CN110703899B - Data center energy efficiency optimization method based on transfer learning - Google Patents
Data center energy efficiency optimization method based on transfer learning Download PDFInfo
- Publication number
- CN110703899B CN110703899B CN201910848273.4A CN201910848273A CN110703899B CN 110703899 B CN110703899 B CN 110703899B CN 201910848273 A CN201910848273 A CN 201910848273A CN 110703899 B CN110703899 B CN 110703899B
- Authority
- CN
- China
- Prior art keywords
- model
- sample
- samples
- training
- list
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
- G06F1/26—Power supply means, e.g. regulation thereof
- G06F1/32—Means for saving power
- G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
- G06F1/3234—Power saving characterised by the action undertaken
- G06F1/3287—Power saving characterised by the action undertaken by switching off individual functional units in the computer system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Hardware Design (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a data center energy efficiency optimization method based on transfer learning, which belongs to the field of data mining and machine learning, and comprises a Model training method and a Model inference decision method, wherein a basic Model is pre-trained by using data samples of all units based on a Base Model, hidden layer parameters learned by the basic Model are transferred to List-wise models of all the units, and fine adjustment is carried out by using a small learning rate, so that the problem of sample loss of part of the units is solved, and the generalization capability of the Model is improved; by designing a multi-task learning model and adding rank constraint to multi-target loss, the over-fitting problem and the model height variance problem caused by noise are solved, and the convergence speed of the model is improved by selecting samples in a random and periodic sliding window mode; the optimal unit control parameters are obtained by using the optimal linear search of the energy consumption prediction task, the control parameters are sequenced by using a sequencing model of the rank prediction task, the optimal control parameters are selected comprehensively, and the accuracy and the robustness of the optimal control parameters are improved.
Description
Technical Field
The invention relates to the technical field of data mining and machine learning, in particular to a data center energy efficiency optimization method based on transfer learning.
Background
Energy conservation and emission reduction are brought forward by the occurrence of energy and environmental problems, a large number of data centers are built by enterprises and governments along with the development of technologies such as cloud service, big data, AI calculation and the like, and the energy consumption of the data centers in China is generally high at present, and the average PUE value is between 2.2 and 3.0. The electricity consumption of the Chinese data center accounts for 3% of the electricity consumption of the whole society, and is predicted to reach 3.3% in 2020.
At present, the research on energy saving of data centers is more, and energy efficiency performances of different schemes are simulated and compared by using energy consumption simulation software, so that design decision and optimization are assisted. Nevertheless, most studies are biased toward studying the energy saving potential of the design phase and do not consider energy efficiency optimization in actual commissioning. For example, an engineer in GOOGLE of 2016 proposes a model predictive control method using deep learning to find the relationship between a control point and a PUE, thereby helping a heating and ventilation engineer to perform control optimization and helping a data center designer to use design experience for reference. However, the modeling method has the following defects: the modeling difficulty is high, the requirements on the data quantity and the data quality are high, overfitting is easy to realize, and the generalization performance is poor.
In view of this, an energy saving decision and optimization method based on migration learning and List-wise model predictive control is explored by taking optimization of energy consumption of a cooling side of a certain data center as an example, so that the defects are overcome, and the energy saving efficiency of a data center heating and ventilation system is improved as much as possible.
Disclosure of Invention
The invention aims to provide a data center energy efficiency optimization method based on transfer learning, and aims to solve the problems in the background technology.
