently use various dimensions information, lead to model not
Can reflect the rule of data well, more additionally, due to cell flow noise, changing rule is more unstable, and base station by
Limitation on some hardware, data storage size is limited, the data of training can be utilized to fall short of, therefore time series analysis
Method is difficult to obtain satisfactory result.Traditional statistical learning method needs manual construction feature, this process needs big
The process of the experience of amount business relevant knowledge and Feature Engineering, manual extraction feature inevitably generates omission, the engineering lacked experience
Teacher had probably both taken a substantial amount of time, and could not also obtain preferable result.Equally, training data is less, noise compared with
In the case of more, traditional deep learning method, such as the full Connection Neural Network of LSTM, depth, in predicted time sequence problem
When using it is same throw the net network be fitted any time data, there are different rules, the nerve nets of shallow-layer for the data of different moments
Network expressive force is poor, can not precisely portray data situation, although complicated data can be presented more by increasing the network number of plies
Strong expressive force, but the parameter in network increases also with the intensification of the number of plies, and excessive parameter means the need to data volume
Ask larger, most probably lead to that over-fitting, especially data volume are smaller when data volume is smaller and when noise is more, the convergence feelings of network
Condition will be extremely unstable, and more difficult acquisition is caused to stablize outstanding prediction result.
In conclusion problem of the existing technology is：When traditional deep learning method is inadequate there are amount of training data
Unstable result causes prediction result inaccurate.
Solve the difficulty and meaning of above-mentioned technical problem：It is difficult to show the distribution of data using simple network structure, make
The rising that will lead to number of parameters with complicated network structure, can not be obtained when data volume is smaller stablize it is outstanding as a result, in number
The result difficulty that obtain outstanding stabilization in the case of inadequate according to amount is higher.Solve the problems, such as this by the time sequence smaller to data volume
Row forecasting problem brings prodigious help, and in cell load predicts scene, the increase of predictablity rate is for the prewired of resource
It is equipped with important guiding effect, user experience can be obviously improved and improve the level of resources utilization.
Invention content
In view of the problems of the existing technology, it is deep based on multiple convolutional neural networks combination frameworks that the present invention provides one kind
Degree study prediction technique.
The invention is realized in this way a kind of being based on multiple convolutional neural networks combination framework deep learning prediction techniques,
Described to transform the data into the dimension for n*m*k based on multiple convolutional neural networks combination framework deep learning prediction techniques, total n is a
Time point, the k dimension datas of m hour before each time point takes construct the matrix of n m*k；Use multiple convolutional neural networks point
The data at other places reason corresponding time point, the rule in data is effectively excavated by the network compared with shallow-layer；Use multiple convolution god
History relatively short-term data information is extracted respectively through network to integrate again.
Further, described to be included the following steps based on multiple convolutional neural networks combination framework deep learning prediction techniques：
Step 1 is n*m*k according to time series period n and data itself dimension transformation input data dimension；
Data are divided into training set, verification collection by step 2 according to the time, and test set is used for model training, assessment and survey
Examination；
Step 3, n identical convolutional neural networks of construction；
Step 4, by the data in step 1 according to its number of which state in the cycle input corresponding convolution
In neural network；
Step 5, by the output result of n convolutional neural networks sequentially in time into rearrangement, by the n after rearrangement
A value inputs the full Connection Neural Network of depth；
Step 6, setting network losses function, optimization algorithm, learning rate, batchsize；
Step 7, using verification collection training pattern in a manner of terminating in advance, training obtains model；
Step 8 predicts test set, and the model obtained according to verification collection accuracy rate situation of change using step 7 is surveyed
Try the prediction result of collection.
Further, in the step 4 by data according to its number of which state in the cycle input corresponding volume
Product neural network in process include：The time series data period is n, then corresponds to n convolutional neural networks, and input dimension is n*
M*k, by it is a kind of it is fixed in a manner of to cycle data number, number data correspond to input with number convolutional neural networks.
Further, in the step 5 by the output result of n convolutional neural networks sequentially in time into rearrangement packet
It includes：Sequence before inputting convolutional neural networks according to data to the data of 1~n of number of convolutional neural networks output is arranged again
Sequence, the time relationship of explicit data and time data to be measured；Data order is (n-p, n-p+ before inputting convolutional neural networks
1, n, 1, n-p-1) and, input convolutional neural networks are inputted according to number correspondence, the number after exporting
According to be rearranged for (n-p, n-p+1, n, 1, n-p-1) input the full Connection Neural Network of depth.
