CN114238054A - Cloud server resource utilization quantity prediction method based on improved TFT - Google Patents
Cloud server resource utilization quantity prediction method based on improved TFT Download PDFInfo
- Publication number
- CN114238054A CN114238054A CN202111547972.9A CN202111547972A CN114238054A CN 114238054 A CN114238054 A CN 114238054A CN 202111547972 A CN202111547972 A CN 202111547972A CN 114238054 A CN114238054 A CN 114238054A
- Authority
- CN
- China
- Prior art keywords
- model
- prediction
- data
- variables
- time
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/30—Monitoring
- G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
- G06F11/3447—Performance evaluation by modeling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/30—Monitoring
- G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
- G06F11/3452—Performance evaluation by statistical analysis
Abstract
The invention provides a method for predicting the resource utilization quantity of a cloud server. The method improves a TFT (time sequence fusion transformer) multi-step time sequence prediction technology and is applied to prediction of the resource utilization quantity of the cloud server. The method comprises the following steps: decomposing characteristic variables of the data set; generating a training set test set; dividing data characteristic variables; designing and improving a TFT prediction model; different data characteristic variables enter corresponding interfaces of the model; training a model; generating a target prediction model; and (5) testing the model. The invention provides a new variable decomposition method, and the advantages of TFT variable classification extraction and multi-step fusion prediction are utilized, so that the model fitting speed is increased after the algorithm is improved, and the prediction accuracy and robustness are improved. The method solves the technical problems that the multi-step prediction accuracy is low due to the fact that characteristic variables of data of the cloud server resource utilization are too few at present, and meets the requirements of a cloud service platform on safe deployment and resource coordination through prediction of the utilization quantity of the server resources.
Description
Technical Field
The invention belongs to the technical field of communication, and particularly relates to a cloud server resource utilization quantity prediction method based on an improved TFT (time sequence fusion transformer model).
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The cloud service platform is a service platform for providing large-scale computing for enterprises by service providers under a certain network environment. Some conventional performance analysis methods cannot well solve the challenges in terms of cloud service platform deployment, resource utilization, architecture, technology and the like, and the problems can be solved to a certain extent through accurate prediction of the utilization quantity of server resources. On one hand, the resource utilization quantity of the cloud platform server can be predicted, and the platform user peak period which is possibly faced in the future can be predicted and deployed in advance, so that the server overload is prevented, and the access safety is ensured; on the other hand, the investment of the server can be reduced in the period of low resource utilization, thereby reducing the resource loss.
However, the cloud server service condition data generated by the cloud platform background only has one column of variable data corresponding to the timestamp, so that the prediction is limited, the characteristics cannot be accurately extracted during the prediction, and the traditional model can learn few things, so that the result is inaccurate. The traditional univariate prediction method generally adopts data of the first steps as model input and data of the next step or the later steps as model labels for training, and the prediction result is not accurate enough due to the lack of characteristic variables. At present, multivariate time series prediction research is still popular, but the multivariate time series prediction research cannot be used in univariate data of cloud server resource utilization quantity. Some conventional predictive models, such as: ARMA (autoregressive moving average model), ARIMA (summation autoregressive moving average model) and LSTM (long short term memory recurrent neural network) are popular prediction technologies for data analysts at present, but these technologies still have the defects of imperfect structure, inaccurate feature extraction, slow model fitting, inaccurate result, large resource consumption and the like, and cannot meet the requirements of the current market.
How to apply the latest multivariate prediction technology to the field of cloud server resource utilization quantity prediction and how to extract features and improve prediction accuracy to meet the requirement of accurate resource deployment is the key point of current research.
Disclosure of Invention
In order to solve the above problems, the present invention provides a method for predicting the resource utilization amount of a cloud server. The embodiment of the invention provides the following technical scheme: performing characteristic variable decomposition on the time stamp of the original data and the corresponding value by applying a time sequence decomposition function of the Prophet to obtain variables such as an integral trend item; manually increasing time items of year, month, week, day, festival and season for expanding an original data set as model input data; dividing data into three variables, inputting the three variables into different interfaces of the TFT model, and further calculating and analyzing the characteristics of the three variables; encoding and decoding the variables using an encoder and a decoder; and improving the attention mechanism to learn the characteristic variables, and finally outputting a prediction result. The improved TFT model greatly improves the model fitting speed and the accuracy of server utilization quantity data prediction.