In order to achieve the purpose, the invention provides the following technical scheme: a data center energy efficiency optimization method based on transfer learning comprises a model training method and a model inference decision method,
the model training method specifically comprises the following steps:
s1: inputting historical original data;
s2: carrying out data preprocessing and characteristic engineering on the original data;
s3: extracting training samples of each unit;
s4: constructing a Base model training sample: combining training samples of all the units, and randomly shuffling;
s5: training a Base model to obtain a prediction model with higher precision and larger variance;
s6: constructing a List-wise model training sample, selecting the training samples of a specified unit, randomly shuffling the training samples and combining the training samples into a List sample;
s7: migrating a Base model pre-training weight, finely tuning a List-wise model, migrating parameters of a pre-trained hidden layer in the Base model to a shared hidden layer of the List-wise model, and finely tuning by adjusting the learning rate of the model to be small;
s8: cutting the Model of S7 into two independent Predictor Model and Rank Model,
the model inference decision method specifically comprises the following steps:
s9: inputting original data on a line;
s10: carrying out data preprocessing and characteristic engineering on the original data;
s11: extracting a prediction sample of each unit;
s12: performing optimized control parameter search solution based on a Predictor Model;
s13: obtaining several groups of optimal control parameters;
s14: sequencing the control parameters by using a Rank Model;
s15: and outputting the optimal control parameters.
Preferably, the Base model training samples are single-target samples, and the operation data of all the units are fused.
Preferably, the Base model is a multilayer perceptron model, and the model aims at predicting the real energy consumption of the sample and initializing the weight for the List-wise model.
Preferably, the List-wise model training samples are multi-target sample pairs, and the weights are initialized by using the weights of the Base model for each unit to be trained independently.
Preferably, the List-wise model is a DNN model designed based on a List-wise structure and a multi-task objective, and the task objective includes predicting actual energy consumption of each sample, predicting a difference between energy consumption of two samples, and predicting a ranking relationship between energy consumption of two samples.
Preferably, the main steps of performing the data preprocessing and feature engineering include:
removing abnormal values, and removing abnormal values larger than three times of variance of the mean value by using a sliding window;
the small window is in sliding average, so that the problem of data fluctuation is solved;
constructing features according to physical characteristics of the device;
performing feature combination and feature intersection;
and (5) performing feature screening according to feature importance.
Preferably, the main steps of constructing the List-wise model sample comprise:
extracting data samples of corresponding units;
sampling the candidate samples to enable the candidate samples to cover the whole characteristic distribution;
searching neighboring samples for the candidate samples, wherein the measurement standard is Euclidean distance between sample vectors under the condition that controlled characteristics are different;
calculating the difference value and the comparison relation between the candidate sample and the adjacent sample as a new target;
the candidate sample and the neighboring sample are grouped into a sample pair.
Compared with the prior art, the invention has the beneficial effects that:
1. a Base Model is pre-trained based on the Base Model using data samples of all units. Hidden layer parameters learned by the basic Model are transferred to the List-wise Model of each unit, fine adjustment is carried out by using a small learning rate, the problem of sample loss of part of units is solved, and meanwhile the generalization capability of the Model is improved;
2. by designing a multi-task learning model and adding rank constraint to multi-target loss, the over-fitting problem and the model height variance problem caused by noise are solved, and meanwhile, the convergence rate of the model can be improved by selecting sample pairs in a random and periodic sliding window mode;
3. the optimal unit control parameters are obtained by using the optimal linear search of the energy consumption prediction task, then the control parameters are sequenced by using a sequencing model of the rank prediction task, and the optimal control parameters are selected comprehensively, so that the accuracy and the robustness of the optimal control parameters are improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of a model training method of the present invention;
FIG. 3 is a flow chart of a model inference decision method of the present invention;
FIG. 4 is a flow chart of a method for performing data preprocessing and feature engineering in accordance with the present invention;
FIG. 5 is a flow chart of a method for constructing a List-wise Model sample according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example one
Referring to fig. 1-2, the present invention provides a technical solution: the model training method in the data center energy efficiency optimization method based on the transfer learning specifically comprises the following steps:
s1: inputting historical original data;
s2: carrying out data preprocessing and characteristic engineering on the original data;
s3: extracting training samples of each unit;
s4: constructing a Base model training sample: combining training samples of all the units, and randomly shuffling;
s5: training a Base model to obtain a prediction model with higher precision and larger variance;
s6: constructing a List-wise model training sample, selecting the training samples of a specified unit, randomly shuffling the training samples and combining the training samples into a List sample;
s7: migrating a Base model pre-training weight, finely tuning a List-wise model, migrating parameters of a pre-trained hidden layer in the Base model to a shared hidden layer of the List-wise model, and finely tuning by adjusting the learning rate of the model to be small;
s8: and (5) Model cutting, namely cutting the Model of the S7 into two independent models of a Predictor Model and a Rank Model.