Another object of the present invention is to provide multiple convolutional neural networks combination framework depth are based on described in a kind of application
Learn the Time Series Analysis System of prediction technique.
In conclusion advantages of the present invention and good effect are：The present invention is more using that can be used in cell flow prediction
The characteristics of time series data, transforms the data into the dimension for n*m*k, total n time point, m small before each time point takes
When k dimension datas, that is, construct n m*k matrix.In view of the data of synchronization often have identical rule, therefore use
Multiple convolutional neural networks handle the data at corresponding time point respectively, the network compared with shallow-layer can be used effectively to excavate by the way
Go out the rule in data, not only used the information in a plurality of time series, moreover it is possible to which same volume is corresponded to by same time point
The mode of product neural network as shown in table 1, is being surveyed using less data, the Web Mining of more shallow-layer to more information
The present invention achieves the accuracy rate not less than conventional method in three cells of examination, and Average Accuracy promotion reaches 5% or more,
Effectively improve prediction accuracy.The equilibrium that the Accurate Prediction of cell future traffic prediction is pre-configured local resource, loads
There is important directive function, the improvement of the promotion and user experience for the level of resources utilization has important meaning, and the prediction is accurate
The promotion of true rate will promote user experience while reducing operator cost.
By the present invention in that the way that history relatively short-term data information is integrated again is extracted respectively with multiple convolutional neural networks,
The rule implied in data can be effectively excavated, while convolutional neural networks output treated result input depth connects god entirely
Through network, the dimension of data is greatly reduced, over-fitting can be effectively relieved, extracts useful information.The present invention uses multiple correspondences
The convolutional neural networks of different moments fully use the information for including in data as far as possible, while coordinating full connection nerve net
Network obtains output to the end, and the process of the full Connection Neural Network of depth is combined to from the output of multiple convolutional neural networks and can also be played
The effective effect for preventing over-fitting.
Description of the drawings
Fig. 1 is provided in an embodiment of the present invention based on multiple convolutional neural networks combination framework deep learning prediction technique streams
Cheng Tu.
Fig. 2 is the deep learning prediction provided in an embodiment of the present invention based on novel multiple convolutional neural networks combination frameworks
Method implementation flow chart.
Fig. 3 is network overall structure figure provided in an embodiment of the present invention.
Fig. 4 is the convolutional neural networks structure chart (convolutional neural networks fractional refinement in Fig. 3) that the present invention uses.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
The present invention is based on the deep learning prediction techniques of novel multiple convolutional neural networks combination frameworks, and this method can be effective
It is not adequately to generate very much stable and accurate result using multidimensional time-series information and in data volume.
As shown in Figure 1, provided in an embodiment of the present invention predicted based on multiple convolutional neural networks combination framework deep learnings
Method includes the following steps：
S101：Data are converted；
S102：Prepare training dataset, validation data set, test data set；
S103：Construction convolutional neural networks and depth fully-connected network simultaneously sequentially input data；
S104：Training is completed using training set and verification collection；
S105：Predict test set result.
It is provided in an embodiment of the present invention based on multiple convolutional neural networks combination framework deep learning prediction techniques include with
Lower step：
Step 1：It is n*m*k according to time series period n and data itself dimension transformation input data dimension.
Step 2：Data are divided into training set, verification collection according to the time, test set is used for model training, assessment and survey
Examination.
Step 3：Construct n identical convolutional neural networks.
Step 4：By the data in step 1 according to its number of which state in the cycle input corresponding convolution
In neural network.
Step 5：By the output result of n convolutional neural networks sequentially in time into rearrangement, by the n after rearrangement
A value inputs the full Connection Neural Network of depth.
Step 6：Network losses function, optimization algorithm, learning rate, batchsize etc. are set.
Step 7：Using verification collection training pattern in a manner of terminating in advance, training obtains model.
Step 8：It predicts test set, using the model obtained according to verification collection accuracy rate situation of change in step 7, obtains
The prediction result of test set.
In step 4 by data according to its number of which state in the cycle input corresponding convolutional neural networks
In process include：
Assuming that the time series data period is n, then n convolutional neural networks are corresponded to, input dimension is n*m*k, with one kind
Fixed mode numbers (such as 24 hours one day numbers) to cycle data, and number data correspond to convolutional Neural of the input with number
Network.