The method comprises the following specific steps:
s 1: the method comprises the steps of obtaining data of the utilization amount of cloud server resources in a period of time to extract and decompose characteristic variables, and enabling the characteristic variables of two columns of original data (time stamps and actual variable values) to be expanded to multiple columns through a prediction model Prophet based on time series decomposition.
s 2: and selectively extracting the value output by the Prophet model, extracting a timestamp and an actual value corresponding to the input value, and acquiring a trend item dynamic variable calculated by the Prophet model to obtain a new data set.
s 3: and further manually expanding time sequence variables for the new data set, increasing the items of year, month, day, week, weekend, festival and season, and finally expanding two columns of original data into multi-column data to form a multi-variable data set as model input.
s 4: and dividing the characteristic variables of the multivariable input data into three types, namely static seasonal variables, past observation variables and known time variables.
s 5: designing and improving a TFT model, improving a GLU (gated linear unit) layer in a variable selection module, and realizing the function of enhancing the selective acquisition and forgetting of past information, wherein the calculation formula is as follows:
the formula I is as follows:
wherein said W1，W2，W3Is a weight parameter to be trained, b, c, d are bias parameters to be trained, gelu () is an activation function gelu,
is calculated for matrix multiplication.
s 6: improving an attention calculation mode in a TFT model, using a double-head attention mechanism, wherein Q, K and V are query, key and value obtained by multiplying input embedding by a weight matrix respectively, sampling K, randomly selecting n K, and obtaining KnTo q is pairediE Q solving the value of M with respect to Kn, the formula is as follows:
the formula II is as follows:
wherein q isi∈Q，kj∈Kn
Find the largest n q of MiForm Qn With respect to KnObtaining A (Q)n,KnV), the formula is as follows:
the formula III is as follows:
wherein QnFor selected n qiComposed matrix, no selected qiInitialization of the average value after A (Q, K, V) is solved to the original QrMatrix, QnUpdating non-0 values to Q in the matrixrObtaining a final Q matrix from the matrix;
the double-headed attention formula is as follows:
the formula four is as follows:
the formula five is as follows:
last output of attention mechanism
The value is obtained.
s 7: establishing a new TFT prediction model according to the improvements in the step s5 and the step s 6.
s 8: and (4) respectively accessing the static seasonal variable, the past observation variable and the time known variable which are divided in the step s4 into the interfaces of the input data of the static covariate, the dynamic time-varying characteristic variable and the dynamic time-invariant characteristic corresponding to the model, so that the model can perform different characteristic selection and calculation on different variables.
s 9: and training the model, setting the number of epochs to be 40 and the decoder prediction time step to be 165, and performing model training and building.
s 10: and testing from the test set to verify the accuracy of the established model.
In a further technical scheme, the Prophet model in the step s2 is only used for data processing, the time series characteristic variable decomposition operation is performed by using the Prophet prediction model, the Prophet prediction model is used as the input of the model, only the nodes in the original data are reserved, and the prediction result of the model is not used.
In a further technical solution, the manually extended time characteristic variable in step s3 may be added with a certain holiday term which affects the data result according to the actual situation of the data.
In a further technical solution, in the step s4, the static seasonal variable is a seasonal feature item of a current time point, the past observed variable is a known observed dynamic variable before the prediction point, and the known time variable is a time variable known in the whole prediction system, such as year, month, week, day, and the like.
In a further technical solution, in the step s5, a gate-controlled linear unit is applied to perform a calculation method for selectively memorizing and forgetting data characteristics.
In a further technical solution, all the formulas in the step s6 need to be used in a matching manner, and are calculated mutually.
In a further embodiment, the seasonal variables are input into the model as static covariates in step s8, so that the seasonal variables can coordinate global features.
The method for predicting the resource utilization quantity of the cloud server meets the requirement that a cloud service platform needs to accurately deploy the server in the future, solves the bottleneck problem of characteristic decomposition, improves the components and the computing mode by taking an improved TFT (time sequence fusion transformer) multi-step time sequence prediction model as a basic framework, and improves the accuracy of the resource utilization quantity prediction of the cloud server. By predicting the use number of the servers, the operation and maintenance efficiency of the cloud service platform is improved, and a platform maintainer can more accurately schedule various resources.