Example two
Referring to fig. 3, the present invention provides a technical solution: the data center energy efficiency optimization method based on the transfer learning further comprises the following steps of:
s9: inputting original data on a line;
s10: carrying out data preprocessing and characteristic engineering on the original data;
s11: extracting a prediction sample of each unit;
s12: executing the search solution of the optimized control parameters based on the Predictor Model of the energy consumption prediction module;
s13: obtaining several groups of optimal control parameters;
s14: sequencing the control parameters by using a sequencing module Rank Model;
s15: and outputting the optimal control parameters.
EXAMPLE III
Referring to fig. 4, the present invention provides a technical solution: the data center energy efficiency optimization method based on the transfer learning further comprises the main steps of performing data preprocessing and feature engineering, and comprises the following steps of:
removing abnormal values, and removing abnormal values larger than three times of variance of the mean value by using a sliding window;
the small window is in sliding average, so that the problem of data fluctuation is solved;
constructing features according to physical characteristics of the device;
performing feature combination and feature intersection;
and (5) performing feature screening according to feature importance.
Example four
Referring to fig. 5, the present invention provides a technical solution: the method for optimizing the energy efficiency of the data center based on the transfer learning further comprises the following main steps of constructing the List-wise model sample:
extracting data samples of corresponding units;
sampling the candidate samples to enable the candidate samples to cover the whole characteristic distribution;
searching neighboring samples for the candidate samples, wherein the measurement standard is Euclidean distance between sample vectors under the condition that controlled characteristics are different;
calculating the difference value and the comparison relation between the candidate sample and the adjacent sample as a new target;
the candidate sample and the neighboring sample are grouped into a sample pair.
The Base model training samples are single-target samples, and the operation data of all the units are fused. The Base model is a multilayer perceptron model, the model target is to predict the real energy consumption of the sample, and weight is initialized for the List-wiseModel.
The List-wise model training samples are multi-target sample pairs, and the weight initialization uses the weight of the Base model aiming at independent training of each unit. The List-wise model is a DNN model designed based on a List-wise structure and a multi-task target, and the task target comprises prediction of real energy consumption of each sample, prediction of energy consumption difference between every two samples and prediction of sequencing relation of energy consumption between every two samples.
Aiming at the existing energy efficiency optimization method of the data center, in order to solve the problems, the invention provides the following steps:
1. the problem of insufficient data is solved by using a multi-system data transfer learning method,
2. a multi-target learning modeling method is used for solving the problems of poor generalization performance of a model and the need of manual intervention in a training process and helping to expand a training set.
3. A constrained optimal control decision problem combines optimization with ranking.
In the description herein, references to the description of "one embodiment," "an example," "a specific example" or the like are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The preferred embodiments of the invention disclosed above are intended to be illustrative only. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise embodiments disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims (7)
1. A data center energy efficiency optimization method based on transfer learning is characterized by comprising the following steps: comprises a model training method and a model inference decision method,
the model training method specifically comprises the following steps:
s1: inputting historical original data;
s2: carrying out data preprocessing and characteristic engineering on the original data;
s3: extracting training samples of each unit;
s4: constructing a Base model training sample: combining training samples of all the units, and randomly shuffling;
s5: training a Base model to obtain a prediction model with higher precision and larger variance;
s6: constructing a List-wise model training sample, selecting the training samples of a specified unit, randomly shuffling the training samples and combining the training samples into a List sample;
s7: migrating a Base model pre-training weight, finely tuning a List-wise model, migrating parameters of a pre-trained hidden layer in the Base model to a shared hidden layer of the List-wise model, and finely tuning by adjusting the learning rate of the model to be small;
s8: cutting the Model of S7 into two independent Predictor Model and Rank Model,
the model inference decision method specifically comprises the following steps:
s9: inputting original data on a line;
s10: carrying out data preprocessing and characteristic engineering on the original data;
s11: extracting a prediction sample of each unit;
s12: performing optimized control parameter search solution based on a Predictor Model;
s13: obtaining several groups of optimal control parameters;
s14: sequencing the control parameters by using a Rank Model;
s15: and outputting the optimal control parameters.
2. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the Base model training sample is a single target sample, and the operation data of all the units are fused.
3. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the Base model is a multilayer perceptron model, the model target is to predict the real energy consumption of the sample, and weight is initialized for the List-wiseModel.
4. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the List-wise model training samples are multi-target sample pairs, and the weight initialization uses the weight of the Base model aiming at independent training of each unit.
5. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the List-wise model is a DNN model designed based on a List-wise structure and a multi-task target, and the task target comprises prediction of real energy consumption of each sample, prediction of energy consumption difference between every two samples and prediction of sequencing relation of energy consumption between every two samples.
6. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the main steps of the data preprocessing and the characteristic engineering comprise:
removing abnormal values, and removing abnormal values larger than three times of variance of the mean value by using a sliding window;
the small window is in sliding average, so that the problem of data fluctuation is solved;
constructing features according to physical characteristics of the device;
performing feature combination and feature intersection;
and (5) performing feature screening according to feature importance.
7. The data center energy efficiency optimization method based on the transfer learning according to claim 1, characterized in that: the main steps for performing the List-wise model sample construction include:
extracting data samples of corresponding units;
sampling the candidate samples to enable the candidate samples to cover the whole characteristic distribution;
searching neighboring samples for the candidate samples, wherein the measurement standard is Euclidean distance between sample vectors under the condition that controlled characteristics are different;
calculating the difference value and the comparison relation between the candidate sample and the adjacent sample as a new target;
the candidate sample and the neighboring sample are grouped into a sample pair.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910848273.4A CN110703899B (en) | 2019-09-09 | 2019-09-09 | Data center energy efficiency optimization method based on transfer learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910848273.4A CN110703899B (en) | 2019-09-09 | 2019-09-09 | Data center energy efficiency optimization method based on transfer learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110703899A CN110703899A (en) | 2020-01-17 |
| CN110703899B true CN110703899B (en) | 2020-09-25 |
Family
ID=69194982
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910848273.4A Active CN110703899B (en) | 2019-09-09 | 2019-09-09 | Data center energy efficiency optimization method based on transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110703899B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111368408B (en) * | 2020-02-27 | 2023-06-23 | 创新奇智(南京)科技有限公司 | Data center energy efficiency optimization continuous decision method based on seq2seq |
| CN111442476A (en) * | 2020-03-06 | 2020-07-24 | 财拓云计算(上海)有限公司 | Method for realizing energy-saving temperature control of data center by using deep migration learning |
| WO2022110213A1 (en) * | 2020-11-30 | 2022-06-02 | 西门子（中国）有限公司 | Method and device for generating prediction model of analysis object, and storage medium |
| US20230289276A1 (en) * | 2022-03-14 | 2023-09-14 | International Business Machines Corporation | Intelligently optimized machine learning models |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102025535A (en) * | 2010-11-17 | 2011-04-20 | 福建星网锐捷网络有限公司 | Virtual machine management method and device and network equipment |
| CN102833171A (en) * | 2012-09-10 | 2012-12-19 | 杭州华三通信技术有限公司 | Data center-cross MAC (Media Access Control) migration processing method and device |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8873813B2 (en) * | 2012-09-17 | 2014-10-28 | Z Advanced Computing, Inc. | Application of Z-webs and Z-factors to analytics, search engine, learning, recognition, natural language, and other utilities |