Include into rearrangement sequentially in time by the output result of n convolutional neural networks in step 5：
To convolutional neural networks output 1~n of number data according to data input convolutional neural networks before sequence
Rearrangement, the time relationship of explicit data and time data to be measured, such as：
Assuming that input convolutional neural networks before data order be (n-p, n-p+1, n, 1, n-p-
1), input convolutional neural networks are inputted according to number correspondence, and the data permutation after exporting is (n-p, n-p+
1, n, 1, n-p-1) and the input full Connection Neural Network of depth.
The present invention uses neural network forecast cell uplink data on flows, and the data that can be used for predicting 6 are tieed up totally, including：Cell
Average user number, cell maximum number of user, uplink traffic, uplink maximum stream flow, downlink traffic and downlink maximum stream flow.Cell number
Strong using day as the characteristic in period according to for hour level data, presenting, i.e. 24 time points are a cycle.
As shown in Fig. 2, provided in an embodiment of the present invention predicted based on multiple convolutional neural networks combination framework deep learnings
Method specifically includes following steps：
Step 1：Input data is transformed to the matrix of 24*6*6, wherein 24 all carry out for 24 time points before the time to be measured
This transformation is derived from the data at this time point forward 6 time points, data itself dimension is 6, therefore is obtained to 24 time points
Input data dimension 24*6*6 after transformation.
Step 2：Data are divided into training set, verification collection according to the time, test set is used for model training, assessment and survey
Examination.
Step 3：24 identical convolutional neural networks are constructed, network structure is as shown in figure 4, parameter configuration is as follows：
(1) input layer is the convolutional layer that convolution kernel is 2*2, totally 64 convolution kernels；
(2) second layer connects the convolutional layer that convolution kernel is 2*2, totally 64 convolution kernels；
(3) third layer is the average pond layer of 2*2, step-length 2；
(4) the 4th layers connect the convolutional layer that convolution kernel is 2*2, totally 16 convolution kernels；
(5) layer 5 connects the convolutional layer that convolution kernel is 2*2, totally 3 convolution kernels；
(6) it is 27 that layer 6, which connects input, exports the full articulamentum for 1, exports the web results；
(7) network parameter is initialized using Xavier initialization modes；
Step 4：By the data in step 1 according to its number of which state in the cycle input corresponding convolution
In neural network, that is, the data dimension of the data at preceding 24 time points, each time point is 6*6, and status number is 1
The data at time point input first convolutional neural networks, second convolutional neural networks of input that status is 2, with such
It pushes away, the 24th convolutional neural networks of output that residing hour is 0.
Step 5：By the output result of 24 convolutional neural networks sequentially in time into rearrangement, after rearrangement
24 value input full Connection Neural Networks of depth, overall network structure is as shown in figure 3, depth every layer of nerve of full Connection Neural Network
First number is followed successively by：24,256,256,128,64,1, initialization mode is to block normal state initialization.
Step 6：Setting network losses function is Absolute Error Loss, and optimization algorithm uses Adam algorithms, learning rate 1
×10-6, be trained using the mode of min-batch, batchsize 1, when training depth full Connection Neural Network last
The dropout that probability is 0.5 is added in layer.
Step 7：Using the mode training pattern terminated in advance, specific implementation mode is that every 400 wheel calculates one-time authentication
Collect accuracy, when continuous 50 records of verification collection accuracy, which are all not above the highest recorded before, once to be recorded, stops instruction
Get model.
Step 8：It predicts test set, using the model obtained according to verification collection accuracy rate situation of change in step 7, obtains
The prediction result of test set.
The application effect of the present invention is described in detail with reference to emulation.
1) data background is emulated
Using 3 cell history numbers of users, flow truthful data, for data from 2017-4-8 to 2017-6-15, there is portion in centre
Divide missing, totally 45 day data, it is expected that it is presetting operator can be helped to carry out rational resource with the load estimation that big data drives
Degree, reduces cost.Specific data include that cell average user number, cell maximum number of user, cell uplink flow, cell are maximum
Uplink traffic, cell downlink flow and cell maximum downstream flow.Simulation and prediction target：3 cells are upper one hour future
Row flow, evaluation index accuracy rate calculation formula are as follows：
Wherein C is cell number to be predicted, and T is period to be predicted, predi,tFor cell i t moment predicted value,
yi,tFor cell i t moment actual value.