Drawings
For a clearer explanation of the embodiments or technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present description.
Fig. 1 is a flowchart of a method for predicting a resource utilization amount of a cloud server according to an embodiment of the present disclosure.
FIG. 2 is a flowchart of a data expansion method according to an embodiment of the present disclosure.
Fig. 3 is a flowchart of data classification according to an embodiment of the present disclosure.
FIG. 4 is a block diagram of model data input in accordance with an embodiment of the present disclosure.
Fig. 5 is a diagram of an internal structure of a TFT model used in an embodiment of the present specification.
Fig. 6 is a diagram illustrating a gated residual network in a model according to an embodiment of the present disclosure.
Detailed Description
The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific embodiments described herein are merely illustrative of implementations of the invention and do not delimit the invention.
Fig. 1 exemplarily shows a flow of a method for predicting the resource utilization amount of a cloud server based on an improved TFT.
Firstly, performing characteristic decomposition on original data, including time decomposition and variable decomposition, and adding corresponding variables and overall trend variables of the data and time characteristic items of year, month, week, day, holiday and the like as data characteristics in an input model.
And secondly, preprocessing data, standardizing and normalizing the data set, and generating a training set and a testing set for training and testing the model.
Thirdly, the data is divided according to characteristics, and according to a unique data input mechanism in the TFT model, the data needs to be divided into static characteristics, dynamic time-varying characteristics and dynamic time-invariant characteristics.
And fourthly, designing and improving the TFT model, changing the calculation of a multi-head attention mechanism and a gate control linear unit on the basis of the original model architecture, and improving the model fitting speed and the prediction accuracy.
Fifthly, the divided data characteristics are respectively accessed into the corresponding model interfaces.
And sixthly, training the model.
And seventhly, generating a training model for predicting the utilization quantity of the cloud server resources.
Eighth, the accuracy of the prediction of the test model is performed in the test set.
Step one, data set characteristic decomposition and pretreatment:
the raw data of the cloud server resource utilization amount comprises two columns including a timestamp and a variable data value every day, and the cloud server resource utilization condition data of a certain platform used by the raw data continuously for nearly 8 years every day is adopted in the embodiment.
Specifically, the original data is subjected to data feature expansion, as shown in fig. 2, which is a flowchart of a data expansion method according to an embodiment of the present disclosure. And (2) selectively extracting the value output by the Prophet model through the Prophet (time decomposition sequence prediction) model, extracting the time stamp and variable data value corresponding to the input value, acquiring the variables such as the trend item calculated by the Prophet, and the like, only reserving the nodes in the original data, and not using the prediction result of the model. The Prophet model is used as a time series prediction model based on mathematical statistics and machine learning, and the embodiment of the specification only uses the characteristic decomposition capability of the Prophet model as the data processing technology of the invention, and does not use the prediction made by the Prophet model. The invention also reasonably and scientifically increases the manual addition on the basis of the characteristic extraction of the Prophet model, increases time items such as year, month, week, day, holiday and the like, generates expanded data, and then preprocesses the expanded data, including the standardization and normalization of the data.
Step two, dividing training set and testing set
The preprocessed data is partitioned into training set test sets, where the example uses 365 data of the last year in the data set as the test set and data of the last few years as the training set.
Step three, data characteristic division
As shown in fig. 3, which is a flow chart of data classification in the embodiment of the present specification, and shows data classification on preprocessed data, the present invention innovatively modifies data classification bases in static features, dynamic time-varying features, and dynamic time-invariant features of data classification required in an original model according to uniqueness of data in the embodiment, and classifies data into static seasonal variables, past observation variables, and known time variables. The static seasonal variable is a seasonal characteristic item of the current time point, and can be used as a static covariate input model because the item is kept unchanged in certain continuous data and has a control effect on the whole, the past observation variable is a known observed dynamic variable before a prediction point, and the known time variable is a known time variable in the whole prediction system, such as year, month, week, day and the like.