| CN106447133A (en) * | 2016-11-03 | 2017-02-22 | 上海交通大学 | Short-term electric load prediction method based on deep self-encoding network |
| CN107169586A (en) * | 2017-03-29 | 2017-09-15 | 北京百度网讯科技有限公司 | Resource optimization method, device and storage medium based on artificial intelligence |
| CN109063911B (en) * | 2018-08-03 | 2021-07-23 | 天津相和电气科技有限公司 | Load aggregation grouping prediction method based on gated cycle unit network |
| CN110163837A (en) * | 2019-01-18 | 2019-08-23 | 腾讯科技（深圳）有限公司 | The recognition methods of video noise, device, equipment and computer readable storage medium |
-
2019
- 2019-09-09 CN CN201910848273.4A patent/CN110703899B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102025535A (en) * | 2010-11-17 | 2011-04-20 | 福建星网锐捷网络有限公司 | Virtual machine management method and device and network equipment |
| CN102833171A (en) * | 2012-09-10 | 2012-12-19 | 杭州华三通信技术有限公司 | Data center-cross MAC (Media Access Control) migration processing method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110703899A (en) | 2020-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110703899B (en) | Data center energy efficiency optimization method based on transfer learning | |
| CN108846526A (en) | A kind of CO2 emissions prediction technique | |
| CN113704956A (en) | Urban road online microscopic simulation method and system based on digital twin technology | |
| CN113762603B (en) | Photovoltaic base station short-term photovoltaic power prediction method based on improved sparrow algorithm optimization | |
| CN109919356B (en) | BP neural network-based interval water demand prediction method | |
| Cheng et al. | Searching toward pareto-optimal device-aware neural architectures | |
| CN105574586B (en) | General-purpose aircraft boat material needing forecasting method based on MPSO-BP networks | |
| CN116402002B (en) | Multi-target layered reinforcement learning method for chip layout problem | |
| CN109816144A (en) | The short-term load forecasting method of distributed memory parallel computation optimization deepness belief network | |
| CN104199870A (en) | Method for building LS-SVM prediction model based on chaotic search | |
| CN111461286A (en) | Spark parameter automatic optimization system and method based on evolutionary neural network | |
| CN102184328A (en) | Method for optimizing land use evolution CA model transformation rules | |
| CN113762387A (en) | Data center station multi-load prediction method based on hybrid model prediction | |
| Shang et al. | Performance of genetic algorithms with different selection operators for solving short-term optimized reservoir scheduling problem | |
| CN111311001B (en) | Bi-LSTM network short-term load prediction method based on DBSCAN algorithm and feature selection | |
| Guo et al. | Applying gated recurrent units pproaches for workload prediction | |
| Tian et al. | Dynamic operation optimization based on improved dynamic multi-objective dragonfly algorithm in continuous annealing process. | |
| Liu et al. | A survey of artificial bee colony algorithm | |
| CN112881869A (en) | Cable joint partial discharge ultrasonic sequence prediction method | |
| CN116503378A (en) | Concentrate grade prediction method for bimodal CNN secondary transfer learning | |
| CN115600492A (en) | Laser cutting process design method and system | |
| Sha et al. | Prediction of commercial building lighting energy consumption based on EPSO-BP | |
| CN110543942A (en) | Multi-space-time long and short memory depth network accurate prediction method | |
| CN113300884A (en) | GWO-SVR-based step-by-step network flow prediction method | |
| Zhang et al. | Optimization study of intelligent decision-making system for coal processing plant based on big data analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20211028 Address after: 569n, building 1, 668 SHANGDA Road, Baoshan District, Shanghai, 201900 Patentee after: Innovation wisdom (Shanghai) Technology Co.,Ltd. Address before: Floor 19, building B, Xingzhi science and Technology Park, No. 6, Xingzhi Road, Nanjing Economic and Technological Development Zone, Jiangsu 210008 Patentee before: AINNOVATION (NANJING) TECHNOLOGY Co.,Ltd. |