2) emulation content and result
The present invention and LSTM and the full Connection Neural Network of depth are subjected to simulation comparison, the two all uses preceding 24 hour data
As feature, LSTM uses one layer, and the full Connection Neural Network of depth uses 6 layer networks, the remainder layer neuron in addition to input layer
Number is set as 300.Parameter of the present invention takes 2017-6-15 as test set.Table 1 shows：In multiple cell test sets, the present invention
Carried algorithm has compared with primal algorithm system performance to be obviously improved, and the present invention accumulates neural network fusion means by multireel,
The preferable information for having excavated multidimensional time-series, improves algorithm capability of fitting, accuracy rate has obviously in multiple cells
Rise, effectively improves the precision of prediction of algorithm.
Table 1
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 etc., should all be included in the protection scope of the present invention made by within refreshing and principle.
Claims (5)
1. one kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques, which is characterized in that described based on more
A convolutional neural networks combination framework deep learning prediction technique transforms the data into the dimension for n*m*k, total n time point, often
The k dimension datas of m hour before a time point takes, construct the matrix of n m*k；It is handled respectively using multiple convolutional neural networks pair
The data for answering time point effectively excavate the rule in data by the network compared with shallow-layer；Use multiple convolutional neural networks point
History relatively short-term data information is indescribably taken to integrate again.
2. being based on multiple convolutional neural networks combination framework deep learning prediction techniques as described in claim 1, feature exists
In described to be included the following steps based on multiple convolutional neural networks combination framework deep learning prediction techniques：
Step 1 is n*m*k according to time series period n and data itself dimension transformation input data dimension；
Data are divided into training set, verification collection by step 2 according to the time, and test set is used for model training, assessment and test；
Step 3, n identical convolutional neural networks of construction；
Step 4, by the data in step 1 according to its number of which state in the cycle input corresponding convolutional Neural
In network；
Step 5, by the output result of n convolutional neural networks sequentially in time into rearrangement, by n value after rearrangement
Input the full Connection Neural Network of depth；
Step 6, setting network losses function, optimization algorithm, learning rate, batchsize；
Step 7, using verification collection training pattern in a manner of terminating in advance, training obtains model；
Step 8 predicts test set, and the model obtained according to verification collection accuracy rate situation of change using step 7 obtains test set
Prediction result.
3. being based on multiple convolutional neural networks combination framework deep learning prediction techniques as claimed in claim 2, feature exists
In, in the step 4 by data according to its number of which state in the cycle input in corresponding convolutional neural networks
Process include：The time series data period is n, then corresponds to n convolutional neural networks, input dimension is n*m*k, solid with one kind
Fixed mode numbers cycle data, and number data correspond to convolutional neural networks of the input with number.
4. being based on multiple convolutional neural networks combination framework deep learning prediction techniques as claimed in claim 2, feature exists
In including into rearrangement sequentially in time by the output result of n convolutional neural networks in the step 5：To convolutional Neural
Network output 1~n of number data according to data input convolutional neural networks before sequence resequence, explicit data with
The time relationship of time data to be measured；Input convolutional neural networks before data order be (n-p, n-p+1, n,
1, n-p-1), input convolutional neural networks are inputted according to number correspondence, and the data permutation after exporting is
(n-p, n-p+1, n, 1, n-p-1) and the input full Connection Neural Network of depth.