Specifically, as shown in fig. 4, a diagram of a model data input structure for the embodiment of the present specification shows how classified data functions in a model. The model is subjected to time series prediction, factors influencing a prediction result not only comprise past known variables, but also comprise known time variables and static seasonal variables of corresponding prediction time points, when the model is trained, the past known variables are required to be input, such as server usage quantity values and trend items of dynamic fluctuation, future known time variables such as annual-year-month-day items of Monday, Tuesday, month-month, February and the like are used as auxiliary characteristics to provide relevant bases for prediction, and seasonal factors influencing the whole are also included. And on the premise of knowing time items such as seasons, years, months, days and the like at the t + h moment, the model finally outputs a predicted data result at the t + h moment.
Step four, designing and improving a TFT prediction model
Fig. 5 is a design diagram of an internal structure of an embodiment of the present specification using an improved TFT model. Static seasonal variables enter a decoder after passing through a feature selection module to generate different data vectors, known time variables enter an encoder after passing through the feature selection module in corresponding models, known dynamic variables enter the encoder and the decoder respectively after feature selection, and outputs of the encoder and the decoder respectively enter an attention mechanism after passing through a GLU (gated linear unit) and a GRU (gated residual error network). In this process, the static seasonal variables again affect the calculation of GRN, enhancing the static features. And finally, inputting the result in the attention mechanism into the GLU, the GRN and the GLU again in sequence, and outputting a predicted value through a full connection layer.
It should be noted that the GLU of this model is a structure for feature extraction, and the inspiration of the structure comes from a gating mechanism of LSTM (long-short term memory recurrent neural network), and can play a role in memorizing past valid information and selectively forgetting invalid information. The calculation formula of the GLU in the original TFT model is as follows:
wherein W1、W2Is the weight parameter to be trained, b, c are the weight parameter to be trained,
is the activation function sigmoid (x). The invention is improved into the following formula, and the calculation capability and the learning accuracy of the model are improved:
wherein W1、W2、W3And b, c and d are bias parameters to be trained, and gelu is an activation function gelu (x).
Specifically, a GRN (gated residual network) structure of the model is shown in fig. 6, and the structure constitutes a feature selection module, and the module realizes a feature selection function by calculating a feature weight, specifically, the GRN module is applied to each feature individually, and then all the features are applied to the GRN module in series, and a weighted sum of each GRN output is generated after obtaining the weight, and the weighted sum is used as the weight of each feature.
It should be noted that, the invention improves the calculation method of the multi-head attention mechanism, the calculation method of the attention parameter in the original transform is too large, when the invention calculates the attention of the multi-head, firstly, the calculation method of the attention of the single head is changed, and when the invention is used for solving the attention of the double heads, the two heads are directly averaged. The concrete implementation is as follows:
improving an attention calculation mode in a TFT (thin film transistor), using a double-head attention mechanism, wherein Q, K and V are query, key and value obtained by multiplying input embedding by a weight matrix respectively, firstly sampling K, randomly selecting n K, and obtaining KnTo q is pairediE Q solving the value of M with respect to Kn, the formula is as follows:
whereinqi∈Q，kj∈Kn。
Find the largest n q of MiForm Qn With respect to KnObtaining A (Q)n,KnV), the formula is as follows:
wherein QnFor selected n qiComposed matrix, no selected qiInitialization of the average value after A (Q, K, V) is solved to the original QrMatrix, QnUpdating non-0 values to Q in the matrixrAnd obtaining a final Q matrix from the matrix.
The double-headed attention formula is as follows:
last output of attention mechanism
The value is obtained.
Step five, different data characteristics enter corresponding model input interfaces:
and (3) completing model creation, putting the training set of the preprocessed related data into the model, and specifically, inputting the data divided in the three steps into the model network structure established in the fourth step respectively.
Step six, model training:
and setting the epoch iteration times and training the model. This embodiment sets the parameters to epoch =40 and predicts a time step of 165, i.e. predicts data for 165 days in the future.
Step seven, generating a target prediction model:
and storing the generated model after the training is finished.
Step eight, model testing:
and carrying out simulation prediction on the test set by using the trained model.
The present embodiment predicts server usage data for each day of the future 165 days.
Experiments prove that the method provided by the invention has obvious improvement on the accuracy of the prediction of the resource utilization quantity of the cloud server. The method for predicting the resource utilization quantity of the cloud server provides an important reference index for the operation and maintenance of the cloud service platform. Resource utilization data generated by the current background are utilized to carry out certain accurate data prediction in the future, so that the platform can carry out resource scheduling and safe deployment according to the prediction data of future resource utilization, and the continuous and healthy development of the cloud service platform is ensured and promoted.
Claims (7)
1. A cloud server resource utilization quantity prediction method based on improved TFT is characterized by comprising the following steps:
s 1: acquiring data of the utilization quantity of cloud server resources in a period of time to extract and decompose characteristic variables, and expanding the characteristic variables of two lines of original data (timestamps and actual variable values) to multiple lines through a prediction model Prophet based on time series decomposition;
s 2: selectively extracting the value output by the Prophet model, extracting a timestamp and an actual value corresponding to an input value, and acquiring a trend item dynamic variable calculated by the Prophet model to obtain a new data set;
s 3: further manually expanding time sequence variables for the new data set, increasing the items of year, month, day, week, weekend, holiday and season, and finally expanding two rows of original data into multi-row data to form a multi-variable data set as model input;
s 4: dividing the characteristic variables of the multivariable input data into three types, namely static seasonal variables, past observation variables and known time variables;
s 5: designing and improving a TFT model, improving a GLU (gated linear unit) layer in a variable selection module, and realizing the function of enhancing the selective acquisition and forgetting of past information, wherein the calculation formula is as follows:
the formula I is as follows:
wherein the W1, W2 and W3 are weight parameters to be trained, the b, c and d are bias parameters to be trained, and the gelu () is an activation function gelu which is matrix multiplication calculation;
s 6: improving an attention calculation mode in a TFT model, using a double-head attention mechanism, wherein Q, K and V are query, key and value obtained by multiplying input embedding by a weight matrix respectively, firstly sampling K, randomly selecting n K to obtain Kn, and solving an M value of qi belonging to Q with respect to Kn by the following formula:
the formula II is as follows:
wherein qi belongs to Q, kj belongs to Kn
Finding the largest n qi in M to form Qn, and solving A (Qn, Kn, V) about Kn, wherein the formula is as follows:
the formula III is as follows:
wherein Qn is a matrix formed by n selected qi, the average value is initialized to an original Qr matrix after A (Q, K, V) is solved by the qi which is not selected, and a non-0 value in the Qn matrix is updated to the Qr matrix to obtain a final Q matrix;
the double-headed attention formula is as follows:
the formula four is as follows:
the formula five is as follows:
last output of attention mechanism
A value;
s 7: building a new TFT prediction model according to the improvements in the step s5 and the step s 6;
s 8: respectively accessing the static seasonal variable, the past observation variable and the time known variable divided in the step s4 to the interfaces of the input data of the static covariate, the dynamic time-varying characteristic variable and the dynamic time-invariant characteristic corresponding to the model, so that the model can perform different characteristic selection and calculation on different variables;
s 9: training a model, setting the number of epochs to be 40 and the decoder prediction time step to be 165, and training and establishing the model;
s 10: and testing from the test set to verify the accuracy of the established model.
2. The method according to claim 1, wherein the Prophet model in step s2 is used only for data processing, the time series characteristic variable decomposition operation is performed by using the Prophet prediction model, only the nodes in the original data are reserved as the input of the model, and the prediction result of the model is not used.
3. The method as claimed in claim 1, wherein the manually extended time characteristic variables in step s3 are added with corresponding holiday terms affecting the result of the data according to the actual condition of the data.
4. The method of claim 1, wherein the static seasonal variables in step s4 are seasonal feature terms of the current time point, the past observed variables are known observed dynamic variables before the prediction point, and the known time variables are time variables known in the whole prediction system, such as year, month, week, day, etc.
5. The method according to claim 1, wherein the step s5 is a calculation method for selective memory and forgetting of data features by using a gated linear unit.