5. pre- based on multiple convolutional neural networks combination framework deep learnings described in a kind of application Claims 1 to 4 any one
The Time Series Analysis System of survey method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810157943.3A CN108446794A (en) | 2018-02-25 | 2018-02-25 | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810157943.3A CN108446794A (en) | 2018-02-25 | 2018-02-25 | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108446794A true CN108446794A (en) | 2018-08-24 |
Family
ID=63192915
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810157943.3A Pending CN108446794A (en) | 2018-02-25 | 2018-02-25 | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108446794A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109460855A (en) * | 2018-09-29 | 2019-03-12 | 中山大学 | A kind of throughput of crowded groups prediction model and method based on focus mechanism |
| CN110018322A (en) * | 2019-04-18 | 2019-07-16 | 北京先见智控科技有限公司 | A kind of Rotating speed measring method and system based on deep learning |
| CN110175711A (en) * | 2019-05-17 | 2019-08-27 | 北京市天元网络技术股份有限公司 | One kind being based on joint LSTM base station cell method for predicting and device |
| CN110264251A (en) * | 2019-05-31 | 2019-09-20 | 杭州博钊科技有限公司 | The data organization form and the prediction technique based on multi-task learning for characterizing cash flow |
| CN111327453A (en) * | 2020-01-19 | 2020-06-23 | 国网福建省电力有限公司经济技术研究院 | Communication bandwidth estimation method considering gridding dynamic and static components |
| CN111445002A (en) * | 2019-01-16 | 2020-07-24 | 中国科学院上海高等研究院 | Energy demand prediction method/system, storage medium, and apparatus |
| CN112001519A (en) * | 2020-05-20 | 2020-11-27 | 国网浙江省电力有限公司 | Power load prediction method based on deep neural network |
| CN112182961A (en) * | 2020-09-23 | 2021-01-05 | 中国南方电网有限责任公司超高压输电公司 | Large-scale fading modeling prediction method for wireless network channel of converter station |
| CN112215390A (en) * | 2019-07-12 | 2021-01-12 | 国际商业机器公司 | Performing multivariate time series prediction using three-dimensional transforms |
| CN112637904A (en) * | 2019-09-24 | 2021-04-09 | 中国移动通信集团陕西有限公司 | Load balancing method and device and computing equipment |
| CN112764932A (en) * | 2021-01-27 | 2021-05-07 | 西安电子科技大学 | Deep reinforcement learning-based calculation-intensive workload high-energy-efficiency distribution method |
| CN112949107A (en) * | 2019-12-10 | 2021-06-11 | 通用汽车环球科技运作有限责任公司 | Composite neural network architecture for stress distribution prediction |
| CN113347657A (en) * | 2021-04-01 | 2021-09-03 | 青岛科技大学 | Method for predicting safety capacity performance of mobile communication system |
| CN113408770A (en) * | 2020-03-16 | 2021-09-17 | 中国人民解放军陆军装甲兵学院 | Equipment maintenance time prediction method based on deep learning |
| CN115050006A (en) * | 2022-06-21 | 2022-09-13 | 荆州职业技术学院 | Lane departure early warning method and system based on convolutional neural network |
| CN115238763A (en) * | 2022-06-10 | 2022-10-25 | 西北工业大学 | Multi-period hydrodynamic experimental data enhancement method of simulated bat aircraft based on deep learning |
-
2018
- 2018-02-25 CN CN201810157943.3A patent/CN108446794A/en active Pending
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109460855A (en) * | 2018-09-29 | 2019-03-12 | 中山大学 | A kind of throughput of crowded groups prediction model and method based on focus mechanism |
| CN111445002A (en) * | 2019-01-16 | 2020-07-24 | 中国科学院上海高等研究院 | Energy demand prediction method/system, storage medium, and apparatus |
| CN111445002B (en) * | 2019-01-16 | 2023-10-31 | 中国科学院上海高等研究院 | Energy demand prediction method/system, storage medium and apparatus |
| CN110018322A (en) * | 2019-04-18 | 2019-07-16 | 北京先见智控科技有限公司 | A kind of Rotating speed measring method and system based on deep learning |
| CN110175711A (en) * | 2019-05-17 | 2019-08-27 | 北京市天元网络技术股份有限公司 | One kind being based on joint LSTM base station cell method for predicting and device |
| CN110264251A (en) * | 2019-05-31 | 2019-09-20 | 杭州博钊科技有限公司 | The data organization form and the prediction technique based on multi-task learning for characterizing cash flow |
| CN110264251B (en) * | 2019-05-31 | 2021-08-10 | 杭州博钊科技有限公司 | Data organization form for representing cash flow and prediction method based on multi-task learning |
| CN112215390A (en) * | 2019-07-12 | 2021-01-12 | 国际商业机器公司 | Performing multivariate time series prediction using three-dimensional transforms |
| CN112637904B (en) * | 2019-09-24 | 2024-04-12 | 中国移动通信集团陕西有限公司 | Load balancing method and device and computing equipment |
| CN112637904A (en) * | 2019-09-24 | 2021-04-09 | 中国移动通信集团陕西有限公司 | Load balancing method and device and computing equipment |
| CN112949107A (en) * | 2019-12-10 | 2021-06-11 | 通用汽车环球科技运作有限责任公司 | Composite neural network architecture for stress distribution prediction |