6. The method of claim 1, wherein all the formulas in step s6 are used together to calculate each other.
7. The method of claim 1, wherein the seasonal variables are input to the model as static covariates in step s8 so that the seasonal variables can reconcile global features.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111547972.9A CN114238054A (en) | 2021-12-17 | 2021-12-17 | Cloud server resource utilization quantity prediction method based on improved TFT |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111547972.9A CN114238054A (en) | 2021-12-17 | 2021-12-17 | Cloud server resource utilization quantity prediction method based on improved TFT |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114238054A true CN114238054A (en) | 2022-03-25 |
Family
ID=80757638
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111547972.9A Pending CN114238054A (en) | 2021-12-17 | 2021-12-17 | Cloud server resource utilization quantity prediction method based on improved TFT |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114238054A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115118602A (en) * | 2022-06-21 | 2022-09-27 | 中船重工信息科技有限公司 | Container resource dynamic scheduling method and system based on usage prediction |
| CN116151459A (en) * | 2023-02-28 | 2023-05-23 | 国网河南省电力公司电力科学研究院 | Power grid flood prevention risk probability prediction method and system based on improved Transformer |
| CN116416479A (en) * | 2023-06-06 | 2023-07-11 | 江西理工大学南昌校区 | Mineral classification method based on deep convolution fusion of multi-scale image features |
-
2021
- 2021-12-17 CN CN202111547972.9A patent/CN114238054A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115118602A (en) * | 2022-06-21 | 2022-09-27 | 中船重工信息科技有限公司 | Container resource dynamic scheduling method and system based on usage prediction |
| CN115118602B (en) * | 2022-06-21 | 2024-05-07 | 中船重工信息科技有限公司 | Container resource dynamic scheduling method and system based on usage prediction |
| CN116151459A (en) * | 2023-02-28 | 2023-05-23 | 国网河南省电力公司电力科学研究院 | Power grid flood prevention risk probability prediction method and system based on improved Transformer |
| CN116416479A (en) * | 2023-06-06 | 2023-07-11 | 江西理工大学南昌校区 | Mineral classification method based on deep convolution fusion of multi-scale image features |
| CN116416479B (en) * | 2023-06-06 | 2023-08-29 | 江西理工大学南昌校区 | Mineral classification method based on deep convolution fusion of multi-scale image features |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114238054A (en) | Cloud server resource utilization quantity prediction method based on improved TFT | |
| CN105956021B (en) | A kind of automation task suitable for distributed machines study parallel method and its system | |
| Zeng et al. | Traffic prediction of wireless cellular networks based on deep transfer learning and cross-domain data | |
| CN108090788B (en) | Advertisement conversion rate estimation method based on time information integration model | |
| CN103336969B (en) | A kind of image, semantic analytic method based on Weakly supervised study | |
| CN112489497B (en) | Airspace operation complexity evaluation method based on deep convolutional neural network | |
| Li et al. | Weather forecasting using ensemble of spatial-temporal attention network and multi-layer perceptron | |
| CN112910711B (en) | Wireless service flow prediction method, device and medium based on self-attention convolutional network | |
| CN105184368A (en) | Distributed extreme learning machine optimization integrated framework system and method | |
| CN106095966A (en) | A kind of user's extendible label for labelling method and system | |
| CN113688253B (en) | Hierarchical perception temporal knowledge graph representation learning method | |
| CN110309907A (en) | It is a kind of based on go tracking self-encoding encoder dynamic missing values complementing method | |
| Kumar et al. | Wind speed prediction using deep learning-LSTM and GRU | |
| CN110750455B (en) | Intelligent online self-updating fault diagnosis method and system based on system log analysis | |
| Kulyasova et al. | The appliance of deep neural networks in the process of managing chemical enterprises | |
| CN111783688B (en) | Remote sensing image scene classification method based on convolutional neural network | |
| CN101930566A (en) | Hydrological experimental simulation system and method based on parallel system | |
| CN116843012A (en) | Time sequence prediction method integrating personalized context and time domain dynamic characteristics | |
| Li et al. | Long-term traffic forecasting based on adaptive graph cross strided convolution network | |
| CN113835964B (en) | Cloud data center server energy consumption prediction method based on small sample learning | |
| CN112000389A (en) | Configuration recommendation method, system, device and computer storage medium | |
| Zhang | Deep learning-based hybrid short-term solar forecast using sky images and meteorological data | |
| Ma et al. | Data model of key indicators of urban architectural design based on long-and short-term convolutional memory network | |
| Dou | Evaluation Method of College English Education Effect Based on Improved Decision Tree Algorithm | |
| Zhang et al. | Traffic Flow Prediction Based on Graph Convolutional Networks |
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 |