| CN111327453A (en) * | 2020-01-19 | 2020-06-23 | 国网福建省电力有限公司经济技术研究院 | Communication bandwidth estimation method considering gridding dynamic and static components |
| CN111327453B (en) * | 2020-01-19 | 2023-04-07 | 国网福建省电力有限公司经济技术研究院 | Communication bandwidth estimation method considering gridding dynamic and static components |
| CN113408770A (en) * | 2020-03-16 | 2021-09-17 | 中国人民解放军陆军装甲兵学院 | Equipment maintenance time prediction method based on deep learning |
| CN112001519A (en) * | 2020-05-20 | 2020-11-27 | 国网浙江省电力有限公司 | Power load prediction method based on deep neural network |
| CN112182961B (en) * | 2020-09-23 | 2023-01-31 | 中国南方电网有限责任公司超高压输电公司 | Converter station wireless network channel large-scale fading modeling prediction method |
| CN112182961A (en) * | 2020-09-23 | 2021-01-05 | 中国南方电网有限责任公司超高压输电公司 | Large-scale fading modeling prediction method for wireless network channel of converter station |
| CN112764932B (en) * | 2021-01-27 | 2022-12-02 | 西安电子科技大学 | Deep reinforcement learning-based calculation-intensive workload high-energy-efficiency distribution method |
| CN112764932A (en) * | 2021-01-27 | 2021-05-07 | 西安电子科技大学 | Deep reinforcement learning-based calculation-intensive workload high-energy-efficiency distribution method |
| CN113347657B (en) * | 2021-04-01 | 2022-04-29 | 青岛科技大学 | Method for predicting safety capacity performance of mobile communication system |
| CN113347657A (en) * | 2021-04-01 | 2021-09-03 | 青岛科技大学 | Method for predicting safety capacity performance of mobile communication system |
| CN115238763A (en) * | 2022-06-10 | 2022-10-25 | 西北工业大学 | Multi-period hydrodynamic experimental data enhancement method of simulated bat aircraft based on deep learning |
| CN115238763B (en) * | 2022-06-10 | 2024-02-13 | 西北工业大学 | Multi-period hydrodynamic experimental data enhancement method of simulated batline aircraft based on deep learning |
| CN115050006A (en) * | 2022-06-21 | 2022-09-13 | 荆州职业技术学院 | Lane departure early warning method and system based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108446794A (en) | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques | |
| CN108846517B (en) | Integration method for predicating quantile probabilistic short-term power load | |
| CN104951425B (en) | A kind of cloud service performance self-adapting type of action system of selection based on deep learning | |
| CN102567391B (en) | Method and device for building classification forecasting mixed model | |
| CN107330514B (en) | Air quality prediction method based on integrated extreme learning machine | |
| CN108764273A (en) | A kind of method, apparatus of data processing, terminal device and storage medium | |
| CN103106535B (en) | Method for solving collaborative filtering recommendation data sparsity based on neural network | |
| CN109492748B (en) | Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network | |
| CN103197983B (en) | Service component reliability online time sequence predicting method based on probability graph model | |
| CN103108382A (en) | Heterogeneous network multi-attribute decision-making method based on network analytic hierarchy process | |
| CN112054943B (en) | Traffic prediction method for mobile network base station | |
| CN109558897A (en) | A kind of more situation extracting methods of Electric Power Network Planning based on data mining | |
| CN105893669A (en) | Global simulation performance predication method based on data digging | |
| CN108984830A (en) | A kind of building efficiency evaluation method and device based on FUZZY NETWORK analysis | |
| CN104050547A (en) | Non-linear optimization decision-making method of planning schemes for oilfield development | |
| CN110866640A (en) | Power load prediction method based on deep neural network | |
| CN113987936B (en) | Equipment test resource overall allocation method based on chaotic genetic algorithm | |
| CN109754122A (en) | A kind of Numerical Predicting Method of the BP neural network based on random forest feature extraction | |
| CN105335800A (en) | Method for forecasting electricity consumption of power consumers based on joint learning | |
| CN103530700B (en) | Urban distribution network saturation loading Comprehensive Prediction Method | |
| CN105260615B (en) | A kind of grain consumption Forecasting Methodology | |
| Lee et al. | Probabilistic wind power forecasting based on the laplace distribution and golden search | |
| CN104834975A (en) | Power network load factor prediction method based on intelligent algorithm optimization combination | |
| CN116894504A (en) | Wind power cluster power ultra-short-term prediction model establishment method | |
| CN116050595A (en) | Attention mechanism and decomposition mechanism coupled runoff amount prediction 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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180824 |